Fault diagnosis method based on sparse heterogeneous data collaboration, medium and product

Abstract

The invention discloses a fault diagnosis method based on sparse heterogeneous data collaboration, a medium and a product, and the method comprises the steps: inputting the collected vibration data of a to-be-diagnosed target into a fault diagnosis network, and outputting a fault diagnosis result; the fault diagnosis network training method comprises the following steps: obtaining feature data after spatial separation processing by using a depth separable convolution module; using an inverted residual module to obtain feature data after feature enhancement; adding a lightweight attention mechanism in the inverted residual module to obtain weighted feature data; respectively calculating time domain branch loss and frequency domain branch loss, balancing the weight of the time domain branch loss and the weight of the frequency domain branch loss, obtaining a fault diagnosis network after preliminary training, carrying out transfer learning, and obtaining pre-training features according to the vibration data; and training the fault diagnosis network by using the plurality of source domain data to obtain a trained fault diagnosis network. According to the invention, the problems of small sample data volume and inaccurate fault identification of the industrial internet in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to a fault diagnosis method, medium and product based on sparse heterogeneous data collaboration, belonging to the technical field of industrial Internet fault diagnosis. Background Art

[0002] In modern industry, machines are becoming more complex and intelligent, and adapting to increasingly harsh operating conditions. Even a slight performance degradation or safety hazard may have serious consequences. If problems are not discovered in time, they may lead to sudden failures or even catastrophic accidents, resulting in huge economic and personnel losses. Therefore, effective fault monitoring is crucial to avoid economic losses caused by equipment failures.

[0003] In order to ensure high reliability and low downtime of mechanical equipment, it is essential to accurately identify the presence and severity of mechanical faults. Fault detection and isolation (FDI) is a subfield of control engineering, and its goal is to accurately identify and locate the type and location of the fault when it occurs. At present, there are two main methods of fault detection and isolation: the first is pattern recognition for sensor faults; the second is to build a model to analyze the difference between the data obtained by the sensor and the predicted ideal value.

[0004] With the rapid development of machine learning methods, mechanical equipment fault diagnosis technology has made significant progress. Some classic machine learning methods have been successfully applied to mechanical fault diagnosis. The fault diagnosis process based on machine learning includes the following steps: first, the collected fault signal is preprocessed, then the processed data is input into the machine learning model, and finally the operation data characteristics of the mechanical equipment are extracted from it, so as to achieve efficient and accurate fault diagnosis. Early machine learning methods mainly include the nearest neighbor method and the support vector machine method. However, these methods cannot directly extract features from the original vibration signal. In addition, the nonlinearity and non-stationarity of the vibration signal limit the application of these methods in mechanical fault diagnosis. At the same time, the real industrial environment is usually complex and accompanied by various noise interferences, and it is difficult to obtain a large amount of effective data.

[0005] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgement or any form of suggestion that the information constitutes the prior art already known to ordinary technicians in this field. Summary of the invention

[0006] The technical problem to be solved by the present invention is: how to improve the accuracy of industrial Internet fault identification and reduce computing delay.

[0007] In order to solve the above technical problems, the present invention is implemented by adopting the following technical solutions.

[0008] In a first aspect, the present invention provides a fault diagnosis method based on sparse heterogeneous data collaboration, comprising:

[0009] Acquire bearing vibration data of the target to be diagnosed;

[0010] The acquired bearing vibration data of the target to be diagnosed is input into the trained fault diagnosis network TF-IRSCNN, and the fault diagnosis result is output;

[0011] The training method of the fault diagnosis network TF-IRSCNN includes:

[0012] The time domain data and frequency domain data of the bearing vibration signal are used as training sets;

[0013] Use the depth-wise separable convolution module to perform spatial separation processing on the training set to obtain feature data after spatial separation processing;

[0014] Through the inverted residual module, feature extraction and enhancement are performed on the feature data after spatial separation processing to obtain feature data after enhanced features;

[0015] Performing weighted processing on the feature data after feature enhancement to obtain weighted feature data;

[0016] Using the dual-branch loss function in the time-frequency domain, respectively calculating the time-domain branch loss and the frequency-domain branch loss of the weighted feature data, and using the gradient normalization method to automatically balance the weights of the time-domain branch loss and the frequency-domain branch loss, obtaining the overall loss function, minimizing the loss value, and obtaining the fault diagnosis network after preliminary training;

[0017] The multi-source domain adaptive adversarial network (MSAAN) is used to perform transfer learning on the fault diagnosis network after preliminary training, and the fault diagnosis network is obtained after fine-tuning with labeled data.

[0018] The network is pre-trained using time-frequency domain features, and pre-trained features are obtained based on the historical bearing vibration data corresponding to the target to be diagnosed;

[0019] Feed the pre-trained features into the fault diagnosis network that has been fine-tuned with labeled data;

[0020] A fault diagnosis network is trained using multiple source domain data to obtain a trained fault diagnosis network TF-IRSCNN, wherein the multiple source data are status data of different faults in the CWRU bearing data set, including bearing operation data of different loads and different fault locations.

[0021] Furthermore, the fault diagnosis network TF-IRSCNN includes:

[0022] The depth-separable convolution module is used to perform spatial separation processing on the training set to obtain feature data after spatial separation processing;

[0023] An inverted residual module with a lightweight attention mechanism is added to extract and enhance the feature data after spatial separation processing to obtain feature data with enhanced features.

[0024] Furthermore, the multi-source domain adaptive adversarial network (MSAAN) is used to perform transfer learning on the fault diagnosis network after preliminary training. After obtaining the fault diagnosis network fine-tuned with labeled data, the fault diagnosis network fine-tuned with labeled data is migrated to the target domain.

[0025] Furthermore, a multi-source domain adaptive adversarial network (MSAAN) is used to perform transfer learning on the fault diagnosis network after preliminary training, and after obtaining the fault diagnosis network fine-tuned with labeled data, a method for migrating the fault diagnosis network fine-tuned with labeled data to a target domain includes:

[0026] Constructing an adversarial transfer learning network MSDA-IRSCNN based on multi-source domain adaptation, extracting features from multiple labeled source domain datasets includes using the private feature extractor and shared feature extractor in the multi-source domain adaptation learning module to extract private features and shared features of the source domain respectively;

[0027] Through adversarial training of multi-domain discriminators, the distribution difference between source domain and target domain data in the feature space is reduced to obtain shared features with domain invariance.

[0028] Using the shared features with domain invariance to train the shared layers of the fault diagnosis network TF-IRSCNN, to obtain the fault diagnosis network fine-tuned with labeled data;

[0029] After completing transfer learning, the fault diagnosis network TF-IRSCNN fine-tuned with labeled data is migrated to the target domain.

[0030] Furthermore, a method for obtaining pre-trained features by using a time-frequency domain feature pre-training network according to bearing vibration history data corresponding to a target to be diagnosed includes:

[0031] Obtain the original time domain data of the bearing vibration signal through the sensor;

[0032] Obtain the frequency domain data of the bearing vibration signal through fast Fourier transform FFT transformation;

[0033] The original time domain data of the bearing vibration signal and the frequency domain data of the bearing vibration signal are input into the time-frequency domain feature pre-training network to obtain pre-trained features.

[0034] Furthermore, a method of using multiple source domain data to train a fault diagnosis network to obtain a trained fault diagnosis network TF-IRSCNN includes:

[0035] Data features are extracted from multi-source domain data, including private data features in each source domain data and shared data features in multi-source domain data. The classification weights of the target domain data are obtained from the perspectives of different source domains through a domain discriminator. The final predicted label of the target domain data is obtained after weighted combination. The network parameters and structure of the fault diagnosis network TF-IRSCNN are adjusted according to the final predicted label to obtain the trained fault diagnosis network TF-IRSCNN.

[0036] Furthermore, using the depthwise separable convolution module, before spatially separating the input data, a width multiplier is introduced into the depthwise separable convolutional neural network SCNN. and resolution multiplier , the width multiplier Acting on the convolution operation, reducing the amount of hyperparameter calculation in the convolution operation, proportionally reducing the number of convolution kernels in each layer of the deep separable convolutional neural network SCNN, the resolution multiplier The input data is compressed to reduce the amount of internal input data of each layer of the deep separable convolutional neural network (SCNN).

[0037] Furthermore, the dual-branch loss function in the time-frequency domain is used to calculate the time-domain branch loss and the frequency-domain branch loss of the weighted feature data respectively, and the gradient normalization method is used to automatically balance the weights of the time-domain branch loss and the frequency-domain branch loss to obtain the overall loss function, which is expressed as:

[0038] ;

[0039] in, and They represent the branch loss in the time domain and frequency domain branch loss The weight of Represents the total loss of the two branches in the time-frequency domain.

[0040] In a second aspect, the present invention further provides a computer-readable storage medium having a computer program / instruction stored thereon, which implements the steps of the above method when the computer program / instruction is executed by a processor.

[0041] In a third aspect, the present invention further provides a computer program product, comprising a computer program / instruction, characterized in that the computer program / instruction implements the steps of the above method when executed by a processor.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] This paper proposes a fault diagnosis method based on sparse heterogeneous data collaboration. The method is based on the fault diagnosis network TF-IRSCNN. Through the dual-branch feature extraction technology in the time-frequency domain, combined with the deep separable convolution module that introduces the width multiplier and the resolution multiplier, the network parameters and input data dimensions are effectively reduced. At the same time, the fault diagnosis network TF-IRSCNN also incorporates the inverted residual module, lightweight attention mechanism and nonlinear activation function, which significantly improves the fault recognition accuracy in the industrial Internet.

[0044] In order to further enhance the ability to utilize cross-domain data, the present invention also constructs an adversarial transfer learning network based on multi-source domain adaptation, MSDA-IRSCNNMSDA-IRSCNN, which uses adversarial transfer learning technology to mine rich private and shared features from multi-source domain data, and assigns reasonable classification weights to different source domain data through domain discriminators, and finally obtains accurate prediction labels for target domain data by weighting. In a sparse data environment, this method can extract richer feature information from time-frequency domain data, which not only greatly improves the accuracy of fault identification, but also significantly reduces the delay in the calculation process, providing an efficient and accurate solution for fault diagnosis of the industrial Internet. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flowchart of a fault diagnosis method based on sparse heterogeneous data collaboration provided by an embodiment of the present invention;

[0046] Figure 2 It is a simulation schematic diagram of time domain signal data of a bearing sample numbered CWRU_097 in a healthy state from a CWRU bearing data set provided by an embodiment of the present invention;

[0047] Figure 3 It is a simulation schematic diagram of time domain signal data of a bearing sample numbered CWRU_105 in a fault state from a CWRU bearing data set provided by an embodiment of the present invention;

[0048] Figure 4 It is a simulation schematic diagram of frequency domain signal data of a bearing sample numbered CWRU_097 in a healthy state from a CWRU bearing data set provided by an embodiment of the present invention;

[0049] Figure 5 It is a simulation schematic diagram of frequency domain signal data of a bearing sample numbered CWRU_105 in a fault state from a CWRU bearing data set provided by an embodiment of the present invention;

[0050] Figure 6 is a schematic diagram comparing the convolution process of a standard convolution module and a depthwise separable convolution module provided in an embodiment of the present invention;

[0051] Figure 7 2 is a schematic diagram comparing the feature extraction process of a traditional residual module and an inverted residual module provided by an embodiment of the present invention;

[0052] Figure 8 Schematic diagram of the feature extraction process of the inverted residual module after adding a lightweight attention mechanism provided by an embodiment of the present invention;

[0053] Fig. 9 1 is a schematic diagram of the network architecture of the time-frequency domain feature pre-training network TF-IRSCNN provided in an embodiment of the present invention;

[0054] Fig.10 is a schematic diagram of the network architecture of a multi-source domain adaptive learning module provided by an embodiment of the present invention;

[0055] Fig.11 It is a schematic diagram of the network architecture based on the multi-source domain adaptive deep adversarial transfer learning network MSDA-IRSCNN provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. The embodiments of the present invention and the technical features in the embodiments may be combined with each other unless there is a conflict.

[0057] Example 1

[0058] This embodiment introduces a fault diagnosis method based on sparse heterogeneous data collaboration, including:

[0059] Acquire bearing vibration data of the target to be diagnosed;

[0060] The acquired bearing vibration data of the target to be diagnosed is input into the trained fault diagnosis network TF-IRSCNN, and the fault diagnosis result is output.

[0061] The training method of the fault diagnosis network TF-IRSCNN includes:

[0062] The time domain data and frequency domain data of the bearing vibration signal are used as training sets;

[0063] Use the deep separable convolution module to perform spatial separation processing on the training set to obtain feature data after spatial separation processing; deep separable convolution effectively extracts the spatial features of the input data by separating the convolution operation on the spatial dimension, while reducing the amount of calculation and the number of parameters. Compared with traditional convolution, deep separable convolution greatly improves computing efficiency while maintaining performance, and is suitable for processing large-scale data in the industrial Internet.

[0064] Through the inverted residual module, the feature data after spatial separation processing is extracted and enhanced to obtain the feature data after enhanced features; the inverted residual module first expands the dimension of the feature data, then performs deep convolution and linear transformation, and finally restores the original size by squeezing the dimension, and enhances the feature representation capability in the above process. By increasing the nonlinear ability of the fault diagnosis network, the fault diagnosis network can capture more complex data patterns and improve the accuracy of fault diagnosis.

[0065] The feature data after enhanced features are weighted to obtain weighted feature data; a dual-branch loss function in the time-frequency domain is used to calculate the losses of the time-domain branch and the frequency-domain branch of the weighted feature data respectively, and the loss value is obtained by fusing them according to the preset weights, and the loss value is minimized.

[0066] The dual-branch loss function in the time-frequency domain is used to calculate the time-domain branch loss and the frequency-domain branch loss of the weighted feature data, and the gradient normalization method is used to automatically balance the weights of the time-domain branch loss and the frequency-domain branch loss to obtain the overall loss function, minimize the loss value, and obtain the fault diagnosis network after preliminary training; the dual-branch loss function in the time-frequency domain evaluates the difference between the output of the fault diagnosis network and the true label from the two dimensions of time and frequency, which helps to capture complex patterns in the data. By calculating the losses in the time and frequency domains separately and fusing them according to the preset weights, a clearer optimization direction is provided for the training of the preliminary trained fault diagnosis network.

[0067] By minimizing the loss value, the initially trained fault diagnosis network can continuously adjust its parameters to better fit the training data, thereby improving its generalization ability on unseen data. As the training progresses, the fault diagnosis network is gradually optimized, and the accuracy and efficiency of fault diagnosis will continue to improve.

[0068] The multi-source domain adaptive adversarial network MSAAN is used to perform transfer learning on the fault diagnosis network after preliminary training to obtain a fault diagnosis network that has been fine-tuned with labeled data; the multi-source domain adaptive adversarial network MSAAN uses the ability of transfer learning to transfer knowledge learned from multiple source domains (i.e., domains with existing labeled data) to the target domain (i.e., domains that need diagnosis but have limited or no labeled data). It can significantly improve the performance on the target domain, especially when the amount of data in the target domain is not enough to train a fault diagnosis network from scratch.

[0069] At the same time, the multi-source domain adaptive adversarial network MSAAN introduces an adversarial training mechanism, which enables the model to reduce the difference between the source domain and the target domain at the feature level, thereby enhancing the generalization ability of the model in the target domain. This helps the fault diagnosis network after initial training to learn more general and robust feature representations. Using labeled data to fine-tune the fault diagnosis network after transfer learning can further optimize the performance of the fault diagnosis network in the target domain, making it more suitable for specific application scenarios.

[0070] The network is pre-trained using time-frequency domain features, and pre-trained features are obtained based on the bearing vibration data of the target to be diagnosed. The pre-trained network can be used to extract useful low-dimensional feature representations, making the features more effective in subsequent classification or diagnosis tasks.

[0071] Feed the pre-trained features into the fault diagnosis network that has been fine-tuned with labeled data;

[0072] A fault diagnosis network is trained using multiple source domain data to obtain a trained fault diagnosis network TF-IRSCNN, wherein the multiple source data are status data of different faults in the CWRU bearing data set, including bearing operation data of different loads and different fault locations.

[0073] Using data from multiple source domains for training can increase the diversity and generalization ability of the fault diagnosis network. Data from different source domains may contain different fault modes and noise characteristics. By learning these data simultaneously, the fault diagnosis network can learn more robust and general fault representations.

[0074] The final fault diagnosis network TF-IRSCNN is a comprehensive fault diagnosis system that integrates time-frequency domain feature extraction, transfer learning, adaptive adversarial training and fine-tuning optimization. It can effectively process bearing vibration signals and achieve high-precision fault diagnosis.

[0075] Example 2

[0076] This embodiment introduces a fault diagnosis method based on sparse heterogeneous data collaboration. Figure 1 As shown, the specific steps include:

[0077] Step 1: Obtain the original time domain data and frequency domain data of the bearing vibration signal and perform data preprocessing.

[0078] Step 1.1: Obtain the original time domain data of the bearing vibration signal through the sensor.

[0079] In order to verify the performance of the fault diagnosis network TF-IRSCNN in diagnosing different faults, the bearing vibration data needs to be preprocessed differently to adapt to different network input requirements. The time domain signal and frequency domain signal of the vibration signal are used as input data respectively. Since the vibration data is collected by a vibration sensor, its original signal is the time domain signal. In order to display the vibration signal information more intuitively, it is plotted as a curve of the change of vibration amplitude over time for visualization.

[0080] In this embodiment, the length of each sample is 1024, and the calculation formula of the total number of samples is expressed as:

[0081] (1);

[0082] in, L represents the length of each bearing vibration signal, represents the total sample, Indicates that data is rounded toward negative infinity.

[0083] exist Figure 2 , Figure 3 The time domain signal data of the bearing sample numbered CWRU_097 in the CWRU bearing dataset in a healthy state and the time domain signal data of the bearing sample numbered CWRU_105 in a faulty state are respectively shown.

[0084] The above data are all collected under the working condition of 0HP load, where the vibration signal amplitude range in the healthy state is between positive and negative 0.2, the vibration signal amplitude range in the fault state is between positive and negative 1, and the time domain data in the fault state has a higher vibration density.

[0085] Due to the occurrence of faults, the vibration signal has more obvious waveform characteristics in the time domain, which causes the bearing fault signal to have a higher mean, variance, standard deviation and peak-to-peak value. Based on these characteristic information, the operating status of the bearing can be effectively distinguished.

[0086] The time domain data of the bearing vibration signal is more sensitive to fault information. The kurtosis of the time domain data can directly reflect the occurrence of the fault. When a local fault occurs, the time domain waveform appears as a periodic impact signal. In addition, the calculation of the time domain characteristics is relatively simple and the computational complexity is low, which is more suitable for real-time monitoring.

[0087] Step 1.2: Obtain the frequency domain data of the bearing vibration signal through fast Fourier transform (FFT).

[0088] The frequency domain data of the bearing vibration signal requires the Fast Fourier Transform (FFT) of the time domain signal of the bearing vibration signal. ) to obtain the frequency domain signal.

[0089] Fast Fourier Transform (FFT) improves the Discrete Fourier Transform (DFT) algorithm and reduces the amount of computation of the Discrete Fourier Transform (DFT).

[0090] The process of discrete Fourier transform DFT calculation is expressed as:

[0091] (2);

[0092] in, It represents the value of the fth frequency domain signal component after discrete Fourier transform DFT. represents the sample value of the i-th time domain signal, is the rotation factor, , j is an imaginary unit, satisfying , N represents the total sampling points, n is the time domain signal number, ranging from 0 to N-1, and f is the frequency domain signal number, ranging from 0 to N-1.

[0093] Fast Fourier Transform (FFT) is an efficient algorithm for implementing DFT. It reduces the time complexity of discrete Fourier transform (DFT) calculation from Reduce to The basic idea of ​​the Fast Fourier Transform (FFT) algorithm is to convert the length The discrete Fourier transform DFT is decomposed into two lengths The discrete Fourier transform DFT is taken is an even number, the original signal can be decomposed into two parts: odd index and even index, which are x [ n ] = x [ 2 m ] and x [ n ] = x [ 2 m + 1 ] ,in is the new index. At this time, the signal that realizes discrete Fourier transform through divide-and-conquer algorithm can be decomposed into:

[0094] (3);

[0095] in, represents the value of the fth frequency domain signal component after discrete Fourier transform by divide-and-conquer algorithm, , are odd-index rotation factors and even-index rotation factors respectively, where, , , j is an imaginary unit, satisfying , N represents the total sampling points, m is the time domain signal number, ranging from 0 to N / 2-1, and f is the frequency domain signal number, ranging from 0 to N-1.

[0096] The sub-discrete Fourier transform DFT formed by the divide-and-conquer algorithm is essentially the discrete Fourier transform DFT of the original signal, but its length is only , the above divide-and-conquer process is recursively repeated until a discrete Fourier transform DFT with a length of 1 is obtained.

[0097] Since the divide-and-conquer algorithm calculates the discrete Fourier transform DFT for each value of the input time domain signal, it is necessary to Sub-real number product and The Fast Fourier Transform (FFT) algorithm uses the periodicity and symmetry of the rotation factors in the Discrete Fourier Transform (DFT) to transform the entire Discrete Fourier Transform (DFT) calculation into a series of Operation, greatly improving the operation process and amount of operation.

[0098] After each sample data is transformed from the time domain to the frequency domain using the Fast Fourier Transform FFT, the data length is halved and the new sample is represented as:

[0099] (4);

[0100] in, represents the value of the fth frequency domain signal component after FFT transformation, the operator Indicates that the time domain signal is converted into Transform to the frequency domain and take the first half of the result.

[0101] Figure 4 , Figure 5 The frequency domain signal data of the bearing sample numbered CWRU_097 in the CWRU bearing data set in a healthy state and the frequency domain signal data of the bearing sample numbered CWRU_105 in a faulty state are respectively displayed.

[0102] Among them, the amplitude range of the vibration signal in the healthy state is between 0 and 20, and the amplitude range of the vibration signal in the fault state is between 0 and 30, and the frequency domain data in the fault state has a higher vibration density. Due to the occurrence of the fault, the vibration signal has a more obvious spectral feature in the frequency domain, resulting in a higher main frequency, spectral bandwidth and spectral kurtosis of the bearing fault signal. Based on these characteristic information, the operating state of the bearing can be effectively distinguished.

[0103] The frequency spectrum information of the bearing vibration signal contains the specific frequency information when the fault occurs. By analyzing and comparing the changes in amplitude at the same frequency or identifying whether there is a new frequency component, the type of equipment fault, the cause of the fault, and the degree of fault degradation can be determined. In order to extract richer data features from the bearing vibration signal, the bearing time domain data is obtained through fast Fourier transform FFT to obtain frequency domain data.

[0104] Step 2: Build the fault diagnosis network TF-IRSCNN.

[0105] Step 2.1: Introduce the depthwise separable convolution module to perform spatial separation processing on the input data to obtain feature data after spatial separation processing.

[0106] Input Data The dimension is ,in, Represents the convolution kernel of the standard convolution module The spatial dimension of Indicates the number of input channels, Indicates the number of output channels, Indicates the input data size, Represents the convolution kernel size.

[0107] Assume the step size is , filled with , when using the convolution kernel of the standard convolution module by When the input data is parameterized by the size and dimensions of , the standard convolution corresponding to each input channel containing a single filter is expressed as:

[0108] (5);

[0109] in, represents the output feature, k is the convolution kernel size, l is the convolution step size, and m is the channel index. The filters, corresponding to the input data Middle channels and produces a filtered output feature map , Indicates size is the depth convolution kernel, b and c are the summation variables of the convolution kernel in the horizontal and vertical convolutions, respectively, indicating the displacement and weighted summation of the input data.

[0110] The computational cost of a standard convolution depends on the number of input channels. , Output Channels , convolution kernel size and feature map size , so the computational cost of the standard convolution is expressed as .

[0111] In pursuit of higher accuracy, neural networks are designed towards deeper and more complex structures, which leads to an increasing demand for GPUs. However, in real life, recognition tasks are often implemented in limited computing environments, and lightweight neural networks have emerged.

[0112] This embodiment constructs a deep separable convolutional neural network (SCNN) by splitting the standard convolution into channel-by-channel convolution and point-by-point convolution, and introduces a deep separable convolution module to achieve lightweight processing of the fault diagnosis network TF-IRSCNN. The convolution process of the standard convolution module and the deep separable convolution module is as follows: Figure 6 shown.

[0113] The deep separable convolutional neural network (SCNN) uses deep separable convolution to separate the connection between the number of output channels and the size of the convolution kernel, where the process of filtering features and combining features can be implemented using separable convolution.

[0114] First, a single filter is assigned to each input channel of the depthwise convolution, and then a linear combination of the output layer is created through point-by-point convolution. The nonlinear activation function ReLU and batch normalization operation BatchNorm are used in these two layers for activation and normalization.

[0115] The computational cost of depthwise separable convolution is expressed as .

[0116] By dividing the standard convolution into two parts: deep filter convolution and point-by-point output combination, the computational cost reduction of depth-wise separable convolution over standard convolution is expressed as:

[0117] (6);

[0118] In order to build a model with shallower layers and lower computational cost, a width multiplier is introduced in the depthwise separable convolutional neural network SCNN. and resolution multiplier , the width multiplier Acting on the convolution operation, reducing the amount of hyperparameter calculation in the convolution operation, proportionally reducing the number of convolution kernels in each layer of the deep separable convolutional neural network SCNN, the resolution multiplier The input data is compressed to reduce the amount of internal input data of each layer of the deep separable convolutional neural network (SCNN).

[0119] In this case, the width multiplier is and the resolution multiplier is The computational cost of the depthwise separable convolution is expressed as ,in, ( α , ρ ) ∈ ( 0 , 1 ] , is the standard convolution process, when The parameters in the convolution process are compressed at this time, which is the reduction parameter.

[0120] Width Multiplier The computational cost and number of parameters are reduced by about , resolution multiplier Usually implicitly set the input dimension of the depthwise separable convolutional neural network SCNN to 224, 192, 160 or 128, which can reduce the computational cost. .

[0121] Step 2.2: Add an inverted residual module to the depthwise separable convolution module, and use the inverted residual module to extract and enhance the feature data after spatial separation processing to obtain feature data after enhanced features.

[0122] By adding an inverted residual connection module between the bottleneck layers of the deep separable convolutional neural network (SCNN), the entire deep separable convolutional neural network (SCNN) can better expand and recompress data, thereby improving the gradient propagation capability between each network layer of the deep separable convolutional neural network (SCNN).

[0123] The traditional residual connection module first passes The convolution operation is used to reduce the dimension, and then The convolution operation is used to extract features, and finally The convolution operation realizes dimensionality increase, and the data transmission forms a structure with large ends and a small middle.

[0124] The inverted residual connection module introduced in this embodiment swaps the order of dimensionality reduction and dimensionality increase. The expansion layer of the convolution maps the low-dimensional data space to the high-dimensional data space, and then extracts features through deep separable convolution, and finally uses The convolutional projection layer maps high-dimensional features to low-dimensional space for data compression, and the data transmission forms a structure with small ends and a large middle.

[0125] Schematic diagram of the feature extraction process comparison between the traditional residual module and the inverted residual module, as shown in Figure 7 shown.

[0126] The input data is The expansion factor is , the convolution kernel is And there is input channels and For a convolution block with output channels, the total number of multiplications and additions required is .

[0127] because Figure 6 The separable convolution process adds an additional The convolutional layer has an extra term compared to the standard convolution computation cost, but the nature of the network allows the data to have smaller input and output dimensions.

[0128] The inverted residual connection module realizes Channels are converted to The process of channels is shown in Table 1. In the convolution operation, the step size of the convolution kernel is , the expansion factor is ;

[0129] Among them, the convolution kernel size of the convolution layer in Table 1 is h×w×k, h is the convolution kernel height, w is the convolution kernel width, k is the number of convolution kernel channels, conv2d is a two-dimensional convolution operation, ReLU6 is the activation function, dwise is a depth-wise separable convolution, and stride is the convolution step size.

[0130] Step 2.3: Add a lightweight attention mechanism after the separable convolution module to perform weighted processing on the feature data after feature enhancement to obtain weighted feature data.

[0131] In order to further improve the calculation speed and classification accuracy, this embodiment adds a lightweight attention mechanism (Squeeze Excitation, SE) to the inverted residual connection module. The lightweight attention mechanism SE can enhance channel features without changing the size of the input feature data. The lightweight attention mechanism SE is placed after the depth-separable convolution module. The Squeeze compression part of the lightweight attention mechanism SE compresses the spatial information of the input feature data, and the Excitation excitation part of the lightweight attention mechanism SE combines the learned channel attention information with the input feature data to obtain feature data with channel attention.

[0132] Before the data is input into the lightweight attention mechanism SE, the dimension of the feature data is After the data is input into the lightweight attention mechanism SE, the feature data is first spatially compressed, and global average pooling is implemented in the spatial dimension to obtain The feature data of size is then learned through the fully connected layer FC to obtain the feature data with channel attention. The channel attention feature data is input into the last fully connected layer FC for restoration to obtain the original input The above operation reduces the computation time and improves the classification accuracy. The inverted residual module after adding the lightweight attention mechanism is shown in Figure 8 shown.

[0133] Step 2.4: Introduce the dual-branch loss function in the time-frequency domain, calculate the losses of the time-domain branch and the frequency-domain branch of the weighted feature data respectively, fuse them according to the preset weights to obtain the loss value, and obtain the fault diagnosis network after preliminary training by minimizing the loss value.

[0134] In general, the same loss function is often used to learn the characteristics of time domain data and frequency domain data, but the time domain mixed frequency domain indicators reflect different characteristics of faults. Time domain data contains direct and easy-to-understand fault information. The kurtosis of time domain data can reflect the signal pulse of the fault, and is more sensitive to fault information, but the direct correlation with the fault type and severity is weak. The kurtosis is very sensitive to the impact pulse that reflects whether the fault has occurred. In the time domain signal of the bearing, the kurtosis range is between plus or minus 0.2 in normal operation, and the kurtosis range is between plus or minus 1.0 when a fault occurs. In addition, when a local fault occurs, the time domain waveform is a periodic impact signal, and the period is the rotation period.

[0135] The spectrum of the vibration signal contains specific frequency information when the fault occurs. By analyzing and comparing the changes in amplitude at the same frequency or identifying whether new frequency components appear, the type of equipment fault, the cause of the fault, and the degree of fault degradation can be determined. Therefore, frequency domain features provide more detailed information than time domain features, so that the type and cause of the fault can be identified with high accuracy. If the time and frequency domain features are learned at the same time, and the differences between time domain and frequency domain features are considered, the network can learn more comprehensive feature information.

[0136] This embodiment uses two different loss functions to extract the mixed features of the vibration signal based on the different characteristics of the time domain and frequency domain indicators. Since the time domain indicators are directly extracted from the original time series, the mean square error (MSE) is used to fit the time domain artificial labels.

[0137] Different from the time domain indicators, the frequency domain indicators are calculated after Fourier transform and other signal decomposition. If MSE is used to fit the frequency domain labels, it will lead to local optimization. Therefore, the relative entropy divergence (Kullback-Leibler, KL) is used to fit the frequency domain indicators, which can make the network converge to the optimal solution. The relative entropy divergence KL is expressed as:

[0138] (7);

[0139] in, Represents the relative entropy between the true distribution p and the approximate distribution q of the frequency domain data. represents the true distribution, represents the predicted distribution, represents the sum of all possible x values, and k represents the number of possible x values.

[0140] Since the time-frequency feature pre-training network needs to process the time domain and frequency domain branches at the same time, the loss functions for the time domain branch and the frequency domain branch are set separately.

[0141] The time domain loss function is expressed as:

[0142] (8);

[0143] in, represents the time domain loss function, represents the total number of samples, represents the number of manual labels of the first branch, represents the true value of the i-th branch and the j-th output channel, Represents the predicted value of the i-th branch and the j-th output channel.

[0144] The frequency domain loss function is expressed as:

[0145] (9);

[0146] in, represents the frequency domain loss function, represents the total number of samples, Indicates the number of manual labels for the second branch, represents the number of manual labels of the first branch, represents the probability of the true probability distribution in the true value of the frequency domain signal of the i-th branch and the j-th output channel, represents the probability of the true probability distribution in the frequency domain signal prediction value of the i-th branch and j-th output channel, Represents the probability of the predicted probability distribution in the true value of the frequency domain signal.

[0147] In addition, the weights of the time domain branch and the frequency domain branch in fault diagnosis are unbalanced, and the magnitudes of different branches are also different. and Adding to obtain the overall loss function of the time-frequency feature pre-trained network is not the best approach.

[0148] In order to reduce the difference in loss functions, this implementation uses the gradient normalization method to automatically balance the weights of the two loss functions and obtain the overall loss function. The overall loss function is expressed as:

[0149] (10);

[0150] in, and They represent the branch loss in the time domain and frequency domain branch loss The weight of Represents the total loss of the two branches in the time-frequency domain.

[0151] The overall loss function is calculated through gradient loss To automatically update the weights, when optimizing the weights of different branches, it is necessary to consider the network training speed and the size of each branch, and then express the gradient norm and average gradient norm of each branch as:

[0152] (11);

[0153] G ¯ w ( t ) = E task [ G w ( t ) i ] (12);

[0154] in, Indicates The branch passes through the deep neural network The actual gradient norm of the weight w obtained from the training, represents the weight gradient vector, is the two-norm operator, Indicates The branch passes through the deep neural network The weights obtained from training, Indicates The branch passes through the deep neural network The loss value obtained from training; Represents the first The average gradient norm of the weight w obtained from training, E task [] Indicates taking the average value.

[0155] In order to balance the training speed of each branch, define The relative inverse training rate of the branches and the The reverse training rate of the branch is The relative inverse training rate of the branches and the The inverse training rate of each branch is expressed as:

[0156] r i ( t ) = L i ( t ) E task [ L i ( t )] (13);

[0157] (14);

[0158] in, Indicates The branch passes through the deep neural network The relative reverse training rate obtained by training. Indicates The branch passes through the deep neural network The reverse training rate obtained by training.

[0159] Therefore, the gradient loss function is expressed as:

[0160] Loss grad = ∑ i | G w ( t ) i − G ¯ w ( t ) [ r i ( t ) o ] | 1 (15);

[0161] in, represents the gradient loss, Indicates The branch passes through the deep neural network The actual gradient norm of the weight w obtained from the training, Indicates The branch passes through the deep neural network After training, we get the average gradient norm of the expected value corresponding to the weight w. represents the minimum absolute error of the gradient vector, o is an additional hyperparameter used to represent the speed balance strength in the dual-branch problem, which can measure the asymmetry between branches, balance the differences between branches through gradient loss, and distinguish the importance of each branch.

[0162] Step 3: Use the multi-source domain adaptive adversarial network MSAAN to perform transfer learning on the fault diagnosis network after preliminary training.

[0163] The steps of using the multi-source domain adaptive adversarial network MSAAN to transfer learning to the fault diagnosis network after preliminary training include:

[0164] Construct an adversarial transfer learning network MSDA-IRSCNN based on multi-source domain adaptation;

[0165] Using the adversarial transfer learning network MSDA-IRSCNN based on multi-source domain adaptation to extract features from multiple labeled source domain datasets includes using the private feature extractor and the shared feature extractor in the multi-source domain adaptation learning module to extract the private features and shared features of the source domain respectively;

[0166] Through adversarial training of multi-domain discriminators, the distribution difference between source domain and target domain data in the feature space is reduced to obtain shared features with domain invariance.

[0167] Using the shared features with domain invariance to train the shared layers of the fault diagnosis network TF-IRSCNN, to obtain the fault diagnosis network fine-tuned with labeled data;

[0168] After completing transfer learning, the fault diagnosis network TF-IRSCNN fine-tuned with labeled data is migrated to the target domain.

[0169] The network architecture of the multi-source domain adaptive deep adversarial transfer learning network MSDA-IRSCNN is as follows: Fig.11 shown.

[0170] In the process of transfer learning, the target domain data is often collected from real industrial scenarios, which are often complex, have unclear features, and lack labels. In this case, more comprehensive features can be extracted from multi-source domain data, and then feature learning can be performed on multi-source domain data through adaptive methods.

[0171] Multi Source Adaptive Adversarial Network (MSAAN) can integrate each source domain data by measuring the similarity between the source domain and target domain data, and use a domain adversarial model to interact with multiple source domain data.

[0172] In multi-source domain adaptation learning MSA, there are A labeled source domain dataset, denoted as ,in , It is the source domain dataset The source domain samples, It is the source domain dataset The source domain The labels of samples, the unlabeled data of the target domain is: , For the The first samples.

[0173] definition , Represent the number of all labeled source data and unlabeled target data, Represents all domain data, is the sample data, is the domain label of the data, where .

[0174] For sample data in the target domain , in order to obtain the decision value of the target label classifier , through Source domain classifier A weighted combination is used to approximate the target classifier. For sample data , approximate the label It is expressed as:

[0175] (16);

[0176] in, Indicates target domain instance to the The weight of the source domain, Indicates target domain instance to the The decision value of the source domain.

[0177] like Fig.10 As shown, the multi-source domain adaptive learning module includes Private feature extractors , a shared feature extractor , a multi-domain label predictor and a multi-domain discriminator , using a shared feature extractor Extract the invariant features of the source domain data using a private feature extractor and multi-domain discriminator To extract the domain feature information of each source domain data.

[0178] In the feature extraction process of multi-source domain adaptive learning, the shared feature extractor Get shared features of all domains , private feature extractor Get private features for each source domain , multi-domain discriminator Used to distinguish the data domain to which the instance data belongs , multi-domain label predictor Get label predictions for the target domain , for the data domain Normalize the data to get the confidence level , and finally predict the label A weighted combination is performed to obtain the predicted label of the target domain data.

[0179] There is a shared feature space in multi-source domain data with small distribution divergence. Adversarial training can reduce the discreteness between different data domains and obtain domain-invariant features. The multi-domain discriminator is used in the multi-source domain adaptive learning module. To conduct adversarial training and extract relevant losses. The target loss L wd It is expressed as:

[0180] (17);

[0181] in, For data The domain label, is the transpose operation, A domain classifier that predicts probabilities for domain labels of output data, is a shared feature extractor, is a private feature extractor. After adversarial training, It is possible to distinguish the data domains from which the data features come, N s Represents all labeled source data, and ln is the logarithm with the constant e as the base.

[0182] Multi-domain label predictor It is possible to predict labels for target domain data based on the connection features in multiple source domains, and to wc It is expressed as:

[0183] (18);

[0184] in, is the class label, is the classifier function, is a shared feature extractor, is a private feature extractor. In the case of multi-source domain adaptation, only using the domain discriminator to constrain the shared feature extractor will cause domain-related feature information to appear in both the shared space and the private space. Therefore, the performance of adaptive learning can be improved by introducing private feature extractors and shared feature extractors. As a constraint, the private features of each source domain are kept in their respective private spaces, and the public feature information is transferred from each source domain to the shared space. Therefore, each private feature extractor The target representation L wp for:

[0185] (19);

[0186] in, is the class label, For data The domain label, is the classifier function, A domain classifier that predicts probabilities for domain labels of output data, is a shared feature extractor, It is a private feature extractor.

[0187] Through adversarial training, the distribution difference between different data domains in the shared feature space is minimized, and the feature extractor is shared. The extracted features are domain invariant and can be shared among all domains, which will be the target L of the shared feature extractor. ws It is expressed as:

[0188] (20);

[0189] in, is the class label, For data The domain label, is the classifier function, A domain classifier that predicts probabilities for domain labels of output data, is a shared feature extractor, is a private feature extractor.

[0190] Goal of sharing feature extractors In Indicates the contribution of private features to the final label prediction, To combat losses, this can enhance The feature extraction capability is thus weakened When the balance point is reached, Will not be able to distinguish The extracted features.

[0191] This embodiment divides the multi-source domain adaptive learning network training process into two stages to update relevant network parameters. In the first stage, the multi-domain discriminator is enhanced by minimizing the adversarial loss. The second stage backtracks the classification error and updates the parameters and .

[0192] Step 4: Train the fault diagnosis network TF-IRSCNN.

[0193] Step 4.1: Build a time-frequency domain feature pre-training network.

[0194] Time domain data and frequency domain data reflect different information about the fault. If only relevant data features are extracted from a single dimension, comprehensive fault information cannot be obtained. In order to obtain the comprehensive features of the vibration signal in the time domain and frequency domain, dual-branch learning can be used. Fig. 9 The figure shows the time-frequency domain feature pre-training network TF-IRSCNN based on dual branches in the time-frequency domain.

[0195] Step 4.2: Input the original time domain data and frequency domain data of the bearing vibration signal into the time-frequency domain feature pre-training network to obtain pre-trained features.

[0196] The input data of the time-frequency domain feature pre-training network TF-IRSCNN includes a time domain branch and a frequency domain branch. First, the time-frequency domain feature pre-training network TF-IRSCNN is pre-trained. The original vibration data is time domain data. The original vibration data is obtained by fast Fourier transform FFT to obtain its frequency domain signal. 10 samples of time domain and frequency domain data are selected as manual labels. The corresponding features are learned from the time domain TD and frequency domain FD respectively through two branches. Then, the time-frequency domain dual-branch loss function is used to optimize the weights of the time-frequency domain feature pre-training network TF-IRSCNN to obtain the pre-training model features integrating time-frequency TF features. .

[0197] Step 4.3: Feed the pre-trained features into the fault diagnosis network that has been fine-tuned on the labeled data.

[0198] Replace all shared layer weights except the last two fully connected layers of the time-frequency domain feature pre-trained network TF-IRSCNN with the pre-trained features , freeze the parameters of all shared layers in the time-frequency domain feature extraction network, and use the pre-trained features Directly fed into a fault diagnosis network that is fine-tuned using labeled data.

[0199] Step 4.4: Use multiple source domain data to train the fault diagnosis network to obtain a trained fault diagnosis network.

[0200] Data features are extracted from multi-source domain data, including private data features in each source domain data and shared data features in multi-source domain data. The classification weights of the target domain data are obtained from the perspectives of different source domains through a domain discriminator. The final prediction label of the target domain data is obtained after weighted combination. The network parameters and structure of the fault diagnosis network are adjusted according to the final prediction label to obtain a trained fault diagnosis network.

[0201] Step 5: Input the acquired bearing vibration data of the target to be diagnosed into the trained fault diagnosis network TF-IRSCNN and output the fault diagnosis result.

[0202] Example 3

[0203] A computer-readable storage medium having computer instructions stored thereon, characterized in that when the computer instructions are executed by a processor, the steps of the method of embodiment 1 or 2 above are implemented.

[0204] Example 4

[0205] The present invention also provides a computer program product, including computer instructions, characterized in that when the computer instructions are executed by a processor, the steps of the method in the above-mentioned embodiment 1 or 2 are implemented.

[0206] In summary, the present invention proposes a fault diagnosis method based on sparse heterogeneous data collaboration, which is based on the fault diagnosis network TF-IRSCNN. Through the dual-branch feature extraction technology in the time-frequency domain, combined with the deep separable convolution module that introduces the width multiplier and the resolution multiplier, the network parameters and input data dimensions are effectively reduced. At the same time, the fault diagnosis network TF-IRSCNN also incorporates the inverted residual module, lightweight attention mechanism and nonlinear activation function, which significantly improves the fault recognition accuracy in the industrial Internet.

[0207] In order to further enhance the ability to utilize cross-domain data, the present invention also constructs an adversarial transfer learning network MSDA-IRSCNN based on multi-source domain adaptation, uses adversarial transfer learning technology to mine rich private and shared features from multi-source domain data, and assigns reasonable classification weights to different source domain data through domain discriminators, and finally obtains accurate prediction labels for target domain data by weighting. In a sparse data environment, this method can extract richer feature information from time-frequency domain data, which not only greatly improves the accuracy of fault identification, but also significantly reduces the delay in the calculation process, providing an efficient and accurate solution for fault diagnosis of the industrial Internet.

[0208] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0209] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0210] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0211] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0212] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

Claims

1. A fault diagnosis method based on sparse heterogeneous data collaboration, characterized in that: include: Acquire bearing vibration data of the target to be diagnosed; The acquired bearing vibration data of the target to be diagnosed is input into the trained fault diagnosis network TF-IRSCNN, and the fault diagnosis result is output; The training method of the fault diagnosis network TF-IRSCNN includes: The time domain data and frequency domain data of the bearing vibration signal are used as training sets; Use the depth-wise separable convolution module to perform spatial separation processing on the training set to obtain feature data after spatial separation processing; Through the inverted residual module, feature extraction and enhancement are performed on the feature data after spatial separation processing to obtain feature data after enhanced features; Performing weighted processing on the feature data after feature enhancement to obtain weighted feature data; Using the dual-branch loss function in the time-frequency domain, respectively calculating the time-domain branch loss and the frequency-domain branch loss of the weighted feature data, and using the gradient normalization method to automatically balance the weights of the time-domain branch loss and the frequency-domain branch loss, obtaining the overall loss function, minimizing the loss value, and obtaining the fault diagnosis network after preliminary training; The multi-source domain adaptive adversarial network (MSAAN) is used to perform transfer learning on the fault diagnosis network after preliminary training, and the fault diagnosis network is obtained after fine-tuning with labeled data. The network is pre-trained using time-frequency domain features, and pre-trained features are obtained based on the historical bearing vibration data corresponding to the target to be diagnosed; Sending the pre-trained features into the fault diagnosis network fine-tuned with the labeled data; A fault diagnosis network is trained using multiple source domain data to obtain a trained fault diagnosis network TF-IRSCNN, wherein the multiple source data are status data of different faults in the CWRU bearing data set, including bearing operation data of different loads and different fault locations.

2. The fault diagnosis method based on sparse heterogeneous data collaboration according to claim 1 is characterized in that: The fault diagnosis network TF-IRSCNN includes: The depth-separable convolution module is used to perform spatial separation processing on the training set to obtain feature data after spatial separation processing; An inverted residual module with a lightweight attention mechanism is added to extract and enhance the feature data after spatial separation processing to obtain feature data with enhanced features.

3. The fault diagnosis method based on sparse heterogeneous data collaboration according to claim 1 is characterized in that: The multi-source domain adaptive adversarial network (MSAAN) is used to perform transfer learning on the fault diagnosis network after preliminary training. After obtaining the fault diagnosis network fine-tuned with labeled data, the fault diagnosis network fine-tuned with labeled data is migrated to the target domain.

4. The fault diagnosis method based on sparse heterogeneous data collaboration according to claim 3 is characterized in that: A method for performing transfer learning on a fault diagnosis network after preliminary training using a multi-source domain adaptive adversarial network MSAAN to obtain a fault diagnosis network fine-tuned with labeled data, and then migrating the fault diagnosis network fine-tuned with labeled data to a target domain includes: Constructing an adversarial transfer learning network MSDA-IRSCNN based on multi-source domain adaptation, extracting features from multiple labeled source domain datasets includes using the private feature extractor and shared feature extractor in the multi-source domain adaptation learning module to extract private features and shared features of the source domain respectively; Through adversarial training of multi-domain discriminators, the distribution difference between source domain and target domain data in the feature space is reduced to obtain shared features with domain invariance. Using the shared features with domain invariance to train the shared layers of the fault diagnosis network TF-IRSCNN, to obtain the fault diagnosis network fine-tuned with labeled data; After completing transfer learning, the fault diagnosis network TF-IRSCNN fine-tuned with labeled data is migrated to the target domain.

5. The fault diagnosis method based on sparse heterogeneous data collaboration according to claim 1 is characterized in that: The method of obtaining pre-trained features by using a time-frequency domain feature pre-training network and according to the bearing vibration history data corresponding to the target to be diagnosed includes: Obtain the original time domain data of the bearing vibration signal through the sensor; Obtain the frequency domain data of the bearing vibration signal through fast Fourier transform FFT transformation; The original time domain data of the bearing vibration signal and the frequency domain data of the bearing vibration signal are input into the time-frequency domain feature pre-training network to obtain pre-trained features.

6. The fault diagnosis method based on sparse heterogeneous data collaboration according to claim 1 is characterized in that: The method of using multiple source domain data to train a fault diagnosis network to obtain a trained fault diagnosis network TF-IRSCNN includes: Data features are extracted from multi-source domain data, including private data features in each source domain data and shared data features in multi-source domain data. The classification weights of the target domain data are obtained from the perspectives of different source domains through a domain discriminator. The final predicted label of the target domain data is obtained after weighted combination. The network parameters and structure of the fault diagnosis network TF-IRSCNN are adjusted according to the final predicted label to obtain the trained fault diagnosis network TF-IRSCNN.

7. The fault diagnosis method based on sparse heterogeneous data collaboration according to claim 1 is characterized in that: Using the depth-wise separable convolution module, before spatially separating the input data, a width multiplier is introduced in the depth-wise separable convolutional neural network (SCNN). and resolution multiplier , the width multiplier Acting on the convolution operation, reducing the amount of hyperparameter calculation in the convolution operation, proportionally reducing the number of convolution kernels in each layer of the deep separable convolutional neural network SCNN, the resolution multiplier The input data is compressed to reduce the amount of internal input data of each layer of the deep separable convolutional neural network (SCNN).

8. The fault diagnosis method based on sparse heterogeneous data collaboration according to claim 1 is characterized in that: Using the dual-branch loss function in the time-frequency domain, the time-domain branch loss and the frequency-domain branch loss of the weighted feature data are calculated respectively, and the gradient normalization method is used to automatically balance the weights of the time-domain branch loss and the frequency-domain branch loss to obtain the overall loss function, which is expressed as: ； in, and They represent the branch loss in the time domain and frequency domain branch loss The weight of Represents the total loss of the two branches in the time-frequency domain.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images