ISEE: Industrial Internet of Things perception in solar cell detection based on edge computing

Design of edge computing unit

In the background of intelligent manufacturing, the proliferation of terminal network equipment brings new challenges to data operation and maintenance, scalability, and reliability. Edge computing addresses these challenges by creating an open platform that integrates core functions like network, computing, storage, and applications. The development of edge computing enables computing to move from the centralized data center to the network’s periphery. It also enables intelligent services close to manufacturing cells to meet critical requirements, such as agile connectivity, data analysis through edge nodes, highly responsive cloud services, and privacy policies. Edge computing can use the embedded computing power of field devices to realize device autonomy based on distributed information processing. Also, edge computing supports the demand of digital manufacturing enterprises to configure intelligent factories quickly. Intelligent factories must adapt to the personalized needs of users and the dynamic changes of production conditions.

The research progress based on edge computing is also developing rapidly. Due to the limitation of a cloud platform, Lopez and others think that edge computing is necessary to move data and services from centralized nodes to the network edge. Condry et al. introduced a model using intelligent edge Internet of things (IoT) devices for safe and fast response of industrial IoT control operation. With the application of the IoT in the industrial field, Georgopoulos and others proposed a manufacturing roadmap of the IoT combined with cloud edge computing. Because edge computing run on the IoT devices’ peripherals, it can support these mobility and geographic independence requirements.

The physical proximity between computing and mobile devices makes it easier to implement low-end delay, high bandwidth, and low jitter services at the network’s edge. The scope of edge computing extends the range of IoT embedded devices with limited storage, cache, and processing capabilities to the edge near the terminal infrastructure. This study found that the edge computing unit has a powerful NPU neural unit for AI hardware acceleration. Under the same circumstances, the NPU’s AI computing power consumption is less than 1% of the GPU, significantly saving the hardware cost in a large amount of data processing in the solar photovoltaic industry.

Edge computing framework based on CNN

We run the convolutional neural networks (CNNs) in real time on edge computing unit to detect solar panels’ defects to transplant quickly. A CNN is a successful algorithm in AI and a powerful candidate for many CV algorithms. ¹⁵ However, the computational complexity of CNN is much higher than that of traditional algorithms. Usually, with the help of GPU acceleration, CNN-based applications are widely deployed on servers. However, the solution based on CNN is still too complex to be applied. To speed up CNN’s development, people have designed the whole design process, deployed quickly, and run efficiently on edge computing units. The system memory requirements, computational complexity, and flexibility for CNN mapping on embedded edge computing units are summarized. Based on these requirements, this article proposes a programmable and flexible CNN edge computing framework. As a feature extraction step, ⁵ CNN has excellent performance, but it is not easy to achieve in daily industrial applications and equipment due to its high computational complexity. For the inference of a 224 × 224 image, a large CNN model can involve 40 g operations (multiplication or addition). ² In practical application, the more considerable image cost is much more than this order of magnitude. Therefore, we usually deploy CNN-based applications on cloud servers. For the industrial application equipment, especially in the solar panel monitoring unit, the traditional CPU platform is challenging to process the CNN model with acceptable processing speed. It is even worse for the target detection tasks that need real-time processing. The high utilization rate of GPU depends on the number of images suitable for parallel processing. However, for the real-time detection of industrial solar cells, mass production is not practical.

This article extends the whole process to a complete design process and maps CNN to the embedded edge computing unit. The process consists of three parts: the strategy of compressing the original network into fixed-point data quantization. Parameterized and run-time configurable hardware architecture supporting various networks and platforms. In this article, we design a compiler to map the CNN model to the hardware architecture. Experiments on edge computing units show that the proposed design process can achieve CNN acceleration with high energy efficiency.

CNN research includes a series of aspects. As the name suggests, the most important layer in CNN is the convolution layer. The full connection layer, nonlinear layer, and convergence layer (downsampling layer) are important parts of the neural network. The convolution layer uses the trained filter to perform two-dimensional convolution on the input feature mapping to extract local features. In order to extract high-level features, it is usually necessary to cascade multiple transformation layers. In the final stage of CNN, it is usually used as a classifier. The correction linear element (ReLU) is one of the most commonly used functions. Hyperbolic tangent function and sigmoid function are often loaded into various neural networks used in the downsampling collection layer. ¹⁶

In this article, as the research model comprises multiple convolution layers and top fully connected layers and contains correlation weight and pooling layer, a CNN is used to intersect other depth networks. ¹⁷

In Figure 8, from left to right, the first image is the original image, and then the image is rotated left and right, up and down, 90° rotation, 180° rotation, 270° rotation, random brightness, and random contrast adjustment and adjustment. Through these technologies, data enhancement is realized to expand the number of samples.

OPEN IN VIEWER

Figure 8.

Actions on images.

When the data set is expanded, all images’ size is cut to 100 × 100, and these images are normalized to facilitate the learning of the neural network.

Algorithm model design process

This article combines CV and CNN technology to detect the defects of solar panels, which uses three layers of convolution layer and three layers of pool layer, two full connection layers and one output layer. The network architecture is shown in Figure 9.

OPEN IN VIEWER

Figure 9.

Algorithm model design process.

Convolution layer: the purpose of this layer is to extract the features of the input image. Each node is just a small piece of the upper neural network. It moves a filter from the upper left corner of the neural network to the lower right corner and calculates each corresponding identity matrix in the movement. The filter size of the convolution layer is 3 * 3. Activation function: a nonlinear activation function is added after each convolution layer. The purpose of adding a nonlinear activation function is to make the neural network more expressive. Because the linear function is used, the model cannot work accurately in many cases. The nonlinear activation function mainly includes the sigmoid function, tanh function, and relu function. This article uses the relu activation function, which can speed up the training speed compared with the previous two functions

Pooling layer: this layer reduces each feature graph’s dimensions while retaining important information in the feature graph. The commonly used pooling methods are average pooling and maximum pooling, and the maximum pooling is used in this article. The filter size of the pooling layer is 2 * 2.

Model neural network structure of the convolution layer and pooling layer parameters of the network are shown in Table 1.

OPEN IN VIEWER

Table 1.

Parameter list of convolution layer and pooling layer.

	Filter size	Number of filters
Conv1	3*3	16
Pooling1	2*2	16
Conv2	3*3	32
Pooling2	2*2	32
Conv3	2*2	64
Pooling3	2*2	64

The main purpose of this article is to extract the full layer connection feature. This article shows 1024 nodes in the Full Connection Layer 1 (FC 1 ) layer and 512 nodes in the Full Connection Layer 2 (FC 2 ) layer. Output layer: the output of FC 2 is two nodes, and the final classification result is obtained by softmax of the output layer.

Dropout: To prevent the model from the over-fitting phenomenon, dropout technology is adopted, which mainly can discard some neurons randomly during training. In this article, keep prob = 0.5 is set.

Batch normalization: in this article, the input image has been normalized in image preprocessing, but the input data distribution in each neural network layer changes during training. This is because updating the front layer’s training parameters will lead to the change of the input data distribution in the later layer, leading to the change of the input data distribution in each layer. Moreover, if there is a slight change in the network’s front layer, the latter will gradually accumulate and enlarge the change. Therefore, this article uses a batch normalization algorithm to solve the network layer’s problem, making each layer’s neural network input keep the same distribution, accelerate the model’s training speed, and improve its performance.

Normalization is used before relu, and the specific implementation process is as follows

where μ B is the batch average, σ B 2 is the batch variance, m is the number of samples in each batch, x ^ i is the output after normalization, γ is the scaling parameter, β is the shift parameter, ε is the smoothing term, which is to prevent the denominator from appearing 0, and y i is the output after batch normalization.

In this article, the size of a single batch is set to 64. First, formula (1) obtains the average value of the current batch, then formula (2) calculates the variance of the current batch, and then uses a formula (3) to get the normalization processing of input data. Finally, the data after batch normalization is obtained by formula (4), which is used as the next layer network input.

L2 regularization: in addition to using dropout technology to prevent over-fitting, this article also adds L2 regularization to the neural network, which can be regarded as a penalty term for the loss function. Assuming that the loss function of the model is, which represents all the parameters of the neural network, after adding L2 regularization, the loss function becomes

In the process of model optimization, it is necessary to optimize the new loss function J ( θ ) + λ R ( w ) . Where λ is the regularization coefficient, R ( w ) = ∑ i | w i 2 | , the sum of squares of all weights.

In this article, adding L2 regularization to the model can limit the weight’s size very well. Making the parameters as close to 0 as possible can prevent the random noise in the model’s training data from fitting arbitrarily to avoid the model’s over-fitting. The coefficient set in this paper is 0.0001.