QRNN with Addax Wolf Bird Optimization-based LPG gas leakage detection and alert system with IoT edge computing

Abstract

Liquefied Petroleum Gas (LPG) is widely used in households, commercial establishments industrial sectors. LPG is highly flammable and poses a risk of explosion if not handled properly. Hence, an early gas leakage detection system is crucial for ensuring safety in various environments. In this article, a Deep Learning (DL)-based LPG Gas leakage detection and Alert system named Quasi-Recurrent Neural Network with Addax Wolf Bird Optimization Algorithm (QRNN_AWBOA) is proposed using the gas leakage data. In this approach, the median normalization method is utilized to normalize the raw data. Then, a fusion model named Deep Neural Network (DNN) with Neyman similarity is utilized for the feature fusion process. Then, the data is augmented using oversampling. Later, the detection process is carried out using the Quasi-Recurrent Neural Network (QRNN) model. The QRNN effectively trained the Addax Wolf Bird Optimization Algorithm (AWBOA). Finally, an alert is sent to the NG112 authorities if the leakage is present. This detection system attained the lowest Mean Squared Error (MSE) of 0.016, Mean Absolute Percentage Error (MAPE) of 0.056, Root Mean Squared Error (RMSE) of 0.128, and Weighted Absolute Percentage Error (WAPE) of 0.057.

Keywords

Liquefied petroleum gas deep learning optimization gas leakage alert system

1 Introduction

In the computing landscape, edge computing emerges as a new significant perception, which benefits the end user by offering a closer and quicker response time application with fast processing utilities of the cloud services.¹ Edge computing has become more viable in bringing computing closer to the network's edge. This approach partially addresses the resource limitations of end devices and effectively reduces the high latency of cloud services.² Edge computing can enhance real-time smart city environments by enabling context awareness, distributed capabilities across geographic locations, low-latency processing, and the migration of computing resources from centralized cloud infrastructure to the network edge.³ Edge Computing⁴ is a service delivery model which offers robust computational resources close to Internet of Things (IoT) devices. In the case of wildfires, a network of high-performance edge servers supports the offloading and processing of computation-intensive tasks, like image recognition, from IoT nodes. This approach reduces energy consumption while ensuring the application meets its strict time constraints.⁵ The frequent and major types of industrial accidents and fire accidents occur due to the presence of subsequent chemical explosives. When a container leaks, flammable and hazardous chemicals can mix with external oxygen. Once the mixture reaches a critical mass for ignition, an exothermic oxidation reaction occurs, leading to a fire. If the liquid or gas leakage is not addressed in a short period, it will progress into an explosion or fire.⁶ Various early fire-detection approaches have been developed to prevent the devastating effects of fires. These approaches typically include traditional infrared and ultraviolet fire detectors, smoke analysis, air transparency testing, relative humidity sampling, temperature sampling, and particle sampling.⁷

In day-to-day life, Liquefied Petroleum Gas (LPG) plays an essential part and is commonly used in service establishments, like hotels, restaurants, factories, and houses, and in all residential areas.⁸ LPG is a widely used fuel across various sectors. It consists mainly of butanes, propylene (propene), and propane, and it remains in liquid form at moderate pressures and ambient temperatures. Similar to natural gas, burning LPG produces greenhouse gases and lower emissions of regulated pollutants. Furthermore, the cost of LPG is comparatively low. Replacing kerosene, diesel, and gasoline with LPG benefits the users by attaining economic and environmental advantages.⁹ However, the leakage of gas poses a critical threat, specifically in infrastructure and buildings, where the consumption of LPG and Compressed Natural Gas (CNG) is very high. Hence, a gas leak detection system is essential to mitigate this dangerous situation.³ In residential, commercial, and industrial environments, the leakage of gas can lead to environmental harm, property damage, and accidents.¹⁰ The impact of LPG leakage is similar to that of greenhouse gas; both cause serious environmental impacts. Moreover, they pose risks of explosions, suffocation, and intoxication, which impact finances, reputation, property, and also human health. Enhancing the safety of gas systems is crucial to prevent this type of outcome. The leaks, when undetected, result in fatalities, business loss, and significant fires. Therefore, timely detection and appropriate action are essential to mitigate these threats.⁸

Economic losses can be prevented by implementing an early detection system for gas leakage. Furthermore, natural gas leaks can worsen health-related issues, including respiratory problems, asthma, pneumonia, pulmonary disease, and hypertension. Hence, the detection approaches are crucial for countries that highly rely on gas.¹¹ In the operation of gas and oil plant industries, the early, accurate, and real-time implementation of leakage detection approaches ensures safety and reliability.¹² The gas leakage detection method is widely used in several industries; these methods are developed based on the manual inspection of vessels and pipelines. As these approaches need considerable labor, time, and investments, they are naturally inefficient.¹³ The IoT-based multivariate time-series analysis has been enabled by Deep Learning (DL) technology based on its efficient representation learning abilities and feature extraction. However, some prevailing studies on time-series analysis have also incorporated unsupervised DL techniques.⁶ The clustering for IoT time series within a unified framework and risk detection based on unsupervised learning is explored.^11,14,15 DL methodologies are commonly used in pipeline leakage detection systems to identify even small leaks by leveraging their effectiveness while processing data in both the frequency and time domains. The feature extraction process is performed by the employment of Convolutional Neural Networks (CNNs) utilizing a series of filters to recognize significant patterns in the data.¹⁶ Additionally, the CNN method eliminates the need for labor-intensive feature extraction and proves its superiority. At the same time, more precise spatial thermal features are extracted by the DL methods, which enhances the classification accuracy.¹⁷

This article proposes a novel DL methodology with a hybrid optimization algorithm named QRNN_AWBOA for the detection of LPG gas leakage. In this system, the raw data from the gas leakage detection dataset is normalized using median normalization. Then, it undergoes a feature fusion process by utilizing a Deep Neural Network (DNN) with Neyman similarity. Then, the fused features are augmented based on oversampling. Lastly, the detection process is carried out using the developed QRNN_AWBOA. If the leakage is detected, the QRNN_AWBOA detection system sends an alert to the NG112 authorities.

The key contribution of this article is detailed as follows,

QRNN_AWBOA for LPG gas leakage detection and alert system: In this work, LPG gas leakage detection is carried out using the DL model named Quasi-Recurrent Neural Network (QRNN). The ORNN model is efficiently optimized by a hybrid algorithm named Addax Wolf Bird Optimization Algorithm (AWBOA). Here, AWBOA is developed by the fusion of the Addax Optimization Algorithm (AOA) and Wolf Bird Optimizer (WBO).

The remaining part of this article is arranged in the following order. Section 2 elaborates on the drawbacks and the challenges faced by the prevailing methods. Section 3 defines the system model. Section 4 details the block diagram and the architecture of the proposed QRNN_AWBOA for LPG gas leakage detection. Section 5 explains the results and the comparative analysis, and Section 6 concludes the research article.

2 Motivation

LPG is highly flammable and poses significant explosion risks. To enhance the safety of an individual and reduce risks associated with gas usage, the early detection of gas leakage and an alert system are required. The prevailing detection system faced difficulties in processing complex features and failed to produce more accurate outcomes. To tackle this drawback, the QRNN_AWBOA is developed for LPG gas leakage detection. The benefits, drawbacks, and challenges encountered by the prevailing detection systems are detailed below.

2.1 Literature review

A smart fire and gas leakage alert system was devised by Maltezos, E., et al.³ based on the Distributed Edge Computing IoT Platform (DECIoT). In the integration of different nodes, the DECIoT method provided potential adaptability, modularity, and flexibility. It also offered an effective incorporation of information from diverse sources. However, the DECIoT method failed to extend the smart building system by incorporating additional edge gateways and sensor nodes, such as camera sensors with video streaming capabilities. The Advantage Actor Critic (AAC) algorithm was introduced by Wei, D., et al.² for the detection of gas pipeline leakage. This approach effectively reduced the overall computational cost and exhibited a faster convergence speed. However, the AAC method failed to consider the collaboration among edge nodes, for lowering the energy consumption of the system, and to optimize the use of idle resources. Dash, A., et al.⁶ proposed an Artificial Intelligence (AI)-Enabled IoT Framework for the leakage of LPG detection and the prediction of its consequences at the time of external transportation. This method was highly compatible with time-sensitive applications and was highly reliable. Nevertheless, frequently transported chemical hazards were not examined to minimize the consequent losses. Wireless Sensor Networks (WSNs) and Mobile Cellular Networks (MCNs) were developed by Djehaiche, R., et al.¹⁸ for IoT/Machine-to-Machine (M2M) smart building systems. This approach attained higher efficiency and produced better outcomes with reduced error rates. However, this approach failed to combine the Machine Learning (ML) strategies to attain robustness.

A Novel Deep Autoencoder and Multivariate Analysis Approach (NDAMA) was developed by Dashdondov, K., et al.,¹¹ for the detection of methane gas leakage. This detection method appeared as the best model and increased the detection accuracy with reduced log loss and variability. However, this method faced more difficulties in interpretation, and it also failed to use huge datasets for effective training. Praveenchandar, J., et al.¹⁹ proposed an IoT-based monitoring and fault detection of harmful toxic gases using the Artificial Neural Network (ANN) and Hidden Markov Model (HMM). This method attained a significantly high fault data detection accuracy with decreased false positive rates. However, this method failed to examine its effectiveness on further datasets with varied parameters to increase its generalization ability. A detection approach for gas leakage was introduced by Kopbayev, A., et al. in¹³ based on a Convolutional Neural Network with a Bi-directional Long Short-Term Memory Layer Network (CNN-BiLSTM). The results demonstrated that the model was able to classify leakage effectively, achieving high accuracy. However, this approach did not incorporate real-world data, nor did it explore how integrating sensor data could impact prediction accuracy and improve overall efficiency. Zuo, Z., et al.¹² devised an LSTM-Autoencoder and One-Class Support Vector Machine (LSTM-AE-OCSVM) for gas leakage detection. The leakage detection technique proved to be more effective and sufficient, delivering improved performance with satisfactory results. Nevertheless, the training phase of the LSTM-AE-OCSVM method failed to consider the dataset with abnormal and little leakage data.

2.2 Challenges

The prevailing LPG Gas leakage detection approaches in Edge computing face the following challenges and issues,

In the AAC² approach, the communication resource allocation, along with the task offloading strategy, was enhanced. However, this approach failed to accurately predict the quantiles, which limited its ability to enhance the understanding of leakage severity and uncertainty.

The AI-enabled IoT framework⁶ predicted the impact of leakage directly at the edge device within the IoT network, bypassing the latency and limitations of cloud-based Computational Fluid Dynamics (CFD) analysis. However, this model failed to capture the complex relationships between multiple features, which hindered its ability to improve accuracy in detecting gas leaks.

The centralized platform provided by the IoT/M2M¹⁸ mechanism efficiently monitored and controlled the services of smart buildings. However, this detection approach did not consider the incorporation of DL strategies to improve robustness and enhance performance.

The NDAMA model¹¹ provided a robust framework for predicting gas leakage. Meanwhile, the NDAMA model failed to prove its effectiveness on different datasets or environments without further tuning and validation, which were essential for the enhanced accurate real-time detection of gas leakage.

In real-time detection, Gas leakage detection systems must process the data to issue immediate alerts. Therefore, rapid data processing and efficient algorithms are crucial for managing large volumes of sensor data.

3 System model

The system model³ addresses various challenges, such as consolidating, refining, and gathering data while interfacing with IoT devices. It also manages the export of data in formats suitable for other platforms, as well as data transformation, processing, temporary local storage, command execution, notifications, system administration, and security. This comprehensive workflow is powered by advanced open-source microservices, optimized for distributed edge IoT architectures. This system model contains various layers, and each layer in this model contains various microservices. The core idea behind microservice architecture is that an application can be structured as a set of independently deployed services, where each service is a self-contained unit responsible for delivering a specific functionality within the application. The communication between these microservices occurs either one-way or two-way. There are four layers in the system model: the Device service layer, the Core service layer, the Support service layer, and the Application service layer. The functionality of the device service layer is to collect data by triggering the device with its command. The device service layer also acts as an interface between both physical and system devices. The MQTT device service is utilized for receiving information from the sensor nodes of LPG. Among the communication between MQTT and sensor nodes, MQTT plays the role of broker. In the center of the system model, the core service layer is presented. This layer consists of three microservices they are the core data service, the command service, and the metadata service. The core data service performs the task of storing data, the command service helps in initiating the triggered commands to the devices, and the metadata service helps in storing the details of the registered devices. Then, a layer called the support service layer comprises the microservices as scheduling services, rules engine services, and logging services. These microservices are used in the typical edge or local analytical application tasks, like data filtering, scheduling, and logging. The application service layer is used to communicate with external applications, like middleware, Representational State Transfer (REST) interface, and Public Safety Answering Point (PSAP). This layer comprises one or more microservices, like middleware application service, NG112 application service, and so on. The NG112 application service, situated in the Application Services layer, operates an MQTT consumer that monitors designated alert topics. Upon receiving an alert, the application verifies whether the same alert has occurred recently within a configurable time period. If it has, the alert is suppressed to avoid overwhelming the system. If not, the alert is sent to both the smart city platform and the REST interface through a REST call. The system model for the LPG gas leakage detection and alert system is depicted in Figure 1.

Figure 1.

System model of LPG gas leakage detection and alert system.

4 Proposed QRNN_AWBOA for LPG gas leakage detection and alert system

The primary intention of this research is to devise a new approach called QRNN_AWBOA for LPG gas leakage detection. Initially, LPG sensor nodes with edge computing are considered first, and the raw data from the sensors are subjected to data normalization, where the data is normalized by employing Median normalization.²⁰ After data normalization, feature fusion is carried out using DNN²¹ with Neyman similarity.²² Following this, data augmentation is performed by utilizing oversampling. At last, LPG gas leakage is detected using a QRNN,²³ which is trained by the AWBOA. Here, AWBOA is formulated by the integration of WBO²⁴ and AOA.²⁵ At last, if the gas leakage is detected, then the alert is sent to the NG112 authorities. The block diagram of QRNN_AWBOA for LPG gas leakage detection is displayed in Figure 2.

Figure 2.

Block diagram of QRNN_AWBOA for LPG gas leakage detection and alert system.

4.1 Gas leakage data acquisition

The fire, smoke, and gas leakage detection dataset F is utilized in the QRNN_AWBOA model for the detection of gas leakage. The leakage detection dataset is formulated as,

\begin{aligned} F = {F_{1}, F_{2}, F_{3}, \dots, F_{q}, \dots, F_{p}} \end{aligned}

(1)

Here, the total amount of data records present in the dataset is denoted by p, and then the $q^{t h}$ data in the dataset is denoted as $F_{q}$ .

4.2 Feature extraction using technical indicators

The required features from the raw data $F_{q}$ are extracted based on technical indicators, like Simple Moving Average (SMA), Triple Exponential Moving Average Oscillator (TRIX), and Weighted Moving Average (WMA). Later, the extracted features are concatenated and form a single feature $c_{E F}$ .

4.2.1 SMA

The SMA-based²⁶ feature extraction helps in filtering the noise from the data and extracting the trends in gas concentration levels from the data. The SMA is used to reduce the short-term fluctuations in time series analysis. The expression of SMA is formulated as,

\begin{aligned} V_{S M A}^{'} = \sum_{k^{'} = 1}^{l} \frac{ε_{k^{'}}}{l^{'}} \end{aligned}

(2)

Here, the total number of moving averages is denoted as $l^{'}$ , and the gas concentration level at each time $k^{'}$ is represented as $ε_{k^{'}}$ . The extracted feature from the data based on SMA is signified as $V_{S M A}^{'}$ .

4.2.2 TRIX

The feature extraction using TRIX²⁶ in gas leakage detection aids in extracting time series data. It is also employed to determine the level of LPG gas concentration from the data. The TRIX-based feature extraction is defined as follows.

\begin{aligned} E_{T R I X}^{'} = \frac{(W 3_{i} - W 3_{i - 1})}{W 3_{i - 1}} \end{aligned}

(3)

Here, $W 3$ symbolizes the third exponential moving average, i denotes the time, and $W 3_{i - 1}$ represents the third exponential moving average at the time $i - 1$ . Finally, the extracted TRIX feature is signified as $E_{T R I X}^{'}$ .

4.2.3 WMA

The WMA-based²⁶ feature extractor calculates the averages in time series by assigning different weights to each data point in the time series. The expression of WMA is formulated as,

\begin{aligned} U_{W M A}^{'} = \frac{i \times A + (i - 1) \times A 1 + (i - 2) \times A 2 + \dots . + A (i - 1)}{1 + 2 + 3 + \dots . i} \end{aligned}

(4)

Here, i signifies the time, and A denotes the weights assigned. The final extracted WMA feature is represented as $U_{W M A}^{'}$ .

Finally, all the features that are extracted based on SMA, TRIX, and WMA are concatenated and form a final feature $c_{E F} = {{U^{'}}_{W M A}, {E^{'}}_{T R I X}, {V^{'}}_{S M A}}$ .

4.3 Data normalization

The data normalization process is used to organize the unstructured format of the data from the dataset. In the gas leakage detection dataset, the median normalization²⁰ technique is used to normalize the extracted feature $c_{E F}$ . In this process, $c_{E F}$ is adjusted by centering it around the median value of the dataset F. This process makes the interpretability of the data easier, faster convergence, and in a more structured format. The use of median normalization is particularly suitable for handling skewed or noisy sensor data, as it is less sensitive to outliers compared to conventional normalization methods. This makes it highly effective in the context of gas sensor readings, where occasional fluctuations or spikes are common, thereby improving the overall stability and reliability of the feature extraction and detection process. The expression for median normalization is formulated as,

\begin{aligned} e_{o}^{'} = \frac{e_{o}}{{\bar{S}}^{'}} \end{aligned}

(5)

where

e_{o}^{'}

denotes the normalized data at the time o,

e_{o}

denotes the original data at the time o, and

\bar{S^{'}}

denotes the median. The normalized data is represented as

I_{d}

4.4 Feature fusion

In the feature fusion process, multiple features of the normalized data $I_{d}$ are combined to form a more informative single representation of the feature set. Thus, a single set of representations can capture a wide range of relationships and information by reducing the overall dimensions. Here, the feature fusion process is carried out using the DNN²¹ with the Neyman similarity²² technique.

The Neyman similarity is a statistical measure that evaluates the closeness or similarity between two data distributions. Initially, the distance between the normalized data and the targeted data is measured based on the Neyman measure, and the features are ranked accordingly. This ensures that highly relevant and informative features are emphasized while redundant or less informative features are suppressed. The Neyman measure is expressed as,

\begin{aligned} h (M, N) = \sum ({(M_{f} - N_{f})}^{2} / M_{f}) \end{aligned}

(6)

here, M denotes the candidate feature, which also symbolizes the normalized data

I_{d}

, N denotes the target feature, h denotes the distance between the normalized data and the targeted data, and f denotes the data point. Then, the sorted features

T_{x}

are fused. The feature fusion expression is formulated as,

\begin{aligned} T_{n}^{N e w} = \sum_{\begin{matrix} f = n \\ f + 1 = n + Z \end{matrix}}^{t} \frac{η_{f}}{f} T_{x} \end{aligned}

(7)

where

T_{n}^{N e w}

symbolizes the fused features, t represents the total number of features, Z represents the number of features to be selected, and

η

represents the coefficient of fusion parameters. The coefficient of fusion parameters

η

is generated by using the DNN. The ground truth for generating

η

is expressed as,

\begin{aligned} η = h (u_{f}, β_{f}) \end{aligned}

(8)

here,

u_{f}

denotes the data record and

β_{f}

denotes the average of

u_{f}

belongs to the class. The final fused feature is represented as T.

Architecture of DNN

The DNN²¹ architecture is utilized for the generation of $η$ in the feature fusion phase. A sub-branch of Artificial Neural Network is DNN, which comprises several layers of neurons for learning the complex patterns in the data. DNNs are well-suited for parallel processing, enabling efficient computation of feature fusion when working with large-scale datasets. Their ability to learn complex, non-linear relationships, handle diverse data types, and improve generalization makes them ideal for feature fusion. The DNN approach takes the input-sorted feature T. The DNN layers consist of an input layer and an output layer. The layers other than the input and output layers are known as hidden layers. The activation function for DNN is the Rectified Linear Unit (ReLU), which makes the DNN architecture highly non-linear. The architecture of DNN is illustrated in Figure 3. The activation value for each node of the hidden layers is represented as,

\begin{aligned} ϖ_{y, z} = R (α_{y, z}) = {\begin{array}{ll} α_{y, z} & i f α_{y, z} \geq 0 \\ 0 & o t h e r w i s e \end{array} \end{aligned}

(9)

here, y signifies the node and z signifies the layer. Then the

y^{t h}

node of z is represented as

κ_{y, z}

. The ReLU activation function is denoted by R. The

ϖ_{y, z}

and

α_{y, z}

signifies the values before and after the activation function. Other than the input layer, the nodes of every layer are linked with its previous layer's node using pre-trained parameters, which are represented as,

\begin{aligned} ϖ_{y, z} = g_{y, z} + \sum_{1 \leq s \leq w_{y - 1}} τ_{y - 1, s, z} . ϖ_{y - 1, s} \end{aligned}

(10)

where the bias for

κ_{y, z}

node is represented as

g_{y, z}

, and the weights for the connection among

κ_{y - 1, s}

and

κ_{y, z}

is represented as

τ_{y - 1, s, z}

. This shows that the DNN architecture poses the expression of both convolutional and fully-connected functions. Finally, the fused feature generated by DNN is exemplified as

η

Figure 3.

Architecture of DNN.

4.5 Data augmentation

The data augmentation process using random oversampling is generally used to reduce the class imbalance issues in the dataset. The increase in the count of instances in the minority class increases the variety and size of the dataset by generating new instances or duplicating the prevailing data points. By using this technique, the diversity of the data is increased. This approach improves class balance during training, helping the model learn equally from both classes and enhancing predictive performance. In the LPG gas leakage detection process, the fused features T are augmented using the oversampling technique. Here, the size of the fused features is increased to resolve the imbalance problem faced during the training phase. The final augmented data is represented as $χ_{a^{'}}$ .

4.6 LPG gas leakage detection

LPG gas plays a crucial role in residential, commercial, and industrial areas. Many households and commercial sectors, like restaurants and food service businesses, highly depend on LPG for cooking. Although LPG gas is useful in many ways, it also poses a high threat of causing fire accidents and explosions. To prevent this accident and casualties, an early gas leakage detection system is required. Hence, the QRNN²³ model with AWBOA is developed for the early detection of LPG gas leakage. Here, the augmented data $χ_{a^{'}}$ is given as input to the QRNN model for leakage detection. The QRNN model is efficiently trained by the AWBOA. The AWBOA is developed by the combination of AOA²⁵ and WBO.²⁴ Finally, the system sends alerts to the NG112 authorities, if the leakage is detected during the detection process.

4.6.1 Architecture of QRNN

A neural hybrid architecture developed by the combination of two neural networks, Long Short-Term Memory (LSTM) and CNN, is named QRNN.²³ The QRNN architecture is illustrated in Figure 4. The QRNN model possesses the capabilities of both CNN and LSTM. The QRNN architecture comprises two subcomponents: pooling layers and the convolutional layers. The pooling components and the convolutional components of the QRNN architecture have properties that are similar to CNN. The pooling components allow fully parallel computation across feature dimensions and minibatches, whereas the convolutional component allows fully parallel computation across spatial dimensions and minibatches. The QRNN model benefits from the parallelization ability of convolutional networks and extracts local features from input sequences. It effectively extracts local spatial features through convolutions while maintaining the ability to capture long-range temporal dependencies via its gating mechanism, similar to that of LSTM. The QRNN model exhibits the property of processing multiple inputs at the same time with less training time. The convolutions in time step with a bank of filters are performed by the convolutional subcomponent of QRNN. The full set of computations in the computational component at each timestep is expressed as,

\begin{aligned} \begin{aligned} O & = \tanh (X_{o} * L) \\ J & = λ (X_{j} * L) \\ B & = λ (X_{b} * L) \end{aligned} \end{aligned}

(11)

where, the forget gate and output gate of the pooling function are denoted by J and B, the update gate is denoted by O. The input augmented data

χ_{a^{'}}

is represented by L, the convolutional filter bank is represented by

X_{o}, X_{j},

and

X_{b}

, and the masked convolution with timestep dimension is denoted by

*

. When the width of the filter becomes 2, equation (11) is reduced to LSTM and it is formulated as,

\begin{aligned} \begin{aligned} o_{ϕ} & = \tanh (X_{o}^{1} υ_{ϕ - 1} + X_{o}^{2} υ_{ϕ}) \\ j_{ϕ} & = λ (X_{j}^{1} υ_{ϕ - 1} + X_{j}^{2} υ_{ϕ}) \\ b_{ϕ} & = λ (X_{b}^{1} υ_{ϕ - 1} + X_{b}^{2} υ_{ϕ}) \end{aligned} \end{aligned}

(12)

where,

o_{ϕ}

refers to the update gate, the output gate, and the forget gate are denoted by

b_{ϕ}

, and

j_{ϕ}

ϕ

signifies the time and

υ

signifies the input. The higher features at each timestep are effectively computed by the convolutional filters with larger widths. The element-wise gates of LSTM are used for the formation of functions suitable for the pooling subcomponent. The functions that are controlled by gates perform independently on every channel of the state vector. The dynamic average pooling utilizes only a forget gate, which is formulated as,

\begin{aligned} δ_{ϕ} = j_{ϕ} ⊙ δ_{ϕ - 1} + (1 - j_{ϕ}) ⊙ o_{ϕ} \end{aligned}

(13)

here, the elementwise multiplication is denoted by

⊙

, and the hidden state of the element is denoted by

δ

. The output gate function is formulated as,

\begin{aligned} \begin{aligned} τ_{ϕ} & = j_{ϕ} ⊙ τ_{ϕ - 1} + (1 - j_{ϕ}) ⊙ o_{ϕ} \\ δ_{ϕ} & = b_{ϕ} ⊙ τ_{ϕ} \end{aligned} \end{aligned}

(14)

here,

τ

signifies the current hidden state. The recurrence relation of the independent forget gate and the input gate is formulated as,

\begin{aligned} \begin{aligned} τ_{ϕ} & = j_{ϕ} ⊙ τ_{ϕ - 1} + μ_{ϕ} ⊙ o_{ϕ} \\ δ_{ϕ} & = b_{ϕ} ⊙ τ_{ϕ} \end{aligned} \end{aligned}

(15)

here,

μ

denotes the input gate. After the gated linear combination of convolutional layers, the single layer of QRNN performs the dependent input pooling. The output produced by the QRNN for gas leakage detection is denoted by

C_{n^{'}}

. Finally, based on the detection result, if the leakage is present, an alert is sent to the NG112 authorities.

Figure 4.

Architecture of QRNN.

4.6.2 Training of QRNN using AWBOA

The LPG gas leakage detection is done by using the QRNN²³ model. The QRNN model is efficiently optimized using an algorithm named AWBOA. The optimization algorithm AWBOA is developed by the combination of AOA²⁵ and WBO.²⁴ AOA is highly utilized for resolving more intricate optimization problems in continuous and discrete domains. To solve this intricate problem, AOA adopts the movement strategies and the foraging behavior of Addax. In a shorter period, AOA identifies an optimal or near-optimal solution by efficiently exploring the solution space. However, the performance of AOA can degrade if its parameters are not properly tuned for a specific problem, often requiring trial-and-error or additional optimization strategies. The WBO resolves the intricate optimization problem by leveraging the adaptive and collaborative behavior of wolves and birds. WBO applies to various optimization problems, like multi-objective, discrete, and continuous problems. By integrating WBO with AOA, AWBOA effectively overcomes the limitations of parameter sensitivity in AOA, resulting in a more robust, scalable, and accurate optimization process. This integration enhances the performance of the QRNN-based detection system, making it more effective for real-time LPG gas leakage detection. The steps followed in the training of QRNN using AWBOA are as follows,

A)
Initialization

In the population $ξ$ , the position of an Addax represents a potential solution to the optimization problem. The initialization is formulated as,
$\begin{aligned} ξ = [\begin{array}{cccccc} ξ_{1, 1} & ξ_{1, 2} & \dots & ξ_{1, l} & \dots & ξ_{1, D i m} \\ ξ_{2, 1} & ξ_{2, 2} & \dots & ξ_{2, l} & \dots & ξ_{2, D i m} \\ ⋮ & ⋮ & ⋱ & ⋮ & ⋱ & ⋮ \\ ξ_{k, 1} & ξ_{k, 2} & \dots & ξ_{k, l} & \dots & ξ_{k, D i m} \\ ⋮ & ⋮ & ⋱ & ⋮ & ⋱ & ⋮ \\ ξ_{v^{'}, 1} & ξ_{v^{'}, 2} & \dots & ξ_{v^{'}, l} & \dots & ξ_{v^{'}, D i m} \end{array}] \end{aligned}$
(16)
where the total count of addax present in a population is denoted by $v^{'}$ , the count of decision variables is denoted by $D i m$ , and the $l^{t h}$ dimension of the $k^{t h}$ addax is denoted by $ξ_{k, l}$ . B)
Fitness Evaluation

The fitness evaluation is used to find the optimal solution to a problem by continuously evaluating and improving the fitness of the addax with respect to the objective function or optimization problem. The evaluation for fitness is represented as,
$\begin{aligned} S_{M S E} = \frac{1}{V} \sum_{n^{'} = 1}^{V} {[C_{n^{'}} * - C_{n^{'}}]}^{2} \end{aligned}$
(17)
here, $S_{M S E}$ refers to the Mean Squared Error (MSE), $C_{n^{'}} $ signifies the expected result, $C_{n^{'}}$ signifies the actual result of QRNN, and the number of training samples is represented as V. C)
Exploration phase: Foraging process

In the exploration phase of AOA, the population members’ positions in the solution space are adjusted and mathematically modeled to reflect the shifts in the addaxes’ positions observed during their foraging behavior. Therefore, this improves the algorithm's ability to explore the search space, enhancing its effectiveness in global search management. For foraging, the addax in AOA arbitrarily selects a designated area. During the foraging process, the addax moves to an appropriate vegetation area, and then a novel position for every member of AOA is designed and it is expressed as,
$\begin{aligned} ξ_{k, l} (m + 1) & = ξ_{k, l} (m) + σ_{k, l} (H P_{k, l} - E_{k, l} ξ_{k, l} (m)) \end{aligned}$
(18)

$\begin{aligned} ξ_{k, l} (m + 1) & = ξ_{k, l} (m) (1 - E_{k, l} * σ_{k, l}) + σ_{k, l} H P_{k, l} \end{aligned}$
(19)
here, m signifies the iteration, $σ_{k, l}$ signifies the random numbers in the range of [0,1], $E_{k, l}$ signifies the numbers that are selected randomly, the selected region for foraging by the addax k is represented as $H P_{k}$ , and $H P_{k, l}$ signifies the $l^{t h}$ dimension of $H P_{k}$ in that area. To enhance the flexibility and robustness of the algorithm, the WOA is incorporated. Further, it enhanced the convergence rate of the technique. The position updating process of WBO is expressed as,
$\begin{aligned} ξ_{k, l} (m + 1) & = ξ_{k, l} (m) + σ_{1} * ξ_{k, l}^{K} (m) - σ_{2} * K G K \end{aligned}$
(20)

$\begin{aligned} ξ_{k, l} (m) & = ξ_{k, l} (m + 1) - σ * ξ_{_{k, l}}^{K} (m) + σ_{2} * K G K \end{aligned}$
(21)
here, $σ_{1}$ and $σ_{2}$ denotes the random numbers in the range [0,1], and $K G K$ signifies the center point of prey. Substituting equation (21) in equation (19),
$\begin{aligned} ξ_{k, l} (m + 1) = (ξ_{k, l} (m + 1) - σ * ξ_{_{k, l}}^{K} (m) + σ_{2} * K G K) (1 - E_{k, l} * σ_{k, l}) + σ_{k, l} H P_{k, l} \end{aligned}$
(22)

$\begin{aligned} ξ_{k, l} (m + 1) - ξ_{k, l} (m + 1) (1 - E_{k, l} * σ_{k, l}) = (- σ * ξ_{_{k, l}}^{K} (m) + σ_{2} * K G K) (1 - E_{k, l} * σ_{k, l}) + σ_{k, l} H P_{k, l} \end{aligned}$
(23)

$\begin{aligned} ξ_{k^{'}, l^{'}} (m + 1) [E_{k^{'}, l^{'}} * σ_{k^{'}, l^{'}}] = (- σ * ξ_{_{k^{'}, l^{'}}}^{K} (m) + σ_{2} * K G K) (1 - E_{k^{'}, l^{'}} * σ_{k^{'}, l^{'}}) + σ_{k^{'}, l^{'}} H P_{k^{'}, l^{'}} \end{aligned}$
(24)

$\begin{aligned} ξ_{k, l} (m + 1) = \frac{1}{E_{k, l} * σ_{k, l}} [σ_{k, l} H P_{k, l} - (σ * ξ_{_{k, l}}^{K} (m) - σ_{2} * K G K) (1 - E_{k, l} * σ_{k, l})] \end{aligned}$
(25)

Equation (25) updates the position of $k^{t h}$ addax which makes the search solution more optimal during the foraging process. D)
Exploitation phase: Digging skill

In AOA, the exploitation phase acts as the second phase. In this phase, also the population members’ positions in the solution space are adjusted and mathematically modeled to reflect the shifts in the addaxes’ positions observed during their sand-digging behavior. In the problem-solving space, slight changes in the positions of the population members occurred during the positional adjustments for digging results. This enhances the algorithm's ability to exploit effectively, improving local search management. The new position for every addax in the population during digging is expressed as,
$\begin{aligned} ξ_{_{k, l}}^{a 2} = ξ_{_{k, l}} + (1 - 2 σ_{_{k, l}}) . \frac{ψ ρ_{l} - ς ρ_{l}}{m} \end{aligned}$
(26)
here, the new position of the $k^{t h}$ addax in $l^{t h}$ dimension is denoted by $ξ_{_{k, l}}^{a 2}$ , m denotes the iteration counter, $ψ ρ_{l}$ signifies the upper bound of the $l^{t h}$ dimension, and $ς ρ_{l}$ signifies the lower bound of the $l^{t h}$ dimension. E)
Feasibility Evaluation

The AWBOA is evaluated for its feasibility by using the MSE from equation (17). This equation ensures that the obtained solution of QRNN after training is optimal or not. F)
Termination

The iteration is repeated until the optimal solution is obtained. The optimal solution found is updated if a better solution is discovered up to a maximum iteration. The pseudocode of the AWBOA is depicted in Algorithm 1.

By incorporating both Wolf and Bird behaviors, AWBOA is able to strike a better balance between exploration and exploitation, improving the algorithm's ability to converge to high-quality solutions. This improves the detection capability of QRNN and produces more effective and accurate detection results.
5 Results and discussion

The effectiveness of the developed QRNN_AWBOA for LPG gas leakage detection is validated using the fire, smoke, and gas leakage detection data. The developed QRNN_AWBOA is compared with the prevailing gas leakage detection system and analyzed for its effectiveness. The following sections detail the experiment setup of the developed QRNN_AWBOA, a description of the utilized fire, smoke, and leakage detection dataset, evaluation metrics, and comparative analysis.

5.1 Experimental set-up

The QRNN-AWBOA for the LPG gas leakage detection is implemented using the Python tool. The parameters and values of the devised QRNN-AWBOA are presented in Table 1.

Table 1.
Parameters and values.

Parameters Values

QRNN

Epoch 50

Batch size 32

Optimizer AWBOA

Kernel size (3,3)

Dropout 0.3

Loss mse

Activation relu

AWBOA

N(population size) 50

M(Decision Variable) 30

Maximum Iteration 200

r random numbers between (0,1)

I random number of (1 or 2)

Parameters	Values
QRNN
Epoch	50
Batch size	32
Optimizer	AWBOA
Kernel size	(3,3)
Dropout	0.3
Loss	mse
Activation	relu
AWBOA
N(population size)	50
M(Decision Variable)	30
Maximum Iteration	200
r	random numbers between (0,1)
I	random number of (1 or 2)

5.2 Dataset description

The proposed LPG gas leakage detection system using QRNN_AWBOA utilized the fire, smoke, and gas leakage detection dataset.²⁷ This dataset is associated with a smart building sensor system and contains a collection of sensor readings and environmental data. This dataset contains the data in the form of sensor type, sensor readings, timestamps, and environmental conditions. The sensor types include LPG/CNG sensors, smoke and CO sensors, flame/IR sensors, and temperature and humidity sensors. The corresponding measurement units are parts per million (ppm) for gas concentrations and degrees Celsius (°C) or percentage (%) for temperature and humidity. The data is collected under normal operation and during controlled physical triggers.

5.3 Evaluation measures

The proposed LPG gas leakage detection system using QRNN_AWBOA is analyzed for its efficiency, using the evaluation metrics, such as MSE, Mean Absolute Percentage Error (MAPE), Root Mean Squared Error (RMSE), and Weighted Absolute Percentage Error (WAPE). The description of the metrics is listed as follows.

1) MSE: The average squared difference between the actual values and the detected value is measured using MSE, and it is represented in equation (17).

2) MAPE: The error percentage of the actual values is measured using MAPE, and it is formulated as,

\begin{aligned} M A P E = \frac{100}{V} \sum_{n^{'} = 1}^{V} | \frac{C_{n^{'}} * - C_{n^{'}}}{C_{n^{'}}} | \end{aligned}

(27)

3) RMSE: The square root of the average of squared differences between actual values and the detected value, and it is expressed as,

\begin{aligned} R M S E = \sqrt{\frac{1}{V} \sum_{n^{'} = 1}^{V} {(C_{n^{'}} * - C_{n^{'}})}^{2}} \end{aligned}

(28)

4) WAPE: It is a variation of MAPE that incorporates weights to measure the error percentage, which is formulated as,

\begin{aligned} W A P E = \frac{\sum_{n^{'} = 1}^{V} ϖ_{n^{'}} | C_{n^{'}} * - C_{n^{'}} |}{\sum_{n^{'} = 1}^{V} ϖ_{n^{'}} | C_{n^{'}} |} \times 100 \end{aligned}

(29)

where,

ϖ_{n^{'}}

signifies weight.

5.4 Prediction analysis

The proposed QRNN_AWBOA for the LPG gas leakage detection and alert is validated based on its prediction and is depicted in Figure 5. The prediction value is compared with the original value over time. The proposed QRNN_AWBOA attained the prediction value of 23.319 on 2022-03-08T12:38, whereas the original value is 2.50E + 01. The predicted value of the proposed QRNN_AWBOA is comparatively lower than the original value. The same continues for time up to 2022-03-08T05:30. The predicted value attained by the proposed QRNN_AWBOA and the original value is almost similar from the time 2022-03-08T05:30 to 2022-03-09T12:14. The proposed QRNN_AWBOA attained the predicted value of 20.00727 at time 2022-03-09T12:45, whereas as the original value is 1.97E + 01 The predicted value of the proposed QRNN_AWBO is comparatively higher than the original value and the trend continues till the time considered.

Figure 5.

Prediction graph analysis.

5.5 Comparative techniques

The QRNN_AWBOA for LPG gas leakage detection is analyzed for its effectiveness by comparing it with the other prevailing detection methods, such as AAC,² AI-enabled IoT framework,⁶ IoT/M2M,¹⁸ NDAMA,¹¹ 1D-CNN,²⁸ BiLSTM,²⁹ QRNN + Adam.^30,31

5.6 Comparative assessment

The comparative assessment of the QRNN_AWBOA is carried out based on the LPG concentration on different days.

5.6.1 Analysis based on LPG-concentration on 2022-03-09

The comparative assessment of the proposed QRNN_AWBOA with other prevailing methods, such as AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam at setup LPG-Concentration on 2022-03-09 is illustrated in Figure 6. The comparative analysis is done based on the evaluation metrics such as MSE, MAPE, RMSE, and WAPE. The analysis of the proposed QRNN_AWBOA based on MSE is depicted in Figure 6(a). It shows the proposed QRNN_AWBOA attained the MSE value of 0.016 at a delay of 4000, and the prevailing methods, like AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam, attained the MSE value of 0.126, 0.098, 0.057, 0.052, 0.047, 0.037, and 0.027. Figure 6(b) presents the analysis of the proposed QRNN_AWBOA based on MAPE. The MAPE value achieved by the QRNN_AWBOA is 0.056 at a delay of 4000, while the existing methods, AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam, produced MAPE values of 0.265, 0.208, 0.146, 0.136, 0.117, 0.101, and 0.099, respectively. The RMSE-based analysis of the proposed QRNN_AWBOA is depicted in Figure 6(c). At delay 4000, the prevailing techniques AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam obtained the RMSE values of 0.356, 0.314, 0.240, 0.228, 0.216, 0.192, and 0.163, respectively, while the QRNN_AWBOA attained a lower RMSE value of 0.128. Figure 6(d) illustrates the analysis of QRNN_AWBOA based on WAPE. The QRNN_AWBOA obtained a WAPE value of 0.057 at a delay of 4000. The WAPE values obtained by the existing methods AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam are 0.217, 0.147, 0.117, 0.110, 0.110, 0.099, and 0.078, respectively.

Figure 6.

Comparative analysis of the proposed QRNN_AWBOA at setup LPG-Concentration on 2022-03-09, considering (a) MSE, (b) MAPE, (c) RMSE, and (d) WAPE.

5.6.2 Analysis based on LPG-concentration on 2022-03-10

Figure 7 depicts the comparative evaluation of the QRNN_AWBOA and the prevailing methods based on MSE, MAPE, RMSE, and WAPE, conducted at the LPG-Concentration setup on 2022-03-10. The MSE-based evaluation of QRNN_AWBOA is presented in Figure 7(a). At a delay of 4000, the QRNN_AWBOA achieves an MSE value of 0.047, while the comparative methods AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam produce MSE values of 0.208, 0.167, 0.108, 0.098, 0.065, 0.060, and 0.054, respectively. Figure 7(b) illustrates the MAPE-based evaluation of QRNN_AWBOA with respect to the comparative approaches, like AAC, AI-enabled IoT framework, IoT/M2M, CNN, NDAMA, BiLSTM, and QRNN + Adam. The QRNN_AWBOA attained the MAPE value of 0.087 at a delay of 4000, whereas the comparative approaches obtained the MAPE values of 0.287, 0.227, 0.176, 0.159, 0.127, 0.101, and 0.099, respectively, for AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam. The RMSE-based analysis of the QRNN_AWBOA is illustrated in Figure 7(c). The QRNN_AWBOA achieved an RMSE value of 0.218, while the comparative methods AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam attained RMSE values of 0.456, 0.409, 0.329, 0.313, 0.256, 0.245, and 0.232 at a delay of 4000. Finally, Figure 7(d) presents the WAPE-based evaluation of QRNN_AWBOA. The QRNN_AWBOA achieved a WAPE value of 0.065 at a delay of 4000, while the comparative methods AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam attained WAPE values of 0.227, 0.176, 0.136, 0.101, 0.098, 0.084, and 0.074, respectively.

Figure 7.

Comparative analysis of the proposed QRNN_AWBOA at setup LPG-Concentration on 2022-03-10 based on (a) MSE, (b) MAPE, (c) RMSE, and (d) WAPE.

5.6.3 Analysis based on LPG-concentration on 2022-03-11

The comparative evaluation of the QRNN_AWBOA against other existing methods, including AAC, AI-enabled IoT framework, IoT/M2M, NDAMA, 1D-CNN, BiLSTM, and QRNN + Adam in the context of the LPG concentration setup on 2022-03-11, is shown in Figure 8. The analysis of QRNN_AWBOA based on MSE, as depicted in Figure 8(a), reveals that the proposed QRNN_AWBOA achieves an MSE value of 0.057 at a delay of 4000. In contrast, the MSE values for the other methods are 0.217 for AAC, 0.176 for the AI-enabled IoT framework, 0.136 for IoT/M2M, 0.087 for NDAMA, 0.109 for 1D-CNN, 0.080 for BiLSTM, and 0.078 for QRNN + Adam. The comparative assessment of QRNN_AWBOA based on MAPE is shown in Figure 8(b). The QRNN_AWBOA achieved a MAPE value of 0.1087 at a delay of 4000, whereas the MAPE value figured by the other methods, like AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam, is 0.298, 0.236, 0.208, 0.158, 0.137, 0.128, and 0.121, respectively. Figure 8(c) illustrates the evaluation of QRNN_AWBOA based on RMSE. At a delay of 4000, the QRNN_AWBOA achieved an RMSE value of 0.240. The other methods, namely AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam, measured higher RMSE values of 0.466, 0.420, 0.369, 0.330, 0.296, 0.283, and 0.279, respectively. Furthermore, the comparative analysis of QRNN_AWBOA based on WAPE is presented in Figure 8(d). The QRNN_AWBOA attained a WAPE value of 0.097 at a delay of 4000, while the other methods, namely AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam, achieved WAPE values of 0.287, 0.246, 0.165, 0.158, 0.137, 0.129, and 0.110, respectively, which is higher than the value figured by the proposed technique.

Figure 8.

Comparative analysis of the proposed QRNN_AWBOA at setup LPG-Concentration on 2022-03-11 with respect to (a) MSE, (b) MAPE, (c) RMSE, and (d) WAPE.

5.7 Algorithmic analysis

Figure 9 demonstrates an algorithmic analysis, in which the proposed QRNN_AWBOA is compared with QRNN_Adam³¹ and QRNN_SGD.³² Figure 9(a) shows the evaluation results based on LPG-Concentration on 2022-03-09. For 200 iterations, the MSE attained by the QRNN_Adam, QRNN_SGD, and QRNN_AWBOA is 0.054, 0.038, and 0.017. Figure 9(b) depicts the assessment based on LPG-Concentration on 2022-03-10. The MSE of the QRNN_AWBOA is 0.048, while the MSE obtained by the QRNN_Adam and QRNN_SGD is 0.099 and 0.066. Figure 9(c) portrays the valuation with respect to LPG-Concentration on 2022-03-11. The proposed QRNN_AWBOA and the compared methods, such as QRNN_Adam and QRNN_SGD, achieved an MSE of 0.058, 0.088, and 0.064. These results demonstrate that QRNN_AWBOA outperforms the existing optimization approaches in minimizing the MSE for LPG leakage detection.

Figure 9.

Algorithmic analysis of the proposed QRNN_AWBOA based on, (a) LPG-Concentration on 2022-03-09, (b) LPG-Concentration on 2022-03-10, and (c) LPG-Concentration on 2022-03-11.

5.8 Analysis based on feature extraction

An analysis based on feature extraction by varying the delay is depicted in Figure 10. In this analysis, the proposed QRNN_AWBOA with SMA, TRIX, and WMA features is compared with QRNN_AWBOA without feature extraction and QRNN_AWBOA with domain features. Figure 10(a) specifies the results using LPG-Concentration on 2022-03-09. When the delay = 4000, the MSE obtained by the QRNN_AWBOA with SMA, TRIX, WMA features is 0.017, whereas the QRNN_AWBOA without feature extraction and QRNN_AWBOA with domain features obtained an MSE of 0.026 and 0.020. Figure 10(b) portrays the evaluation based on LPG-Concentration on 2022-03-10. The QRNN_AWBOA with SMA, TRIX, and WMA features, QRNN_AWBOA without feature extraction and QRNN_AWBOA with domain features achieved an MSE of 0.048, 0.100, and 0.067. Figure 10(c) shows the valuation with respect to LPG-Concentration on 2022-03-11. The MSE gained by the QRNN_AWBOA without feature extraction, QRNN_AWBOA with domain features, and QRNN_AWBOA with SMA, TRIX, and WMA features is 0.110, 0.078, and 0.058. These results indicate that using SMA, TRIX, and WMA features significantly improves the accuracy across different delay settings.

Figure 10.

Feature extraction based analysis of the QRNN_AWBOA based on, (a) LPG-Concentration on 2022-03-09, (b) LPG-Concentration 0n 2022-03-10, and (c) LPG-Concentration on 2022-03-11.

5.9 Analysis based on feature fusion

Figure 11 portrays a feature fusion-based analysis by evaluating the concatenation used by the QRNN_AWBOA with weighted fusion. Figure 11(a) establishes the valuation concerning LPG-Concentration on 2022-03-09. The MSE attained by concatenation is 0.017, while the MSE of the weighted fusion is 0.037 for a delay of 4000. In Figure 11(b), the estimation based on LPG-Concentration on 2022-03-10 is depicted. The MSE gained by the weighted fusion and concatenation are 0.067 and 0.048. Figure 11(c) demonstrated the assessment with respect to LPG-Concentration on 2022-03-11. The MSE achieved by the concatenation is 0.058, and the weighted fusion is 0.067. These results demonstrate that concatenation consistently outperforms weighted fusion in minimizing prediction error across different delay settings.

Figure 11.

Feature fusion based analysis of the QRNN_AWBOA based on, (a) LPG-Concentration on 2022-03-09, (b) LPG-Concentration 0n 2022-03-10, and (c) LPG-Concentration on 2022-03-11.

5.10 Analysis of computational time, inference time, end-to-end alert latency, and resource utilization

Table 2 presents an analysis of the computational time, Inference Time, End-to-End Alert Latency, and Resource usage of the devised QRNN_AWBOA and the compared methods, namely AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam. He evaluation is performed based on LPG concentration measurements on 2022-03-09 to 2022-03-11. The table shows that the proposed QRNN_AWBOA consistently outperforms all other methods across all metrics. For instance, the LPG-Concentration on 2022-03-09, QRNN_AWBOA achieves a computational time of 10.887 s, an inference time of 2.958 s, an end-to-end alert latency of 0.589 s, and a resource utilization of 92.998%, which are significantly better than the compared approaches. These results demonstrate that the proposed QRNN_AWBOA provides faster computations, lower latency, and efficient resource utilization, making it highly effective for LPG leakage detection.

Table 2.
Computational time, inference time, end-to-end alert latency, and resource utilization of the proposed QRNN_AWBOA and the compared methods.

AAC AI-enabled IoT framework IoT/M2M 1D-CNN NDAMA BiLSTM QRNN + Adam Proposed QRNN_ AWBOA

Computational time (sec)

LPG-Concentration on 2022-03-09 38.999 33.988 31.000 28.988 25.988 19.888 15.988 10.887

LPG-Concentration on 2022-03-10 43.999 39.999 35.999 32.990 29.990 23.999 17.989 14.888

LPG-Concentration on 2022-03-11 36.880 32.998 28.999 24.999 20.888 19.888 14.999 12.888

Inference time (sec)

LPG-Concentration on 2022-03-09 10.598 9.236 8.424 7.877 7.062 5.404 4.345 2.958

LPG-Concentration on 2022-03-10 11.956 10.869 9.782 8.965 8.149 6.521 4.888 4.046

LPG-Concentration on 2022-03-11 10.022 8.967 7.880 6.793 5.676 5.404 4.076 3.502

End-to-end alert latency (sec)

LPG-Concentration on 2022-03-09 3.699 2.999 2.479 1.989 1.690 1.379 0.999 0.589

LPG-Concentration on 2022-03-10 4.009 3.879 3.368 2.899 2.379 1.988 1.378 0.989

LPG-Concentration on 2022-03-11 2.889 2.468 1.998 1.689 1.479 1.279 0.888 0.779

Resource usage (%)

LPG-Concentration on 2022-03-09 80.998 82.988 84.999 86.999 88.990 89.989 90.998 92.998

LPG-Concentration on 2022-03-10 83.999 86.999 87.990 89.998 90.999 91.887 92.888 94.887

LPG-Concentration on 2022-03-11 86.990 87.990 89.988 90.879 92.888 93.999 94.999 96.777

	AAC	AI-enabled IoT framework	IoT/M2M	1D-CNN	NDAMA	BiLSTM	QRNN + Adam	Proposed QRNN_ AWBOA
Computational time (sec)
LPG-Concentration on 2022-03-09	38.999	33.988	31.000	28.988	25.988	19.888	15.988	10.887
LPG-Concentration on 2022-03-10	43.999	39.999	35.999	32.990	29.990	23.999	17.989	14.888
LPG-Concentration on 2022-03-11	36.880	32.998	28.999	24.999	20.888	19.888	14.999	12.888
Inference time (sec)
LPG-Concentration on 2022-03-09	10.598	9.236	8.424	7.877	7.062	5.404	4.345	2.958
LPG-Concentration on 2022-03-10	11.956	10.869	9.782	8.965	8.149	6.521	4.888	4.046
LPG-Concentration on 2022-03-11	10.022	8.967	7.880	6.793	5.676	5.404	4.076	3.502
End-to-end alert latency (sec)
LPG-Concentration on 2022-03-09	3.699	2.999	2.479	1.989	1.690	1.379	0.999	0.589
LPG-Concentration on 2022-03-10	4.009	3.879	3.368	2.899	2.379	1.988	1.378	0.989
LPG-Concentration on 2022-03-11	2.889	2.468	1.998	1.689	1.479	1.279	0.888	0.779
Resource usage (%)
LPG-Concentration on 2022-03-09	80.998	82.988	84.999	86.999	88.990	89.989	90.998	92.998
LPG-Concentration on 2022-03-10	83.999	86.999	87.990	89.998	90.999	91.887	92.888	94.887
LPG-Concentration on 2022-03-11	86.990	87.990	89.988	90.879	92.888	93.999	94.999	96.777

5.11 Analysis based on detection

Figure 12 illustrates the detection analysis of the devised QRNN_AWBOA model based on various metrics by varying the training data. Figure 12(a) shows the precision-based evaluation. The precision of the proposed method and the existing models, namely AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam, is 96.347% 89.888%, 90.888%, 91.879%, 92.887%, 93.989%, 94.777%, and 95.877%, and for 90% of training data. Figure 12(b) depicts the assessment based on recall. The QRNN_AWBOA obtained a recall of 97.378%, whereas the compared methods, like AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam, achieved a recall of 87.888%, 89.887%, 90.777%, 91.666%, 93.000%, 94.888%, and 95.555%. Figure 12(c) demonstrates the F1-score results. The proposed QRNN_AWBOA and the previous methods achieved an F1-score of 88.877%, 90.385%, 91.325%, 92.272%, 93.492%, 94.832%, 95.716%, and 96.860%. In Figure 12(d) the assessment with respect to false alarm rate is portrayed. The false alarm rate of the devised QRNN_AWBOA is 2.268%, whereas the compared methods, including AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam, gained a false alarm rate of 10.887%, 8.689%, 7.579%, 6.179%, 5.778%, 3.776%, and 3.479%. Figure 12(e) presents the valuation based on miss rate. The proposed method and the prior methods achieved a miss rate of 5.112%, 13.256%, 11.866%, 11.368%, 10.257%, 9.327%, 8.369%, and 8.091%. Figure 12(f) demonstrates the estimation concerning detection latency. The detection latency gained by the developed QRNN_AWBOA and the AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam is 0.369 s, 0.738 s, 0.700 s, 0.659 s, 0.598 s, 0.548 s, 0.480 s, and 0.399 s Figure 12(g) signifies the evaluation with respect to ROC_AUC. When the False Positive Rate (FPR) is 80%, the True Positive Rate (TPR) attained by the proposed model is 97.099%, while the TPR of the compared methods, such as AAC, AI-enabled IoT framework, IoT/M2M, 1D-CNN, NDAMA, BiLSTM, and QRNN + Adam, is 88.988%, 89.888%, 91.999%, 93.000%, 93.888%, 94.888%, and 96.889%. These findings collectively confirm that the QRNN_AWBOA model offers superior detection performance, minimal latency, and strong generalization across varying training conditions compared to existing intrusion detection techniques.

Figure 12.

Detection analysis of the proposed QRNN_AWBOA, (a) precision, (b) recall, (c) F1-score, (d) false alarm rate, (e) miss rate, (f) detection latency, and (g) ROC_AUC.

5.12 Analysis based on mean

\pm

confidence interval (CI)

Table 3 presents the mean values and corresponding 95% CIs of the devised QRNN_AWBOA framework for key performance metrics MSE, MAPE, RMSE, and WAPE. The MSE obtained a mean of 0.073 with a 0.05 CI lower and 0.10 CI upper. Similarly, the MAPE and WAPE recorded mean values and CI values of 0.136 [0.10, 0.18] and 0.123 [0.09, 0.16], respectively. Furthermore, the RMSE achieved mean and CI values of 0.270 [0.22, 0.32]. Overall, the narrow CI ranges across all metrics demonstrate that the proposed QRNN_AWBOA achieves reliable and statistically consistent performance.

Table 3.
Mean $\pm$ CI of the proposed QRNN_AWBOA.

Metric Mean 95% CI Lower 95% CI Upper Formatted

MSE 0.073 0.05 0.10 0.073 [0.05, 0.10]

MAPE 0.136 0.10 0.18 0.136 [0.10, 0.18]

RMSE 0.270 0.22 0.32 0.270 [0.22, 0.32]

WAPE 0.123 0.09 0.16 0.123 [0.09, 0.16]

Metric	Mean	95% CI Lower	95% CI Upper	Formatted
MSE	0.073	0.05	0.10	0.073 [0.05, 0.10]
MAPE	0.136	0.10	0.18	0.136 [0.10, 0.18]
RMSE	0.270	0.22	0.32	0.270 [0.22, 0.32]
WAPE	0.123	0.09	0.16	0.123 [0.09, 0.16]

5.13 Analysis of model's prediction vs. ground truth and alert threshold

Figure 13 represents the timeline plot of the QRNN_AWBOA in detecting LPG leakage. The Y-axis signifies the LPG leakage, while the X-axis designates the time period from March 10th to March 11th, 2022. The graph shows that the Ground Truth (actual leakage) suddenly increases after some time. The model's Prediction tracks this rise rapidly, resulting in an Alert. The predicted leakage value immediately exceeds the threshold of 800,000 and remains consistently high, demonstrating that the model successfully provided a fast and timely alert for a significant leakage event.

Figure 13.

Timelines showing model predictions vs ground truth with the alert threshold.

5.14 Analysis based on calibration plot

Figure 14 depicts the calibration plot of the devised QRNN_AWBOA. This plot shows the relationship between the predicted probabilities generated by a model and the true observed frequencies of the event. This graph is plotted against the predicted probability and the observed frequency. The dashed diagonal line signifies perfect calibration, while the blue line denotes the calibration of the QRNN_AWBOA model. The results designate that the model is well-calibrated, particularly at the lower (<0.3) and higher (0.9) probability ranges, representing strong reliability between predicted and observed outcomes.

Figure 14.

Calibration plot of the QRNN_AWBOA.

5.15 Analysis based on convergence vs standard deviation

Figure 15 illustrates the training and testing loss of the QRNN_AWBOA by varying the epochs, highlighting the convergence behavior and associated standard deviation. Initially, when the epoch is 0, the training and testing losses of the devised model are 0.645 and 0.689, respectively. As the number of epochs increases to 50, the training and testing losses decrease significantly to 0.078 and 0.116, demonstrating that the model converges efficiently while maintaining low variability. This analysis demonstrates the stability and robustness of the proposed model during the training and testing process.

Figure 15.

Convergence vs standard deviation.

5.16 Comparative discussion

The comparative analysis of the proposed QRNN_AWBOA for LPG gas leakage detection, alongside existing detection methods, such as AAC, AI-enabled IoT framework, IoT/M2M, NDAMA, 1D-CNN, BiLSTM, and QRNN + Adam, based on various LPG concentration setups is presented in Table 4. The proposed QRNN_AWBOA achieved MSE of 0.016, MAPE of 0.056, RMSE of 0.128, and WAPE of 0.057 at the LPG concentration setup on 2022-03-09. In comparison, the AAC method yielded the MSE of 0.126, MAPE of 0.265, RMSE of 0.356, and WAPE of 0.217. The AI-enabled IoT framework produced MSE of 0.098, MAPE of 0.208, RMSE of 0.314, and WAPE of 0.147, IoT/M2M achieved MSE of 0.057, MAPE of 0.146, RMSE of 0.240, and WAPE of 0.117, NDAMA showed MSE of 0.046, MAPE of 0.116, RMSE of 0.216, and WAPE of 0.109, 1D-CNN gained MSE of 0.052, MAPE of 0.136, RMSE of 0.228, and WAPE of 0.110, BiLSTM achieved MSE of 0.037, MAPE of 0.101, RMSE of 0.192, and WAPE of 0.099, and QRNN + Adam obtained MSE of 0.027, MAPE of 0.099, RMSE of 0.163, and WAPE of 0.078. These results demonstrate that the proposed QRNN_AWBOA is more efficient than the prevailing gas leakage detection systems, as it consistently achieved the lowest error rates for LPG gas leakage detection across all evaluation metrics. The QRNN effectively modeled the complex, nonlinear relationships in the data, resulting in reduced error and enhanced generalizability. The AWBOA enhanced convergence rate and robustness, thus the QRNN_AWBOA achieved effective results.

Table 4.
Comparative discussion.

Methods

Variations Metrics AAC AI-enabled IoT framework IoT/M2M 1D-CNN NDAMA BiLSTM QRNN + Adam Proposed QRNN_ AWBOA

LPG-Concentration on 2022-03-09 MSE 0.126 0.098 0.057 0.052 0.047 0.037 0.027 0.016

MAPE 0.265 0.208 0.146 0.136 0.117 0.101 0.099 0.056

RMSE 0.356 0.314 0.240 0.228 0.216 0.192 0.163 0.128

WAPE 0.217 0.147 0.117 0.110 0.110 0.099 0.078 0.057

LPG-Concentration 0n 2022-03-10 MSE 0.208 0.167 0.108 0.098 0.065 0.060 0.054 0.047

MAPE 0.287 0.227 0.176 0.159 0.127 0.101 0.099 0.087

RMSE 0.456 0.409 0.329 0.313 0.256 0.245 0.232 0.218

WAPE 0.227 0.176 0.136 0.101 0.098 0.084 0.074 0.065

LPG-Concentration on 2022-03-11 MSE 0.217 0.176 0.136 0.109 0.088 0.080 0.078 0.057

MAPE 0.298 0.236 0.208 0.158 0.137 0.128 0.121 0.108

RMSE 0.466 0.420 0.369 0.330 0.296 0.283 0.279 0.240

WAPE 0.287 0.246 0.165 0.158 0.137 0.129 0.110 0.097

		Methods
LPG-Concentration on 2022-03-09	MSE	0.126	0.098	0.057	0.052	0.047	0.037	0.027	0.016
MAPE	0.265	0.208	0.146	0.136	0.117	0.101	0.099	0.056
RMSE	0.356	0.314	0.240	0.228	0.216	0.192	0.163	0.128
WAPE	0.217	0.147	0.117	0.110	0.110	0.099	0.078	0.057
LPG-Concentration 0n 2022-03-10	MSE	0.208	0.167	0.108	0.098	0.065	0.060	0.054	0.047
MAPE	0.287	0.227	0.176	0.159	0.127	0.101	0.099	0.087
RMSE	0.456	0.409	0.329	0.313	0.256	0.245	0.232	0.218
WAPE	0.227	0.176	0.136	0.101	0.098	0.084	0.074	0.065
LPG-Concentration on 2022-03-11	MSE	0.217	0.176	0.136	0.109	0.088	0.080	0.078	0.057
MAPE	0.298	0.236	0.208	0.158	0.137	0.128	0.121	0.108
RMSE	0.466	0.420	0.369	0.330	0.296	0.283	0.279	0.240
WAPE	0.287	0.246	0.165	0.158	0.137	0.129	0.110	0.097

Bold values represent the best performance.

6 Conclusion

With the development of more advanced technologies, an early gas leakage detection system is crucial for ensuring safety in various environments. In this article, a DL model with an IoT edge computing system named QRNN_AWBOA is proposed for LPG gas leakage detection and alert. In this system, the gas leakage detection data is normalized using median normalization, then the normalized data is fused using a DNN model with Neyman similarity. Then, the fused features are augmented using the oversampling technique. Then, augmented data is fed into the QRNN model for the detection process. The developed QRNN model is efficiently trained by a hybrid optimization algorithm named AWBOA. This effectively reduces the error percentage and improves the generalizability. The effectiveness of the proposed QRNN_AWBOA is validated based on MSE, MAPE, RMSE, and WAPE. The QRNN_AWBOA attained an MSE of 0.016, MAPE of 0.056, RMSE of 0.128, and WAPE of 0.057, which illustrates the superiority of the detection model. In the future, a hybrid DL method will be utilized for more enhanced prediction results and to further minimize the error rate.

Footnotes

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

ORCID iDs

Pravin Narayan Kathavate

Santosh Ashokrao Darade

Jayashree Pradip Tamkhade

References

Khan

Ahmed

Hakak

, et al. Edge computing: a survey. Future Gener Comput Syst 2019; 97: 219–235.

Wei

Wang

Xia

, et al. Edge computing offloading method based on deep reinforcement learning for gas pipeline leak detection. Mathematics 2022; 10: 4812.

Maltezos

Petousakis

Dadoukis

, et al. A smart building fire and gas leakage alert system with edge computing and NG112 emergency call capabilities. Information 2022; 13: 164.

Shi

Dustdar

. The promise of edge computing. Computer (Long Beach Calif) 2016; 49: 78–81.

Rana

Singh

Vijaya

. Recent trust management models for secure IoT ecosystem. Int J Intell Syst Appl Eng 2022; 10: 23–33.

Dash

Bandopadhay

Samal

, et al. AI-enabled IoT framework for leakage detection and its consequence prediction during external transportation of LPG. Sensors 2023; 23: 6473.

Chen

Chiou

. An early fire-detection method based on image processing. In: Proceedings of 2004 international conference on image processing, 2004. ICIP'04, vol. 3, October 2004, pp.1707–1710: IEEE.

Salameh

HAB

Dhainat

Benkhelifa

. An end-to-end early warning system based on wireless sensor network for gas leakage detection in industrial facilities. IEEE Syst J 2020; 15: 5135–5143.

Morganti

Foong

Brear

, et al. The research and motor octane numbers of liquefied petroleum gas (LPG). Fuel 2013; 108: 797–811.

10.

Babu

Nair

Kishore

, et al. Enhancing gas leak detection with IoT technology: an innovative approach. Procedia Comput Sci 2024; 235: 961–969.

11.

Dashdondov

Kim

. NDAMA: a novel deep autoencoder and multivariate analysis approach for IOT-based methane gas leakage detection. IEEE Access 2023; 11: 140740–140751.

12.

Zuo

Liang

, et al. A semi-supervised leakage detection method driven by multivariate time series for natural gas gathering pipeline. Process Saf Environ Prot 2022; 164: 468–478.

13.

Kopbayev

Khan

Yang

, et al. Gas leakage detection using spatial and temporal neural network model. Process Saf Environ Prot 2022; 160: 968–975.

14.

Cao

Ding

Wang

, et al. Enhancing physical-layer security for IoT with nonorthogonal multiple access assisted semi-grant-free transmission. IEEE Internet Things J 2022; 9: 24669–24681.

15.

Cheng

Wang

Zhao

, et al. Situation-aware dynamic service coordination in an IoT environment. IEEE/ACM Trans Networking 2017; 25: 2082–2095.

16.

Spandonidis

Theodoropoulos

Giannopoulos

. A combined semi-supervised deep learning method for oil leak detection in pipelines using IIoT at the edge. Sensors 2022; 22: 4105.

17.

Xiong

Jiang

Pan

, et al. Deep learning approach for detection of underground natural gas micro-leakage using infrared thermal images. Sensors 2022; 22: 5322.

18.

Djehaiche

Aidel

Sawalmeh

, et al. Adaptive control of IoT/m2m devices in smart buildings using heterogeneous wireless networks. IEEE Sensors J 2023; 23: 7836–7849.

19.

Praveenchandar

Vetrithangam

Kaliappan

, et al. Iot-based harmful toxic gases monitoring and fault detection on the sensor dataset using deep learning techniques. Sci Program 2022; 2022: 7516328.

20.

Lima

Souza

. A large comparison of normalization methods on time series. Big Data Res 2023; 34: 100407.

21.

Sun

Huang

Kroening

, et al. Testing deep neural networks. arXiv preprint arXiv:1803.04792. 2018.

22.

Cha

. Comprehensive survey on distance/similarity measures between probability density functions. City 2007; 1: 1.

23.

Bradbury

Merity

Xiong

, et al. Quasi-recurrent neural networks. arXiv preprint arXiv:1611.01576. 2016.

24.

Azizi

Shishehgarkhaneh

Basiri

, et al. Wolf-Bird Optimizer (WBO): a novel metaheuristic algorithm for Building Information Modeling-based resource tradeoff. J Eng Res 2023; 13: 763–785.

25.

Hamadneh

Kaabneh

Alssayed

, et al. Addax optimization algorithm: a novel nature-inspired optimizer for solving engineering applications. Int J Intell Eng Syst 2024; 17: 732–743.

26.

The technical indicators are taken from, https://library.tradingtechnologies.com/trade/chrt-technical-indicators.html (accessed October 2024).

27.

Fire, smoke and gas leakage detection Dataset is taken from, https://zenodo.org/records/6616632 (accessed August 2024).

28.

Song

Yan

, et al. Research on oil and gas pipeline leak detection method based on 1DCNN-DBO-LSTM. Meas Sci Technol 2024; 36: 016024.

29.

Yang

Zhao

. A BiLSTM based pipeline leak detection and disturbance assisted localization method. IEEE Sensors J 2021; 22: 611–620.

30.

Wang

Han

, et al. Quantum recurrent neural networks for sequential learning. Neural Netw 2023; 166: 148–161.

31.

Bock

Goppold

Weiß

. An improvement of the convergence proof of the ADAM-Optimizer. arXiv preprint arXiv:1804.10587. 2018.

32.

Newton

Yousefian

Pasupathy

. Stochastic gradient descent: recent trendsRecent advances in optimization and modeling of contemporary problems. Catonsville, MD: INFORMS, 2018, pp.193–220.

		Methods
Variations	Metrics	AAC	AI-enabled IoT framework	IoT/M2M	1D-CNN	NDAMA	BiLSTM	QRNN + Adam	Proposed QRNN_ AWBOA
LPG-Concentration on 2022-03-09	MSE	0.126	0.098	0.057	0.052	0.047	0.037	0.027	0.016
	MAPE	0.265	0.208	0.146	0.136	0.117	0.101	0.099	0.056
	RMSE	0.356	0.314	0.240	0.228	0.216	0.192	0.163	0.128
	WAPE	0.217	0.147	0.117	0.110	0.110	0.099	0.078	0.057
LPG-Concentration 0n 2022-03-10	MSE	0.208	0.167	0.108	0.098	0.065	0.060	0.054	0.047
	MAPE	0.287	0.227	0.176	0.159	0.127	0.101	0.099	0.087
	RMSE	0.456	0.409	0.329	0.313	0.256	0.245	0.232	0.218
	WAPE	0.227	0.176	0.136	0.101	0.098	0.084	0.074	0.065
LPG-Concentration on 2022-03-11	MSE	0.217	0.176	0.136	0.109	0.088	0.080	0.078	0.057
	MAPE	0.298	0.236	0.208	0.158	0.137	0.128	0.121	0.108
	RMSE	0.466	0.420	0.369	0.330	0.296	0.283	0.279	0.240
	WAPE	0.287	0.246	0.165	0.158	0.137	0.129	0.110	0.097

QRNN with Addax Wolf Bird Optimization-based LPG gas leakage detection and alert system with IoT edge computing

Abstract

Keywords

1 Introduction

2 Motivation

2.1 Literature review

2.2 Challenges

3 System model

4.2.1 SMA

4.6 LPG gas leakage detection

4.6.1 Architecture of QRNN

5.1 Experimental set-up

Table 1. Parameters and values. Parameters Values QRNN Epoch 50 Batch size 32 Optimizer AWBOA Kernel size (3,3) Dropout 0.3 Loss mse Activation relu AWBOA N(population size) 50 M(Decision Variable) 30 Maximum Iteration 200 r random numbers between (0,1) I random number of (1 or 2)

5.3 Evaluation measures

5.6 Comparative assessment

5.6.1 Analysis based on LPG-concentration on 2022-03-09

Table 3. Mean ± CI of the proposed QRNN_AWBOA. Metric Mean 95% CI Lower 95% CI Upper Formatted MSE 0.073 0.05 0.10 0.073 [0.05, 0.10] MAPE 0.136 0.10 0.18 0.136 [0.10, 0.18] RMSE 0.270 0.22 0.32 0.270 [0.22, 0.32] WAPE 0.123 0.09 0.16 0.123 [0.09, 0.16]

Footnotes

Funding

Declaration of conflicting interests

ORCID iDs

References

Table 1.
Parameters and values.

Parameters Values

QRNN

Epoch 50

Batch size 32

Optimizer AWBOA

Kernel size (3,3)

Dropout 0.3

Loss mse

Activation relu

AWBOA

N(population size) 50

M(Decision Variable) 30

Maximum Iteration 200

r random numbers between (0,1)

I random number of (1 or 2)

Table 3.
Mean $\pm$ CI of the proposed QRNN_AWBOA.

Metric Mean 95% CI Lower 95% CI Upper Formatted

MSE 0.073 0.05 0.10 0.073 [0.05, 0.10]

MAPE 0.136 0.10 0.18 0.136 [0.10, 0.18]

RMSE 0.270 0.22 0.32 0.270 [0.22, 0.32]

WAPE 0.123 0.09 0.16 0.123 [0.09, 0.16]