An end-to-end pulse neural network industrial process fault diagnosis method

Abstract

The application discloses an end-to-end pulse neural network industrial process fault diagnosis method, and belongs to the field of fault diagnosis and brain-like computing. The method is characterized in that the end-to-end fault diagnosis is performed through a pulse coding module, a multi-scale feature extraction module, a parallel excitation-inhibition module and a pulse decoding module. First, the input process variables are encoded into pulse sequences by using convolution layers and pulse neuron layers with different receptive field sizes. Then, multi-scale high-dimensional features related to faults are extracted by using multi-scale convolution layers, and the extracted multi-scale features are selectively focused by using a channel attention mechanism. Subsequently, the features related to faults are further extracted by using parallel excitatory pulse pathways and inhibitory pulse pathways. Finally, the extracted features are decoded to obtain fault diagnosis results by using convolution layers and fully connected layers. The method can accurately diagnose the faults occurring in industrial processes, and has lower energy consumption compared with traditional methods.

Description

Technical Field

[0001] The present invention belongs to the fields of fault diagnosis and brain-like computing, and specifically relates to an end-to-end industrial process fault diagnosis method based on a pulse neural network. Background Art

[0002] In automated industrial production processes, timely and accurate fault diagnosis can effectively reduce economic losses and safety incidents caused by system failures. Thanks to the rapid development of hardware equipment performance and artificial intelligence technology, data-driven fault diagnosis methods have achieved excellent results in industrial processes. Data-driven methods mainly include those based on multivariate statistical analysis, machine learning, and deep learning. Among them, deep learning-based methods, which use deep artificial neural networks (ANNs), achieve better fault diagnosis results than the other two methods. However, deep ANNs require extremely high computing resources for training and fault diagnosis, which is not conducive to their deployment in real industrial environments.

[0003] Spiking Neural Networks (SNNs) are inspired by biological neural systems and use discrete, asynchronous action potentials (pulses) to transmit information. Pulse signals that are either 0 or 1 can be calculated solely through addition operations, thus avoiding heavy multiplication operations. This makes SNNs have much lower power consumption than ANNs when deployed on hardware. In addition, spiking neurons are always iterating between the three states of charging, discharging, and resetting. The time-related internal dynamics within spiking neurons give SNNs excellent temporal feature extraction capabilities. The low power consumption and temporal feature extraction capabilities of SNNs have attracted widespread attention in fields such as computer vision, natural language processing, and reinforcement learning.

[0004] Currently, some methods have been developed to apply SNNs to fault diagnosis. However, the application of SNNs in fault diagnosis is still in its early stages. Existing methods use shallow SNNs (single-layer or two-layer fully connected networks) or residual network structures for fault diagnosis, which does not fully utilize the potential of SNNs. In addition, existing methods use SNNs to diagnose faults in bearing equipment or sensor equipment, but do not apply SNNs to fault diagnosis in other equipment or scenarios. Therefore, the architecture of SNNs in fault diagnosis still needs to be improved, and the scope of application needs to be broadened. Summary of the Invention

[0005] In order to improve the performance of SNN in fault diagnosis and broaden the application scope of SNN in fault diagnosis, the present invention provides an end-to-end pulse neural network industrial process fault diagnosis method, which uses pulse neural networks for industrial process fault diagnosis; the present invention uses a pulse encoding module, a multi-scale feature extraction module, a parallel excitation-inhibition module and a pulse decoding module to diagnose fault conditions based on process variables monitored by sensors in the industrial process. The multi-scale feature extraction module can effectively extract abstract features related to the fault, and the parallel excitation-inhibition module models the dynamic balance of excitability and inhibition of the biological nervous system, further improving the network's characterization capabilities. The technical problem proposed by the present invention is solved as follows:

[0006] An end-to-end spiking neural network industrial process fault diagnosis method includes the following steps:

[0007] Step 1: The process variables collected by multiple sensors are normalized to the maximum and minimum values ​​and then input into the spiking neural network;

[0008] Step 2: The pulse encoding module in the spiking neural network performs preliminary filtering on the input process variable X and converts it into a pulse sequence S. The obtained pulse sequence S is concatenated with the process variable X in the channel dimension and then passed as input to the multi-scale feature extraction module.

[0009] Step 3: The multi-scale feature extraction module uses three parallel feature extraction pathways to extract fault-related features. The features extracted by the three parallel pathways are concatenated in the channel dimension, and the channel attention mechanism is used to selectively focus on the concatenated features.

[0010] In step 4, the result obtained in step 3 is passed as input to the parallel excitation-inhibition module. The parallel excitation-inhibition module uses the parallel excitatory pulse path and the inhibitory pulse path to further extract fault-related features. The features extracted from the two paths are adaptively weighted and fused and then added to the result obtained in step 3.

[0011] Step 5: pass the result obtained in step 4 as input to the pulse decoding module to obtain the fault diagnosis result;

[0012] Step 6: Repeat steps 1 to 5 for a total of T times, and take the average result generated over multiple times as the final fault diagnosis result.

[0013] Step 1 of the present invention comprises:

[0014] Step 1-1: Get the maximum value of each process variable based on historical sensor data and minimum value

[0015] Step 1-2, the maximum value of each process variable obtained according to step 1-1 and minimum value Perform maximum-minimum normalization on each process variable:

[0016]

[0017] in represents the original process variable, X i Represents the normalized process variable.

[0018] Step 2 of the present invention comprises:

[0019] In step 2-1, a convolutional layer with a convolution kernel size of 1×5 is used to perform a convolution operation on the input process variable X to obtain the shallow feature X1 after preliminary filtering:

[0020] X1=conv 1×5 (X),

[0021] Where conv represents the convolution operation;

[0022] Step 2-2: Input the shallow feature X1 obtained in step 2-1 into the spiking neuron layer, which generates a pulse sequence S1:

[0023] S1=sn(X1),

[0024] Where sn represents the spiking neuron layer;

[0025] Step 2-3, use a convolution layer with a convolution kernel size of 1×3 to perform a convolution operation on the pulse sequence S1 to obtain the further filtered feature X2:

[0026] X2=conv 1×3 (S1);

[0027] Step 2-4: Input the feature X2 obtained in step 2-3 into the spiking neuron layer, which generates a pulse sequence S:

[0028] S = sn(X2);

[0029] Step 2-5: Concatenate the pulse sequence S obtained in step 2-4 with the input process variable X in the channel dimension:

[0030]

[0031] in Represents a splicing operation.

[0032] Step 3 of the present invention comprises:

[0033] In step 3-1, three parallel feature extraction pathways use a convolutional layer-spiking neuron layer-convolutional layer-spiking neuron layer structure to extract fault-related features:

[0034]

[0035] The convolution kernel sizes in the three feature extraction pathways are 1×1, 1×3, and 1×5 respectively;

[0036] Step 3-2: Concatenate the features extracted by the three feature extraction paths in step 3-1 in the channel dimension to obtain X ms :

[0037]

[0038] Step 3-3, for the spliced ​​feature X ms Use the channel attention mechanism for selective attention and use the global average pooling operation to obtain the feature X ms The global descriptor g(X ms ), and then use the convolution kernel size of 1×5 and the sigmoid function to get the feature X ms The channel-wise attention score d ms , channel-wise attention score d ms and feature X ms Multiply to get the features after selective attention

[0039]

[0040] Where f(·) represents the ms Perform a global average pooling operation, where L is the size of a single channel feature and i is the index of the feature in the spatial dimension. Represents the sigmoid function, and ⊙ represents a multiplication operation with a broadcast mechanism.

[0041] Step 4 of the present invention comprises:

[0042] Step 4-1, features obtained in step 3 As input, it is passed to the parallel excitatory pulse pathway and inhibitory pulse pathway. The excitatory pulse pathway and inhibitory pulse pathway are composed of convolution layer-spiking neuron layer-convolution layer-spiking neuron / inhibitory spike neuron layer. The two pathways generate excitatory pulse outputs X e and inhibitory pulse output X i :

[0043]

[0044] where sn IIndicates the inhibitory spiking neuron layer. When the neuron's membrane potential reaches the firing threshold, sn I Generates an inhibitory pulse, i.e. "-1";

[0045] Step 4-2: Output X generated by the excitatory pulse pathway and inhibitory pulse pathway obtained in step 4-1 e and X i Perform adaptive weighting:

[0046] X ei =αX e +(1-α)X i

[0047] Among them, α∈(0,1) is a parameter that changes adaptively as the network trains, obtained by applying the sigmoid function to the learnable parameter x;

[0048] Step 4-3, the output X obtained in step 4-2 ei and the output produced by step 3 Element-wise addition:

[0049]

[0050] in, is the output produced by step 4.

[0051] Step 5 of the present invention comprises:

[0052] Step 5-1, obtained in step 4 The input is passed to the pulse decoding module, which uses the convolution layer-spiking neuron layer-convolution layer-spiking neuron layer pair. Decoding is performed, where the convolution kernel size is 1×3;

[0053] Step 5-2, use global average pooling to process the decoded result obtained in step 5-1 and change its dimension to

[0054] In step 5-3, the fully connected layer is used to process the pooled result of step 5-2, and the output of the network is obtained as the fault diagnosis result.

[0055] The present invention proposes an end-to-end pulse neural network industrial process fault diagnosis method, which has the following beneficial effects:

[0056] The present invention combines a multi-scale feature extraction module and a parallel excitation-inhibition module, and can perform fault diagnosis based on process variables collected by sensors in an end-to-end manner. Compared with existing methods, the method proposed in the present invention has higher diagnostic accuracy when performing fault diagnosis. The present invention uses a low-power spiking neural network for industrial process fault diagnosis, which is more suitable for deployment in real-world scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 The accompanying drawings of the patent abstract of the present invention and the overall flow chart of the method of the present invention;

[0058] Figure 2 This is a structural diagram of a pulse coding module according to an embodiment of the present invention;

[0059] Figure 3 This is a structural diagram of a multi-scale feature extraction module according to an embodiment of the present invention;

[0060] Figure 4 1 is a structural diagram of a parallel excitation-inhibition module according to an embodiment of the present invention;

[0061] Figure 5 This is a structural diagram of a pulse decoding module according to an embodiment of the present invention;

[0062] Figure 6 Schematic diagram of the accuracy change during the training process of an embodiment of the present invention; DETAILED DESCRIPTION

[0063] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0064] The overall process of the method proposed by the present invention is as follows Figure 1 As shown, the Leaky Integrate-and-Fire (LIF) neuron model is used as the spiking neuron, the total time step T is 10, that is, the average result of 10 runs is taken as the final diagnosis result, the initial discharge threshold of the spiking neuron is set to 1.0, and the network is tested on the Tennessee-Eastman (TE) process data. The network is trained for a total of 100 epochs. The specific steps for constructing image recognition in the embodiment of the present invention are as follows:

[0065] Step 1: The process variables collected by 52 sensors in the TE process are normalized to the maximum and minimum values ​​and then input into the spiking neural network;

[0066] Step 2: The pulse encoding module in the pulse neural network performs preliminary filtering on the input process variable X and converts it into a pulse sequence S. The obtained pulse sequence S is concatenated with the process variable X in the channel dimension and then passed as input to the multi-scale feature extraction module, as shown in the schematic diagram. Figure 2 As shown;

[0067] Step 3: The multi-scale feature extraction module uses three parallel feature extraction pathways to extract fault-related features. The features extracted by the three parallel pathways are spliced ​​in the channel dimension, and the channel attention mechanism is used to selectively focus on the spliced ​​features. The schematic diagram is shown in the figure. Figure 3 As shown;

[0068] Step 4: The result obtained in step 3 is passed as input to the parallel excitation-inhibition module. The parallel excitation-inhibition module uses the parallel excitatory pulse path and the inhibitory pulse path to further extract fault-related features. The features extracted from the two paths are adaptively weighted and fused and then added to the result obtained in step 3. The schematic diagram is shown in FIG. Figure 4 As shown;

[0069] Step 5: The result obtained in step 4 is passed as input to the pulse decoding module to obtain the fault diagnosis result, as shown in the schematic diagram. Figure 5 As shown;

[0070] Step 6: Repeat steps 1 to 5 10 times, and take the average result as the final fault diagnosis result.

[0071] Step 1 of the present invention comprises:

[0072] Step 1-1: Based on the sensor data of the TE process, obtain the maximum value of 52 process variables and minimum value

[0073] Step 1-2, the maximum value of the 52 process variables obtained according to step 1-1 and minimum value Perform maximum-minimum normalization on each process variable:

[0074]

[0075] in represents the original process variable, X i Represents the normalized process variable.

[0076] Step 2 of the present invention comprises:

[0077] In step 2-1, a convolutional layer with a convolution kernel size of 1×5 is used to perform a convolution operation on the input process variable X to obtain the shallow feature X1 after preliminary filtering:

[0078] X1=conv 1×5 (X)

[0079] Where conv represents the convolution operation;

[0080] Step 2-2: Input the shallow feature X1 obtained in step 2-1 into the spiking neuron layer, which generates a pulse sequence S1:

[0081] S1=sn(X1)

[0082] Where sn represents the spiking neuron layer;

[0083] Step 2-3, use a convolution layer with a convolution kernel size of 1×3 to perform a convolution operation on the pulse sequence S1 to obtain the further filtered feature X2:

[0084] X2=conv 1×3 (S1);

[0085] Step 2-4: Input the feature X2 obtained in step 2-3 into the spiking neuron layer, which generates a pulse sequence S:

[0086] S = sn(X2);

[0087] Step 2-5: Concatenate the pulse sequence S obtained in step 2-4 with the input process variable X in the channel dimension:

[0088]

[0089] in Represents a splicing operation.

[0090] Step 3 of the present invention comprises:

[0091] In step 3-1, three parallel feature extraction pathways use a convolutional layer-spiking neuron layer-convolutional layer-spiking neuron layer structure to extract fault-related features:

[0092]

[0093] The convolution kernel sizes in the three feature extraction pathways are 1×1, 1×3, and 1×5 respectively;

[0094] Step 3-2: Concatenate the features extracted by the three feature extraction paths in step 3-1 in the channel dimension to obtain X ms :

[0095]

[0096] Step 3-3, for the spliced ​​feature X ms Use the channel attention mechanism for selective attention and use the global average pooling operation to obtain the feature X ms The global descriptor g(X ms ), and then use the convolution kernel size of 1×5 and the sigmoid function to get the feature X msThe channel-wise attention score d ms , channel-wise attention score d ms and feature X ms Multiply to get the features after selective attention

[0097]

[0098] d ms =σ(conv 1×5 (g(X ms ))),

[0099]

[0100] Where f(·) represents the ms Perform a global average pooling operation, where L is the size of a single channel feature and i is the index of the feature in the spatial dimension. Represents the sigmoid function, and ⊙ represents a multiplication operation with a broadcast mechanism.

[0101] Step 4 of the present invention comprises:

[0102] Step 4-1, features obtained in step 3 As input, it is passed to the parallel excitatory pulse pathway and inhibitory pulse pathway. The excitatory pulse pathway and inhibitory pulse pathway are composed of convolution layer-spiking neuron layer-convolution layer-spiking neuron / inhibitory spike neuron layer. The two pathways generate excitatory pulse outputs X e and inhibitory pulse output X i :

[0103]

[0104] where sn I Indicates the inhibitory spiking neuron layer. When the neuron's membrane potential reaches the firing threshold, sn I Generates an inhibitory pulse, i.e. "-1";

[0105] Step 4-2: Output X generated by the excitatory pulse pathway and inhibitory pulse pathway obtained in step 4-1 e and X i Perform adaptive weighting:

[0106] X ei =αX e +(1-α)X i

[0107] Among them, α∈(0,1) is a parameter that changes adaptively as the network trains, obtained by applying the sigmoid function to the learnable parameter x;

[0108] Step 4-3, the output X obtained in step 4-2 ei and the output produced by step 3 Element-wise addition:

[0109]

[0110] in is the output produced by step 4;

[0111] Step 5 of the present invention comprises:

[0112] Step 5-1, obtained in step 4 The input is passed to the pulse decoding module, which uses the convolution layer-spiking neuron layer-convolution layer-spiking neuron layer pair. Decoding is performed, where the convolution kernel size is 1×3;

[0113] Step 5-2, use global average pooling to process the decoded result obtained in step 5-1 and change its dimension to

[0114] In step 5-3, the fully connected layer is used to process the pooled result of step 5-2, and the output of the network is obtained as the fault diagnosis result.

[0115] The present invention realizes end-to-end industrial process fault diagnosis using a pulse neural network through a pulse encoding module, a multi-scale feature extraction module, a parallel excitation-inhibition module and a pulse decoding module, which not only reduces the energy consumption required for fault diagnosis but also improves the diagnostic performance of the fault diagnosis method.

[0116] The present invention provides an end-to-end pulse neural network industrial process fault diagnosis method. The industrial process data set, pulse neuron type and specific parameter values ​​used in the above embodiments are specific implementation methods combined with the principles of the present invention. The scope of protection of the present invention is not limited to such specific implementation methods.

Claims

1. An end-to-end pulse neural network industrial process fault diagnosis method, characterized by: The following steps are involved: Step 1: The process variables collected by multiple sensors are normalized to the maximum and minimum values ​​and then input into the spiking neural network; Step 2: The pulse encoding module in the spiking neural network performs preliminary filtering on the input process variable X and converts it into a pulse sequence S. The obtained pulse sequence S is concatenated with the process variable X in the channel dimension and then passed as input to the multi-scale feature extraction module. In step 3, the multi-scale feature extraction module uses three parallel feature extraction pathways to extract fault-related features. The features extracted by the three parallel feature extraction pathways are concatenated in the channel dimension, and the channel attention mechanism is used to selectively focus on the concatenated features. The specific process is shown in steps 3-1 to 3-3: In step 3-1, three parallel feature extraction pathways use a convolutional layer-spiking neuron layer-convolutional layer-spiking neuron layer structure to extract fault-related features: Where conv represents the convolutional layer, sn represents the spike neuron layer, and X a It is the result of concatenating the pulse sequence S obtained in step 2 with the input process variable X in the channel dimension. The convolution kernel sizes in the three feature extraction paths are 1×1, 1×3, and 1×5 respectively. Step 3-2: Concatenate the features extracted by the three feature extraction paths in step 3-1 in the channel dimension to obtain X ms : Step 3-3, for the spliced ​​feature X ms Use the channel attention mechanism for selective attention and use the global average pooling operation to obtain the feature X ms The global descriptor g(X ms ), and then use the convolution kernel size of 1×5 and the sigmoid function to get the feature X ms The channel-wise attention score d ms , channel-wise attention score d ms and feature X ms Multiply to get the features after selective attention d ms =σ(conv 1×5 (g(X ms ))), Where f(·) represents the ms Perform a global average pooling operation, where L is the size of a single channel feature and i is the index of the feature in the spatial dimension. represents the sigmoid function, ⊙ represents the multiplication operation with broadcast mechanism; In step 4, the result obtained in step 3 is passed as input to the parallel excitation-inhibition module. The parallel excitation-inhibition module uses the parallel excitatory pulse path and the inhibitory pulse path to further extract fault-related features. The features extracted from the two paths are adaptively weighted and fused and then added to the result obtained in step 3. Step 5: pass the result obtained in step 4 as input to the pulse decoding module to obtain the fault diagnosis result; Step 6: Repeat steps 1 to 5 for a total of T times, and take the average result of T times as the final fault diagnosis result.

2. The method according to claim 1, characterized in that Step 2 includes: Step 2-1: Use a convolution layer with a convolution kernel size of 1×5 to perform a convolution operation on the input process variable X to obtain the shallow feature X1 after preliminary filtering; Step 2-2, input the shallow feature X1 obtained in step 2-1 into the spiking neuron layer, and the spiking neuron layer generates a pulse sequence S1; Step 2-3, use a convolution layer with a convolution kernel size of 1×3 to perform a convolution operation on the pulse sequence S1 to obtain the further filtered feature X2; In step 2-4, the feature X2 obtained in step 2-3 is input to the spiking neuron layer, and the spiking neuron layer generates a pulse sequence S. In step 2-5, the pulse sequence S obtained in step 2-4 is concatenated with the input process variable X in the channel dimension: in Represents a splicing operation.

3. The method according to claim 1, characterized in that Step 4 includes: Step 4-1, features obtained in step 3 As input, it is passed to the parallel excitatory pulse pathway and inhibitory pulse pathway. The excitatory pulse pathway and inhibitory pulse pathway are composed of convolution layer-spiking neuron layer-convolution layer-spiking neuron / inhibitory spike neuron layer. The two pathways generate excitatory pulse outputs X e and inhibitory pulse output X i : where sn I Indicates the inhibitory spiking neuron layer. When the neuron's membrane potential reaches the firing threshold, sn I Generates an inhibitory pulse, i.e. "-1"; Step 4-2: Output X generated by the excitatory pulse pathway and inhibitory pulse pathway obtained in step 4-1 e and X i Perform adaptive weighting: X ei =αX e +(1-a)X i , Among them, α∈(0,1) is a parameter that changes adaptively as the network trains, obtained by applying the sigmoid function to the learnable parameter x; Step 4-3, the output X obtained in step 4-2 ei and the output produced by step 3 Element-wise addition: in, is the output produced by step 4.

4. The method according to claim 1, characterized in that Step 5 includes: Step 5-1, obtained in step 4 The input is passed to the pulse decoding module, which uses the convolution layer-spiking neuron layer-convolution layer-spiking neuron layer pair. Decode; Step 5-2, use global average pooling to process the decoded result obtained in step 5-1 and change its dimension to In step 5-3, the fully connected layer is used to process the pooled result of step 5-2, and the output of the network is obtained as the fault diagnosis result.

Images