Fault diagnosis of elevator systems based on multidomain feature extraction and SHAP feature selection

Abstract

In smart buildings, elevator faults may affect the traveling efficiency and the safety of passengers. It is important to find a quick and accurate method for fault diagnosis to minimize disruption time. Most existing fault diagnosis methods are for gear systems, which ignore the characteristics of elevators and need adaptations before applying them to the elevators. This study presents a novel multidomain feature extraction method for elevator fault diagnosis. It calculates the time domain features and wavelet packet energy and then extracts the elevator frequency features. These multidomain features are utilized as the input of a support vector machine to diagnose elevator faults. Subsequently, SHapley Additive exPlanations is used to select features to remove redundant features without reducing diagnosis accuracy. Comparative experiments show that multidomain feature extraction based on triaxial data improves diagnosis accuracy. Additionally, the elevator frequency feature extraction method further enhances diagnosis accuracy. In addition, experiments demonstrate that SHapley Additive exPlanations feature selection can be effectively applied to remove redundant features in different fault diagnoses, which improves the training efficiency and reduces the risk of overfitting.

Practical applications

In this paper, we present a practical solution for elevator fault diagnosis in intelligent buildings. By the proposed method, maintainers can find elevator faults based on objective evidence rather than intuitive feeling. We utilize a phone instead of the vulnerable and expensive equipment to collect and analyze signals, which makes the cost for this solution inexpensive and propagable. Thus, this solution can be applied to most practical elevator maintenance to reduce the elevator downtime.

Keywords

Elevators fault diagnosis feature extraction feature selection

Introduction

A smart building is a comprehensive system that integrates and applies advanced technologies for building management and operation. Through intelligent control and management, it can improve building energy efficiency, enhance occupant comfort, strengthen safety and optimize resource allocation.¹ In smart buildings, elevators, as important vertical transportation facilities, are endowed with intelligent functions. Intelligent elevator systems² can optimize elevator scheduling and operational efficiency through monitoring and data analysis, reducing waiting time and energy consumption.³

Elevators are essential for modern urban life. Their breakdowns cause inconveniences to residents.⁴ Fault diagnosis of elevators has also become a key part of this intelligent system. Through sensors, the elevator system can monitor the operating status and parameters. When an abnormality occurs, the system automatically analyzes and diagnoses the cause of the fault, and can even predict potential problems in advance. This intelligent fault diagnosis improves the reliability and safety of elevators. It ensures the travel convenience and safety of people in the buildings. Locating the cause of faults is a challenge due to the complexity of an elevator system. Conventional elevator maintenance is based on schedules, which means that maintenance staff overhaul elevators according to a predetermined maintenance plan, regardless of the actual operating conditions.⁵ By analyzing vibration signals, maintenance staff can obtain more information about the elevator, which aids in identifying and locating faults.⁶ As known, faults are primarily diagnosed based on practical experience. Maintenance technicians locate faults by using the elimination method,⁷ which involves analyzing vibration data and considering previous maintenance experience. The accuracy of fault localization depends on the experience of the maintenance technicians.

In order to solve the problem of low efficiency and poor accuracy of fault diagnosis relying only on the experience of maintenance workers, the Internet of Things (IoT) technology⁸ is leveraged for monitoring and collecting a variety of data to quickly identify and locate faults. At the same time, the application of machine learning models can further improve the intelligence of fault diagnosis. Machine learning models can learn and extract fault patterns from a large amount of historical data. Compared with the traditional method of relying on the experience of maintenance personnel, machine learning models can reduce human error, deal with complex and multi-dimensional data and provide fault analysis. Therefore, IoT technology and machine learning⁹ gradually replace traditional experience-based maintenance methods and reduce maintenance costs.

The current fault diagnosis methods are mainly based on feature extraction and machine learning, such as multilayer perceptron,¹⁰ support vector machine (SVM),¹¹ and bayesian networks.¹² In some studies, neural networks are introduced for feature extraction. To extract vibration signal features, a Smart Bearing Diagnostic System (SBDS) is proposed.¹³ In that research, convolution extracts the vibration features, and the classification is performed by the neuro-fuzzy network. Jia et al.¹⁴ introduce an elevator fault diagnosis method utilizing one-dimensional convolutional neural networks. Mishra and Hutala¹⁵ apply a deep autoencoder¹⁶ to extract features from raw sensor vibration data of elevators. The imbalance between positive and negative samples, caused by the lack of data on elevator faults, results in unsatisfactory outcomes in practical applications.

This imbalance has less impact on feature extraction through signal processing methods. Jia et al.¹⁷ process and assemble various sensor signals into feature vectors after decomposing the acceleration signal using wavelet packet decomposition and reconstruction. Wan et al.¹⁸ employ the optimal wavelet packet for extracting frequency domain features from vibration signals. To determine the fault cause, they used the resulting fault feature vector as input of least squares support vector machines,¹⁹ with parameters optimized using k-fold cross-validation.²⁰ Due to the non-stationary nature of vibration signals, Zheng and Zhao²¹ utilize wavelet packet decomposition to extract frequency domain features. Subsequently, a neural network employed these features to classify between normal and faulty samples. Mo et al.²² establish a dimensionless index sensitive to bearing fault characteristics by processing noise-reduced data, thereby accomplishing the extraction of fault features from the bearing signals.

Most of the existing fault diagnosis methods of elevators are directly migrated from bearing fault diagnosis methods and are not redesigned for elevator systems. The elevator system involves a variety of complex components. It is difficult for a bearing fault diagnosis method to cover all potential faults of the elevators. In order to improve the accuracy of elevator fault diagnosis, simply introducing a large number of features is undesirable.²³ Increasing the number of features may bring a series of negative effects. Firstly, too many features may degrade the generalization ability of the model.²⁴ In addition, an increase in the number of features can significantly extend the training time of the model. Finally, processing and computing these redundant features increase the computational burden of the system, which may lead to the waste of computational resources. Therefore, in fault diagnosis, a balance between feature selection and model complexity is required to ensure diagnostic accuracy.

To address these issues, this work analyzes the elevator structure, with a focus on extracting elevator frequency features from vibration signals. Subsequently, SVM is employed for fault diagnosis. To reduce redundant features, SHapley Additive exPlanations (SHAP)²⁵ is used to select features²⁶ for elevator fault diagnosis. The main findings of this research can be encapsulated as follows:

(1) Compared with the traditional feature extractions of vertical signals, the features based on triaxial signals keep more temporal and spatial information, which is more consistent with the practical application.

(2) Combining frequency features of the elevator model improves the accuracy of fault diagnosis. These features vary with the elevator parameters, which facilitates the migration of feature extraction and fault diagnosis algorithms to different types of elevators.

(3) Due to the load limit, the computational capacity provided by the edge devices is barely enough. We use SHAP feature selection to remove redundant features such that the diagnosis method can maintain high accuracy with less computational cost. It effectively avoids the problem of high-dimensional data and reduces the chance of overfitting.

Methodology

Data processing

Data collection

The PMT EVA-625-FD device is commonly used for elevator vibration signal acquisition. However, with limited memory capacity, it can only acquire data for 700 seconds at one time. In addition, the device is bulky, which affects its portability. In contrast, mobile phones offer significant advantages in terms of size and portability. They also have large memory capacity, which enables them to store more data. Importantly, both the mobile phone and the PMT EVA-625-FD device have the same accuracy in terms of data collection.

In our experiments, elevator vibration data are collected by an Honor 60 phone instead of an edge device in the IoT, which has a built-in triaxial accelerometer with a sampling frequency of 500 Hz. It is positioned at the center of the car ground. The collected vibration signals are triaxial, including X-axis direction, Y-axis direction and Z-axis direction. The X-axis direction is perpendicular to the car door plane, the Y-axis direction is parallel to the car door, and the Z-axis direction is perpendicular to the ground.

The decision to place the device inside the elevator car is based on several factors:

(1) Convenience and Practicality: Our primary goal is to provide a monitoring tool that is convenient and easy for maintenance personnel to use. Placing the data collection device inside the elevator car is more practical for routine maintenance operations compared to positioning it on the crosshead. Therefore, we use data collection inside the car.

(2) Noise Characteristics: Noise is a significant factor in data collection. We conducted experiments showing that noise levels at the crosshead position are more complex. As shown in the figures below, Figure 1(a) represents vibration data collected inside the car, while Figure 1(b) shows data collected simultaneously at the crosshead on the same elevator. The crosshead is more affected by wind noise and shaft disturbances, leading to more complex and variable noise.

(3) Feature Extraction: Data collected inside the car allows us to extract key features used for fault diagnosis effectively. Despite the damping effects between the car and the frame, our data analysis demonstrates that the vibration signals inside the car contain rich fault-related features. These features are successfully utilized in our diagnostic model, showing that data collected inside the car is sufficiently sensitive for fault detection.

Figure 1.

Data collected at different locations. (a) Data collected inside the car. (b) Data collected on the crosshead.

For convenience and practicality, placing the equipment inside the car is more suitable for daily maintenance operations. At the same time, most of the elevator operation indicators are also based on the situation in the car. Therefore, equipment placed in the car for fault diagnosis has a reference for maintenance work. Secondly, the noise outside the car is more complicated. Inside the car, vibration data is more stable and has less noise interference. Finally, despite the presence of rubber between the car and the frame, the vibration signal inside the car still contains fault-related features, which can effectively support fault diagnosis. Therefore, we choose to collect data inside the car to ensure signal quality and improve diagnostic efficiency at this time.

Data preprocessing

Before feature extraction, some processes are executed to ensure data quality and usability.

(1) Data cleaning

To ensure that the data used for training is complete and contains all features, incomplete samples, as illustrated in Figure 2, are removed. A complete sample contains an operation cycle of the elevator, including the stationary state, accelerated state, constant speed state, decelerated state, and stop.

Figure 2.

Z-axis vibration data of complete and incomplete samples. (a) Complete sample. (b) Incomplete sample.

(2) Segmentation of Operational stages

The vibration characteristics of the elevator during the acceleration and deceleration stages differ from those during the constant velocity stage. Therefore, each stage requires separate feature extraction. To streamline subsequent feature extraction calculations, the entire elevator operation is divided into three stages based on acceleration: acceleration stage, constant velocity stage, and deceleration stage.

Multidomain feature extraction

Time domain features

The vibration signal is denoted by

x = \{x_{1}, x_{2}, \dots, x_{n}\}

. The time domain features^27,21,28 for the X-axis direction, Y-axis direction, and Z-axis direction at each operational stage are shown in Table 1:

Table 1.

Time domain features.

Features	Equation
Variance	$\frac{1}{n} \sum_{i = 1}^{n} {(x_{i} - \bar{x})}^{2}$
Peak-to-peak	$\max (x) - \min (x)$
Root mean square (RMS)	$\sqrt{\frac{1}{n} \sum_{i = 1}^{n} {x_{i}}^{2}}$
Kurtosis factor	$\frac{\frac{1}{N} \sum_{i = 1}^{N} {(x_{i} - \bar{x})}^{4}}{{(\sqrt{\frac{1}{N} \sum_{i = 1}^{N} {(x_{i} - \bar{x})}^{2}})}^{4}}$
Skewness factor	$\frac{\frac{1}{N} \sum_{i = 1}^{N} {(x_{i} - \bar{x})}^{3}}{{(\sqrt{\frac{1}{N} \sum_{i = 1}^{N} {(x_{i} - \bar{x})}^{2}})}^{3}}$
Crest factor	$\frac{\max \|x\|}{\sqrt{\frac{1}{N} \sum_{i = 1}^{N} {(x_{i} - \bar{x})}^{2}}}$
Shape factor	$\frac{R M S}{\frac{1}{N} \sum_{i = 1}^{N} \|x_{i}\|}$
Margin factor	$\frac{\max \|x\|}{{(\frac{1}{N} \sum_{i = 1}^{N} \sqrt{\|x_{i}\|})}^{2}}$
Impulse factor	$\frac{\max \|x\|}{\frac{1}{N} \sum_{i = 1}^{N} \|x_{i}\|}$

Elevator frequency features

In consideration of the elevator structures and their operational characteristics, we introduce the frequency features of the elevator. The structures of the elevator²⁹ are illustrated in Figure 3. The components we focus on include the traction wheel, traction motor, guide wheel, car pulley, and rope strands. The related frequencies are:

(1) Rotational frequency of the traction wheel ( f_s):

f_{s} = \frac{V \cdot I}{π \cdot D_{s}},

(1)

where V is the velocity, I is the hanging ratio, and D_s is the diameter of the elevator traction wheel.

(2) Pulse frequency of the traction motor ( f_m):

f_{m} = \frac{f_{s} \cdot P}{2},

(2)

where P is the number of motor poles.

(3) Rotational frequency of the guide wheel ( f_bp):

f_{b p} = \frac{V \cdot I}{π \cdot D_{b p}},

(3)

where D_bp is the diameter of the guide wheel.

(4) Rotating frequency of the car pulley ( f_cp):

f_{c p} = \frac{V \cdot I}{π \cdot D_{c p}},

(4)

where D_cp is the diameter of the car pulley.

(5) Frequency of rope strands ( f_rp):

f_{r p} = \frac{V}{p r},

(5)

where pr is the distance between two strands of the rope.

Figure 3.

Elevator structures.

In the analysis of elevator fault vibration, we find that the frequency multipliers of these frequencies are also part of the fault features. Therefore, we also pay attention to the frequencies that are two and three times the values of the elevator frequency mentioned above, as depicted in Figure 4.

Figure 4.

Elevator frequency bands for feature extraction.

Wavelet packet energy

Wavelet Packet Decomposition (WPD) decomposes an original vibration signal into several packets. For each packet, its energy is used as a feature of the signal.³⁰ After the decomposition, eight packets in the third layer are obtained, denoted as $P_{j}^{3}$ , where j = 0, 1, 2, 3, 4, 5, 6, 7. The energy of each packet is calculated:

E_{3 j} = \sum_{i = 1}^{N} {(x_{i j})}^{2},

(6)

where i serves as the subscript variable, j indicates the sequential number of the packets, N denotes the length of the data, and x_ij represents the amplitude of discrete points of packet

P_{j}^{3}

. The total energy E₃ of the signal equals the cumulative energy sum, i.e.,

E_{3} = \sum_{j = 0}^{7} E_{3 j} .

(7)

After normalizing the energy, the frequency domain eigenvector X_f of the original vibration signal is obtained as follows:

X_{f} = \frac{[E_{30}, E_{31}, E_{32}, E_{33}, E_{34}, E_{35}, E_{36}, E_{37}]}{E_{3}} .

(8)

During the process of feature extraction, a total of 165 features are obtained. Among them, a total of 81 time domain features of triaxial vibration data in acceleration, deceleration, and constant velocity stages are extracted according to Table 1. At the same time, the elevator frequency of the elevator model and its frequency multipliers are calculated according to (1) – (5). Based on these frequencies, a total of 12 elevator features in the frequency are extracted. Wavelet packet energy values are extracted on each axis according to (6) – (8) at each operational stage for a total of 72 features. Normalization of the extracted features aids in expediting gradient descent, leading to faster convergence of the training network. The normalized data are subsequently utilized as input of the SVM classifier in the process of fault diagnosis.

Support vector machine

The SVM builds a classification model by transforming the initial nonlinear data space into a high dimensional space and identifying the optimal hyperplane using support vectors. SVM algorithm utilizes the nonlinear function $φ (x)$ to map the training set x into a high dimensional linear space, enabling linear modeling to estimate the decision function. The classifier plane is represented as:

w * φ (x) + b = 0,

(9)

where w is the weight vector that determines the direction and orientation of the hyperplane; function

φ (x)

is the feature mapping that projects the original input data x into a higher-dimensional feature space, enabling linear separation; symbol b is the bias that determines the offset of the hyperplane from the origin; and ∗ denotes the dot product operation. The decision function is given by

f (x) = sign (w * φ (x) + b = 0)

. This method allows a linear method to effectively solve nonlinear problems.

In the case of nonlinearity, the optimization problem for the classified hyperplane is expressed as follows:

\begin{aligned} \min_{α} \frac{1}{2} w^{T} w - C \sum_{i = 1}^{n} ξ_{i}, \\ s.t. y_{i} (w^{T} * x_{i} + b) \geq 1 - ξ_{i}, \end{aligned}

(10)

where ξ_i ≥ 0, i = 1, 2, …, n, C is a penalty parameter indicating the degree of tolerance to error, and ξ_i is a non-negative slack variable. By introducing the Lagrangian function, the problem undergoes transformation into:

\begin{aligned} \min_{α} \frac{1}{2} \sum_{i = 1}^{n} \sum_{j = 1}^{n} y_{i} y_{j} α_{i} α_{j} K (x_{i}, x_{j}) - \sum_{i = 1}^{n} α_{i}, \\ s.t. \sum_{i = 1}^{n} α_{i} y_{i} = 0,0 \leq α_{i} \leq C, i = 1,2, \dots, n, \end{aligned}

(11)

where α_i is the support vector value, y_i and y_j are the classification labels. If 0 ≤ α_i ≤ C, x_i is the support vector. The kernel function is defined as

K (x_{i}, x_{j}) = ϕ (x_{i}) * ϕ (x_{j})

SHapley Additive exPlanations

SHAP, a popular method for interpreting the relationship between input and output, provides insights into how each feature contributes to model predictions for specific instances. It is grounded in cooperative game theory and Shapley values.

Shapley value

The Shapley value, rooted in cooperative game theory principles,³¹ quantifies the contribution of each feature in the outcome. It is calculated as the average marginal contribution across all subsets of features. The Shapley value (ϕ_i) of the ith feature is expressed as follows:

ϕ_{i} = \sum_{S \subseteq N \ {i}} \frac{| S |! \cdot (n - | S | - 1)!}{n!} [f (S \cup {i}) - f (S)],

(12)

where n represents the total number of features, N denotes the set of all features, and S signifies a subset of features excluding the ith feature. The function

f (\cdot)

represents the predicted value of the model given a subset of features.

Local interpretable model-agnostic explanations

Ribeiro et al. propose Local Interpretable Model-agnostic Explanations (LIME) to interpret predictions from a single model through a local approximation around the specific prediction point.³² This method is intended to explain individual predictions, especially in cases where the model’s behavior is complex or less interpretable. LIME aims to generate simple and understandable local explanations, even when the underlying model is intricate. LIME uses the term “interpretable inputs” to denote simplified inputs x′ . It employs the mapping x = h_x (x′) to transform a binary vector of these inputs into the original input space, with different mappings h_x used for various input spaces. To find ϕ, the goal of LIME is to minimize the objective function as follows:

ξ = \underset{g \in G}{\arg \min} L (f, g, π_{x^{'}}) + Ω (g) .

(13)

The fidelity of the explanation model g (z′) to the original model $f (h_{x} (z^{'})$ relies on minimizing the loss function L across a sample set in the simplified input space.²⁵ This minimization is weighted by the local kernel $π_{x}^{'}$ .The complexity of g is penalized by the term Ω , where each weight in the linear model g corresponds to the Shapley value ϕ.

Kernel SHAP

Kernel SHAP is an extension of the SHAP framework designed to interpret complex machine learning models. It is especially useful for models that cannot be easily explained using traditional linear methods. Kernel SHAP expands SHAP values to handle models that incorporate kernels, such as SVM and XGBoost. In the case of complex models, it is not feasible to compute Shapley values exactly. Instead, Kernel SHAP employs an approximation method similar to LIME. This method involves generating a weighted dataset of perturbed instances and fitting an interpretable model (e.g., linear regression) to this dataset.

Lundberg and Lee²⁵ assign specific values (π_x′, L, Ω) to predefined forms. The process involves fitting a linear model, where each weight determined by the final linear model corresponds to the Shapley value.

Ω (g) = 0,

(14)

π_{x}^{SHAP} (z^{'}) = \frac{(n - 1)}{(\begin{matrix} n \\ |z^{'}| \end{matrix}) |z^{'}| (p - |z^{'}|)},

(15)

L (f, g, π_{x}) = \sum_{z^{'} \in Z^{'}} π_{x}^{SHAP} (z^{'}) \cdot {[f (h_{x} (z^{'})) - g (z^{'})]}^{2},

(16)

where

|z^{'}|

is the number of non-zero elements in z′.

The proposed model

The model structure for elevator fault diagnosis is shown in Figure 5. The overall process can be divided into feature extraction, feature selection, model training and retraining, and final fault diagnosis. The specific process is as follows:

Figure 5.

Fault diagnosis model.

First, the model collects multi-dimensional data during elevator operation in real time via a mobile device, and preprocesses them to clean and normalize the data in preparation for the subsequent multi-domain feature extraction. First, The model collects multi-dimensional data during elevator operation in real time via a mobile device, and preprocesses them to clean and normalize the data in preparation for the subsequent multi-domain feature extraction. In the feature extraction stage, based on the elevator operating parameters (e.g., maximum speed, suspension rate, car pulley diameter, etc.), the system extracts key features from the preprocessed data.

Next, these features are fed into the SVM classification model for initial training to generate a set of preliminary fault classification models (Model 1 to Model N). To improve the accuracy and interpretability of the models, the system performs feature selection based on SHAP. By calculating the contribution of each feature to the model output, the SHAP method selects the feature sets to remove redundant or invalid features. The model is then retrained based on these refined feature sets to further improve the accuracy of fault classification. The retrained model is saved and used for fault diagnosis. In the fault diagnosis stage, the system extracts features from the real-time elevator data and inputs them into the retrained SVM model for classification, and finally outputs diagnostic results that point out the specific type of fault of the elevator.

Experiment

Elevator parameters and experimental setup

Elevator parameters

Table 2 shows the parameters of the experimental elevator used for feature extraction. Equations (1) – (5) are calculated according to these parameters. Based on the results, the frequency features of the elevator in the vibration signals are extracted.

Table 2.

Elevator parameters.

Parameter	Symbolic	Value	Unit
Max velocity	V	1.75	m/s
Hanging ratio	I	2	—
Motor poles	P	32	—
Diameter of traction wheel	D _s	0.4	m
Diameter of guide wheel	D _bp	0.4	m
Diameter of car pulley	D _cp	0.4	m
Distance between two strands of the rope strands	pr	52	mm

Elevator fault simulation

We simulate a total of five elevator faults on an experimental elevator with six floors. These faults are unbalanced loads, uneven rail joint, uneven rope tension, surface wear of the guide shoe, and uneven surface in the car pulley rope groove.

(1) Unbalanced loads (UL)

The loads are placed on the back side of the car to simulate the condition of UL, as shown in Figure 6(a).

Figure 6.

Experimental setup. (a) Experimental setup for unbalanced loads. (b) Experimental setup for uneven rail joint. (c) Experimental setup for uneven rope tension. (d) Experimental setup for surface wear of the guide shoe. (e) Experimental setup for uneven surface in car pulley rope groove.

(2) Uneven rail joint (URJ)

The simulation of URJ is achieved by pasting a tape on the rail joint, as illustrated in Figure 6(b). The red mark indicates the area where the tape is.

(3) Uneven rope tension (URT)

The URT fault is simulated by adjusting the head springs, as depicted in Figure 6(c). The graph shows that the two head springs expand more than the other springs.

(4) Surface wear of the guide shoe (SWGS)

As the guide shoe wears, the gap between it and the rail increases. Therefore, the fault of guide shoe wear is simulated by adjusting the distance of the guide shoe from the guide rail, as shown in Figure 6(d).

(5) Uneven surface in car pulley rope groove (USCP)

The uneven surface in the car pulley rope groove is simulated by artificially pasting tapes, as shown in Figure 6(e).

Verification accuracy

Vibration data from five different faults are collected as training data. In total, the model is tested for its diagnostic effectiveness for five different faults. The model performance is assessed using both accuracy and F1 score, with the F1 score being a metric that combines precision and recall into a single measure. This metric is calculated as:

F 1 S c o r e = 2 \frac{(p r e c i s i o n * r e c a l l)}{(p r e c i s i o n + r e c a l l)},

(17)

where precision is the ratio of correct positive predictions to total positive predictions, and recall is the ratio of correct positive predictions to total actual positives.

Table 3 presents the results of accuracy verification experiments for different types of faults. The diagnosis of SWGS achieves perfect performance, with both accuracy and F1 score at 100%. The evaluation of detecting an uneven surface in the car pulley rope groove yielded excellent results, achieving an accuracy rate of 96.67% and an F1 score of 96.43%. Tests on UL and URJ also showed good results with 93.33% and 96.67% accuracy and F1 scores of 94.87% and 96.77% respectively. The results for URT is comparatively lower, with the accuracy of 90% and the F1 score of 89.29%. Overall, the model shows high effectiveness and reliability in diagnosing various faults.

Table 3.

Results of accuracy verification experiments.

Fault	Fault data	Accuracy (%)	F1 score (%)
UL	101	93.33	94.87
URJ	109	96.67	96.77
URT	98	90	89.29
SWGS	93	100	100
USCP	102	96.67	96.43

Feature extraction based on single-axis data with triaxial data

As shown in Figure 7, the triaxial vibration feature extraction method generally outperforms feature extraction based on single-axis data. One reason is that the elevator exhibits vibration characteristics in different axes when different faults occur. Another reason is that the vibration characteristics of some faults are not apparent in a single axis.

Figure 7.

Feature extraction based on single-axis data and triaxial data.

The vibration characteristics of the elevator are manifested in different axes when different faults occur. The uneven joints of the guideway increase the magnitude of the front and rear swaying of the car, which is in the X-axis direction. Therefore, the result based on vibration data in the X-axis is much better than the result obtained by the vibration data in the other two axes. Note that both feature extraction methods based on only the Y-axis data and triaxial feature extraction achieve 100% accuracy when guide shoe surface wears. If only based on single-axis data, the performance of the model degrades. We simulate the SWGS by adjusting the spacing of the guide shoe with the guide rail, resulting in an increased vibration amplitude in the Y-axis direction. Therefore, even if only Y-axis vibration data features are extracted, the fault diagnosis accuracy is high. When different faults occur, the vibration characteristics of the elevator appear along different axes. Therefore, in order to diagnose different faults, we combine the characteristics of the three axes to ensure that important features are extracted.

There are certain faults whose vibration characteristics are not apparent on a single axis. When UL occurs, the elevator car is tilted and the friction changes, and the guide shoe is close to the guide rail. The vibration in the front-to-back and up-to-down directions is significantly different from normal vibration. When the car is unbalanced, the car vibrates abnormally in both the X-axis direction and Z-axis direction. The URT and USCP do not exhibit prominent characteristics. These two types of elevator fault diagnosis relying only on single-axis data have the same level of accuracy. The diagnostic effectiveness of the combined triaxial vibration data is more than 2% higher than that based on single-axis data only.

Overall, the triaxial vibration feature extraction method shows higher accuracy and stability in fault diagnosis. The feature extraction methods, which only use single-axis data, perform well under some specific faults, but the overall performance is not as good as the triaxial feature extraction method. The advantage of the triaxial vibration feature extraction method is mainly reflected in its ability to synthesize the vibration information in all directions, thus extracting fault features more comprehensively. This comprehensive information extraction method significantly improves the accuracy and reliability of fault detection.

Feature extraction with and without elevator frequency features

Traditional feature extraction methods ignore elevator frequency features. We compare these methods with the proposed method. The fault for comparison is USCP. Results of the experiment are shown in Table 4. The conventional fault diagnosis method for fault data of USCP has 93.33% accuracy. The multidomain feature extraction method incorporates the frequency features of the elevator. Equation (4) calculates the rotational frequency of the car pulley. These features increase the fault diagnosis accuracy to 96.67%. The reported method achieves a higher accuracy rate in comparison without elevator frequency features. This experiment shows that the extraction of the frequency features of the elevator improves the fault diagnosis accuracy of the model.

Table 4.

Result of frequency feature extraction.

Fault	Method	Accuracy (%)	F1 score (%)
USCP	Before	93.33	93.10
USCP	After	96.67	96.43

Multiple faults

In order to assess the diagnostic ability of our model in the presence of multiple faults at the same time, we conducted a comprehensive experiment using data collected from an office building elevator. This elevator type is different from the elevator type used in training, which adds to the complexity of evaluating the robustness of our model. The faults we collected are UL and USCP faults at the same time.

In total, we collect 143 fault data samples under these conditions. These data samples are input into our trained model to assess its ability to accurately diagnose the multiple faults. The results of this testing phase are presented in Table 5, highlighting the performance metrics of our model.

Table 5.

Result of multiple faults.

Total	Multiple faults	Only UL	Only USCP
143	125	18	0

Out of a total of 143 data samples, our model successfully diagnosed 125 samples with an accuracy of 87.41%. This high level of accuracy demonstrates that our model effectively recognizes and diagnoses faults, even when multiple issues occur concurrently. Among the diagnosed faults, 18 instances are incorrectly predicted as containing only UL fault, indicating some limitations in distinguishing between overlapping fault signals.

Our model maintains relatively high accuracy of multiple faults. Moreover, the robustness of our model in diagnosing faults from an elevator model not included in the training data highlights its potential for generalization. It indicates that the model can adapt to new data while maintaining its diagnostic accuracy. This adaptability is vital for practical applications, ensuring that the model can be deployed in various environments and on different elevator systems without requiring extensive retraining.

SHAP features selection

SHAP ranks the features for each diagnosis model based on SHAP values and selects them in descending order. The SHAP feature selection method aims for high accuracy and selects the fewest features. Table 6 illustrates the number of selected features and the percentage of redundant features removed across different fault diagnostic models. The vibration characteristics of SWGS are relatively simple, with distinct vibration characteristics in a single axis. High accuracy of SWGS fault diagnosis is achieved with a few features. Since UL and USCP vibration characteristics are more complex, more features are required.

Table 6.

Result of SHAP features selection.

Fault	Quantity	Accuracy (%)	Proportion of redundant features (%)
UL	50	95	70
URJ	9	100	95
URT	24	91.67	85
SWGS	1	100	99
USCP	66	96.67	60

The multidomain feature extraction method extracts 165 features, in which more than 60% of the features are eliminated after SHAP feature selection. It drastically reduces the computational burden on the edge devices. With SHAP feature selection, the diagnostic accuracy of some faults, such as Ul, URJ, and URT, increases by about 2%. SHAP feature selection also reduces the risk of overfitting.

Conclusion

The existing fault diagnosis methods for elevators ignore the vibration signal characteristics caused by the elevator mechanical structures and usually diagnose faults based on only single-axis data, which often leads to poor diagnostic results. We propose a multidomain feature extraction method grounded in the structural model of the elevator and triaxial vibration data. Abnormal vibrations caused by the structural components of elevators are linked to specific frequency features. These elevator frequency features, based on elevator structural information, improve the suitability of the model for elevator fault diagnosis. The primary advantage of this triaxial vibration feature extraction method lies in its ability to integrate vibration information from all directions, enabling a more comprehensive extraction of fault features. According to the experimental results, the multidomain feature extraction method improves the accuracy of SVM using existing feature extraction of the elevator in five given kinds of faults. Meanwhile, we utilize SHAP for feature selection to reduce redundant features, which reduces the computational burden of the device, prevents high-dimensional data issues and minimizes the likelihood of overfitting. We propose a vibration-based fault diagnosis method for elevators, which achieves good results in single fault and multiple faults detection. The method has been validated on multiple elevators, including cases with multiple faults, achieving an accuracy of 87.41%. However, further testing and verification are necessary to generalize these results to elevators from different manufacturers or in different environments. Our approach incorporates domain-specific parameters into the feature extraction process to enhance adaptability. While these parameters are designed to accommodate variations across elevators, further research is necessary to verify the robustness of the model across diverse elevator architectures and operating conditions. We aim to develop more sophisticated models capable of handling complex multiple faults scenarios. We hope to improve the practical applicability of the proposed method and make it more adaptable to the differences between different elevator systems.

Footnotes

Declaration of conflicting interests

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

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Science and Technology Development Fund, FDCT, Macau SAR, under grant 0101/2022/A and the Guangzhou Innovation and Entrepreneurship Leading Team Project Funding under grant No. 202009020008.

ORCID iD

Li Yin

References

Elnour

Ahmad

Abdelkarim

, et al. Empowering smart cities with digital twins of buildings: applications and implementation considerations of data-driven energy modelling in building management. Build Serv Eng Res Technol 2024; 45: 475–498. DOI: 10.1177/01436244241239290.

Al-Kodmany

. Tall buildings and elevators: a review of recent technological advances. Buildings 2015; 5(3): 1070–1104.

Qing

Zhou

Wang

. Intelligent guidance method for elevator emergency treatment based on automatic recommendation and fault prediction. 2022 IEEE 22nd International Conference on Software Quality, Reliability, and Security Companion (QRS-C). IEEE, 2022, pp. 462–468.

Zhang

Zhao

Yan

, et al. An image recognition measurement approach and uncertainty for overspeed governors employed in elevator safety. Build Serv Eng Res Technol 2023; 44(3): 247–268. DOI: 10.1177/01436244231159022.

Park

Yang

. An implementation of risk-based inspection for elevator maintenance. J Mech Sci Technol 2010; 24: 2367–2376.

Niu

Guo

, et al. Preventive maintenance period decision for elevator parts based on multi-objective optimization method. J Build Eng 2021; 44: 102984.

Alassafi

Al-Gahtani

Almohsen

, et al. The causal factors of elevator maintenance: a perspective from Saudi Arabia healthcare facility management. Buildings 2022; 12(5): 532.

Patruni

Saraswathi

. Securing internet of things devices by enabling ethereum blockchain using smart contracts. Build Serv Eng Res Technol 2022; 43(4): 473–484. DOI: 10.1177/01436244221078933.

Kumar

Mishra

Dash

. An IOT and semi-supervised learning-based sensorless technique for panel level solar photovoltaic array fault diagnosis. IEEE Trans Instrum Meas 2023; 72: 1–12.

10.

Gardner

Dorling

. Artificial neural networks (the multilayer perceptron)—a review of applications in the atmospheric sciences. Atmos Environ 1998; 32(14-15): 2627–2636.

11.

Cortes

Vapnik

. Support-vector networks. Mach Learn 1995; 20: 273–297.

12.

Lam

Segre

. A distributed learning algorithm for Bayesian inference networks. IEEE Trans Knowl Data Eng 2002; 14(1): 93–105.

13.

Lin

Jhang

. Bearing fault diagnosis using a grad-cam-based convolutional neuro-fuzzy network. Mathematics 2021; 9(13): 1502.

14.

Jia

Gao

, et al. Elevator running fault monitoring method based on vibration signal. Shock Vib 2021; 2021: 1–10.

15.

Mishra

Huhtala

. Elevator fault detection using profile extraction and deep autoencoder feature extraction for acceleration and magnetic signals. Appl Sci 2019; 9(15): 2990.

16.

Albuquerque

Amador

Ferreira

, et al. Learning to rank with deep autoencoder features. 2018 International Joint Conference on Neural Networks (IJCNN). IEEE, 2018, pp. 1–8.

17.

Jia

Zhang

Chen

, et al. Fault diagnosis analysis of angle grinder based on ACD-De and SVM hybrid algorithm. Mathematics 2022; 10(18): 3279.

18.

Wan

, et al. Diagnosis of elevator faults with LS-SVM based on optimization by K-CV. J Electr Comput Eng 2015; 2015(1): 1–8.

19.

Suykens

Vandewalle

. Least squares support vector machine classifiers. Neural Process Lett 1999; 9: 293–300.

20.

Rodriguez

Perez

Lozano

. Sensitivity analysis of k-fold cross validation in prediction error estimation. IEEE Trans Pattern Anal Mach Intell 2009; 32(3): 569–575.

21.

Zheng

Zhao

. Wavelet packet decomposition and neural network based fault diagnosis for elevator excessive vibration. 2019 Chinese Automation Congress (CAC). IEEE, 2019, pp. 5105–5110.

22.

Han

Liu

, et al. Research on SVM-based bearing fault diagnosis modeling and multiple swarm genetic algorithm parameter identification method. Mathematics 2023; 11(13): 2864.

23.

Wang

Xue

Liang

, et al. Differential evolution with duplication analysis for feature selection in classification. IEEE Trans Cybern 2022; 53(10): 6676–6689.

24.

Liu

Yang

, et al. Semifree: semisupervised feature selection with fuzzy relevance and redundancy. IEEE Trans Fuzzy Syst 2023; 31(10): 3384–3396.

25.

Lundberg

Lee

. A unified approach to interpreting model predictions. In: Guyon

Luxburg

Bengio

,(eds). Advances in Neural Information Processing Systems 30 (NIPS 2017), Advances in Neural Information Processing Systems, 2017, vol 30.

26.

Hancock

Bauder

Khoshgoftaar

. A model-agnostic feature selection technique to improve the performance of one-class classifiers. 2023 IEEE 35th International Conference on Tools with Artificial Intelligence (ICTAI). IEEE, 2023, pp. 92–98.

27.

Chae

Lee

. Fault diagnosis of elevator doors using control state information. IEEE Access 2022; 10: 7207–7222.

28.

Hui

Ooi

Lim

, et al. An improved wrapper-based feature selection method for machinery fault diagnosis. PLoS One 2017; 12(12): e0189143.

29.

Niu

, et al. Rapid fault diagnosis method of elevator system based on multiattribute decision making. Shock Vib 2021; 2021: 1–9.

30.

Huang

Zhang

Guo

, et al. Wavelet packet decomposition-based multiscale cnn for fault diagnosis of wind turbine gearbox. IEEE Trans Cybern 2023; 53(1): 443–453.

31.

Strumbelj

Kononenko

. An efficient explanation of individual classifications using game theory. J Mach Learn Res 2010; 11: 1–18.

32.

Ribeiro

Singh

Guestrin

. “Why should I trust you?” Explaining the predictions of any classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016, pp. 1135–1144.