Cloud-edge collaborative bearing fault detection method and related product

Abstract

The invention relates to the technical field of bearing fault detection, in particular to a cloud-edge collaborative bearing fault detection method and a related product, and the method comprises the steps: carrying out the preprocessing of collected signal data of a bearing in different states; performing feature extraction on the preprocessed signal data by using CNN, inputting the extracted features into a GRU model, and generating a preliminary fault detection result based on the output of the GRU model; according to the cloud detection method, fusion processing and feature extraction are carried out on collected signal data, a local binary residual network model is established, result conversion is carried out on an output result of the local binary residual network model through a multi-layer perceptron and a hidden layer, and a final bearing fault detection result is generated. The bearing fault detection method effectively solves the technical problems that the maintenance time is too long, the production process is influenced and the maintenance cost is increased due to the fact that the bearing fault detection method in the prior art cannot judge the bearing fault type.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault detection, and in particular to a cloud-edge collaborative bearing fault detection method and related products. Background Art

[0002] Bearings are components that fix and reduce the load friction coefficient during mechanical transmission. Their main function is to support mechanical rotating bodies, reduce the friction coefficient during their movement, and ensure their rotation accuracy. The quality of bearings directly affects the performance of related equipment and components. However, existing bearing detection methods have low accuracy and high costs.

[0003] In this regard, a patent document with application publication number CN110823569A discloses a bearing fault detection method, in which the inner ring and outer ring of the bearing are respectively connected to a transmission shaft and a housing, and the transmission shaft and the housing are electrically connected to a detection circuit. The detection circuit is used to detect whether the oil film resistance of the bearing is maintained in a stable state, and the transmission shaft is rotated to drive the inner ring to rotate. When the oil film resistance is maintained in a stable state, the electrical signal displayed by the display device connected to the detection circuit is maintained in a stable state, and it is determined that the bearing has not failed. When the oil film resistance changes, the electrical signal displayed by the display device changes suddenly, and it is determined that the bearing has failed.

[0004] It can be seen that the above method determines whether the bearing is faulty by detecting the oil film resistance. The steps are simple and the requirements for the detection equipment are low, which reduces the cost of bearing fault detection. However, the above method can only determine whether the bearing is faulty, but cannot determine what kind of fault the bearing has. Maintenance personnel are required to inspect the bearing after the equipment is shut down to determine the type and cause of the fault, and then perform maintenance, resulting in a long downtime of the equipment, which will not only affect the production plan and progress, but also increase additional maintenance costs. Summary of the invention

[0005] The present invention provides a cloud-edge collaborative bearing fault detection method and related products to solve the technical problem that the bearing fault detection method in the prior art cannot determine the type of bearing fault, resulting in too long maintenance time, affecting the production process and increasing maintenance costs.

[0006] To solve the above problems, the cloud-edge collaborative bearing fault edge detection method provided by the present invention adopts the following technical solutions: A cloud edge bearing fault edge end detection method, comprising: The edge end collects signal data of the bearing in normal and fault conditions through multiple sensors installed on the bearing, and pre-processes the collected signal data; CNN is used to extract features from the preprocessed signal data and the extracted features are input into the GRU model to learn the temporal dynamic features of the signal data; Based on the output of the GRU model, preliminary fault detection results are generated.

[0007] The beneficial effect of the cloud-edge collaborative bearing fault edge detection method provided by the present invention is: by collecting signal data from multiple sensors installed on the tested bearing in normal states and multiple fault states, and preprocessing the collected signal data, the fault detection results are output based on the preprocessed signal data, CNN and GRU models. It combines the advantages of CNN in feature extraction and the ability of the GRU model in processing time series data, which can not only capture the local features of the signal, but also understand the changes of these features over time, so that different fault states of the bearing can be effectively and accurately identified.

[0008] Furthermore, the edge end collects signal data of the bearing in normal state and fault state through sensors installed on the bearing, and pre-processes the collected signal data, including: The signal data transmitted by the sensor is divided into training data, verification data and test data in a ratio of 8:1:1; The signal data in the data set is segmented, and the signal data in the original training data and test data are divided into subsequences of fixed length. At the same time, the overlapping sampling strategy is used to enhance the training samples, and each segment of the signal data is normalized to obtain the normalized signal data.

[0009] Furthermore, CNN is used to extract features from the preprocessed data and the extracted features are input into the GRU model, including: The preprocessed signal data is used as the input of the first-layer wide convolutional neural network to obtain the feature sequence of the training data; The fault feature sequence obtained after the downsampling operation is input into the GRU model for further feature extraction.

[0010] Beneficial effects: Using a wide convolutional neural network can retain more features contained in the data, deepen the number of network layers, and enhance the generalization ability of the model.

[0011] Furthermore, based on the output of the GRU model, a preliminary fault detection result is generated, including: classifying the features extracted by the GRU model and using them as the input of the fully connected layer for fully connected output, the hidden layer of the GRU model uses the ReLU activation function, and the final fully connected layer uses the Softmax activation function.

[0012] Furthermore, the GRU model includes an update gate and a reset gate. The update gate is used to control the amount of information that the current state needs to retain from the historical state and the amount of new information that needs to be received from the candidate state. The reset gate is used to control whether the calculation of the candidate state depends on the state at the previous moment.

[0013] Beneficial effects: GRU is used to improve the long short-term memory network (LSTM). The forget gate and input gate in the traditional LSTM are merged into an update gate, and the output gate is used as the reset gate. No additional internal state is introduced, and a linear dependency relationship is directly introduced between the current state and the historical state. GRU has fewer parameters, faster convergence speed, and less time, which can speed up the iteration process.

[0014] Furthermore, the operation formula of the update gate is: , the operation formula of the reset gate is: , where: - Updated doors, - activation function, - Update the gate weight matrix 1, - input vector, - Update the gate weight matrix 2, -history status, - Update the gate bias vector, - the activation vector of the forget gate, -The weight matrix of the forget gate is 1, -The weight matrix of the forget gate 2, -Bias vector of forget gate; candidate state The calculation formula is: , current status The calculation formula of the hidden state is: , where - Candidate status, - hyperbolic tangent function, -The weight matrix of the current state is 1, - input vector, -The weight matrix 2 of the current state, - Updated doors, -history status, - the bias vector of the current state, - Current status, - Activation vector of the forget gate.

[0015] To solve the above problems, the cloud-edge collaborative bearing fault edge detection system provided by the present invention adopts the following technical solutions: A cloud-edge collaborative bearing fault edge detection system, comprising: Edge processor; An edge-end memory stores computer program instructions, which implement the cloud-edge collaborative bearing fault edge-end detection method when the computer program instructions are executed by the edge-end processor.

[0016] The beneficial effect of the cloud-edge collaborative bearing fault edge detection system provided by the present invention is: by adopting the above-mentioned technical solution, the above-mentioned cloud-edge collaborative bearing fault edge detection method is generated into a computer program and stored in the edge memory, so that it can be loaded and executed by the edge processor, thereby making a terminal device based on the edge memory and the edge processor, which is convenient to use.

[0017] To solve the above problems, the cloud-edge collaborative bearing fault cloud detection method provided by the present invention adopts the following technical solutions: A cloud-edge collaborative bearing fault cloud detection method, comprising: The cloud receives the signal data and preliminary fault detection results from the edge, and performs fusion processing on the signal data; The circular LBP operator is used to extract features from the fused data to capture deeper feature information and establish a local binary residual network model; Perform residual skip connection and pooling operations on the two residual blocks consisting of a BN layer, a ReLU activation layer, and a convolutional layer to enhance the learning ability of the local binary residual network model for fault features; The output results of the local binary residual network model are converted through a multi-layer perceptron and a hidden layer to generate the final bearing fault detection results; Outputting the fault detection result through the Softmax function and comparing it with the preliminary fault detection result to generate a fault analysis report; The local binary residual network model is trained and optimized, and the optimized model parameters are synchronized to the GRU model at the edge to update the GRU model.

[0018] The beneficial effects of the cloud-edge collaborative bearing fault cloud detection method provided by the present invention are: by fusing the signal data received at the edge, and combining the processed signal data with the local binary pattern and the residual network for feature extraction, jump connection and pooling operations, the learning ability of the model established in the edge for fault characteristics is enhanced, which can effectively reduce the cost of the enterprise's own server construction and model training and data storage. The trained model with relatively simple input data type (1D) is stored in the local data center for preliminary diagnosis and alarm, and the 2D model is trained and stored in the cloud. The abnormal data in the preliminary diagnosis results transmitted from the edge is received for further precise diagnosis and fault classification, which can greatly improve the accuracy of fault status judgment.

[0019] Furthermore, the signal data is fused and the circular LBP operator is used to extract features from the fused data to capture deeper feature information. The local binary residual network model is established, including: Perform normalization preprocessing on the signal data and convert the one-dimensional time domain signal into a two-dimensional grayscale image; Exploiting LBP8 1 The improved operator is used as a preprocessing method for feature extraction. A circle with a variable radius is used to encode neighboring pixels. The LBP feature map is obtained after calculation for each pixel. The calculated LBP feature map is used as the input of the subsequent residual convolutional neural network to establish a local binary residual network model.

[0020] Beneficial Effects: LBP8 with round shape 1 operator, extending the traditional neighborhood to any neighborhood and replacing the square neighborhood with a circular neighborhood, so that the improved LBP8 1 The operator can have any number of pixels within the circular neighborhood.

[0021] To solve the above problems, the cloud-edge collaborative bearing fault cloud detection system provided by the present invention adopts the following technical solutions: A cloud-edge collaborative bearing fault cloud detection system, comprising: Cloud processor; A cloud storage device stores computer program instructions, which, when executed by the cloud processor, implements the cloud-edge collaborative bearing fault cloud detection method.

[0022] The beneficial effect of the cloud-edge collaborative bearing fault cloud detection system provided by the present invention is: by adopting the above-mentioned technical solution, the above-mentioned cloud-edge collaborative bearing fault cloud detection method is generated into a computer program and stored in the cloud storage device to be loaded and executed by the cloud processor, thereby making a terminal device based on the cloud storage device and the cloud processor for easy use.

[0023] The beneficial effects of the cloud-edge collaborative bearing fault detection method and related products provided by the present invention are: by using common sensors in the prior art to collect signal data of bearings in normal states and multiple abnormal states, and processing and analyzing the signal data through CNN, GRU models, ReLU activation function, Softmax function, maximum pooling operation and residual jump connection, and optimizing the GRU model, it can accurately determine whether the bearing is in an abnormal state and what abnormal state it is in. Enterprises do not need to purchase too much additional mechanical equipment to detect bearing fault status, while ensuring the timeliness in actual fault diagnosis, reducing the company's detection costs.

[0024] In summary, the present invention effectively solves the technical problem that the bearing fault detection method in the prior art cannot determine the bearing fault type, resulting in excessively long maintenance time, affecting the production process and increasing maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein: Figure 1 Flowchart of the cloud-edge collaborative bearing fault edge detection method and cloud-based detection method provided by the present invention. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Those skilled in the art should know that the embodiments described below are part of the embodiments of the present disclosure, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0027] It should be noted that the main idea of ​​the present invention is to collect signal data of the bearing in normal and abnormal states through common sensors such as acceleration sensors, temperature sensors and acoustic emission sensors set on the tested bearing, and analyze and process the collected signal data through the edge detection system and the cloud detection system, and output a detection report, optimize the GRU model established on the edge, so as to improve the accuracy of fault detection, so that timely measures can be taken according to the fault detection results, reduce the probability of bearing failure, and extend the service life of the bearing.

[0028] After introducing the basic principle of the present invention, various non-limiting embodiments of the present invention are described in detail below. The number of any element in the drawings is for illustration and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0029] The principle and spirit of the present invention are explained in detail below with reference to several representative embodiments of the present invention.

[0030] Embodiment 1 of the cloud-edge collaborative bearing fault edge detection method provided by the present invention: like Figure 1 As shown, the Yunbian bearing fault detection method includes: The edge end collects signal data of the bearing in normal and fault conditions through multiple commonly used sensors such as acceleration sensors, temperature sensors and acoustic emission sensors installed on the bearing, and pre-processes the collected signal data; CNN is used to extract features from the preprocessed signal data and the extracted features are input into the GRU model to learn the temporal dynamic features of the signal data; Based on the output of the GRU model, preliminary fault detection results are generated.

[0031] Furthermore, the edge end collects signal data of the bearing in normal state and fault state through sensors installed on the bearing, and pre-processes the collected signal data, including: The signal data transmitted by the sensor is divided into training data, verification data and test data in a ratio of 8:1:1; The signal data in the data set is segmented, and the signal data in the original training data and test data are divided into subsequences of fixed length. At the same time, the overlapping sampling strategy is used to enhance the training samples, and each segment of the signal data is normalized to obtain the normalized signal data.

[0032] Specifically, the process of preprocessing the collected signal data includes three processing steps: data segmentation, data enhancement and normalization.

[0033] First, the signal data in the signal data set is divided into subsequences of length 1024; Then, the overlapping sampling strategy is used to enhance the training data set. That is, when collecting training samples from the original signal, each signal segment overlaps with the next signal segment. This method can use more training samples and enhance the generalization ability of the machine learning model. Finally, the labels of the training set and the test set are one-hot encoded to convert the categorical signal data into numerical data.

[0034] The processing formula used to normalize the signal data is: , where: -Normalized signal data, - original signal data, - minimum value of signal data, - Maximum value of the signal data.

[0035] Furthermore, CNN is used to extract features from the preprocessed data and the extracted features are input into the GRU model, including: The preprocessed signal data is used as the input of the first-layer wide convolutional neural network to obtain the feature sequence of the training data. The CNN part extracts the features of the signal data through a one-dimensional convolutional network, uses the ReLU activation function, and performs downsampling operations with maximum pooling. The fault feature sequence obtained after the downsampling operation is input into the GRU model for further feature extraction.

[0036] Among them, the convolution kernel of the first layer of wide convolutional neural network is a large convolution kernel of 64×1, and the convolution kernel size of other convolutional layers is 3×1. In addition, a BN layer is added after each convolution layer, which makes the loss function space of the neural network smoother and less prone to gradient explosion and gradient disappearance.

[0037] Furthermore, based on the output of the GRU model, a preliminary fault detection result is generated, including: classifying the features extracted by the GRU model and using them as the input of the fully connected layer for full connection output, the hidden layer of the GRU model uses the ReLU activation function, and the final fully connected layer uses the Softmax activation function to convert the input neurons into a probability distribution of 0 and 1, and output the category with the highest probability as the classification result.

[0038] Furthermore, the GRU model includes an update gate and a reset gate. The update gate is used to control the amount of information that the current state needs to retain from the historical state and the amount of new information that needs to be received from the candidate state. The reset gate is used to control whether the calculation of the candidate state depends on the state at the previous moment.

[0039] Furthermore, the operation formula of the update gate is: , the operation formula of the reset gate is: , where: - Updated doors, - activation function, - Update the gate weight matrix 1, - input vector, - Update the gate weight matrix 2, -history status, - Update the bias vector of the gate, - the activation vector of the forget gate, -The weight matrix of the forget gate is 1, -The weight matrix of the forget gate 2, -Bias vector of forget gate; candidate state The calculation formula is: , current status The calculation formula of the hidden state is: , where - Candidate status, - hyperbolic tangent function, -The weight matrix of the current state is 1, - input vector, -The weight matrix 2 of the current state, - Updated doors, -history status, - the bias vector of the current state, - Current status, - Activation vector of the forget gate.

[0040] In the operation of the GRU model, the derivative of the objective function with respect to the last layer of neurons is first solved, and the derivative of the objective function with respect to all the weights is calculated layer by layer from back to front through the chain rule. The loss function of the model adopts the cross entropy loss function, and the gradient descent adopts the Adam optimizer.

[0041] It should be noted that, in this embodiment, CNN represents a convolutional neural network, GRU represents a gated recurrent unit, the BN layer is a normalization layer, the ReLU activation function is a rectified linear unit, and the Softmax function is a normalization function.

[0042] When collecting signal data of bearings in normal operating and fault conditions, based on the structure and operating characteristics of the rolling bearings, the vibration sensor should be installed on the bearing seat or bearing cover, the temperature sensor should be located in an easily measurable position near the bearing, and the acoustic emission sensor can be installed inside or near the bearing.

[0043] To ensure that the sensor signal is stable and can be reliably transmitted to the data acquisition system, shielded cables are used for vibration sensors and acoustic emission sensors to reduce external interference.

[0044] During the detection process, the edge data acquisition and preprocessing module is used to collect vibration, temperature and acoustic emission signals in real time to ensure that the system has sufficient sampling rate and resolution to capture the details of the bearing operation. The edge data acquisition and preprocessing module is connected to an ESP32 high-performance microcontroller at the end close to the sensor. The ESP32 high-performance microcontroller is connected to the vibration sensor, temperature sensor and acoustic emission sensor and processes the data detected by the three sensors. The Wi-Fi module or Bluetooth module of the ESP32 high-performance microcontroller is used to connect to the Internet or other devices, and the read data is simply processed, such as converting the read data into a recognizable format or applying a filtering algorithm to remove noise. Then, the network library of the ESP32 high-performance microcontroller is used to send the processed data to the edge through the HTTP or MQTT protocol.

[0045] When using the cloud-edge collaborative bearing fault edge detection method for detection, the bearing vibration data used comes from the open dataset of Case Western Reserve University (CWRU) in the United States. The data selected from this dataset covers ten fault categories including inner race fault (IRF), rolling element fault (BF), outer race fault (ORF), etc. Each fault diameter covers 0 inches, 0.014 inches and 0.021 inches, a total of nine damage states, and the normal operation state at 0 horsepower motor load and 1797 speed, as shown in Table 1: Table 1

[0046] For the CNN processing part, the first convolution layer uses a large convolution kernel for convolution operation, and then passes through the ReLU activation function to become a set of feature vectors, and then passes through the maximum pooling for downsampling operation. The second layer of convolution uses a small convolution kernel to repeat the above operation once. The structural parameters of the convolutional neural network are shown in Table 2. The feature vector output by the second pooling layer is directly passed to the GRU model, eliminating the process of flattening and reshaping the data dimension.

[0047] Table 2

[0048] For the GRU model, the input data is processed step by step and passes through multiple GRU networks. The output of each stage will be affected by the output of the previous stage. In this way, the GRU network can effectively utilize the time information in the signal and deeply explore the intrinsic characteristics of the signal data. Among them, the hidden state of the first layer of GRU is used as the input of the second layer of GRU, and the hidden state of the second layer of GRU is used as the input of the third layer of GRU. The last layer of GRU does not output the hidden state.

[0049] Afterwards, the neural network is trained by updating parameters through error back propagation. The main approach is: first solve the derivative of the objective function with respect to the last layer of neurons, and then calculate the value of the objective function with respect to all values ​​layer by layer from back to front through the chain rule. Jupiter's loss function uses the cross entropy loss function, and the gradient descent uses the Adam optimizer.

[0050] Next, the cloud-edge collaborative bearing fault edge detection method is further explained through specific data. The CNN part of the following model consists of two convolutional layers and two pooling layers. The specific structural parameters of the model are shown in Table 2. Eight groups of experiments were carried out using the cloud-edge collaborative bearing fault edge detection method. The number of iterations for each group was set to fifty times. The experimental results are shown in Table 3.

[0051] Table 3

[0052] It can be seen that with the increase in the number of training iterations, the loss rate shows a downward trend, while the accuracy gradually increases. The highest accuracy of this model can reach 100%, and the average accuracy is as high as 99.873%, which has excellent performance.

[0053] The model (WDC-GRU) provided by the present invention and four commonly used models, including support vector machine (SVM), one-dimensional convolutional neural network (1D-CNN), two-dimensional convolutional neural network (2D-CNN) and recurrent neural network (RNN), are trained and tested on the CWRU bearing dataset. The results are shown in Table 4.

[0054] Table 4

[0055] It can be seen from Table 4 that the model provided by the present invention exhibits good generalization ability. Its advantage is that it can directly process the original signal without complicated preprocessing and feature extraction process. This design enables the model to better retain the information in the original data, thereby improving the accuracy of fault detection.

[0056] Embodiment 1 of the cloud-edge collaborative bearing fault edge detection system provided by the present invention: A cloud-edge collaborative bearing fault edge detection system includes: an edge processor and an edge memory, the edge memory stores computer program instructions, and when the computer program instructions are executed by the edge processor, a cloud-edge collaborative bearing fault edge detection method is implemented.

[0057] Embodiment 1 of the cloud-edge collaborative bearing fault cloud detection method provided by the present invention: like Figure 1 As shown, the cloud-edge collaborative bearing fault cloud detection method includes: The cloud receives the signal data and preliminary fault detection results from the edge, and performs fusion processing on the signal data; The circular LBP operator is used to extract features from the fused data to capture deeper feature information and establish a local binary residual network model; Perform residual skip connection and pooling operations on the two residual blocks consisting of a BN layer, a ReLU activation layer, and a convolutional layer to enhance the learning ability of the local binary residual network model for fault features; The output results of the local binary residual network model are converted through a multi-layer perceptron and a hidden layer to generate the final bearing fault detection results; The fault detection result is output through the Softmax function, and compared with the preliminary fault detection result output by the above-mentioned cloud-edge collaborative bearing fault edge detection method to generate a fault analysis report; The local binary residual network model is trained and optimized, and the optimized model parameters are synchronized to the GRU model at the edge to update the GRU model.

[0058] Furthermore, the signal data is fused and the circular LBP operator is used to extract features from the fused data to capture deeper feature information. The local binary residual network model is established, including: Perform normalization preprocessing on the signal data and convert the one-dimensional time domain signal into a two-dimensional grayscale image; Exploiting LBP8 1 The improved operator is used as a preprocessing method for feature extraction. A circle with a variable radius is used to encode neighboring pixels. The LBP feature map is obtained after calculation for each pixel. The calculated LBP feature map is used as the input of the subsequent residual convolutional neural network to establish a local binary residual network model.

[0059] In LBP8 1 The improved operator is used as a preprocessing method for feature extraction. When a circle with a variable radius is used to encode neighboring pixels, the original LBP operator is defined in a 3×3 neighborhood window, with the gray value of the pixel at the center of the window as the center. As the threshold, the gray values ​​of the eight adjacent pixels are Compare with it and set the indicator function , the calculation formula of LBP operator is: , where: -Circular LBP eigenvalues; - Gray value of the center pixel; -No. The gray value of each sampling point; -Indicator function, when hour, ,otherwise, .

[0060] When performing residual skip connections and pooling operations to extract features, it is necessary not only to consider the global sample characteristics but also to prevent data loss during training. Therefore, it is necessary to use the residual network for feature extraction. Suppose the initial optimal mapping is , while in the case of nonlinear layers and other mappings After integration, the optimal mapping becomes , even when the residual is 0, the stacking layer only performs identity mapping, which will not cause the network performance to degrade.

[0061] When using the cloud-edge collaboration strategy to optimize the GRU model on the edge, the following operations are required: First, we conduct weight importance evaluation and cloud-edge collaboration preparation. We perform weight significance calculation for each convolution layer in CNN, sort the weights of each convolution kernel, and set a pruning threshold to retain the most important 50% weights in each convolution layer. For GRU, we set the pruning threshold to 60%. At the same time, we prepare a cloud-edge collaboration framework to ensure that the communication protocol between edge devices and cloud servers is consistent, thereby ensuring smooth communication between modules. Then, the cloud-edge collaborative pruning operation is performed. According to the weight importance evaluation results, the pruning operation is performed on the edge, and the weights below the set threshold are reset to zero. The pruning operation is performed synchronously on the cloud. Then fine-tune the model and optimize the cloud-edge collaboration. After the pruning operation, the pruned model is fine-tuned to restore the model performance. By using a smaller learning rate and combining some of the original training data, the model is trained for several rounds to adjust the remaining weights and compensate for the performance loss caused by pruning. In the cloud-edge collaboration framework, the pruning ratio of model parameters and the allocation of computing tasks are dynamically adjusted; Finally, iterative pruning and performance monitoring are performed to re-evaluate the importance of weights and repeat the pruning and fine-tuning process until the size and performance of the model reach the target balance. At the same time, performance indicators on the edge and cloud are monitored in real time for further pruning and optimization.

[0062] It should be noted that, in this embodiment, CNN represents a convolutional neural network, GRU represents a gated recurrent unit, the LBP operator is a visual description operator for texture classification, the BN layer is a normalization layer, the ReLU activation function is a rectified linear unit, and the Softmax function is a normalization function.

[0063] Next, we will further explain the cloud-edge collaborative bearing fault cloud detection method through specific data.

[0064] When normalizing the signal data and converting the one-dimensional time domain signal into a two-dimensional grayscale image, the size of the two-dimensional grayscale image depends on the duration of the signal. , considering The application of the value is specific and needs to be set to ensure that all possible fault vibrations are included in each signal image and that each image contains at least one complete fault vibration. The reasonable minimum value is set to ,in Is the lowest possible frequency of mechanical vibration.

[0065] Assuming that from running The signal data acquired within seconds contains L samples. The first step of the conversion process is to standardize the L samples so that their values ​​are between 0 and the maximum grayscale pixel intensity. Then, these L samples are converted into an M×N grayscale image, where M and N are integers, M×N=L. Among the L samples, the first N samples constitute the first row of the image; next, N samples constitute the second row, and so on, eventually forming a two-dimensional grayscale image.

[0066] In using LBP8 1 The improved operator is used as a preprocessing method for feature extraction. A circle with a variable radius is used to encode neighboring pixels. In the process of calculating each pixel to obtain the LBP feature map, the encoding formula used is: , , where ( , ) is the neighborhood center point, ( , ) is a sampling point, is the circle radius, Indicates a specific point or stage in the current period or cycle. Indicates the total number of cycles or loops.

[0067] After that, the pixel value of the changed point is obtained by bilinear interpolation: The LBP feature map obtained after calculating each pixel point is used as the input of the subsequent convolutional neural network, and on this basis, the features of the nine fault types contained in the test data and the bearing signals in normal state are extracted.

[0068] When the calculated LBP feature map is used as the input of the subsequent residual convolutional neural network to establish a local binary residual network model, the calculated LBP feature map is first used as the input of the subsequent residual convolutional neural network. After two jump connections between two residual blocks, the convolution classification result is finally obtained. The process includes: first, the obtained LBP feature map is input into a residual block composed of two BN layers, two ReLU activation functions and two convolution layers with a kernel of 3×3 for residual feature extraction. After that, two deep convolutional layers and two maximum pooling layers are linked to perform multiple feature extraction, combination and dimensionality reduction operations to extract the hierarchical features of the image and reduce the amount of data while retaining important information. Considering efficiency, robustness and difficulty of implementation, the model uses the Adam optimizer for parameter update and bias correction, sets the optimizer learning rate to 0.0005, and iterates 350 times.

[0069] After passing through the residual block of the network and deep processing, the features are flattened into multi-dimensional features, and then the fully connected layer and the Softmax layer are linked. The fully connected layer is used to integrate and classify the features, and finally obtain the fault type. The Softmax layer converts a dimensional vector into another dimensional real vector, where each element ranges between (0,1), and the sum of the elements in the compressed vector is 1. This can transform the vector into a probability distribution. For any vector The calculation formula is: , where is the base of natural logarithms, Represents the first elements, for a vector Each element in , the Softmax function will calculate a corresponding value.

[0070] When training and optimizing the local binary residual network model, multiple sets of training data are used to detect the model recognition effect, perform hyperparameter tuning and model performance evaluation. Rolling bearings will inevitably experience load changes in actual operation, so the fault diagnosis method must have good generalization ability. For the experimental data set, data sets A, B, C, and D with loads of 0hp, 1hp, 2hp, and 3hp are constructed. Through experimental comparison, it is found that the experimental results under different loads are higher than the other three methods, and the overall accuracy remains above 90%. This is because the addition of the LBP operator and jump connection enables the model to fully extract fault feature information, thereby significantly improving the fault diagnosis accuracy, which also shows that this method has stronger generalization ability.

[0071] The method of combining the LBP operator with the residual convolutional neural network is used to optimize the limitations of the convolutional neural network in bearing fault identification. By adding Gaussian white noise with different signal-to-noise ratios (SNR) to the original fault signal, an experimental comparison is carried out under a load of 1hp. Table 5 shows the recognition accuracy of different methods under different noise conditions. It can be found from Table 5 that with the decrease of the signal-to-noise ratio, the traditional convolutional neural network has an insufficient effect on the processing of noise. The reason is that the feature extraction ability of the fuzzy signal is insufficient. However, the present invention still has a good noise reduction ability and can still maintain an 83.5% recognition accuracy under -3dB noise interference.

[0072] Table 5

[0073] Embodiment 1 of the cloud-edge collaborative bearing fault cloud detection system provided by the present invention: A cloud-edge collaborative bearing fault cloud detection system includes a cloud processor and a cloud storage device. The cloud storage device stores computer program instructions. When the computer program instructions are executed by the cloud processor, a cloud-edge collaborative bearing fault cloud detection method is implemented.

[0074] According to the above description of this specification, those skilled in the art may also understand that the terms used below, such as "up", "down", "front", "back", "left", "right", "width", "horizontal", "top", "bottom", "inside", "outside" and the like, which indicate orientation or positional relationships, are based on the orientation or positional relationships shown in the drawings of this specification, and are only for the purpose of facilitating the explanation of the scheme of the present invention and simplifying the description, rather than explicitly or implicitly indicating that the devices or elements involved must have the specific orientation, be constructed and operated in a specific orientation. Therefore, the above-mentioned orientation or positional relationship terms cannot be understood or interpreted as limitations on the scheme of the present invention.

[0075] In addition, in the description of this specification, “plurality” means at least two, for example, two, three or more, etc., unless otherwise clearly and specifically defined.

Claims

1. A cloud-edge collaborative bearing fault edge detection method, characterized in that: include: The edge end collects signal data of the bearing in normal and fault conditions through multiple sensors installed on the bearing, and pre-processes the collected signal data; CNN is used to extract features from the preprocessed signal data and the extracted features are input into the GRU model to learn the temporal dynamic features of the signal data; Based on the output of the GRU model, preliminary fault detection results are generated.

2. The cloud-edge collaborative bearing fault edge detection method according to claim 1 is characterized in that: The edge end collects signal data of the bearing in normal and fault conditions through sensors installed on the bearing, and pre-processes the collected signal data, including: The signal data transmitted by the sensor is divided into training data, verification data and test data in a ratio of 8:1:1; The signal data in the data set is segmented, and the signal data in the original training data and test data are divided into subsequences of fixed length. At the same time, the overlapping sampling strategy is used to enhance the training samples, and each segment of the signal data is normalized to obtain the normalized signal data.

3. The cloud-edge collaborative bearing fault edge detection method according to claim 1 is characterized in that: Use CNN to extract features from the preprocessed data and input the extracted features into the GRU model, including: The preprocessed signal data is used as the input of the first-layer wide convolutional neural network to obtain the feature sequence of the training data; The fault feature sequence obtained after the downsampling operation is input into the GRU model for further feature extraction.

4. The cloud-edge collaborative bearing fault edge detection method according to claim 1 is characterized in that: Based on the output of the GRU model, generate preliminary fault detection results, including: The features extracted by the GRU model are classified and used as the input of the fully connected layer for full connection output. The hidden layer of the GRU model uses the ReLU activation function, and the final fully connected layer uses the Softmax activation function.

5. The cloud-edge collaborative bearing fault edge detection method according to claim 1 is characterized in that: The GRU model includes an update gate and a reset gate. The update gate is used to control the amount of information that the current state needs to retain from the historical state and the amount of new information that needs to be received from the candidate state. The reset gate is used to control whether the calculation of the candidate state depends on the state at the previous moment.

6. The cloud-edge collaborative bearing fault edge detection method according to claim 5 is characterized in that: The calculation formula of the update gate is: , the operation formula of the reset gate is: , where: - Updated doors, - activation function, - Update the gate weight matrix 1, - input vector, - Update the gate weight matrix 2, -history status, - Update the bias vector of the gate, - the activation vector of the forget gate, -The weight matrix of the forget gate is 1, -The weight matrix of the forget gate 2, - Bias vector of forget gate; Candidate Status The calculation formula is: , current status The calculation formula of the hidden state is: , where - Candidate status, - hyperbolic tangent function, -The weight matrix of the current state is 1, - input vector, -The weight matrix 2 of the current state, - Updated doors, -history status, - the bias vector of the current state, - Current status, - Activation vector of the forget gate.

7. A cloud-edge collaborative bearing fault edge detection system, characterized in that: include: Edge processor; An edge-side memory storing computer program instructions, which, when executed by the edge-side processor, implements the cloud-edge collaborative bearing fault edge-side detection method according to any one of claims 1 to 6.

8. A cloud-edge collaborative bearing fault cloud detection method, characterized in that: include: The cloud receives the signal data and preliminary fault detection results from the edge, and performs fusion processing on the signal data; The circular LBP operator is used to extract features from the fused data to capture deeper feature information and establish a local binary residual network model; Perform residual skip connection and pooling operations on the two residual blocks consisting of a BN layer, a ReLU activation layer, and a convolutional layer to enhance the learning ability of the local binary residual network model for fault features; The output results of the local binary residual network model are converted through a multi-layer perceptron and a hidden layer to generate the final bearing fault detection results; Outputting the fault detection result through the Softmax function and comparing it with the preliminary fault detection result to generate a fault analysis report; The local binary residual network model is trained and optimized, and the optimized model parameters are synchronized to the GRU model at the edge to update the GRU model.

9. The cloud-edge collaborative bearing fault cloud detection method according to claim 8 is characterized in that: The signal data is fused and the circular LBP operator is used to extract features from the fused data to capture deeper feature information. The local binary residual network model is established, including: Perform normalization preprocessing on the signal data and convert the one-dimensional time domain signal into a two-dimensional grayscale image; Exploiting LBP8 1 The improved operator is used as a preprocessing method for feature extraction. A circle with a variable radius is used to encode neighboring pixels. The LBP feature map is obtained after calculation for each pixel. The calculated LBP feature map is used as the input of the subsequent residual convolutional neural network to establish a local binary residual network model.

10. A cloud-edge collaborative bearing fault cloud detection system, characterized in that: include: Cloud processor; A cloud storage device stores computer program instructions, which, when executed by the cloud processor, implements the cloud-edge collaborative bearing fault cloud detection method according to claim 8 or 9.

Images