Fault prediction method and system for virtual-real migration equipment based on SNN network

Abstract

The invention discloses a virtual-real migration equipment fault prediction method and system based on an SNN network, and belongs to the technical field of mechanical equipment fault diagnosis. Real fault data are deficient, and virtual and real data have distribution differences. According to the SNN network-based virtual-real migration equipment fault prediction method and system, a source domain adopts a sufficient fault number generated by a simulation model, a target domain adopts a small amount of real fault data, and isometric samples are generated through sliding segmentation; sNN network parameters are optimized through a genetic algorithm, and classification loss and domain loss are alternately minimized, so that the model reduces inter-domain distribution difference while extracting fault features. And inputting the target domain state sequence into the trained model, and outputting a fault type prediction result through the classification layer. Aiming at health management requirements of key parts such as rolling bearings, the equipment fault diagnosis accuracy is improved through fusion of simulation data and transfer learning of real data.

Description

TECHNICAL FIELD

[0001] The application relates to a mechanical equipment fault diagnosis method, in particular to a virtual-real migration equipment fault prediction method and system based on an SNN network. BACKGROUND

[0002] In the field of equipment fault diagnosis, a data-driven method relies on a large amount of labeled real fault data, but in an actual scene, such data is high in acquisition cost and scarce in samples, leading to insufficient model generalization ability. Although a traditional transfer learning method can utilize source domain data to assist target domain diagnosis, simulation data and real data are significantly different in feature space distribution due to amplitude difference and different noise distribution, so that common features cannot be effectively extracted. When processing mechanical vibration data, an existing domain adaptation technology has limited diagnosis performance improvement because the signal pulse characteristics are not fully considered. Therefore, a virtual-real migration method combining data characteristics and a biologically inspired neural network is needed to reduce the difference between domains and improve diagnosis accuracy. SUMMARY

[0003] The application aims to solve the problems of lack of real fault data of equipment and distribution difference between virtual and real data, and proposes a virtual-real migration equipment fault prediction method and system based on an SNN network.

[0004] The above object is achieved by the following technical scheme: A virtual-real migration equipment fault prediction method based on an SNN network, the method comprising the following steps: The step of constructing a data set refers to constructing a source domain data set containing simulation fault data and a target domain data set containing a small amount of real fault data. The simulation fault data is generated based on a fault mechanism model and contains various fault types. The real data comes from production equipment life cycle monitoring data. The step of state sequence conversion refers to generating a time-frequency matrix by performing short-time Fourier transform on the vibration signals of the source domain and the target domain respectively, clustering the time-frequency matrix array into two categories of random steady state and repeatable transient state through a frequency spectrum clustering algorithm, and generating a binary state sequence. The step of constructing an SNN network refers to building a pulse neural network containing a feature extraction module, a classification module and a domain adaptation module. The feature extraction module adopts a pulse convolution layer and a pooling layer of a leaky integral discharge neuron to extract time sequence pulse features of the state sequence layer by layer. The step of training domain adaptation refers to quantifying the feature distribution difference between the source domain and the target domain through a JS divergence loss function, constructing a total loss function combined with a cross-entropy classification loss, optimizing network parameters through adversarial training, and making the feature extractor learn domain-invariant fault features. The step of fault diagnosis reasoning refers to inputting the target domain state sequence into the trained network, and outputting a fault type prediction result through the classification module.

[0005] Further, the process of generating a binary state sequence by clustering the time-frequency matrix column vectors into two categories of random steady state and repeatable transient state through a spectral clustering algorithm is specifically, using K-means++ to initialize the cluster center, taking each column vector of the time-frequency matrix as input, setting the cluster number to 2, generating a 0-1 binary state sequence, retaining the fault frequency characteristics while reducing the data dimension.

[0006] Further, in the SNN network, the structure of the feature extraction module includes three levels of pulse convolution layers, two levels of pulse pooling layers and a pulse full connection layer, wherein the three levels of pulse convolution layers include: S-Conv1 adopts a 5x1 convolution kernel, 64 channels, and LIF neuron activation; S-Conv2 adopts a 3x1 convolution kernel and 128 channels; S-Conv3 adopts a 3x1 convolution kernel and 256 channels; The two levels of pulse pooling layers both adopt 2x1 maximum pooling operations; The classification module outputs classification probabilities through the pulse full connection layer, and adopts a cross-entropy loss function to optimize the classification performance; The domain adaptation module introduces JS divergence as a domain loss function to quantify the feature distribution difference between the source domain and the target domain, and forces the feature extractor to learn domain-invariant features through adversarial training.

[0007] Further, in the process of constructing the total loss function by combining the cross-entropy classification loss, the total loss function expression is: L = L CS + L CT + λ LD wherein, L CS and L CT are the cross-entropy classification losses of the source domain and the target domain respectively, LD is the JS divergence domain loss, λ is a domain adaptation weight parameter, and the adversarial training is realized through a gradient reversal layer.

[0008] A system of a virtual-real transfer equipment fault prediction method based on an SNN network, the system comprising: a data layer, a model layer, and an application layer; The data layer is used to collect vibration data, construct a source domain simulation data set and a target domain real data set, and convert them into binary state sequences through spectral clustering; The model layer is used for deploying the SNN network, and comprises a feature extraction module, a classification module and a domain adaptation module, the feature extraction module adopts pulse convolution and a pooling layer of LIF neurons, and the domain adaptation module realizes feature distribution alignment based on JS divergence; The application layer is used for receiving target domain state sequence input, outputting fault diagnosis results and supporting real-time early warning.

[0009] Further, the state sequence conversion module of the data layer comprises an STFT transformation unit and a K-means++ clustering unit, the STFT window length is set to 512, the sliding amount is set to 256, and the clustering output binary state sequence.

[0010] Further, the SNN network parameter optimization method of the model layer adopts a genetic algorithm.

[0011] The present application has the following beneficial effects: The present application is a virtual-real migration fault prediction method and system based on an SNN network, which realizes efficient migration of simulation data knowledge to real scenes through collaborative optimization of state sequence conversion and pulse neural networks.

[0012] The present application is used for solving the equipment fault diagnosis problem in the real fault data scarce scene, especially for the health management demand of key components such as rolling bearings, and improves the diagnosis accuracy through the transfer learning of fusion simulation data and real data. Specific embodiments

[0013] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below in combination with the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Specific embodiment one

[0015] The method comprises the following steps: The step of constructing the data set refers to constructing a source domain data set containing a sufficient amount of simulation fault data and a target domain data set containing a small amount of real fault data, wherein the simulation fault data is generated based on a fault mechanism model and contains various fault types, and the real data comes from the whole life cycle monitoring data of a production device. The step of state sequence conversion refers to generating a time-frequency matrix by performing short-time Fourier transform (STFT) on the vibration signals of the source domain and the target domain respectively based on the periodic pulse characteristics of the fault vibration signals, extracting time-frequency domain distribution characteristics, clustering the time-frequency matrix array vectors into two categories of random steady state and repeatable transient state by a frequency spectrum clustering algorithm, generating a binary state sequence, and eliminating the amplitude difference between the simulation data and the real data. The step of constructing the SNN network refers to building a pulse neural network containing a feature extraction module, a classification module, and a domain adaptation module, wherein the feature extraction module adopts a pulse convolution layer and a pooling layer of a leaky integrate-and-fire neuron (LIF) to extract time sequence pulse characteristics of the state sequence layer by layer. The step of training domain adaptation refers to quantifying the feature distribution difference between the source domain and the target domain by a JS divergence loss function, constructing a total loss function in combination with a cross-entropy classification loss, optimizing network parameters through adversarial training, and enabling the feature extractor to learn domain-invariant fault features. The step of fault diagnosis reasoning refers to inputting the state sequence of the target domain into the trained network and outputting a fault type prediction result by the classification module. Specific implementation method two: The method of the embodiment is different from the specific implementation method one in that the process of generating a binary state sequence by clustering the time-frequency matrix array vectors into two categories of random steady state and repeatable transient state by a frequency spectrum clustering algorithm is specifically as follows: K-means++ is used to initialize the cluster center, each column vector of the time-frequency matrix is taken as input, the cluster number is set to 2, a 0-1 binary state sequence is generated, the amplitude difference between the simulation data and the real data is eliminated, the fault frequency characteristics are retained while the data dimension is reduced, input encoding with pulse characteristics is provided for the SNN, and information loss of traditional encoding is avoided. Specific implementation method three: The method of the embodiment is different from the specific implementation method one or two in that, in the construction of the SNN network, the structure of the feature extraction module includes three-level pulse convolution layers (S-Conv1 / 2 / 3), two-level pulse pooling layers (S-Pool1 / 2), and a pulse fully connected layer (S-Fc1 / 2), wherein the three-level pulse convolution layers include: S-Conv1 uses a 5x1 convolution kernel, 64 channels, LIF neuron activation; S-Conv2 uses a 3x1 convolution kernel, 128 channels; S-Conv3 uses a 3x1 convolution kernel, 256 channels; Both secondary pulse pooling layers use 2x1 max pooling operations; The classification module outputs classification probabilities through a pulse fully connected layer (S-Fc3) and uses a cross-entropy loss function to optimize classification performance. The domain adaptation module introduces JS divergence as a domain loss function, quantifies the feature distribution difference between the source domain (simulation data) and the target domain (real data), and forces the feature extractor to learn domain-invariant features through adversarial training. Specific implementation four: The virtual-real transfer equipment fault prediction method based on the SNN network of the embodiment is different from the specific implementation three. In the process of constructing the total loss function combined with the cross-entropy classification loss, the total loss function expression is: L = L CS + L CT + λ LD Among them, L CS and L CT are the cross-entropy classification losses of the source domain and the target domain, respectively, LD JS divergence domain loss, λ is the domain adaptation weight parameter, and the adversarial training is realized through the gradient reversal layer. Specific implementation five: The virtual-real transfer equipment fault prediction system based on the SNN network of the embodiment, the system comprises: a data layer, a model layer, and an application layer; The data layer is used to collect vibration data, construct a source domain simulation data set and a target domain real data set, and convert them into binary state sequences through spectral clustering; The model layer is used to deploy the SNN network, which includes a feature extraction module, a classification module, and a domain adaptation module. The feature extraction module uses pulse convolution and pooling layers of LIF neurons, and the domain adaptation module realizes feature distribution alignment based on JS divergence; The application layer is used to receive target domain state sequence input, output fault diagnosis results, and support real-time early warning. Specific implementation six: The SNN network-based virtual-real migration equipment fault prediction system of the embodiment is different from the specific embodiment five in that the state sequence conversion module of the data layer comprises an STFT transformation unit and a K-means++ clustering unit, the STFT window length is set to 512, the sliding amount is set to 256, and the clustering output binary state sequence. Specific embodiment seven: The SNN network-based virtual-real migration equipment fault prediction system of the embodiment is different from the specific embodiment five or six in that the SNN network parameter optimization method of the model layer adopts a genetic algorithm. Embodiment

[0022] First, state sequence conversion is performed. The source domain adopts a sufficient number of faults generated by a simulation model, and the target domain adopts a small amount of real fault data, and equal-length samples are generated through sliding segmentation. The source domain simulation data and the target domain real data are respectively subjected to STFT transformation to generate time-frequency matrices.

[0023] Each column vector of the time-frequency matrix is subjected to K-means++ clustering, the clustering number is set to 2, and a binary state sequence (0 represents a random steady state, and 1 represents a repeatable transient state) is generated.

[0024] Then, the SNN network training parameters are set. Network layer configuration: S-Conv1: convolution kernel, channel, and activation function is a LIF neuron.

[0025] S-Pool1: maximum pooling;

[0026] S-Conv2: convolution kernel, channel; S-Conv3: convolution kernel, channel.

[0027] S-Pool2: maximum pooling; S-Fc1: neuron; S-Fc2: neuron; S-Fc3: output dimension is the number of fault categories.

[0028] The embodiments of the application disclose the preferred embodiments, but are not limited thereto, and those skilled in the art can easily understand the spirit of the application according to the above-mentioned embodiments, and make different inferences and changes, as long as they do not deviate from the spirit of the application, and are within the protection scope of the application.

Claims

1. A method for predicting faults in virtual-to-real migration equipment based on SNN networks, characterized in that: The method includes the following steps: The steps of constructing the dataset refer to constructing a source domain dataset containing simulated fault data and a target domain dataset containing a small amount of real fault data. The simulated fault data is generated based on a fault mechanism model and includes various fault types. The real data comes from the monitoring data of the entire life cycle of production equipment. The steps of state sequence transformation refer to performing short-time Fourier transform on the vibration signals in the source domain and the target domain respectively to generate time-frequency matrices, and then using a spectral clustering algorithm to cluster the column vectors of the time-frequency matrices into two categories: random steady state and repeatable transient state, thereby generating a binary state sequence. Constructing an SNN network refers to building a spiking neural network that includes a feature extraction module, a classification module, and a domain adaptation module. The feature extraction module uses spiking convolutional layers and pooling layers of leaky integral discharge neurons to extract temporal spiking features of the state sequence layer by layer. Training domain adaptation refers to quantifying the difference in feature distribution between the source and target domains using the JS divergence loss function, constructing a total loss function by combining cross-entropy classification loss, and optimizing network parameters through adversarial training so that the feature extractor learns domain-invariant fault features. Fault diagnosis reasoning refers to inputting the target domain state sequence into a trained network and outputting a fault type prediction result through a classification module.

2. The method for predicting faults in virtual-to-real migration equipment based on SNN networks according to claim 1, characterized in that, The process of clustering the column vectors of the time-frequency matrix into two categories, random steady state and repeatable transient state, using the spectral clustering algorithm to generate a binary state sequence is as follows: K-means++ is used to initialize the cluster centers, each column vector of the time-frequency matrix is ​​used as input, the number of clusters is set to 2, and a 0-1 binary state sequence is generated, which retains the fault frequency characteristics while reducing the data dimensionality.

3. A method for predicting faults in virtual-to-real migration equipment based on an SNN network according to claim 1 or 2, characterized in that, In the constructed SNN network, the feature extraction module structure includes a three-level pulsed convolutional layer, a two-level pulsed pooling layer, and a pulsed fully connected layer. The three-level pulsed convolutional layer includes: S-Conv1 uses a 5×1 convolutional kernel, 64 channels, and LIF neuron activation; S-Conv2 uses a 3×1 convolution kernel and has 128 channels; S-Conv3 uses a 3×1 convolution kernel and 256 channels; The second-level pulse pooling layers all employ 2×1 max pooling operation; The classification module outputs classification probabilities through a pulsed fully connected layer and optimizes classification performance using a cross-entropy loss function. The domain adaptation module introduces JS divergence as a domain loss function to quantify the difference in feature distribution between the source and target domains, and forces the feature extractor to learn domain-invariant features through adversarial training.

4. The method for predicting faults in virtual-to-real migration equipment based on SNN networks according to claim 3, characterized in that, In the process of constructing the total loss function by combining cross-entropy classification loss, the expression of the total loss function is as follows: L = L CS + L CT + λ LD in, L CS and L CT The cross-entropy classification loss represents the source and target domains, respectively. LD For JS divergence domain loss, λ To adapt the weight parameters to the domain, adversarial training is achieved through a gradient inversion layer.

5. A prediction system for a fault prediction method for virtual-to-real migration equipment based on an SNN network as described in any one of claims 1-4, characterized in that, The system comprises: a data layer, a model layer, and an application layer; The data layer is used to collect vibration data, construct source domain simulation datasets and target domain real datasets, and convert them into binary state sequences through spectral clustering; The model layer, used to deploy the SNN network, includes a feature extraction module, a classification module, and a domain adaptation module. The feature extraction module uses pulsed convolution and pooling layers of LIF neurons, and the domain adaptation module aligns the feature distribution based on JS divergence. The application layer is used to receive the target domain state sequence input, output fault diagnosis results, and support real-time early warning.

6. The fault prediction system for virtual-real migration equipment based on SNN network according to claim 5, characterized in that, The state sequence conversion module of the data layer includes an STFT transformation unit and a K-means++ clustering unit. The STFT window length is set to 512, the slip is set to 256, and the clustering output is a binary state sequence.

7. A fault prediction system for virtual-to-real migration equipment based on an SNN network according to claim 5 or 6, characterized in that, The SNN network parameter optimization method for the model layer uses a genetic algorithm.