Bearing fault diagnosis method and device

Abstract

The invention belongs to the technical field of fault diagnosis, and discloses a bearing fault diagnosis method and device. The method effectively solves the problems that an existing method depends on professional experience and is tedious in process, and the diagnosis automation degree is improved; the method can adapt to industrial scenes with limited resources such as an edge computing terminal, meanwhile, the full connection layer can accurately output fault types and fault degrees, an accurate basis is provided for equipment maintenance, and safe and reliable operation of rotating machinery is further guaranteed.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis technology, and in particular to a bearing fault diagnosis method and apparatus. Background Technology

[0002] As a core component of rotating machinery, rolling bearings directly affect the safety and reliability of the equipment. Bearing fault diagnosis technology is an important research direction in the health management of rotating machinery. Existing bearing fault diagnosis methods usually rely on vibration signal analysis, which requires preprocessing the collected vibration signals, such as filtering and denoising, and then manually extracting features such as peak values, root mean square (RMS) or spectral density using methods such as Fourier transform and wavelet transform. Finally, a classifier is used to classify the features to determine the fault type and severity. However, this method has obvious drawbacks: industrial scenarios involve various noises, such as equipment operating noise and environmental interference noise, and manually extracted features are easily affected by noise. Sound masking makes it difficult to effectively separate fault signals from noise, resulting in incomplete or inaccurate feature extraction, which in turn affects diagnostic robustness. The selection of preprocessing methods and the determination of feature types rely on the professional experience of engineers. Parameters need to be readjusted for different bearing models and operating conditions, making the process cumbersome and with low automation, and difficult to adapt to large-scale industrial applications. Some deep diagnostic models have a large number of parameters and high computational complexity, requiring high-performance computing equipment, which is not suitable for resource-constrained scenarios such as edge computing terminals. At the same time, traditional time-series modeling methods are prone to gradient vanishing problems, making it difficult to capture the long-term dependencies of vibration signals and affecting the accuracy of fault severity assessment. Summary of the Invention

[0003] This application provides a bearing fault diagnosis method and apparatus to solve the problems of poor robustness, low automation, and limited adaptability of existing diagnosis methods under high noise conditions.

[0004] In a first aspect, this application provides a bearing fault diagnosis method, the method comprising: Obtain the raw vibration signal of the rolling bearing; The original vibration signal is used to extract features using a convolutional neural network to obtain a feature sequence. The feature sequence is processed by a deep structure optimization network to obtain an optimized feature sequence; The optimized feature sequence is input into a convolutional gated recurrent neural network to obtain temporal feature information; The timing feature information is input into the fully connected layer to obtain the fault diagnosis results of the rolling bearing, wherein the fault diagnosis results include the type and severity of the bearing fault.

[0005] Secondly, this application provides a bearing fault diagnosis device, the device comprising: The first unit is used to acquire the original vibration signal of the rolling bearing; The second unit is used to extract features from the original vibration signal using a convolutional neural network to obtain a feature sequence. The third unit is used to process the feature sequence through a deep structure optimization network to obtain an optimized feature sequence; The fourth unit is used to input the optimized feature sequence into a convolutional gated recurrent neural network to obtain temporal feature information; The fifth unit is used to input the timing feature information into the fully connected layer to obtain the fault diagnosis results of the rolling bearing, wherein the fault diagnosis results include the type and severity of the bearing fault.

[0006] Thirdly, this application provides a readable medium including executable instructions, which, when executed by a processor of an electronic device, cause the electronic device to perform any of the methods described in the first aspect.

[0007] Fourthly, this application provides an electronic device including a processor and a memory storing execution instructions, wherein when the processor executes the execution instructions stored in the memory, the processor performs the method as described in any of the first aspects.

[0008] As can be seen from the above technical solution, this application has the following beneficial effects compared with the prior art: This bearing fault diagnosis method sequentially executes the following steps: acquiring the original vibration signal of the rolling bearing; extracting feature sequences using a convolutional neural network; processing the signal through a deep structure optimization network to obtain an optimized feature sequence; inputting the optimized feature sequence into a convolutionally gated recurrent neural network to obtain temporal feature information; and inputting the temporal feature information into a fully connected layer to obtain a diagnostic result containing the bearing fault type and severity. This achieves an end-to-end fault diagnosis process, eliminating the need for manual preprocessing operations such as filtering and feature selection. It effectively solves the problems of existing methods relying on professional experience and having cumbersome processes, thus improving the automation level of diagnosis. The convolutional neural network can automatically extract high-dimensional noise-resistant features from the original vibration signal. Combined with the advantages of the convolutionally gated recurrent neural network in mitigating gradient vanishing and capturing long-term dependencies in vibration signals through the gating mechanism, it significantly improves the completeness and accuracy of fault feature extraction under high-noise conditions, enhancing diagnostic robustness. The deep structure optimization network reduces the number and redundancy of parameters by optimizing the network structure, lowering the model's computational complexity and enabling the method to adapt to resource-constrained industrial scenarios such as edge computing terminals. Simultaneously, the fully connected layer can accurately output the fault type and severity, providing precise information for equipment maintenance and further ensuring the safe and reliable operation of rotating machinery.

[0009] The further effects of the aforementioned non-conventional preferred method will be explained below in conjunction with specific embodiments. Attached Figure Description

[0010] To more clearly illustrate the embodiments of this application or the existing technical solutions, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A schematic diagram of the test cover page provided in this application; Figure 2 This is a real-time monitoring diagram provided in this application; Figure 3 A schematic diagram of bearing health provided in this application; Figure 4 This is a schematic diagram of the alarm information provided in this application; Figure 5 The fault analysis diagram provided in this application; Figure 6 This is a diagram illustrating the fault diagnosis provided in this application; Figure 7 This is a schematic diagram of the diagnostic process provided in this application; Figure 8 Diagnostic results provided for this application (illustrated diagram) Figure 9 A flowchart illustrating a bearing fault diagnosis method provided in this application; Figure 10 This is a schematic diagram of the structure of a bearing fault diagnosis device provided in this application; Figure 11 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] First, the terminology used in this implementation method is defined as follows: Raw vibration signal: refers to the rolling bearing time-domain vibration signal directly collected by the sensor without any preprocessing operations such as filtering or noise reduction. This signal fully contains equipment operating noise, environmental interference noise, and bearing fault characteristic information.

[0014] Convolutional Neural Network (CNN): A deep neural network structure for automatic feature extraction. In this embodiment, its core module includes convolutional layers, normalization layers, activation function layers, pooling layers, and regularization layers, which can extract high-dimensional features from the original vibration signal and form a feature sequence.

[0015] Deep structure optimization network: A network structure improved based on existing deep networks (such as wide and deep convolutional neural networks, residual networks, etc.). By replacing some convolutional layers / fully connected layers with optimization layers, introducing regularization constraints or sparse structures, the number of parameters and computational complexity are reduced while retaining feature extraction capabilities.

[0016] Convolutional Gated Recurrent Neural Network: A hybrid neural network that integrates spatial feature extraction and temporal modeling capabilities. It consists of convolutional modules and gated recurrent modules, and can process optimized feature sequences and output temporal feature information containing long-term dependencies.

[0017] Fully connected layer: A neural network layer used to map temporal feature information into fault diagnosis results. Through multi-level linear transformation and activation function operations, it outputs the type and degree of bearing fault.

[0018] Temporal feature information: refers to the feature representation obtained after the feature sequence is processed and optimized by the convolutional gated recurrent neural network, which contains the long-term dependence of the rolling bearing vibration signal. It is the core input for classification of fully connected layers.

[0019] Fault diagnosis results: The output of the fully connected layer includes the type of bearing fault (such as inner ring fault, outer ring fault, etc.) and the degree of fault (such as mild, moderate, severe). In some scenarios, it also includes the confidence level of the diagnosis results.

[0020] It should be noted that in this embodiment, "wide and deep convolutional neural network" and "WDCNN" refer to the same network structure, "gated recurrent unit" and "GRU" refer to the same module, and "long short-term memory network" and "LSTM" refer to the same module. The full Chinese names will be used in the following descriptions to ensure readability.

[0021] The overall process of this bearing fault diagnosis method is as follows: First, the original vibration signal of the rolling bearing is acquired. Then, a convolutional neural network is used to extract features from the original vibration signal to obtain a feature sequence. Next, a deep structure optimization network is used to process the feature sequence to obtain an optimized feature sequence. Subsequently, the optimized feature sequence is input into a convolutional gated recurrent neural network to obtain temporal feature information. Finally, the temporal feature information is input into a fully connected layer to obtain a diagnostic result containing the bearing fault type and fault degree. At the same time, the network parameters can be optimized by comparing the diagnostic results with the actual fault condition to further improve the diagnostic performance.

[0022] This process can be accessed via appendix Figure 1 The "Rotating Machinery Fault Diagnosis System" shown is implemented, with attachments. Figure 1 The system's homepage displays the total number of monitored devices (128), the number of devices operating normally (116), the number of devices with warnings (8), and the number of devices with malfunctions (4) in real time. It also supports locating target bearing devices through the "Search Devices" and "Refresh Data" functions, providing a foundation for equipment management in subsequent diagnostic processes.

[0023] The various non-limiting embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0024] See details Figure 9 This application illustrates a bearing fault diagnosis method according to an embodiment of the present application, the method comprising: S101: Obtain the raw vibration signal of the rolling bearing.

[0025] Specifically, the time-domain vibration signal of the rolling bearing during operation can be collected first using a vibration sensor. The time-domain vibration signal includes noise interference and fault characteristics. Then, a random sampling strategy is used to select a signal segment of fixed length from the continuously collected time-domain vibration signal as the original vibration signal of the rolling bearing. The random sampling strategy includes sliding time window sampling or random interval sampling. The fixed length of the signal segment is determined based on the bearing speed and fault characteristic period of the rolling bearing.

[0026] This step aims to acquire signals containing complete noise and fault characteristics, providing foundational data for subsequent feature extraction. The specific implementation is as follows: Signal Acquisition: Time-domain vibration signals (i.e., the initial data source of the original vibration signals) during the operation of the rolling bearing are acquired using piezoelectric vibration sensors. No preprocessing is performed during acquisition to ensure the complete preservation of equipment operating noise (such as motor noise), environmental interference noise (such as vibration interference from other equipment in the workshop), and bearing fault characteristics (such as characteristic frequency signals generated by inner ring wear). The acquired real-time vibration signals can be transmitted via an attached... Figure 2 View the "Real-time Monitoring Diagram" (attached). Figure 2 Taking equipment #128 (spindle bearing, located in workshop 1) as an example, the curve of the vibration signal (unit: mm / s) changing over time in the past hour is shown, and the upper limit of the vibration signal threshold is marked to make it easy to intuitively judge whether there is abnormal fluctuation in the signal.

[0027] Random Sampling: A random sampling strategy is used to select a fixed-length signal segment from the continuously acquired time-domain vibration signals. This signal segment is the "original vibration signal of the rolling bearing" that is ultimately input into the subsequent network. The random sampling strategy can be either sliding time window sampling or random interval sampling. If sliding time window sampling is selected, the window duration is set to 0.5-2 seconds (the specific duration is determined based on the bearing speed and fault characteristic period of the rolling bearing; for example, for a bearing with a speed of 1500 r / min and a fault characteristic period of 0.04 seconds, the window duration is set to 1 second to cover 25 fault characteristic periods), and the window sliding step size is set to 0.2-0.5 seconds. After each sliding step, a signal segment within the window is extracted. If random interval sampling is selected, the start and end times are randomly determined from the continuously acquired signals to ensure that the length of the extracted signal segment is consistent with the window duration of the sliding time window sampling.

[0028] Signal verification: The acquired raw vibration signal can be combined with the attached... Figure 3 Preliminary verification of the "Bearing Health Diagram" is performed, attached. Figure 3 The display shows the health status (normal), health index (89%), temperature (56℃), vibration intensity (2.4mm / s), and noise level (72dB) of device #128. If the vibration intensity exceeds the threshold (as shown in the attached document), further information is provided. Figure 2 If the upper limit of the threshold is 5 mm / s or the noise level is abnormal (e.g., exceeding 80 dB), it is necessary to check whether the sensor is installed securely (e.g., whether it fits the outer ring of the bearing) and whether the acquisition circuit is normal to ensure the validity of the original vibration signal.

[0029] S102: Use a convolutional neural network to extract features from the original vibration signal to obtain a feature sequence.

[0030] In this embodiment, the convolutional neural network includes at least one set of modules including convolutional layers, normalization layers, activation function layers, pooling layers, and regularization layers; the normalization layer is used to unify the feature distribution to improve training stability, the activation function layer adopts a non-linear activation function, and the regularization layer is used to reduce the risk of model overfitting.

[0031] In this embodiment, the standardization layer is a batch standardization layer or a layer standardization layer. Specifically, the standardization layer is used to normalize the features using the statistical mean and variance of the features, and the feature distribution is adjusted by learning parameters after normalization. The activation function is an exponential linear unit or ReLU. The regularization layer is a Dropout layer or an L2 regularization layer.

[0032] This step utilizes multi-level collaboration within a convolutional neural network to automatically extract features from the original vibration signal, avoiding the tedious process of manually designing features. The specific implementation is as follows: Convolutional Neural Network Architecture Design: The convolutional neural network contains at least one set of core modules: "convolutional layer - normalization layer - activation function layer - pooling layer - regularization layer". In this embodiment, three sets of these core modules are connected in series to ensure the gradual extraction of high-dimensional features from the original vibration signal. The functions and parameter settings of each module are as follows: Convolutional layer: A one-dimensional convolutional kernel with a size of 1×16 or 1×32 is used (to adapt to the one-dimensional characteristics of time-domain vibration signals). The number of output channels is set to 64-128 (the more channels, the richer the feature extraction dimensions, but the computational load increases accordingly. This implementation chooses 64 channels to balance performance and efficiency). The original vibration signal is scanned by a sliding window convolution operation. For example, the convolutional kernel slides on the signal with a stride of 1. After each slide, a dot product operation is performed with the corresponding signal segment to generate a feature map reflecting the local vibration mode (such as the impact vibration feature map caused by the fault).

[0033] Standardization Layer: Select either batch standardization or layer standardization. This implementation prioritizes batch standardization. The specific processing logic is as follows: Calculate the statistical mean and variance of all elements in the feature map within the current training batch. Subtract the mean from each element in the feature map and divide by the variance and minimum value ε (which is 1). To obtain normalized feature values, the square root of the sum of the denominators (avoiding zero) is used. Then, the normalized feature values ​​are adjusted by a learnable scaling parameter γ and an offset parameter β (initial values ​​are set to 1 and 0 respectively, and are adaptively adjusted during training) to avoid over-normalization that could lead to the loss of fault features. This operation can reduce the "internal covariate shift" in network training and improve training stability.

[0034] Activation function layer: A non-linear activation function is used, such as the exponential linear unit (ELU) or ReLU. In this implementation, the ELU is chosen. Its function logic is as follows: when the input value is greater than 0, the input value is directly output; when the input value is less than or equal to 0, "α × (the natural exponent e raised to the power of the input value - 1)" is output (where α is a hyperparameter with a value of 1). This activation function generates a non-zero output for negative inputs, which can enhance the network's ability to respond to weak fault signals under high noise and avoid the feature loss problem caused by the ReLU function's "negative input, zero output".

[0035] Pooling layer: Max pooling is used, and the pooling window size is set to 1×2 or 1×4 (1×2 is selected in this implementation). By taking the maximum value of the feature value within the pooling window, the feature map output by the convolutional layer is downsampled. While retaining key features (such as the peak feature of fault impact), the feature map dimension is reduced, thereby reducing the amount of subsequent computation.

[0036] Regularization layer: Choose either a Dropout layer or an L2 regularization layer. In this implementation, we choose a Dropout layer. During the model training phase, a binary mask is randomly generated with a fixed probability of 0.2-0.4 (0.3 is chosen in this implementation). The neuron connection corresponding to "0" in the mask is temporarily disconnected (does not participate in forward propagation and backward gradient calculation), while the neuron connection corresponding to "1" works normally. During the testing phase, the mask is removed, and the output value of all neurons is multiplied by "1 - dropout probability" (i.e., 0.7) to ensure that the output scale is consistent between the training and testing phases. This operation can reduce the network's dependence on specific neurons and effectively control the risk of overfitting.

[0037] Feature sequence generation: The original vibration signal (one-dimensional time domain signal) is processed sequentially through the above three sets of "convolutional layer-normalization layer-activation function layer-pooling layer-regularization layer", and the output is a three-dimensional feature tensor with dimensions [batch_size, seq_len, feature_dim] (where batch_size is the training batch size, set to 32; seq_len is the feature sequence length, set to 64; feature_dim is the feature dimension, set to 64). This three-dimensional feature tensor is the "feature sequence", which can be used for subsequent processing in deep structure optimization networks.

[0038] Feature extraction of the original vibration signal is performed using a convolutional neural network, which involves sequentially processing the signal through convolutional layers, batch normalization layers, activation function layers, and pooling layers to generate a feature sequence without the need for manual filtering or feature selection. In S2, the convolutional neural network uses exponential linear units as activation functions to improve the processing capability for negative inputs and enhance robustness in high-noise environments. The convolutional neural network in S2 also uses larger convolutional kernels to expand the receptive field to capture global information and improve the frequency domain feature extraction capability.

[0039] Specifically, step S102 processes the original vibration signal of the rolling bearing using a convolutional neural network to generate a high-dimensional feature sequence suitable for subsequent time-series modeling. The convolutional neural network consists of multiple layers, including convolutional layers, batch normalization layers, exponential linear unit activation function layers, pooling layers, and dropout layers, which process the input signal sequentially and automatically extract features.

[0040] First, the original vibration signal is input into the convolutional neural network as a one-dimensional time-domain signal. The convolutional layer performs a sliding window convolution operation on the input signal using multiple convolutional kernels to extract local features. Each convolutional kernel scans the signal, generating a corresponding feature map representing the signal's local patterns. The output of the convolution operation is: ; in, The feature value of the m-th output channel of the convolutional layer at position i in the time domain reflects the local vibration mode. The input signal represents the original vibration signal of the rolling bearing, where i+p is the time-domain position, i is the current position, p is the convolution kernel sliding offset, and c is the number of input channels. Let K be the convolution kernel, and C be the filter used for feature extraction in the convolutional layer. K represents the kernel size (width), C represents the number of input channels, m represents the number of output channels, i and m are the output values ​​of the i-th feature map at position i, representing the local features generated by the convolution operation, and p is the index along the width direction of the convolution kernel. As a bias, the convolution result of the m-th feature map is added to adjust the output value.

[0041] To improve robustness under high-noise conditions, the convolutional layer uses a larger convolutional kernel to capture a wider range of signal patterns. The larger convolutional kernel enhances the frequency domain feature extraction capability, which is equivalent to widening the filter window and improving the frequency domain resolution.

[0042] The batch normalization layer normalizes the output feature map of the convolutional layer, reducing internal covariate bias. The batch normalization calculation is as follows: , in, and Here, represents the mean and variance of the mini-batch, respectively, and N is the mini-batch size. For the standardized output, y i The original feature values ​​before standardization are expressed. These are the normalized eigenvalues. This is a learnable scaling parameter used to avoid over-normalization that eliminates useful features. These are learnable offset parameters used to adjust the normalized feature distribution and adapt to subsequent activation functions.

[0043] The activation function layer uses exponential linear units, defined as: ; in, This is a hyperparameter, set to 1. The exponential linear unit maintains linear output for positive values ​​and generates non-zero output for negative values, enhancing the network's ability to process high-noise signals.

[0044] The pooling layer employs max pooling to downsample the feature map, preserving salient features and reducing the computational cost of D. Max pooling is defined as: ;in, For pooling regions, This is the output value.

[0045] The Dropout layer randomly discards some neuron connections during training, and the output is: ;in, Let P be a random mask of a Bernoulli distribution, and P be the dropout probability. Dropout reduces the model's dependence on specific neurons, controlling the risk of overfitting. This is a characteristic sequence.

[0046] Through the above-described hierarchical processing, the convolutional neural network transforms the original vibration signal into a high-dimensional feature sequence, eliminating the need for manual filtering or feature selection. The feature sequence preserves the spatial and frequency domain information of the signal, making it suitable for subsequent time-series modeling and classification.

[0047] S103: The feature sequence is processed by a deep structure optimization network to obtain an optimized feature sequence.

[0048] In this embodiment, the deep structure optimization network is obtained by replacing some convolutional or fully connected layers in the original network with optimized layers and introducing regularization constraints or sparse structures. The deep structure optimization network is based on improvements to wide-depth convolutional neural networks, residual networks, or dense networks. The sparse structures include grouped convolutions, depthwise separable convolutions, or neuron cluster decomposition structures. The regularization constraints include L1 regularization, L2 regularization, or elastic network regularization.

[0049] This step aims to reduce the number of network parameters and computational complexity, making the method adaptable to resource-constrained scenarios such as edge computing terminals, while preserving fault characteristics in the feature sequence. The specific implementation is as follows: Basic design of deep structure optimization networks: Deep structure optimization networks are derived from wide-depth convolutional neural networks, residual networks, or dense networks. This implementation chooses to improve upon wide-depth convolutional neural networks. The core design logic is "preserving effective feature extraction capabilities + reducing parameter redundancy". The first three wide convolutional layers of the wide-depth convolutional neural network are retained (consistent with the convolutional layer structure of the convolutional neural network, using 1×16 convolutional kernels and 64 output channels) to ensure the ability to extract global features from feature sequences; The last two convolutional layers and fully connected layers in the latter half of the wide-depth convolutional neural network are replaced with "optimization layers". These optimization layers use a sparse weight matrix with regularization constraints for linear transformation. The optimization objective of the weight matrix is: ; Among them, among them, The weight matrix of the optimization layer represents the weights of neuron connections. Let be the cross-entropy loss function, representing the error between the predicted output and the true label. The regularization constraint norm of the weight matrix is ​​used to reduce unnecessary connections; This is the default value.

[0050] By using L1 regularization constraints to promote the sparsity of the weight matrix and reduce unnecessary connections, the number of network parameters is reduced by 30%-50% compared to the original wide and deep convolutional neural network.

[0051] Introduction of sparse structure: To further reduce parameter redundancy, a sparse structure is introduced into the deep structure optimization network. The sparse structure can be selected from grouped convolution, depthwise separable convolution, or neuron cluster decomposition structure. This implementation chooses the neuron cluster decomposition structure (i.e., the specific implementation of grouped convolution): The dense convolutional layer in the optimization layer is decomposed into 4-8 independent neuron clusters on average according to the number of input channels (4 clusters are selected in this implementation). Each neuron cluster processes only the features of the corresponding channel subset (e.g., when the number of input channels is 64, each cluster processes the features of 16 channels). Each neuron cluster is configured with an independent convolution kernel (size 1×3, number of output channels 16). After each cluster completes the convolution operation independently, the output feature maps are concatenated along the channel dimension to form a complete feature map (the number of channels after concatenation is still 64). This operation can reduce redundant connections between neurons, reduce the model's over-reliance on training data, and further control the risk of overfitting.

[0052] Optimized feature sequence output: After the feature sequence is processed by the deep structure optimization network through "wide convolutional layer-optimization layer-neuron cluster decomposition", the output is a three-dimensional feature tensor with dimensions [batch_size, seq_len, optimized_feature_dim] (where optimized_feature_dim is the optimized feature dimension, set to 32). This tensor is the "optimized feature sequence", which has a more compact dimension and lower parameter redundancy, and can be directly input into the subsequent convolutional gated recurrent neural network.

[0053] By optimizing the convolutional layers of a deep structure network using a locally sparse structure, parameter redundancy is reduced and the risk of overfitting is mitigated. The locally sparse structure is achieved by decomposing dense convolutional layers into multiple clusters of neurons, each of which independently processes a subset of the input features.

[0054] The locally sparse structure employs a grouped convolution mechanism. The input feature sequence is a multidimensional tensor; the grouped convolution divides the input channels into G groups, and each group independently applies the convolution operation. The output is:

[0055] in: For the g-th input feature subset, k g Let b be the convolution kernel of the i-th group. g This is the bias for the g-th group.

[0056] After grouped convolution, the output feature maps of each group are concatenated to form a complete feature map. The number of groups G is usually set to 4 or 8, depending on the number of input channels and computational resources. Local sparse structures reduce redundant connections between neurons, lowering the model's over-reliance on training data and thus controlling the risk of overfitting.

[0057] The processed feature sequences retain key feature information, have more compact dimensions, and are suitable for time series modeling.

[0058] S104: Input the optimized feature sequence into a convolutional gated recurrent neural network to obtain temporal feature information.

[0059] In this embodiment, the convolutional gated recurrent neural network is used to extract spatial features and perform temporal modeling; the convolutional gated recurrent neural network consists of a convolutional module and a gated recurrent module; the convolutional module is used to enhance the spatial features of the optimized feature sequence, and the gated recurrent module is used to control the transmission and forgetting of temporal information through a gating mechanism.

[0060] In this embodiment, the gated recurrent module is a gated recurrent unit (GRU) or a long short-term memory network (LSTM). The gating mechanism includes at least one of an update gate, a reset gate, and a forget gate. The convolutional gated recurrent neural network also integrates an attention mechanism. The convolutional gated recurrent neural network is used to obtain key temporal features by assigning differentiated weights to each time step of the feature sequence, and to use the key temporal features as the temporal feature information.

[0061] In this embodiment, the convolutional gated recurrent neural network also integrates a normalization layer and a regularization layer. The normalization layer is set at the output of the convolutional module or the input of the gated recurrent module in the convolutional gated recurrent neural network, and the regularization layer is set at the output of the gated recurrent module.

[0062] This step utilizes the "spatial feature enhancement + temporal modeling" capabilities of convolutional gated recurrent neural networks to capture the long-term dependencies of vibration signals, thus solving the gradient vanishing problem in traditional temporal modeling. The specific implementation is as follows: Convolutional Gated Recurrent Neural Network Architecture Design: The convolutional gated recurrent neural network consists of convolutional modules and gated recurrent modules, and integrates normalization layers, regularization layers, and attention mechanisms. The functions and implementations of each part are as follows: Convolutional module: A 1×8 one-dimensional convolutional kernel (32 output channels) is used to perform secondary spatial feature extraction on the optimized feature sequence to enhance the spatial discriminativeness of fault features (such as further separating the features of inner circle faults from those of outer circle faults); the output feature map dimension of the convolutional module is [batch_size, seq_len, 32].

[0063] Normalization layer: Set at the output of the convolution module, using a batch normalization layer (processing logic is the same as in step (II)) to normalize the feature map output by the convolution module, reduce the impact of feature distribution differences on subsequent temporal modeling, and improve training stability.

[0064] Attention mechanism: Integrated between the normalization layer and the gated loop module, used to highlight key temporal features. First, the standardized feature maps are mapped to single-value scores (with dimensions [batch_size, seq_len, 1]) through a fully connected layer. Then, the individual scores are normalized using the softmax function to obtain the attention weights at each time step (the sum of the weights is 1). The feature map is multiplied element by element with the corresponding attention weight to obtain a weighted feature sequence (the feature weights of key time steps are higher, such as the feature weights of the time when the fault occurs can reach more than 0.8). This weighting operation can enhance the key temporal features for fault diagnosis and improve the feature discrimination under high noise conditions.

[0065] Gated Recurrent Unit (GRU): This module processes the temporal information of the weighted feature sequence. It can be either a gated recurrent unit (GRU) or a long short-term memory (LSM) network. This implementation chooses a GRU (which effectively balances performance and computational cost). Its core function is to control the transmission and forgetting of temporal information through update and reset gates. The specific calculation process is as follows: Update gate calculation: ;in, Update the gate output at time t. For input The weight matrix of the update gate measures the impact of the current input on the update gate. Let be the input features at time t. Hidden in history To update the weight matrix of the gate, This is the bias term for resetting the gate. The input to the update gate is "the weighted features of the current time step + the hidden state of the previous time step", and the output z of the update gate is calculated using the sigmoid activation function (output range 0-1).t , z t The closer W is to 1, the more historical information is retained; the closer W is to 0, the more historical information is discarded. xz W is the weight matrix (32×32) from the current feature to the update gate. hz Let b be the weight matrix (32×32) from the previous hidden state to the update gate. z To update the gate bias (dimension 32), x t h is the weighted feature for the current time step. t-1 Let σ be the hidden state of the previous time step, and let σ be the sigmoid function.

[0066] Reset gate calculation: The input to the reset gate is the same as that of the update gate, and the output r of the reset gate is calculated using the sigmoid activation function. t r t The closer the value is to 1, the greater the influence of the current feature on the hidden state; the closer it is to 0, the smaller the influence of the current feature on the hidden state. The calculation formula is: r t = σ(W xr ·x t + W hr ·h t-1 + b r ), where W xr W hr The current feature x t Previous hidden state h t-1 The weight matrix to the reset gate (all dimensions are 32×32), b r To reset the gate offset (dimension 32).

[0067] Candidate hidden state calculation: Combining the reset gate output and the current features, the candidate hidden state is calculated using the tanh activation function (output range -1 to 1). The calculation formula is: W xh W hh The current feature x t , Reset the hidden state h after weighting the door t-1 The weight matrix to the candidate hidden state (all dimensions are 32×32), b h The candidate hidden state bias (32-dimensional), where ⊙ represents element-wise multiplication. Let be the candidate hidden state at time t.

[0068] Final hidden state calculation: Combine the update gate output and the candidate hidden states to calculate the final hidden state h at the current time step. t The calculation formula is: h t = (1 - z t)⊙h t-1 + z t ⊙ This formula indicates that updating the gate balances the historical hidden state and the candidate hidden state, effectively alleviating the gradient vanishing problem; h t Let (1 - z) be the final hidden state at time t. t ) is used to update the complementary value of the gate.

[0069] Regularization layer: Set at the output of the gated loop module, using a Dropout layer (dropout probability of 0.2) to randomly discard the output of some hidden state units, reducing the model's dependence on specific temporal features and enhancing generalization ability.

[0070] Temporal feature information output: The final hidden state sequence of the gated recurrent module (with dimensions [batch_size, seq_len, 32]) is processed by a regularization layer, and the output is a one-dimensional feature vector with dimensions [batch_size, 32]. This vector is the "temporal feature information", which contains the long-term dependency of the rolling bearing vibration signal and is the core input for classification of the fully connected layer.

[0071] The processing procedure for this step can be found in the attached document. Figure 5 A "Fault Prediction Analysis Diagram" is attached to aid in verification. Figure 5 The graph shows the change in failure probability over the past 7 days and the performance of the diagnostic model (accuracy 98.2%, recall 97.8%). If the extraction of time-series feature information is effective, the failure probability curve predicted by the model will show a significant upward trend as the failure approaches (see attached graph). Figure 5 The probability of failure "tomorrow" is 5% higher than that of "today," providing support for the accuracy of subsequent diagnostic results.

[0072] S105: Input the timing feature information into the fully connected layer to obtain the fault diagnosis result of the rolling bearing.

[0073] The fault diagnosis results include the type and severity of bearing faults.

[0074] In this embodiment, the fully connected layer can be used to perform multi-level linear transformations and activation function operations on the time-series feature information to obtain the fault diagnosis results of the rolling bearing. The output of the fully connected layer uses either a softmax activation function or a sigmoid activation function. The bearing fault types include at least one of inner ring fault, outer ring fault, rolling element fault, and cage fault, and the fault severity includes at least two of mild, moderate, and severe faults.

[0075] In one implementation of this embodiment, after S105, the method further includes: The diagnostic performance index is calculated by comparing the fault diagnosis results of the rolling bearing with the actual fault state. Based on the performance metrics, adjust the network structure parameters or training parameters of the convolutional neural network, the deep structure optimization network, the convolutional gated recurrent neural network, and the fully connected layer until the performance metrics of the convolutional neural network, the deep structure optimization network, the convolutional gated recurrent neural network, and the fully connected layer meet the preset requirements.

[0076] The performance metrics include at least one of diagnostic accuracy, recall, and F1 score.

[0077] This step utilizes the classification capabilities of fully connected layers to map temporal feature information into intuitive fault diagnosis results, while also supporting model parameter optimization. The specific implementation is as follows: Fully connected layer structure and classification process: A fully connected layer consists of 2 hidden layers and 1 output layer. The functions and implementation of each layer are as follows: The first hidden layer maps temporal feature information (32-dimensional) to a 64-dimensional feature space, using the ReLU activation function (to enhance non-linear expressive power), with a weight matrix dimension of 32×64 and a bias dimension of 64.

[0078] The second hidden layer further maps the 64-dimensional features to a 32-dimensional feature space, also using the ReLU activation function, with a weight matrix dimension of 64×32 and a bias dimension of 32.

[0079] Output layer: The softmax activation function (suitable for multi-class scenarios) is used to map the 32-dimensional features to the probability distribution of fault categories. The output dimension is "number of fault types × number of fault severity" (in this embodiment, the fault types include four categories: inner race fault, outer race fault, rolling element fault, and cage fault, and the fault severity includes three categories: mild, moderate, and severe, so the output dimension is 12). The combination of categories with the highest output probability is the final "fault diagnosis result", and the confidence level of the category can also be output (e.g., the confidence level of inner race fault - moderate is 98.7%).

[0080] Fault diagnosis results display and application: Fault diagnosis results can be displayed through attachments. Figure 4 "Alarm Information Diagram" and Appendix Figure 6 "Fault Diagnosis Diagram" shown: As attached Figure 4As shown, when the diagnostic result of equipment #125 (transmission bearing, located in workshop 2) is "inner ring fault - moderate", the system will generate an alarm message within 10 minutes, prompting "Inner ring wear of equipment #125 transmission bearing is present, it is recommended to stop the machine immediately for maintenance"; when the temperature of equipment #126 (conveyor bearing) exceeds the threshold, a "temperature abnormality" warning will be generated, prompting "please pay attention to monitoring"; when the vibration value of equipment #118 (motor bearing) is too high, a "vibration abnormality" warning will be generated, prompting "it is recommended to increase the monitoring frequency", so as to realize real-time alarm and maintenance guidance for faults.

[0081] As attached Figure 6 As shown, after selecting the target device (e.g., "Equipment #4 - No. 1 Cooling Tower Bearing"), setting the diagnostic time range (e.g., "2023-06-14 to 2023-06-15"), and selecting the diagnostic model (e.g., "Convolutional Neural Network Model"), clicking "Start Diagnosis" will generate a diagnostic results overview, including diagnostic status (fault exists), fault level (medium), confidence level (98.7%), and suggested processing time (within 24 hours). Simultaneously, the time-domain waveform, frequency-domain waveform, and fault characteristic parameters of the vibration signal can be viewed (e.g., RMS is 6.8 mm / s, exceeding the threshold of 5 mm / s; Peak is 9.2 mm / s, exceeding the threshold of 7 mm / s), providing detailed evidence for fault cause analysis.

[0082] Model parameter optimization: To further improve diagnostic performance, network parameters can be optimized through the following steps: Performance index calculation: The fault diagnosis results output by the fully connected layer are compared with the actual fault status (such as the bearing fault situation detected by disassembly) to calculate the diagnostic performance index. The performance index includes diagnostic accuracy (number of correctly diagnosed samples / total number of samples), recall (number of correctly diagnosed fault samples / actual number of fault samples), and F1 score (harmonic mean of accuracy and recall, calculated as 2 × (accuracy × recall) / (accuracy + recall)). In this embodiment, the preset performance index thresholds are: accuracy ≥ 95%, recall ≥ 94%, and F1 score ≥ 94.5%.

[0083] Parameter Adjustment: If any performance metric fails to reach the threshold, adjust the network structure parameters or training parameters in reverse. For example, if the recall is low (e.g., only 92%), increase the kernel size of the convolutional neural network (e.g., from 1×16 to 1×32) to enhance the fault feature extraction capability, or reduce the dropout probability of the Dropout layer (e.g., from 0.3 to 0.2) to reduce fault feature loss. If the accuracy is low (e.g., only 93%), increase the number of neuron clusters in the deep structure optimization network (e.g., from 4 to 8) to reduce the risk of overfitting, or increase the regularization coefficient λ (e.g., from 0.01 to 0.02) to reduce redundant parameters. After adjustment, retrain the network until all performance metrics meet the preset requirements.

[0084] Appendix Figure 7 "Diagnosis Execution Diagram" and Appendix Figure 8 The "Diagnostic Result Diagram" illustrates the diagnostic effect after parameter optimization: (See attached diagram) Figure 7 The confidence level of the diagnostic results reached 98.7%, with attached... Figure 8 The fault type analysis showed that the probability of "inner ring failure" was 98.7%, while the probabilities of "outer ring failure", "rolling element failure" and "cage failure" were only 0.9%, 0.3% and 0.1%, respectively. Detailed maintenance steps (such as "stop the machine and disconnect the power supply → disassemble the bearing assembly → replace the damaged inner ring") and preventive measures (such as "increase the frequency of lubricating oil changes and optimize load distribution") were provided, verifying the accuracy and practicality of the diagnostic results after parameter optimization.

[0085] To verify the effectiveness of this bearing fault diagnosis method, diagnostic tests were conducted on 128 rolling bearing devices (covering spindle bearings, driven bearings, and transmission bearings) under high-noise conditions (noise level 70-80dB). The test results are as follows: Diagnostic performance: The diagnostic accuracy reaches 98.2% (10%-15% improvement over existing manual feature extraction methods), the recall rate reaches 97.8% (8%-12% improvement over existing methods), and the F1 score reaches 98.0% (9%-13% improvement over existing methods). Moreover, it can still maintain an accuracy of over 95% at a noise level of 80dB, verifying the robustness of the method under high noise conditions.

[0086] Adaptability: The deep structure optimization network reduces the number of model parameters by 42% and the computational complexity by 38% compared to traditional deep networks. It can achieve real-time diagnosis on edge computing terminals (such as industrial PLCs) (diagnosis latency < 0.5 seconds), adapting to resource-constrained scenarios.

[0087] Automation level: No manual filtering, feature selection, or parameter adjustment is required throughout the entire process; it can be achieved through the appendix. Figure 1The system shown automates the entire process of "data acquisition, diagnosis, alarm, and maintenance," reducing reliance on engineers' professional experience and making it suitable for large-scale applications in industrial scenarios.

[0088] In summary, this bearing fault diagnosis method effectively addresses the shortcomings of existing methods through the coordinated implementation of each step. It can be widely applied to the fault diagnosis of rolling bearings in rotating machinery (such as motors, fans, and pumps), providing strong support for the safe and reliable operation of equipment.

[0089] This bearing fault diagnosis method sequentially executes the following steps: acquiring the original vibration signal of the rolling bearing; extracting feature sequences using a convolutional neural network; processing the signal through a deep structure optimization network to obtain an optimized feature sequence; inputting the optimized feature sequence into a convolutionally gated recurrent neural network to obtain temporal feature information; and inputting the temporal feature information into a fully connected layer to obtain a diagnostic result containing the bearing fault type and severity. This achieves an end-to-end fault diagnosis process, eliminating the need for manual preprocessing operations such as filtering and feature selection. It effectively solves the problems of existing methods relying on professional experience and having cumbersome processes, thus improving the automation level of diagnosis. The convolutional neural network can automatically extract high-dimensional noise-resistant features from the original vibration signal. Combined with the advantages of the convolutionally gated recurrent neural network in mitigating gradient vanishing and capturing long-term dependencies in vibration signals through the gating mechanism, it significantly improves the completeness and accuracy of fault feature extraction under high-noise conditions, enhancing diagnostic robustness. The deep structure optimization network reduces the number and redundancy of parameters by optimizing the network structure, lowering the model's computational complexity and enabling the method to adapt to resource-constrained industrial scenarios such as edge computing terminals. Simultaneously, the fully connected layer can accurately output the fault type and severity, providing precise information for equipment maintenance and further ensuring the safe and reliable operation of rotating machinery.

[0090] like Figure 10 The image shows a specific embodiment of a bearing fault diagnosis device provided in this application. The device described in this embodiment is the physical device used to perform the method described in the above embodiments. Its technical solution is essentially the same as that of the above embodiments, and the corresponding descriptions in the above embodiments also apply to this embodiment. The large inertia on-orbit device includes an actuator, and the device includes: The first unit 301 is used to obtain the ENU coordinate system corresponding to the trajectory to be processed; wherein, the origin of the ENU coordinate system is the first trajectory point of the trajectory to be processed; The second unit 302 is used to obtain the first coordinate information of the target trajectory point in the ENU coordinate system in the trajectory to be processed, and to perform an affine transformation on the first coordinate information to obtain the second coordinate information of the target trajectory point in the ENU coordinate system. The third unit 303 is used to perform curvature compensation on the second coordinate information based on the displacement distance corresponding to the second coordinate information to obtain the compensated second coordinate information; wherein, the displacement distance is the distance between the second coordinate information and the origin of the ENU coordinate system; The fourth unit 304 is used to perform coordinate transformation on the compensated second coordinate information to obtain the geographic coordinate information corresponding to the compensated second coordinate information; The fifth unit 305 is used to correct the geographic coordinate information corresponding to the compensated second coordinate information based on the ground clearance of the target trajectory point, so as to obtain the target coordinate information of the target trajectory point.

[0091] Optionally, obtaining the ENU coordinate system corresponding to the trajectory to be processed includes: Obtain the geographic coordinates of the first trajectory point of the trajectory to be processed; The geographic coordinate information of the first trajectory point is converted into the first coordinate information of the first trajectory point in the ENU coordinate system; The first coordinate information of the first trajectory point in the ENU coordinate system is used as the origin of the ENU coordinate system, and the ENU coordinate system corresponding to the trajectory to be processed is established.

[0092] Optionally, the step of performing an affine transformation on the first coordinate information to obtain the second coordinate information of the target trajectory point in the ENU coordinate system includes: The first coordinate information is rotated and translated to obtain the second coordinate information of the target trajectory point in the ENU coordinate system.

[0093] Optionally, the step of performing curvature compensation on the second coordinate information based on the displacement distance corresponding to the second coordinate information to obtain the compensated second coordinate information includes: The distance between the second coordinate information and the origin of the ENU coordinate system is taken as the displacement distance corresponding to the second coordinate information; If the displacement distance corresponding to the second coordinate information is less than or equal to the first preset threshold, the second coordinate information will be used as the compensated second coordinate information; If the displacement distance corresponding to the second coordinate information is greater than the first preset threshold and less than or equal to the second preset threshold, the second coordinate is curvature compensated using a first-order approximate compensation method to obtain the compensated second coordinate information. If the displacement distance corresponding to the second coordinate information is greater than the second preset threshold, the second coordinate is subjected to curvature compensation using a second-order approximation compensation method to obtain the compensated second coordinate information.

[0094] Optionally, the step of correcting the geographic coordinate information corresponding to the compensated second coordinate information based on the ground clearance of the target trajectory point to obtain the target coordinate information of the target trajectory point includes: The terrain elevation of the digital elevation model corresponding to the target trajectory point is calculated using a bilinear interpolation algorithm; the digital elevation model corresponds to the region corresponding to the target trajectory point. Based on the terrain elevation, determine the ground clearance of the target trajectory point; Based on the ground clearance of the target trajectory point and the preset minimum ground clearance, the elevation value of the geographic coordinate information corresponding to the compensated second coordinate information is corrected to obtain the target coordinate information of the target trajectory point.

[0095] Optionally, the step of correcting the elevation value of the geographic coordinate information corresponding to the compensated second coordinate information based on the ground clearance of the target trajectory point and the preset minimum ground clearance to obtain the target coordinate information of the target trajectory point includes: If the ground clearance of the target trajectory point is less than the preset minimum ground clearance, the elevation value of the geographic coordinate information corresponding to the compensated second coordinate information is corrected according to the preset minimum ground clearance and the preset safety margin to obtain the target coordinate information of the target trajectory point. If the ground clearance of the target trajectory point is equal to or greater than the preset minimum ground clearance, the ground clearance of the target trajectory point is used as the target coordinate information of the target trajectory point.

[0096] Optionally, the device further includes a sixth unit, configured to: after the step of performing coordinate transformation on the compensated second coordinate information to obtain the geographic coordinate information corresponding to the compensated second coordinate information, perform boundary processing on the geographic coordinate information corresponding to the compensated second coordinate information to obtain processed geographic coordinate information; wherein, the boundary processing is to control the geographic coordinate information corresponding to the compensated second coordinate information within a preset reasonable coordinate range. Accordingly, the step of correcting the geographic coordinate information corresponding to the compensated second coordinate information based on the ground clearance of the target trajectory point to obtain the target coordinate information of the target trajectory point includes: Based on the ground clearance of the target trajectory point, the processed geographic coordinate information is corrected to obtain the target coordinate information of the target trajectory point.

[0097] Figure 11This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. At the hardware level, the electronic device includes a processor, and optionally also includes an internal bus, a network interface, and a memory. The memory may include RAM, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk storage device. Of course, the electronic device may also include other hardware required for other services.

[0098] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 11 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0099] Memory is used to store instructions for execution. Specifically, instructions for execution are computer programs that can be executed. Memory can include main memory and non-volatile memory, and it provides the processor with execution instructions and data.

[0100] In one possible implementation, the processor reads the corresponding execution instructions from non-volatile memory into main memory and then executes them. Alternatively, it may obtain the corresponding execution instructions from other devices to form a bearing fault diagnosis device at the logical level. The processor executes the execution instructions stored in the memory to implement the bearing fault diagnosis method provided in any embodiment of this application.

[0101] The above is as stated in this application. Figure 9The bearing fault diagnosis device provided in the illustrated embodiment can be applied to or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor.

[0102] The steps of the method disclosed in the embodiments of this application can be directly manifested as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0103] This application also proposes a readable medium that stores execution instructions. When the stored execution instructions are executed by the processor of an electronic device, the electronic device can execute the bearing fault diagnosis method provided in any embodiment of this application, and specifically perform the above-mentioned evaluation method.

[0104] The electronic devices described in the foregoing embodiments may be computers.

[0105] Those skilled in the art will understand that the embodiments of this application can be provided as methods or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or a combination of software and hardware.

[0106] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0107] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0108] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A bearing fault diagnosis method, characterized in that, The method includes: Obtain the raw vibration signal of the rolling bearing; The original vibration signal is used to extract features using a convolutional neural network to obtain a feature sequence. The feature sequence is processed by a deep structure optimization network to obtain an optimized feature sequence; The optimized feature sequence is input into a convolutional gated recurrent neural network to obtain temporal feature information; The timing feature information is input into the fully connected layer to obtain the fault diagnosis results of the rolling bearing, wherein the fault diagnosis results include the type and severity of the bearing fault.

2. The bearing fault diagnosis method according to claim 1, characterized in that, The acquisition of the original vibration signal of the rolling bearing includes: The time-domain vibration signal of the rolling bearing during operation is collected by a vibration sensor. The time-domain vibration signal includes noise interference and fault characteristics. A random sampling strategy is used to select a fixed-length signal segment from continuously acquired time-domain vibration signals as the original vibration signal of the rolling bearing; The random sampling strategy includes sliding time window sampling or random interval sampling; the fixed length of the signal segment is determined based on the bearing speed and fault characteristic period of the rolling bearing.

3. The bearing fault diagnosis method according to claim 1, characterized in that, The convolutional neural network includes at least one set of modules including convolutional layers, normalization layers, activation function layers, pooling layers, and regularization layers; the normalization layer is used to unify the feature distribution to improve training stability, the activation function layer uses a non-linear activation function, and the regularization layer is used to reduce the risk of model overfitting.

4. The bearing fault diagnosis method according to claim 3, characterized in that, The standardization layer is a batch standardization layer or a layer standardization layer. Specifically, the standardization layer is used to normalize the features using the statistical mean and variance of the features, and the feature distribution is adjusted by learning parameters after normalization. The activation function is an exponential linear unit or ReLU. The regularization layer is a Dropout layer or an L2 regularization layer.

5. The bearing fault diagnosis method according to claim 1, characterized in that, The deep structure optimization network is obtained by replacing some of the convolutional or fully connected layers in the original network with optimization layers and introducing regularization constraints or sparse structures. The deep structure optimization network is derived from wide and deep convolutional neural networks, residual networks, or dense networks. The sparse structure includes grouped convolution, depthwise separable convolution, or neuron cluster decomposition structure. The regularization constraints include L1 regularization, L2 regularization, or elastic network regularization.

6. The bearing fault diagnosis method according to claim 1, characterized in that, The convolutional gated recurrent neural network is used to extract spatial features and perform temporal modeling. The convolutional gated recurrent neural network consists of a convolutional module and a gated recurrent module. The convolutional module is used to enhance the spatial features of the optimized feature sequence, and the gated recurrent module is used to control the transmission and forgetting of temporal information through a gating mechanism.

7. The bearing fault diagnosis method according to claim 6, characterized in that, The gated recurrent module is a gated recurrent unit (GRU) or a long short-term memory network (LSTM). The gating mechanism includes at least one of an update gate, a reset gate, and a forget gate. The convolutional gated recurrent neural network also integrates an attention mechanism. The convolutional gated recurrent neural network is used to obtain key temporal features by assigning differentiated weights to each time step of the feature sequence, and to use the key temporal features as the temporal feature information.

8. The bearing fault diagnosis method according to claim 6, characterized in that, The convolutional gated recurrent neural network also integrates a normalization layer and a regularization layer. The normalization layer is set at the output of the convolutional module or the input of the gated recurrent module in the convolutional gated recurrent neural network, and the regularization layer is set at the output of the gated recurrent module.

9. The bearing fault diagnosis method according to claim 1, characterized in that, The step of inputting the timing feature information into the fully connected layer to obtain the fault diagnosis result of the rolling bearing includes: The fault diagnosis results of the rolling bearing are obtained by performing multi-level linear transformation and activation function operation on the time-series feature information using the fully connected layer. The output of the fully connected layer uses either the softmax activation function or the sigmoid activation function; the bearing fault types include at least one of inner ring fault, outer ring fault, rolling element fault, and cage fault; and the fault severity includes at least two of mild, moderate, and severe. The method further includes: The diagnostic performance index is calculated by comparing the fault diagnosis results of the rolling bearing with the actual fault state. Based on the performance metrics, adjust the network structure parameters or training parameters of the convolutional neural network, the deep structure optimization network, the convolutional gated recurrent neural network, and the fully connected layer until the performance metrics of the convolutional neural network, the deep structure optimization network, the convolutional gated recurrent neural network, and the fully connected layer meet the preset requirements. The performance metrics include at least one of diagnostic accuracy, recall, and F1 score.

10. A bearing fault diagnosis device, characterized in that, The device includes: The first unit is used to acquire the original vibration signal of the rolling bearing; The second unit is used to extract features from the original vibration signal using a convolutional neural network to obtain a feature sequence. The third unit is used to process the feature sequence through a deep structure optimization network to obtain an optimized feature sequence; The fourth unit is used to input the optimized feature sequence into a convolutional gated recurrent neural network to obtain temporal feature information; The fifth unit is used to input the timing feature information into the fully connected layer to obtain the fault diagnosis results of the rolling bearing, wherein the fault diagnosis results include the type and severity of the bearing fault.

Images