Fault diagnosis method based on cloud edge intelligent cooperation mechanism and lightweight module design

Abstract

The invention provides a fault diagnosis method based on a cloud edge intelligent cooperation mechanism and lightweight module design. Firstly, a multi-scale depth separable convolutional neural network model fusing a bidirectional gating circulation unit and an ICA attention mechanism is constructed, effective filtering of high-frequency noise signals is achieved, and the method is suitable for various industrial equipment fault diagnosis tasks. Then, on the basis of a cloud edge intelligent cooperation mechanism, a federated learning framework based on heterogeneous network optimization and ECC elliptic curve encryption and a model compression module are adopted to optimize the whole model training and reasoning process, and a lightweight fault diagnosis module capable of being migrated and deployed is formed; according to the method, effective real-time fault diagnosis analysis is realized through combination of cloud edge cooperative calculation and powerful feature self-learning capability of deep learning, so that potential faults can be found and solved in time, the dead time of a production line is reduced, and the production efficiency and the equipment availability are improved.

Description

Technical Field

[0001] The present invention relates to the fields of industrial Internet, fault diagnosis, deep learning, and cloud-edge collaborative computing, and specifically to a fault diagnosis method based on cloud-edge intelligent collaborative mechanism and lightweight module design. Background Art

[0002] In today's industrial workshops, the efficient operation of industrial equipment plays a vital role in the automated operation of production lines. However, traditional fault diagnosis methods are often limited by limited monitoring methods and response speed, and cannot meet the needs of rapidly changing and highly automated production environments. Therefore, to achieve the transformation of intelligent workshops, real-time fault diagnosis of industrial equipment has become a top priority.

[0003] Currently, deep learning-based fault diagnosis methods have become the mainstream solution. However, noisy production environments and variable operating conditions pose challenges to the design and application of deep learning models. In deep convolutional networks, increasing module depth to extract more comprehensive fault features is an effective solution for improving model generalization and robustness. However, industrial field fault diagnosis services are typically deployed at the edge, close to the equipment. In the limited resource environment of edge nodes, increasing model depth leads to excessive parameter and computational complexity, which can easily lead to excessive node load and increased queuing delays, which are detrimental to real-time equipment fault diagnosis. Summary of the Invention

[0004] In view of this, the present invention proposes a fault diagnosis method based on cloud-edge intelligent collaborative mechanism and lightweight module design, constructs a multi-scale deep separable convolutional neural network model that integrates bidirectional gated recurrent units and ICA attention mechanism, realizes effective filtering of high-frequency noise signals, and is suitable for various industrial equipment fault diagnosis tasks. Based on the cloud-edge-end intelligent collaborative mechanism, federated learning and model compression modules are used to optimize the entire model training and reasoning process to form a lightweight fault diagnosis module that can be migrated and deployed. Cloud-edge collaborative computing combined with the powerful feature self-learning ability of deep learning realizes effective fault diagnosis and analysis, thereby timely discovering and resolving potential faults, reducing the downtime of the production line, and improving production efficiency and equipment availability.

[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a fault diagnosis method based on a cloud-edge intelligent collaborative mechanism and a lightweight module design, in which the model training task is deployed in a federated learning mode, model training is performed on each edge node, and the gradient parameters of the model training of each edge node are aggregated by the cloud to obtain the final inference model, and then sent to each edge node for fault diagnosis; wherein, the model training includes the following steps:

[0006] Collect the original time domain signal and preprocess the original time domain signal to obtain time series data;

[0007] The time series data is overlapped and sampled using a sliding window, and the one-dimensional time series data is converted into a two-dimensional grayscale image as the input sample;

[0008] Build a BiGRU-CA-DSCNN fault diagnosis model;

[0009] The BiGRU-CA-DSCNN fault diagnosis model is trained and verified using samples.

[0010] The construction of the BiGRU-CA-DSCNN fault diagnosis model includes the following steps:

[0011] The multi-scale depth-wise separable convolution module obtains spatial features by constructing two parallel convolution modules of different scales to achieve the fusion of multi-scale information;

[0012] The ICA attention module outputs weights in the vertical and horizontal directions according to the spatial features, and multiplies them with the original input features to obtain the feature information of the fused attention weights as the output features;

[0013] The BiGRU module extracts sequence context temporal features based on the output features of the ICA attention module, fuses them with the spatial features extracted by the multi-scale depthwise separable convolution module, and finally performs dimensionality reduction by a global average pooling layer;

[0014] Softmax classifier, obtains fault classification results based on the features after dimensionality reduction.

[0015] The construction of the multi-scale depth-separable convolution module includes the following steps:

[0016] Construct two parallel convolution modules of different scales in the multi-scale depth-wise separable convolution module;

[0017] The first convolution module uses a deeper scale convolution, including four sequentially connected convolutions of different sizes and a pooling layer; the second convolution module uses a shallower scale convolution, including two sequentially connected depthwise separable convolutions of different sizes and a pooling layer; AdaBN normalization is used after each convolution layer, and MetaAconC is used as the activation function;

[0018] The different scale features output by the two parallel convolution modules are multiplied in the feature fusion layer to obtain spatial features;

[0019] Among them, the depth-wise separable convolution consists of two parts: depth-wise convolution and point-by-point convolution; the depth-wise convolution adopts grouped convolution to divide the input feature vector into n groups, let the number of feature input channels and the number of output channels be n, use a convolution kernel for each of the n channels to obtain a feature vector with n channels of 1, and then splice these n feature vectors in sequence to obtain an output feature vector with N channels; the point-by-point convolution adopts 1×1 convolution to perform linear combination between channels on the output of the depth-wise convolution, so that the depth-wise separable convolution module can freely change the number of output channels.

[0020] The construction of the ICA attention module includes the following steps:

[0021] Based on the adaptive global maximum pooling operation, the spatial features along the vertical and horizontal directions are aggregated into two separate one-dimensional vectors respectively; then the two one-dimensional vectors are spliced ​​in the spatial dimension, the channels are compressed through the convolution layer, and the spatial information in the vertical and horizontal directions is encoded through AdaBN normalization and Meta-ACON activation function;

[0022] Then, the Split operation is performed to obtain the horizontal and vertical feature information, and the attention weights in the two directions are obtained through two convolutional layers respectively;

[0023] Finally, the attention weights in the two directions are multiplied by the input spatial features to obtain the output features.

[0024] The construction of the multi-scale depth-separable convolution module and the ICA attention module is performed by a lightweight module;

[0025] The cloud aggregates the gradient parameters of the edge node model training to obtain the final inference model, and sends it to each edge node for fault diagnosis, including the following steps:

[0026] The edge node and the cloud establish a communication connection based on the key generation algorithm;

[0027] In each round of communication between the edge node and the cloud, the heterogeneous network optimized federated learning algorithm is used to train the final gradient parameters, and then the final inference model is obtained and sent to each edge node for fault diagnosis.

[0028] The key generation algorithm comprises the following steps:

[0029] For edge node j, j = 1, ... K, construct the elliptic curve E based on the given a, b values, and then select a random number v j As the private key, then choose a point G on the elliptic curve E j As a generator, generate the public key Q j =v j Gj ; Edge node j will send the public key group {E:y 3 =x 3 +ax+b,Q j ,G j} is sent to the cloud, where x and y represent the horizontal and vertical coordinates of the elliptic curve E, and a and b are coefficients;

[0030] For the cloud, based on the given a, b, p values, an elliptic curve E is constructed, a random number v0 is selected as the private key, and a point G0 is selected on the elliptic curve E as the generator to generate the public key Q0 = v0G0. The cloud server will 3 =x 3 +ax+b,Q0,G0} is sent to each edge node.

[0031] The heterogeneous network optimization federated learning algorithm includes the following steps:

[0032] ① The cloud first selects K edge nodes according to the probability P k , randomly select N clients, the set is S t ;

[0033] ② The cloud's communication parameter w for this round t The code is M, a random number I is selected, and M is encrypted for each edge node j; the encrypted ciphertext group is (R j ,S j ), where R j =IS j ,S j =M+IQ j ,Q j The public key generated for edge node j sends the encrypted ciphertext group to the selected edge nodes in turn;

[0034] ③Edge node j uses the local private key to encrypt the ciphertext group (R j ,S j ) decrypted to get M, M = S j -v j R j , then decode M to get the original gradient parameters, where v j is the private key;

[0035] ④ Each edge node obtains w by optimizing the objective function k t+1 :

[0036] w k t+1 =argmin w h k (w; w t)=F k (w)+μ / 2||ww t ||2

[0037] Among them, F k (w) is the objective function in the fedavg algorithm, h k (w; w t ) is the objective function in the fedfrox algorithm, μ is the proportional coefficient, w represents the gradient parameter obtained in this round of training, w k t+1 is to optimize h k (w; w t ) and the solution obtained after

[0038] ⑤ Each edge node will get w k t+1 The data is sent back to the cloud, where the cloud aggregates the w of each edge node using a weighted average method. k t+1 Get the initial parameter w for the next round t+1 , return to step ②;

[0039] ⑥ When the set number of iterations is reached, the obtained w T As the final gradient parameter, it is used to update the BiGRU-CA-DSCNN fault diagnosis model to obtain the final inference model.

[0040] The fault diagnosis is specifically as follows:

[0041] The collected original time-domain vibration signal is input into the final reasoning model to obtain the fault diagnosis classification.

[0042] The fault diagnosis system based on cloud-edge intelligent collaboration mechanism and lightweight module design includes:

[0043] Edge nodes are used for model training and fault diagnosis based on the final inference model sent from the cloud.

[0044] The cloud is used to aggregate the gradient parameters of the model training of each edge node and obtain the final inference model, which is then sent to each edge node for fault diagnosis.

[0045] The present invention has the following beneficial effects and advantages:

[0046] 1. Achieve a balance between algorithm accuracy and computational efficiency. The multi-scale depthwise separable convolution module proposed in this paper fully balances model accuracy and computational complexity. At shallower scales, all standard convolutions are replaced with depthwise separable convolutions, reducing the number of parameters and computational complexity while retaining good local feature information. At deeper scales, standard convolutions are retained to avoid the loss of fine-grained global features caused by the introduction of depthwise separable convolutions, which would cause a sharp drop in model accuracy.

[0047] 2. Effectively filter noise features to avoid the impact of invalid information. This paper introduces the ICA attention mechanism module to achieve effective filtering of high-frequency noise by calculating the attention weight of each feature.

[0048] 3. Multi-dimensional feature extraction: The multi-scale depthwise separable convolution module uses two convolution kernels of different sizes to extract global and local features, respectively. This preserves both edge local information of low-level features and high-level abstract semantic information. The fused extracted features significantly increase the amount of effective information retained. The bidirectional gated recurrent unit comprehensively extracts temporal features from the sequence context, which are then fused with the spatial features extracted by the convolution module. The global pooling module then provides a compact feature representation for the classifier.

[0049] 4. Lightweight model inference: Minimize the complexity of the inference model through model compression and TensorRT lightweight deployment. Deliver pod deployment services to edge devices through cloud-based Kubernetes, ensuring that edge nodes function normally with limited computing resources.

[0050] 5. Privacy and security: When using a distributed machine learning model, data communication between nodes may result in privacy leakage. Based on the federated learning mechanism proposed in this invention, each node uses a local data set for training data, and at the same time encrypts the model parameters of cloud-edge interaction using an elliptic curve encryption algorithm, thereby continuously improving the overall model performance while protecting data privacy. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is the overall flow chart of the present invention;

[0052] Figure 2 Schematic diagram of a multi-scale deep separable convolutional neural network model that integrates bidirectional gated recurrent units and ICA attention mechanism;

[0053] Figure 3 It is a multi-scale depth-separable convolution module diagram;

[0054] Figure 4 This is the ICA attention mechanism module diagram. DETAILED DESCRIPTION

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0056] A fault diagnosis method based on cloud-edge intelligent collaboration mechanism and lightweight module design, including:

[0057] 1) A multi-scale deeply separable convolutional neural network model that integrates a bidirectional gated recurrent unit with an ICA attention mechanism. First, the one-dimensional time series data is converted into a two-dimensional grayscale image using the Gram Angular Field (GASF). A multi-scale deeply separable convolutional module that integrates the attention mechanism is designed to comprehensively extract multi-scale local receptive field features of different scales that incorporate attention weights, while significantly reducing computational and parameter complexity. Secondly, a bidirectional gated recurrent unit (BiGRU) is introduced to extract the time series features of the vibration signal, achieving spatiotemporal feature fusion. In addition, to meet the requirements of lightweight and generalization, the training strategy is optimized, using Meta-ACON as the activation function, label smoothing (LSR) as the loss function, and AdamP as the optimizer.

[0058] A cloud-edge intelligent collaboration mechanism based on federated learning and model compression optimization is implemented, and multi-terminal online training of the central cloud and edge nodes is achieved based on federated learning. The central cloud aggregates the trained models to form a global model, optimizes it through model pruning and knowledge distillation, and parses the model into a lightweight inference engine based on TensorRT, which is then delivered to the edge through Kubernetes as a Pod, enabling edge devices to perform efficient model inference locally while retaining offline autonomy and business migration capabilities at the edge.

[0059] like Figure 1 As shown in the figure, a fault diagnosis method based on cloud-edge intelligent collaboration mechanism and lightweight module design includes the following contents:

[0060] Figure 1 The following figure shows an overall flow chart of a fault diagnosis method based on cloud-edge intelligent collaboration mechanism and lightweight module design, which includes the following steps:

[0061] Step 1: collect the original time domain vibration signal and perform data preprocessing such as normalization and outlier detection on the original time series data;

[0062] Step 2: To ensure the generalization ability of the training model, Gaussian white noise with a specified signal-to-noise ratio is added to the original time series data;

[0063] Step 3: Use a sliding window to perform overlapping sampling on the time series data after adding noise, and convert the original one-dimensional data into a two-dimensional grayscale image using the Gram angle field (GASF) as the input sample;

[0064] Step 4 is to divide the input samples into training set, validation set and test set according to a certain ratio;

[0065] Step 5: Build the BiGRU-CA-DSCNN fault diagnosis model, randomly initialize the model weights and biases, and set hyperparameters such as batch size, learning rate, and optimizer.

[0066] Step 6 uses the training set to train the model, extracting fault feature information. Model weights and biases are gradually corrected based on the gradient function. Backpropagation is used to monitor the learning process and fine-tune the parameters of each layer to calculate the error loss for each network iteration. Using the LSR label smoothing regularization loss function provides better generalization capabilities in noise-sensitive environments. The calculation formula is as follows:

[0067] Step 7: Use the validation set to verify the model training effect and adjust the model parameters based on the validation results.

[0068] Step 8 determines whether the model has converged or reached the preset maximum number of training iterations. If so, proceed to the next step; otherwise, repeat step 6.

[0069] In step 9, if the task is a centralized machine learning task, the trained model performance is tested locally using a test set, the accuracy is calculated, and the fault diagnosis results are output. If the task is a distributed machine learning task, the model training task is deployed in a federated learning mode based on the cloud-edge intelligent collaboration mechanism. The cloud aggregates the weights of each node and iteratively trains them to obtain the final inference model, which is then pruned and compressed. The central cloud then centrally manages and distributes Pods to edge nodes, enabling real-time inference at the edge.

[0070] The details are as follows:

[0071] (1) A multi-scale deep separable convolutional neural network model integrating bidirectional gated recurrent units and ICA attention mechanism

[0072] The present invention uses multi-scale depth separable convolution module, ICA attention module and BiGRU module as the main sub-network to build a network model. The model architecture is as follows: Figure 2 As shown in the figure, the model's input is a 40×40×1 grayscale image sample constructed from a one-dimensional vibration signal. First, two parallel convolution modules of different scales are constructed within the multi-scale depthwise separable convolution module. The deeper convolution module sequentially connects two standard convolutions of sizes 6×6 and 2×2, a 2×2 pooling layer, two standard convolutions of size 2×2, and a 2×2 pooling layer. The shallower convolution module sequentially connects two depthwise separable convolutions of different sizes and a 2×2 pooling layer. The depthwise convolutions in the two depthwise separable convolutions are 10×10 and 2×2, respectively, and the pointwise convolutions are all 1×1. AdaBN normalization is used after each convolution module to accelerate model training and prevent overfitting. MetaAconC is used as the activation function to adaptively select whether to activate neurons. Experimental results show that it can improve accuracy by 1-2 percentage points in time series tasks.

[0073] The different-scale features output by the two parallel convolutions are multiplied in the feature fusion layer to achieve multi-scale information fusion. The fused features are directly input into the ICA attention module, which outputs vertical and horizontal weights. These are multiplied with the original input features to obtain the feature information of the fused attention weights. The fused attention weight features are then input into the bidirectional gated recurrent unit layer to further extract global feature information in the temporal dimension. The output multi-dimensional features are then input into the global average pooling layer for dimensionality reduction and information integration. Finally, a softmax classifier is used to determine the fault classification result.

[0074] 1) Multi-scale Depthwise Separable Convolutional Module. Traditional convolutional neural networks (CNNs) typically use a fixed-size receptive field to capture features. This fixed receptive field size is often insufficient to encompass semantic information at different scales, resulting in degraded model performance when processing multi-scale information. To increase the receptive field to encompass a wider range of contextual information, CNN models often employ deeper network structures and pooling operations, which can lead to a significant loss of spatial resolution and affect the accurate grasp of details and local structures. Furthermore, deeper network structures require a greater amount of computation and parameters, which is disastrous for edge devices with limited computing power in industrial sites.

[0075] To address the above problems, the present invention is based on the Inception model multi-branch computing and multi-scale fusion feature extraction method. It uses convolution kernels of different sizes to extract features at different spatial levels and splice them together, while retaining the edge local information of low-level features and high-level abstract semantic information. The amount of effective information retained by the extracted features after fusion is significantly increased.

[0076] In order to reduce the computational complexity of the model, depth-wise separable convolution is introduced. The specific structure is as follows: Figure 3 As shown. Depthwise separable convolution consists of two parts: depthwise convolution and pointwise convolution. The structural diagram is shown in Figure 3 As shown in Figure 1. The depthwise convolution uses grouped convolution to divide the input feature map into N groups. The number of input and output channels of the feature map is set to N. A convolution kernel is applied to each of the N channels to obtain a feature map with N channels of 1. These N feature maps are then concatenated in sequence to obtain an output feature map with N channels. Pointwise convolution uses 1×1 convolution to perform linear combinations between channels on the output of the D depthwise convolution, allowing the depthwise separable convolution module to freely change the number of output channels.

[0077] The multi-scale depthwise separable convolution module proposed in this paper strikes a balance between model accuracy and computational complexity. At shallower scales, all standard convolutions are replaced with depthwise separable convolutions, reducing the number of parameters and computational complexity while retaining good local feature information. At deeper scales, standard convolutions are retained to avoid the loss of fine-grained global features caused by the introduction of depthwise separable convolutions, which would result in a sharp drop in model accuracy.

[0078] 2) ICA Attention Module

[0079] This paper proposes an ICA attention mechanism, the network structure is as follows Figure 4 By embedding the position information into the channel attention, the edge network can obtain information from a larger area without introducing large overhead.

[0080] First, based on an adaptive global max pooling operation, the input vectors in the vertical and horizontal directions are aggregated into two separate one-dimensional vectors. The two one-dimensional vectors are then concatenated in the spatial dimension. A convolutional layer with a 1×1 kernel size is used to compress the channels. AdaBN and Meta-ACON are used to encode the vertical and horizontal spatial information. A split operation is then performed to obtain horizontal and vertical feature information. Attention weights for each direction are then obtained through two convolutional layers with a 1×1 kernel size. Finally, the attention weights for each direction are multiplied by the input features to obtain the output features. The ICA attention mechanism considers both inter-channel relationships and positional information. It not only captures cross-channel information but also incorporates direction- and position-sensitive information, enabling more accurate target region localization in time series processing and analysis. The adaptive pooling operation and the choice of smaller convolutional kernels ensure that the ICA attention mechanism is lightweight and can be embedded in a wider range of network architectures.

[0081] 3) Bidirectional Gated Recurrent Unit (BiGRU) and Global Average Pooling Module

[0082] Recurrent neural networks (RNNs) are sensitive to time series and use state variables to store past information and current inputs to calculate current outputs. Traditional RNNs are prone to vanishing or exploding gradients when processing long sequences, making it difficult to capture long-term dependencies within the sequence. Bidirectional long short-term memory networks and bidirectional gated recurrent units are two RNN variants that address this issue by introducing gating structures that allow the model to selectively remember and forget information. Bidirectional propagation simultaneously considers the forward and backward flow of information in the sequence, thus addressing these issues. Compared to the two, BiCRU further optimizes the gating structure, reducing the number of parameters and accelerating computational speed.

[0083] BiGRU introduces an update gate and a reset gate. The update gate is responsible for determining when to update the hidden state, and the reset gate is responsible for determining when to allow past information to affect the current hidden state.

[0084] The present invention selects BiGRU to comprehensively extract sequence context temporal features, which are then integrated with the spatial features extracted by the multi-scale depthwise separable convolution module mentioned above. Finally, a global average pooling layer is used to achieve dimensionality reduction, thereby retaining important sequence information while reducing the number of parameters and providing a compact feature representation for subsequent classifiers.

[0085] (2) Cloud-edge intelligent collaboration mechanism based on federated learning and model compression optimization

[0086] 1) Federated learning mechanism based on heterogeneous network optimization and ECC encryption algorithm

[0087] The cloud-edge intelligent collaboration model, based on the FedProx federated learning model, addresses device and data heterogeneity in the federated learning process in a cloud-edge environment. Combined with the ECC elliptic curve cryptography algorithm, it ensures secure parameter transmission while preventing excessive resource consumption due to encryption and decryption operations, thereby achieving a highly real-time and fault-tolerant distributed learning architecture. The specific algorithm process is as follows:

[0088] Set the total number of clients K, the number of communication rounds T, μ, γ, the server initialization parameter w0, the number of selected clients N, a certain client k=1,......N, the probability P of the kth client being selected k ,a,b,p.

[0089] Algorithm 1: Key Generation Algorithm

[0090] ①Edge node j (j=1,...K) constructs elliptic curve E based on given a, b, p values, and then selects a random number v j As the private key, then choose a point G on the elliptic curve E j As a generator (G j The order is a prime number), generate the public key Q j =v j G j . Edge node j will send the public key group {E:y 3 =x 3 +ax+b,Q j ,G j}Sent to the cloud server.

[0091] ② Similarly, the cloud server constructs the elliptic curve E based on the given a, b, p values, and then selects a random number v0 as the private key. Then, it selects a point G0 on the elliptic curve E as the generator (the order of G0 is a prime number) to generate the public key Q0=v0G0. The cloud server takes the public key group {E:y 3 =x 3 +ax+b,Q0,G0} is sent to each edge node.

[0092] Algorithm 2: Heterogeneous Network Optimization Federated Learning Algorithm

[0093] For each round of communication t=0,......T

[0094] ① The cloud server first selects K clients according to the probability P k , randomly select N clients, whose set is S t .

[0095] ② The cloud server sets the communication parameter w for this round t The code is M, a random number I is selected, and M is encrypted for each client j (j=1,...K). The encrypted ciphertext group is (R j ,S j ), where R j =IS j ,S j =M+IQ j ,Q j The public key generated for client j sends the encrypted ciphertext group to the selected clients in turn.

[0096] ③ Each selected client j (j=1,...K) uses the local private key v j For the ciphertext group (R j ,S j ) decrypted to get M, M = S j -v j R j , and then decode M to get the original parameters.

[0097] ④ Each client needs to find a w k t+1 , w k t+1 =argmin w h k (w; w t )=F k (w)+μ / 2||ww t || 2 . Among them F k (w) is the objective function in the fedavg algorithm, h k (w; w t ) is the objective function in the fedfrox algorithm, w t is the original parameter obtained by decoding in step ③, w is the parameter obtained in this round of training, and w k t+1 is to optimize h k (w; w t ) is the solution obtained after .

[0098] ⑤ Each client will get w k t+1 The parameters are encrypted using the public key Q0 generated by the server and passed back to the cloud server. The cloud server decrypts the parameters using the local private key v0 and uses the weighted average method to aggregate these parameters to obtain the initial parameters for the next round.

[0099] 2) Model compression optimization and lightweight deployment strategy

[0100] Since edge inference models need to be customized and developed to adapt to limited resource environments and strike a balance between inference accuracy and execution latency, the cloud-edge intelligent collaborative model proposes three solutions: model pruning, knowledge distillation, and lightweight deployment.

[0101] Model pruning trains a deep neural network model in the cloud to obtain benchmark performance, analyzes the model's weights on edge devices, and sets pruning thresholds based on a hybrid standard of gradient and weight size. This approach considers both the importance of parameters and the gradient information during training. It uses a pruning standard based on weight size for fully connected layers and a pruning standard based on gradient magnitude for convolutional layers to more comprehensively evaluate the importance of parameters and perform more refined pruning operations on the edge side to remove connections and neurons deemed unimportant.

[0102] Knowledge distillation, based on the HKDA knowledge distillation architecture, builds a more compact model on the edge while still maintaining high performance. By training a complex large model in the cloud, the model introduces a self-attention mechanism to better capture the long-range dependencies of the input data. The generation of soft labels not only takes into account the prediction probability of the large model, but also utilizes the self-attention weight information in the large model. On the edge device, by building a lightweight model, the self-attention mechanism is also introduced and trained using soft labels. HKDA further uses the attention matrix of the large model as an additional supervisory signal to guide the small model to better simulate the focus of the large model on the input. This method is committed to improving the performance of the small model on long-range dependencies through the cross-entropy loss and attention-based loss of knowledge distillation. Evaluation on edge devices shows that the HKDA method can effectively improve the performance of the small model, enabling it to better adapt to complex tasks in resource-constrained environments.

[0103] Lightweight deployment: Use TensorRT to deploy lightweight models. Export the Pytorch-trained model as an ONNX graph representation supported by TensorRT. Use the ONNX2TensorRT tool to convert the generated ONNX model into a TensorRT engine. Use the TensorRT engine to optimize the model, including precision reduction, layer fusion, dynamic graph optimization, pipeline parallelization, and asynchronous inference, to improve inference performance. Consider using low-precision quantization techniques and selecting an inference engine suitable for the target hardware, such as FP16 or INT8. Deploy the optimized model using the Standalone TensorRT runtime API and deliver it to edge devices through Kubernetes in the cloud.

[0104] 3) Edge offline autonomy

[0105] In the cloud-edge intelligent collaboration mechanism, the central cloud Kubernetes sends lightweight models to the edge nodes in the form of pods for execution. After the edge nodes lose connection with the central cloud, the edge offline autonomous mechanism realizes the decoupling between cloud-edge nodes, and independently schedules and manages them to ensure the normal operation of edge inference tasks. At the same time, metadata such as containers and service configurations are stored locally and persistently. After the connection is restored, the local metadata can be used to synchronize the cloud-edge status in time, and the latest model iterations can be updated to the edge nodes.

[0106] 4) Edge business migration

[0107] If the central cloud fails to receive a heartbeat response from an edge node within a specific time period, the system will determine that edge node to be in the NotReady state. This situation typically has two possible causes: one is that the edge node itself is down, and the other is that the edge node is operating normally but cannot communicate with the central cloud. To address these issues, we have designed an edge business migration strategy based on the Kubernetes task orchestration and scheduling function. The specific steps are as follows:

[0108] When the current edge node status is NotReady, the central cloud will first use feedback information from other edge nodes to determine whether the model inference task corresponding to the current node is still running normally.

[0109] If the model inference task runs normally, it means that there is a problem with the cloud-edge network communication. This will not affect edge production operations, so task migration is not performed.

[0110] If the model inference task does not run properly, the system determines that the edge node has failed. In this case, the central cloud Kubernetes executes the business migration strategy and uses the container eviction function (Taint & Toleration) to migrate the model inference container on the failed node to another edge node on the same production line, ensuring the continuous operation of the model inference task.

[0111] The edge service migration function can realize unified scheduling and management of edge nodes, and has a certain buffering effect on single-node failures, thereby ensuring the continuity of edge services to a certain extent.

Claims

1. A fault diagnosis method based on cloud-edge intelligent collaboration mechanism and lightweight module design, characterized by: The model training task is deployed in a federated learning mode. The model training is performed on each edge node. The cloud aggregates the gradient parameters of the model training on each edge node and obtains the final inference model. The final model is then sent to each edge node for fault diagnosis. The model training includes the following steps: Collect the original time domain signal and preprocess the original time domain signal to obtain time series data; The time series data is overlapped and sampled using a sliding window, and the one-dimensional time series data is converted into a two-dimensional grayscale image as the input sample; Build a BiGRU-CA-DSCNN fault diagnosis model; The BiGRU-CA-DSCNN fault diagnosis model is trained and verified using samples.

2. The fault diagnosis method based on cloud-edge intelligent collaboration mechanism and lightweight module design according to claim 1 is characterized in that: The construction of the BiGRU-CA-DSCNN fault diagnosis model includes the following steps: The multi-scale depth-wise separable convolution module obtains spatial features by constructing two parallel convolution modules of different scales to achieve the fusion of multi-scale information; The ICA attention module outputs weights in the vertical and horizontal directions according to the spatial features, and multiplies them with the original input features to obtain the feature information of the fused attention weights as the output features; The BiGRU module extracts sequence context temporal features based on the output features of the ICA attention module, fuses them with the spatial features extracted by the multi-scale depthwise separable convolution module, and finally performs dimensionality reduction by a global average pooling layer; Softmax classifier, obtains fault classification results based on the features after dimensionality reduction.

3. The fault diagnosis method based on cloud-edge intelligent collaboration mechanism and lightweight module design according to claim 2 is characterized in that: The construction of the multi-scale depth-separable convolution module includes the following steps: Construct two parallel convolution modules of different scales in the multi-scale depth-wise separable convolution module; The first convolution module uses a deeper scale convolution, including four sequentially connected convolutions of different sizes and a pooling layer; the second convolution module uses a shallower scale convolution, including two sequentially connected depthwise separable convolutions of different sizes and a pooling layer; AdaBN normalization is used after each convolution layer, and MetaAconC is used as the activation function; The different scale features output by the two parallel convolution modules are multiplied in the feature fusion layer to obtain spatial features; Among them, the depth-wise separable convolution consists of two parts: depth-wise convolution and point-by-point convolution; the depth-wise convolution adopts grouped convolution to divide the input feature vector into n groups, let the number of feature input channels and the number of output channels be n, use a convolution kernel for each of the n channels to obtain a feature vector with n channels of 1, and then splice these n feature vectors in sequence to obtain an output feature vector with N channels; the point-by-point convolution adopts 1×1 convolution to perform linear combination between channels on the output of the depth-wise convolution, so that the depth-wise separable convolution module can freely change the number of output channels.

4. The fault diagnosis method based on cloud-edge intelligent collaboration mechanism and lightweight module design according to claim 2 is characterized in that: The construction of the ICA attention module includes the following steps: Based on the adaptive global maximum pooling operation, the spatial features along the vertical and horizontal directions are aggregated into two separate one-dimensional vectors respectively; then the two one-dimensional vectors are spliced ​​in the spatial dimension, the channels are compressed through the convolution layer, and the spatial information in the vertical and horizontal directions is encoded through AdaBN normalization and Meta-ACON activation function; Then, the Split operation is performed to obtain the horizontal and vertical feature information, and the attention weights in the two directions are obtained through two convolutional layers respectively; Finally, the attention weights in the two directions are multiplied by the input spatial features to obtain the output features.

5. The fault diagnosis method based on cloud-edge intelligent collaboration mechanism and lightweight module design according to claim 1 is characterized in that: The cloud aggregates the gradient parameters of the edge node model training to obtain the final inference model, and sends it to each edge node for fault diagnosis, including the following steps: The edge node and the cloud establish a communication connection based on the key generation algorithm; In each round of communication between the edge node and the cloud, the heterogeneous network optimized federated learning algorithm is used to train the final gradient parameters, and then the final inference model is obtained and sent to each edge node for fault diagnosis.

6. The fault diagnosis method based on cloud-edge intelligent collaboration mechanism and lightweight module design according to claim 5 is characterized in that: The key generation algorithm comprises the following steps: For edge node j, j = 1, ... K, construct an elliptic curve E based on the given a, b values, and then select a random number v j As the private key, then choose a point G on the elliptic curve E j As a generator, generate the public key Q j =v j G j ; Edge node j will send the public key group {E:y 3 =x 3 +ax+b,Q j ,G j } is sent to the cloud, where x and y represent the horizontal and vertical coordinates of the elliptic curve E, and a and b are coefficients; For the cloud, based on the given a, b, p values, an elliptic curve E is constructed, a random number v0 is selected as the private key, and a point G0 is selected on the elliptic curve E as the generator to generate the public key Q0 = v0G0. The cloud server will 3 =x 3 +ax+b,Q0,G0} is sent to each edge node.

7. The fault diagnosis method based on cloud-edge intelligent collaboration mechanism and lightweight module design according to claim 5 is characterized in that: The heterogeneous network optimization federated learning algorithm includes the following steps: ① The cloud first selects K edge nodes according to the probability P k , randomly select N clients, the set is S t ; ② The cloud's communication parameter w for this round t The code is M, a random number I is selected, and M is encrypted for each edge node j; the encrypted ciphertext group is (R j ,S j ), where R j =IS j ,S j =M+IQ j ,Q j The public key generated for edge node j sends the encrypted ciphertext group to the selected edge nodes in turn; ③Edge node j uses the local private key to encrypt the ciphertext group (R j ,S j ) decrypted to get M, M = S j -v j R j , then decode M to get the original gradient parameters, where v j is the private key; ④ Each edge node obtains w by optimizing the objective function k t+1 : In k t+1 =argmin w h k (w;w t )=F k (w)+μ / 2||ww t ||2 Among them, F k (w) is the objective function in the fedavg algorithm, h k (w; w t ) is the objective function in the fedfrox algorithm, μ is the proportional coefficient, w represents the gradient parameter obtained in this round of training, w k t+1 is to optimize h k (w; w t ) and the solution obtained after ⑤ Each edge node will get w k t+1 The data is sent back to the cloud, where the cloud aggregates the w of each edge node using a weighted average method. k t+1 Get the initial parameter w for the next round t+1 , return to step ②; ⑥ When the set number of iterations is reached, the obtained w T As the final gradient parameter, it is used to update the BiGRU-CA-DSCNN fault diagnosis model to obtain the final inference model.

8. The fault diagnosis method based on cloud-edge intelligent collaboration mechanism and lightweight module design according to claim 1 is characterized in that: The fault diagnosis is specifically as follows: The collected original time-domain vibration signal is input into the final reasoning model to obtain the fault diagnosis classification.

9. The fault diagnosis system based on cloud-edge intelligent collaboration mechanism and lightweight module design is characterized by: include: Edge nodes are used for model training and fault diagnosis based on the final inference model sent from the cloud. The cloud is used to aggregate the gradient parameters of the model training of each edge node and obtain the final inference model, which is then sent to each edge node for fault diagnosis.

Images