Mechanical fault diagnosis method and system based on spiking neural network

Abstract

The invention relates to the technical field of mechanical fault diagnosis, in particular to a mechanical fault diagnosis method and system based on a pulse neural network. The method comprises the following steps: constructing a fault diagnosis network, wherein the fault diagnosis network comprises an improved gating attention coding module and a pulse residual network; the method comprises the following steps: replacing a common convolution layer of a gating attention coding module with a one-dimensional sparse convolution layer, and replacing a two-dimensional batch normalization layer with a one-dimensional batch normalization layer to obtain an improved gating attention coding module; setting a pulse neuron in the fault diagnosis network as a PSN neuron, and embedding a time step shrinkage layer in the pulse residual network; and inputting the mechanical vibration signal of the current sampling period into a fault diagnosis network, and outputting a predicted fault category label of the current sampling period. According to the invention, high-efficiency and high-precision mechanical fault intelligent diagnosis is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and in particular to a mechanical fault diagnosis method and system based on pulse neural network. Background Art

[0002] Mechanical systems are an important part of industrial automation. Once a failure occurs, it will affect the economic benefits of enterprise production and pose a safety hazard. Therefore, the fault diagnosis of mechanical systems is of great significance. Existing fault diagnosis methods are roughly divided into methods based on mathematical modeling and data-driven methods. Methods based on mathematical modeling have many limitations, such as requiring a lot of prior knowledge, difficulty in modeling complex systems, and high resource consumption in the maintenance process. With the development of artificial intelligence and big data technology, data-driven methods have received widespread attention in recent years, mainly including signal processing methods and machine learning methods. Among them, signal processing and traditional machine learning methods usually require certain prior knowledge and theoretical knowledge of statistical learning. Correspondingly, deep learning in machine learning can learn autonomously, providing a powerful support tool for analyzing large-capacity data for fault diagnosis. With the efforts of researchers, many deep learning network models have emerged in recent years, such as convolutional neural networks, long short-term memory networks, generative adversarial networks, graph neural networks, and Transformers. The above network models have been widely used in fields such as image recognition and intelligent fault diagnosis and have achieved excellent results. However, as a black box model, deep learning has excellent performance but lacks certain interpretability.

[0003] Based on the demand for interpretability, the spiking neural network (SNN), known as the "third generation neural network", was proposed to imitate the information transmission method of biological neurons. The spiking neural network not only has strong biological interpretability, but its discrete binary characteristics also enable it to save a certain amount of energy when deployed on hardware. Therefore, the spiking neural network (SNN) quickly became a research hotspot in the field of deep learning. When the spiking neural network (SNN) initially entered the field of fault diagnosis, researchers first built a simple single-layer spiking neural network (SNN) to diagnose rolling bearing faults. The experimental results confirmed the feasibility of the spiking neural network (SNN) for mechanical equipment fault diagnosis and its unique advantages in interpretability. The current research on spiking neural networks (SNN) is divided into multiple aspects, covering encoding methods, training algorithms, neuron models, network architectures, etc. The existing commonly used spiking neural network (SNN) pulse encoding method is gated attention encoding, which is used for two-dimensional image processing. It uses an ordinary two-dimensional convolutional layer to extract input features, and after the batch normalization processing layer and the spiking neuron, it is multiplied element by element with the output of the gated attention unit. The pulse encoding result of the gated attention encoding is obtained in the above way. The gated attention unit completes the fusion of temporal attention and spatial attention. The addition of the attention mechanism can better guide the process of pulse encoding. Therefore, the existing gated attention encoding method can perform pulse encoding well. Unlike traditional artificial neural networks, spiking neural networks (SNNs) have the concept of time and incorporate the time step parameter. Researchers have found that the setting of the time step greatly affects the training effect of spiking neural networks (SNNs). Specifically, too large a time step will lead to insufficient memory and slow training speed, while too small a time step will lead to the loss of the temporal characteristics of the spiking neural network (SNN) and poor training effect; existing spiking neural networks (SNNs) usually choose a fixed moderate value, such as 4 or 8, to avoid the above problems. For spiking neurons, the most commonly used one is the leaky integral firing (LIF) neuron, because this neuron model better balances biological characteristics and computational complexity. The leaky integral firing (LIF) neuron performs a serial iterative process through its dynamic learning charge and discharge and trigger reset operations to complete the transmission of spiking neural network (SNN) information and achieve good training results.

[0004] However, the existing gated attention coding module is used to process the two-dimensional structure and spatial correlation of images. It is only applicable to two-dimensional images, but not suitable for processing one-dimensional time series signals. Therefore, it cannot be directly used for pulse coding of mechanical vibration signals. At the same time, the gated attention coding module uses ordinary convolution, which involves all input features or pixels in each convolution operation. It does not fully consider which areas are more important for feature extraction, which reduces the accuracy and efficiency of the model. In addition, the existing pulse neural network sets the time step to a fixed moderate value. Although it can alleviate the adverse effects of the time step on network training to a certain extent, it cannot adapt to the dynamic changes of different input data and may lead to unnecessary calculations. Especially when the input information is simple or redundant, too long a time step will waste computing resources, and too short a time step may cause information loss, which in turn affects the learning rate and convergence effect of the model. In addition, fault diagnosis of mechanical equipment usually requires real-time processing of a large amount of data to achieve rapid evaluation and prediction of the health status of the equipment. The leaky integral firing (LIF) neuron model is a serial transmission mode. Each neuron state update usually depends on the state of the previous time step. This gradual update feature limits the efficiency of parallel computing, especially in the application of large-scale neural networks, which may lead to high time costs. Summary of the invention

[0005] To this end, the technical problem to be solved by the present invention is to overcome the defects that the existing gated attention encoding module is not suitable for one-dimensional signals and has low sparsity, the time step is set to a fixed value and has limitations, and the serial transmission mode of the pulse neuron is inefficient, resulting in poor accuracy in mechanical fault diagnosis.

[0006] In order to solve the above technical problems, the present invention provides a mechanical fault diagnosis method based on a pulse neural network, comprising the following steps: Constructing a fault diagnosis network, the fault diagnosis network comprising: an improved gated attention coding module and a pulse residual network; wherein the improved gated attention coding module is obtained by replacing the ordinary convolution layer of the gated attention coding module with a one-dimensional sparse convolution layer, and replacing the two-dimensional batch normalization layer with a one-dimensional batch normalization layer; setting the pulse neurons in the fault diagnosis network as PSN neurons, and embedding a time step shrinkage layer in the pulse residual network; The mechanical vibration signal of the current sampling period is input into the improved gated attention coding module for pulse coding to obtain the pulse sequence of the current sampling period; The pulse sequence of the current sampling period is input into the pulse residual network, and the predicted fault category label of the current sampling period is output.

[0007] Preferably, the step of inputting the mechanical vibration signal of the current sampling period into the improved gated attention coding module for pulse coding to obtain the pulse sequence of the current sampling period includes: The mechanical vibration signal of the current sampling period is sequentially passed through a one-dimensional sparse convolution layer, a one-dimensional batch normalization layer, and a PSN neuron to generate a pulse output of the current sampling period; The mechanical vibration signal of the current sampling period is input into the gated attention unit to generate the attention weight map of the current sampling period; The pulse output of the current sampling period is Hadamard-producted with the attention weight map to obtain the pulse sequence of the current sampling period.

[0008] Preferably, the gated attention unit includes: a parallel temporal attention channel and a spatial attention channel; the temporal attention channel includes: a maximum pooling layer, an average pooling layer and a shared multi-layer perceptron; the spatial attention channel includes: a sequence conversion function.

[0009] Preferably, the step of inputting the mechanical vibration signal of the current sampling period into the gated attention unit to generate the attention weight map of the current sampling period includes: The mechanical vibration signal of the current sampling period is input into the time attention channel, and the time attention features of the current sampling period are output, including: The mechanical vibration signal of the current sampling period is sequentially passed through the maximum pooling layer and the shared multi-layer perceptron to obtain the first mechanical vibration feature of the current sampling period; The mechanical vibration signal of the current sampling history period is sequentially passed through the average pooling layer and the shared multi-layer perceptron to obtain the second mechanical vibration feature of the current sampling period; Adding the first mechanical vibration feature and the second mechanical vibration feature of the current sampling period to obtain the time attention feature of the current sampling period; The mechanical vibration signal of the current sampling period is input into the spatial attention channel, and the spatial attention map of the current sampling period is output; After multiplying the temporal attention features of the current sampling period by the spatial attention map element by element, the attention weight map of the current sampling period is obtained by processing it through the Sigmoid function.

[0010] Preferably, the time step shrinkage layer includes: a fully connected layer, a maximum pooling layer and a Softmax classifier connected in sequence.

[0011] Preferably, a time step shrinkage layer is embedded between two convolutional layers where the number of channels changes in the pulse residual network.

[0012] Preferably, any one of SEW-ResNet18, MS-ResNet18, and SEW-ResNet50 is used as the basic network, a time step shrinkage layer is embedded between two basic layers in which the number of convolution channels in the basic network is changed, and the pulse neurons in the basic network are set to PSN neurons to construct a pulse residual network.

[0013] Preferably, a pulse residual network is constructed based on SEW-ResNet18, and the pulse residual network includes: a downsampling layer, a first basic layer, a time step contraction layer, a second basic layer, a time step contraction layer, a third basic layer, a time step contraction layer, a fourth basic layer and a fully connected layer connected in sequence, and the number of channels of the four basic layers are different; Among them, the downsampling layer includes: a convolutional layer, a batch normalization layer, a PSN neuron, and a maximum pooling layer connected in sequence. Each basic layer consists of two basic blocks with the same number of channels. Each basic block includes: a 3×3 convolution, a batch normalization layer, a PSN neuron, a 3×3 convolution, a batch normalization layer and a PSN neuron connected in sequence.

[0014] Preferably, the dynamics formula of the PSN neuron is: , in, is the input sequence of PSN neurons, is the learnable weight matrix, is the hidden state sequence, B is the learnable threshold, is a unit step function, is the binary output pulse train of the PSN neuron.

[0015] The present invention also provides a mechanical fault diagnosis system based on a pulse neural network, comprising: A model building module is used to build a fault diagnosis network, wherein the fault diagnosis network includes: an improved gated attention coding module and a pulse residual network; wherein the improved gated attention coding module is obtained by replacing the ordinary convolution layer of the gated attention coding module with a one-dimensional sparse convolution layer, and replacing the two-dimensional batch normalization layer with a one-dimensional batch normalization layer; the pulse neurons in the fault diagnosis network are set as PSN neurons, and a time step shrinkage layer is embedded in the pulse residual network; A pulse sequence acquisition module is used to input the mechanical vibration signal of the current sampling period into the improved gated attention encoding module, perform pulse encoding, and obtain the pulse sequence of the current sampling period; The prediction module is used to input the pulse sequence of the current sampling period into the pulse residual network and output the predicted fault category label of the current sampling period.

[0016] The above technical solution of the present invention has the following beneficial effects compared with the prior art: The present invention discloses a mechanical fault diagnosis method and system based on a pulse neural network, by replacing the ordinary convolution layer of the gated attention coding module with a one-dimensional sparse convolution layer, and replacing the two-dimensional batch normalization layer with a one-dimensional batch normalization layer; since the mechanical vibration signal is usually a one-dimensional time series, the timing characteristics of the signal are very important. The one-dimensional sparse convolution layer can better standardize the time series data and only focus on the important parts of the mechanical vibration signal, thereby improving the sparsity and computational efficiency of the network, avoiding the limitations of the traditional image processing network structure in processing mechanical vibration signals, and can better decode and process the timing characteristics of these pulse signals, thereby improving the diagnostic accuracy of mechanical faults.

[0017] The time step shrinkage layer is embedded in the pulse residual network; the time step shrinkage layer includes a fully connected layer, a maximum pooling layer and a Softmax classifier connected in sequence. The time step shrinkage layer receives the time step of the previous layer and the time step size to be converted, and smoothly completes the conversion of the time step through two channels without losing information in other dimensions. It can not only effectively extract features related to the time step, but also perform feature aggregation and classification. The maximum pooling layer helps to extract important spatiotemporal features, while the Softmax classifier can achieve efficient classification output. The time step shrinkage layer dynamically adjusts the time step, so that the network can automatically adjust the time scale of the calculation according to the characteristics of the input data, which can more accurately capture the important features in the signal, avoid redundant calculations, and improve the accuracy and efficiency of mechanical fault diagnosis.

[0018] In addition, traditional neuron models usually rely on iterative summation, which leads to complex calculation process and large time cost. The present invention sets the pulse neuron as PSN neuron. PSN neuron can make full use of the parallelism of matrix operations and can process data of multiple time steps at the same time without relying on the sequential update of time series, which significantly improves the processing speed of fault diagnosis network. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein: Figure 1 It is a training diagram of the fault diagnosis network.

[0020] Figure 2 It is a test diagram of the fault diagnosis network.

[0021] Figure 3 This is a t-SNE dimension reduction visualization of the classification results of the gearbox small bearing test data sample.

[0022] Figure 4It is a confusion matrix diagram of the classification results of the gearbox small bearing test data sample. DETAILED DESCRIPTION

[0023] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0024] Embodiment 1 of the present invention provides a mechanical fault diagnosis method based on a pulse neural network, comprising the following steps: Step S1: constructing a fault diagnosis network, the fault diagnosis network comprising: an improved gated attention coding module and a pulse residual network; wherein the improved gated attention coding module is obtained by replacing the ordinary convolution layer of the gated attention coding module with a one-dimensional sparse convolution layer, and replacing the two-dimensional batch normalization layer with a one-dimensional batch normalization layer; setting the pulse neurons in the fault diagnosis network as PSN neurons, and embedding a time step shrinkage layer in the pulse residual network; In this embodiment, preferably, the time step shrinkage layer includes: a fully connected layer, a maximum pooling layer and a Softmax classifier connected in sequence.

[0025] Embedding the time step shrinkage layer in the pulse residual network can not only effectively extract the features related to the time step, but also perform feature aggregation and classification. The maximum pooling layer helps to extract important spatiotemporal features, while the Softmax classifier can achieve efficient classification output. The time step shrinkage layer dynamically adjusts the time step, allowing the network to automatically adjust the time scale of calculation according to the characteristics of the input data, which can more accurately capture the important features in the signal, avoid redundant calculations, and improve the accuracy and efficiency of mechanical fault diagnosis.

[0026] The time-step shrinkage layer has the feature of plug-and-play and can be freely embedded between the layers of the Spike Residual Network.

[0027] In this embodiment, preferably, a time step contraction layer is embedded between two layers where the number of convolution channels changes in the pulse residual network, so as to gradually shrink the value of the time step during the training process. The time step contraction layer achieves the compression and refinement of information between the two layers where the number of convolution channels changes by introducing the maximum pooling layer and the fully connected layer. This structure helps to reduce the amount of calculation of subsequent layers, because the pooling layer can reduce the size of the feature map, and the fully connected layer reduces the number of weights that need to be trained by parameter sharing. By adding this layer at the position where the number of convolution channels changes, the amount of calculation and storage requirements of each convolution operation can be effectively reduced, especially when the number of convolution channels increases, the computational cost may increase sharply. At the end of the time step contraction layer, the introduction of the Softmax classifier helps to perform the final classification processing on the extracted features. By embedding the Softmax classifier between the two layers where the number of convolution channels changes, it can be ensured that the network can make corresponding decisions at each time step after the feature conversion. This not only helps the end-to-end optimization of the model, but also further improves the classification accuracy of the network on time series tasks.

[0028] In this embodiment, specifically, the dynamics formula of the PSN neuron is: , in, is the input sequence of PSN neurons, is the learnable weight matrix, is the hidden state sequence, B is the learnable threshold, is a unit step function, is the binary output pulse train of the PSN neuron.

[0029] The present invention sets the pulse neuron as a PSN neuron. The PSN neuron can fully utilize the parallelism of matrix operations and can process data of multiple time steps simultaneously without relying on the sequential gradual update of the time series, thereby significantly improving the processing speed of the fault diagnosis network.

[0030] Step S2: inputting the mechanical vibration signal of the current sampling period into the improved gated attention coding module, performing pulse coding, and obtaining the pulse sequence of the current sampling period; Specifically, the mechanical vibration signal of the current sampling period is input into the improved gated attention coding module for pulse coding to obtain the pulse sequence of the current sampling period, including: Step S21: The mechanical vibration signal of the current sampling period is sequentially passed through a one-dimensional sparse convolution layer, a one-dimensional batch normalization layer, and a PSN neuron to generate a pulse output of the current sampling period; Step S22: inputting the mechanical vibration signal of the current sampling period into the gated attention unit to generate an attention weight map of the current sampling period; In this embodiment, specifically, the gated attention unit includes: a parallel temporal attention channel and a spatial attention channel; the temporal attention channel includes: a maximum pooling layer, an average pooling layer and a shared multi-layer perceptron; the spatial attention channel includes: a sequence conversion function.

[0031] The step of inputting the mechanical vibration signal of the current sampling period into the gated attention unit to generate an attention weight map of the current sampling period includes: Step S221: Input the mechanical vibration signal of the current sampling period into the time attention channel, and output the time attention features of the current sampling period, including: Step S222: The mechanical vibration signal of the current sampling period is sequentially passed through the maximum pooling layer and the shared multi-layer perceptron to obtain the first mechanical vibration feature of the current sampling period; Step S223: The mechanical vibration signal of the current sampling history period is sequentially passed through the average pooling layer and the shared multi-layer perceptron to obtain the second mechanical vibration feature of the current sampling period; Step S224: adding the first mechanical vibration feature and the second mechanical vibration feature of the current sampling period to obtain the time attention feature of the current sampling period; Step S225: inputting the mechanical vibration signal of the current sampling period into the spatial attention channel, and outputting the spatial attention map of the current sampling period; Step S226: After multiplying the temporal attention features of the current sampling period by the spatial attention map element by element, the attention weight map of the current sampling period is obtained by processing through the Sigmoid function.

[0032] Step S23: Perform a Hadamard product of the pulse output of the current sampling period and the attention weight map to obtain the pulse sequence of the current sampling period.

[0033] Step S3: Input the pulse sequence of the current sampling period into the pulse residual network, and output the predicted fault category label of the current sampling period.

[0034] In this embodiment, specifically, any one of SEW-ResNet18, MS-ResNet18, and SEW-ResNet50 is used as the basic network, a time step shrinkage layer is embedded between two basic layers in which the number of convolution channels in the basic network is changed, and the pulse neurons in the basic network are set to PSN neurons to construct a pulse residual network.

[0035] Preferably, a pulse residual network is constructed based on SEW-ResNet18, and the pulse residual network includes: a downsampling layer, a first basic layer, a time step contraction layer, a second basic layer, a time step contraction layer, a third basic layer, a time step contraction layer, a fourth basic layer and a fully connected layer connected in sequence, and the number of channels of the four basic layers are different; Among them, the downsampling layer includes: a convolutional layer, a batch normalization layer, a PSN neuron, and a maximum pooling layer connected in sequence. Each basic layer consists of two basic blocks with the same number of channels. Each basic block includes: a 3×3 convolution, a batch normalization layer, a PSN neuron, a 3×3 convolution, a batch normalization layer and a PSN neuron connected in sequence.

[0036] Therefore, the number of channels will change between each basic layer, and the time step shrinkage layer is embedded between two adjacent basic layers. Finally, it is classified through the maximum pooling layer and the fully connected layer to form an 18-layer network depth.

[0037] Specifically, a first time step shrinkage layer is embedded between the first basic layer and the second basic layer of the pulse residual network to change the initial time step from 32 to 16; a second time step shrinkage layer is embedded between the second basic layer and the third basic layer of the pulse residual network to change the time step from 16 to 8; a third time step shrinkage layer is embedded between the third basic layer and the fourth basic layer of the pulse residual network to change the time step from 8 to 4.

[0038] In this embodiment, preferably, the training process of the fault diagnosis network is: Acquire a mechanical failure data set and perform preprocessing to divide the mechanical failure data set into a training set and a test set; wherein the mechanical failure data set includes mechanical vibration signals of multiple historical sampling periods and their true labels; In this embodiment, specifically, preprocessing the mechanical fault data set includes: intercepting mechanical vibration signals of multiple sampling periods as data samples, unifying the sample lengths, and normalizing the sample amplitudes to a range of [0, 1].

[0039] The training set is used to train the fault diagnosis network, and the test data set does not participate in model training and is only used to test the accuracy of the model results. The mechanical working conditions and fault categories contained in the training data set and the test data set are the same.

[0040] like Figure 1 As shown, Figure 1 The training diagram of the fault diagnosis network is shown in Figure 2. The training set is input into the fault diagnosis network to obtain the predicted fault category label of each historical sampling period. The fault diagnosis network is trained using an optimization algorithm to obtain a trained fault diagnosis network. like Figure 2 As shown, Figure 2This is a test diagram of the fault diagnosis network. The performance of the fault diagnosis network is evaluated through the test set to obtain the target fault diagnosis network; The optimization algorithm is any one of a root mean square transfer algorithm, a stochastic gradient descent method, and an adaptive moment estimation algorithm.

[0041] This embodiment 1 improves the one-dimensional applicability of the gated attention coding module and the sparsity of the pulse coding process by improving the existing coding method; at the same time, a time step contraction layer is designed to smooth the time step conversion process, so as to improve the convergence speed of the model while retaining information; further, the processing speed and efficiency of network transmission are improved by using parallel processing PSN pulse neurons. Finally, a fault diagnosis network is formed that can extract deep features of mechanical vibration signals. The fault diagnosis network proposed in this embodiment 1 has high sparsity in extracting features, low energy consumption in hardware, high accuracy in intelligent diagnosis of mechanical faults, and fast processing speed.

[0042] Based on the first embodiment, the second embodiment of the present invention uses a small bearing fault vibration data set of a high-speed railway gearbox as an example to perform intelligent fault diagnosis, including: Taking a small bearing fault vibration dataset of a high-speed railway gearbox as an example, the dataset contains five types of health status data: normal state (N), rolling element crack fault (B), outer ring crack fault (O), outer ring corrosion fault (OC) and outer ring rolling element crack composite fault (OB). The fault category labels are represented by 0, 1, 2, 3 and 4 respectively. The data sampling frequency is 10.24kHz. The experiment contains data collected at two different speeds, positive and negative. The description of the gearbox small bearing fault vibration data used in this embodiment 2 is shown in Table 1. Table 1 is a description of the gearbox small bearing fault vibration dataset.

[0043] Table 1

[0044] In order to highlight the anti-interference ability of the fault diagnosis network proposed in the present invention and test whether the fault diagnosis network can correctly classify fault categories under different working conditions, the data under positive and negative speeds are packaged into a category label in the same proportion in the second embodiment. In the preprocessing, 1000 data points are used as a sample, and the number of samples in each sub-state is 320, so the number of samples for each category label is 640, and the training set and the test set are divided at a ratio of 4:1, that is, each category has 512 samples for training and 128 samples for testing.

[0045] The gearbox vibration signal is intercepted into a data sample with a length of 1000 points, and then the sample amplitude is normalized to the range of [0, 1], and the data set is divided into a training data set and a test data set.

[0046] The preprocessed time domain signal data samples are used as the input of the fault diagnosis network. The pulse encoding is performed through the improved gated attention encoding module to output a pulse sequence with a length of 1000. Among them, the input channel number of the improved gated attention encoding module is 1, the output channel number is 3, the convolution kernel size is 3, and the time step is 32.

[0047] In the second embodiment, the time step value of the PSN neuron is set to 32, and a pulse residual network is constructed based on SEW-ResNet18.

[0048] In the second embodiment, the fault diagnosis network is trained using the training set, and the mean square error between the predicted fault category label output by the fault diagnosis network and its true label is minimized to optimize the training. The optimization algorithm uses the adaptive moment estimation algorithm (Adam), the learning rate is 0.001, the momentum is 0.9, and the loss function tends to balance after 16 iterations, completing the training of the fault diagnosis network. The test set is input into the trained fault diagnosis network to perform intelligent fault diagnosis, and the test accuracy result is calculated from the predicted fault category label and the true label obtained by inputting the test sample into the fault diagnosis network.

[0049] like Figure 3 As shown, Figure 3 This is a t-SNE dimension reduction visualization diagram of the classification results of the gearbox small bearing test data sample. Figure 3 It can be seen that the mechanical fault diagnosis method based on pulse neural network proposed in the present invention can effectively cluster samples of the same category and make the features of samples of different categories have more obvious boundaries, which proves that the present invention can learn the deep features in the gearbox small bearing data set and improve the intra-class clustering and inter-class separability of fault categories.

[0050] like Figure 4 As shown, Figure 4 This is the confusion matrix diagram of the classification results of the gearbox small bearing test data sample. Figure 4 It can be seen that the mechanical fault diagnosis method based on pulse neural network proposed in the present invention has a diagnostic accuracy of 99.84%, and only one classification error occurred in the 640 test samples in five health states, demonstrating the superior diagnostic performance and generalization ability of the method of the present invention.

[0051] In summary, the present invention establishes a fault diagnosis network, performs sparse pulse coding through an improved gated attention coding module, and replaces the commonly used serial LIF neurons with parallel PSN neurons by embedding a time step shrinkage layer between the levels of the pulse residual network. This allows the model to efficiently extract deep-level features of the gearbox small bearing fault data set and achieve high-precision, high-efficiency and anti-working condition interference intelligent fault diagnosis.

[0052] Embodiment 3 of the present invention provides a mechanical fault diagnosis system based on a pulse neural network, comprising: A model building module is used to build a fault diagnosis network, wherein the fault diagnosis network includes: an improved gated attention coding module and a pulse residual network; wherein the improved gated attention coding module is obtained by replacing the ordinary convolution layer of the gated attention coding module with a one-dimensional sparse convolution layer, and replacing the two-dimensional batch normalization layer with a one-dimensional batch normalization layer; the pulse neurons in the fault diagnosis network are set as PSN neurons, and a time step shrinkage layer is embedded in the pulse residual network; A pulse sequence acquisition module is used to input the mechanical vibration signal of the current sampling period into the gated attention encoding module, perform pulse encoding, and obtain the pulse sequence of the current sampling period; The prediction module is used to input the pulse sequence of the current sampling period into the pulse residual network and output the predicted fault category label of the current sampling period.

[0053] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt 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 codes.

[0054] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes 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 generate 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.

[0055] 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.

[0056] 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.

[0057] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A mechanical fault diagnosis method based on pulse neural network, characterized in that: The following steps are involved: Constructing a fault diagnosis network, the fault diagnosis network comprising: an improved gated attention coding module and a pulse residual network; wherein the improved gated attention coding module is obtained by replacing the ordinary convolution layer of the gated attention coding module with a one-dimensional sparse convolution layer, and replacing the two-dimensional batch normalization layer with a one-dimensional batch normalization layer; setting the pulse neurons in the fault diagnosis network as PSN neurons, and embedding a time step shrinkage layer in the pulse residual network; The mechanical vibration signal of the current sampling period is input into the improved gated attention coding module for pulse coding to obtain the pulse sequence of the current sampling period; The pulse sequence of the current sampling period is input into the pulse residual network, and the predicted fault category label of the current sampling period is output.

2. A mechanical fault diagnosis method based on pulse neural network according to claim 1, characterized in that: The mechanical vibration signal of the current sampling period is input into the improved gated attention coding module for pulse coding to obtain the pulse sequence of the current sampling period, including: The mechanical vibration signal of the current sampling period is sequentially passed through a one-dimensional sparse convolution layer, a one-dimensional batch normalization layer, and a PSN neuron to generate a pulse output of the current sampling period; The mechanical vibration signal of the current sampling period is input into the gated attention unit to generate the attention weight map of the current sampling period; The pulse output of the current sampling period is Hadamard-producted with the attention weight map to obtain the pulse sequence of the current sampling period.

3. A mechanical fault diagnosis method based on pulse neural network according to claim 2, characterized in that: The gated attention unit includes: parallel temporal attention channel and spatial attention channel; the temporal attention channel includes: maximum pooling layer, average pooling layer and shared multi-layer perceptron; the spatial attention channel includes: sequence conversion function.

4. A mechanical fault diagnosis method based on pulse neural network according to claim 3, characterized in that: The step of inputting the mechanical vibration signal of the current sampling period into the gated attention unit to generate an attention weight map of the current sampling period includes: The mechanical vibration signal of the current sampling period is input into the time attention channel, and the time attention features of the current sampling period are output, including: The mechanical vibration signal of the current sampling period is sequentially passed through the maximum pooling layer and the shared multi-layer perceptron to obtain the first mechanical vibration feature of the current sampling period; The mechanical vibration signal of the current sampling history period is sequentially passed through the average pooling layer and the shared multi-layer perceptron to obtain the second mechanical vibration feature of the current sampling period; Adding the first mechanical vibration feature and the second mechanical vibration feature of the current sampling period to obtain the time attention feature of the current sampling period; The mechanical vibration signal of the current sampling period is input into the spatial attention channel, and the spatial attention map of the current sampling period is output; After multiplying the temporal attention features of the current sampling period by the spatial attention map element by element, the attention weight map of the current sampling period is obtained by processing it through the Sigmoid function.

5. The mechanical fault diagnosis method based on pulse neural network according to claim 1 is characterized in that: The time step shrinkage layer includes: a fully connected layer, a maximum pooling layer and a Softmax classifier connected in sequence.

6. A mechanical fault diagnosis method based on pulse neural network according to claim 5, characterized in that: A time-step shrinkage layer is embedded between two convolutional layers with changed number of channels in the Spike Residual Network.

7. A mechanical fault diagnosis method based on pulse neural network according to claim 6, characterized in that: Take any one of SEW-ResNet18, MS-ResNet18, and SEW-ResNet50 as the base network, embed a time step shrinkage layer between the two basic layers in which the number of convolution channels in the base network changes, and set the spiking neurons in the base network to PSN neurons to construct a spiking residual network.

8. The mechanical fault diagnosis method based on pulse neural network according to claim 7 is characterized in that: A pulse residual network is constructed based on SEW-ResNet18. The pulse residual network includes: a downsampling layer, a first basic layer, a time step contraction layer, a second basic layer, a time step contraction layer, a third basic layer, a time step contraction layer, a fourth basic layer and a fully connected layer connected in sequence. The number of channels in the four basic layers is different. Among them, the downsampling layer includes: a convolutional layer, a batch normalization layer, a PSN neuron, and a maximum pooling layer connected in sequence. Each basic layer consists of two basic blocks with the same number of channels. Each basic block includes: a 3×3 convolution, a batch normalization layer, a PSN neuron, a 3×3 convolution, a batch normalization layer and a PSN neuron connected in sequence.

9. The mechanical fault diagnosis method based on pulse neural network according to claim 1 is characterized in that: The dynamics formula of PSN neurons is: ， in, is the input sequence of PSN neurons, is the learnable weight matrix, is the hidden state sequence, B is the learnable threshold, is a unit step function, is the binary output pulse train of the PSN neuron.

10. A mechanical fault diagnosis system based on pulse neural network, characterized in that: include: A model building module is used to build a fault diagnosis network, wherein the fault diagnosis network includes: an improved gated attention coding module and a pulse residual network; wherein the improved gated attention coding module is obtained by replacing the ordinary convolution layer of the gated attention coding module with a one-dimensional sparse convolution layer, and replacing the two-dimensional batch normalization layer with a one-dimensional batch normalization layer; the pulse neurons in the fault diagnosis network are set as PSN neurons, and a time step shrinkage layer is embedded in the pulse residual network; A pulse sequence acquisition module is used to input the mechanical vibration signal of the current sampling period into the improved gated attention encoding module, perform pulse encoding, and obtain the pulse sequence of the current sampling period; The prediction module is used to input the pulse sequence of the current sampling period into the pulse residual network and output the predicted fault category label of the current sampling period.

Images