A high-pass filter fault feature extraction method based on PSO-DBN

Abstract

The present invention takes a typical high-pass filter as the research object and provides a high-pass filter fault feature extraction method based on PSO-DBN, which is a prediction method based on PSO-DBN. First, a circuit model of a high-pass filter is built, and fault injection is performed using PSpice software, and data collection is performed on the output response under different fault modes. Secondly, through Monte Carlo simulation, a certain amount of sample data is obtained for each fault mode, and the normalized data is input into the DBN deep belief network with different RBM stacking numbers to extract fault features, and the simulation time and feature extraction effect are compared to determine the optimal number of hidden layers. Finally, the PSO particle swarm optimization algorithm is used to determine the number of hidden layer neurons, to optimize the DBN model structure and further improve the model feature extraction performance. After comparison with other models, the prediction effect of the PSO-DBN model is significantly better than other models.

Description

Technical Field

[0001] The present invention belongs to the field of circuit fault feature extraction methods, and in particular relates to a high-pass filter fault feature extraction method based on a deep belief network. Background Art

[0002] A high-pass filter, also known as a low-cut filter or low-resistance filter, allows frequencies above a certain cutoff frequency to pass through while significantly attenuating low frequencies. It removes unnecessary low-frequency components from the signal. Its applications are extensive, encompassing signal processing, audio processing, image processing, and other fields. For example, in signal processing, high-pass filters can be used to filter out low-frequency noise; in audio processing, they can be used to enhance high-frequency signals; and in image processing, they can be used to extract edge information from images.

[0003] In summary, as one of the common parts in electronic systems, ensuring the reliable operation of high-pass filter circuits is of great significance to reducing equipment maintenance costs. Summary of the Invention

[0004] 1. Purpose: The present invention takes a typical high-pass filter as the research object and provides a high-pass filter fault feature extraction method based on a deep belief network. It is a prediction method based on PSO-DBN. PSO means: particle swarm optimization algorithm; DBN means: deep belief network. First, a circuit model of a high-pass filter is built, and fault injection is performed using PSpice software, and data collection is performed on the output response under different fault modes. Secondly, through Monte Carlo simulation, a certain amount of sample data is obtained under each fault mode, and the normalized data is input into a DBN deep belief network with different numbers of RBM (restricted Boltzmann machines) stacked layers for fault feature extraction. The simulation time and feature extraction effect are compared to determine the optimal number of hidden layers. Finally, the particle swarm optimization algorithm PSO is used to determine the number of hidden layer neurons, optimize the DBN model structure, and further improve the model feature extraction performance.

[0005] 2. Technical solution:

[0006] First, introduce a few definitions:

[0007] Definition 1: A Deep Belief Network (DBN) is a neural network composed of several neurons, whose components are Restricted Boltzmann Machines (RBMs). DBN training is layer-by-layer. In each layer, the data vector is used to infer the hidden layer, which is then used as the data vector for the next layer.

[0008] Definition 2: Restricted Boltzmann Machine (RBM) is the basic component of Deep Belief Network (DBN), which includes the visual layer v = {v1, v2, ..., v D}∈{0,1} and hidden layer h={h1,h2,…,h k}∈{0,1}. The two-layer random neural network RBM model can be represented by a bipartite undirected graphical model. The basic structure is as follows Figure 1 As shown. i Represents the state of the i-th visual layer neuron, h i represents the state of the i-th hidden layer neuron, w ij Indicates v i With h j The connection weight between i and b i denote the bias vectors of the visual layer and the hidden layer respectively.

[0009] The present invention provides a high-pass filter fault feature extraction method based on PSO-DBN, and the specific method comprises the following steps:

[0010] Step 1: High-pass filter modeling and fault injection based on PSpice:

[0011] PSpice software was used to build a circuit model for a high-pass filter. Different fault modes were injected by changing the values ​​of key components. The software was then used to collect data on the output under different fault modes.

[0012] Step 2: Select the number of hidden layers of the DBN-based fault feature extraction model:

[0013] Monte Carlo simulations were used to obtain a certain amount of sample data for each failure mode, which was then normalized. The number of RBM stacks in the DBN model was set to 1, 2, 3, 4, or 5, and feature extraction was performed on each model. Feature evaluation metrics (i.e., the average ratio of intra-class distance to inter-class distance) and runtime were compared to determine the optimal number of RBM stacks (i.e., the number of hidden layers).

[0014] Step 3: Determine the number of neurons in the DBN hidden layer based on PSO:

[0015] After determining the number of hidden layers in step 2, the particle swarm optimization algorithm (PSO) is used to calculate and obtain the change curve of the fitness function, determine the number of iterations, and then determine the number of hidden neurons in each layer, and finally obtain a complete high-pass filter fault feature extraction model.

[0016] The advantages and benefits of this invention lie in its ability to learn the essential characteristics of data from a limited sample set, enabling low-level to high-level feature representation and extraction of measured data. From the perspective of minimizing reconstruction error, heuristic algorithms, including the particle swarm optimization algorithm, are utilized to optimize and constrain the number of model layers and model parameter training, enabling more efficient feature extraction and providing a foundation for subsequent fault diagnosis and prediction research. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is the basic structure of RBM.

[0018] Figure 2 It is a feature extraction method based on the PSO-DBN model.

[0019] Figure 3 It is a four-op-amp biquad high-pass filter circuit.

[0020] Figure 4 is the output response under different fault modes.

[0021] Figure 5 is the changing curve of the fitness function.

[0022] Table 1 shows the circuit excitation signal parameters.

[0023] Table 2 shows the fault settings for the four-op-amp biquad high-pass filter circuit.

[0024] Table 3 shows the DBN feature extraction effect with different numbers of hidden layers.

[0025] Table 4 shows the evaluation indicators of wavelet packet transform feature extraction. DETAILED DESCRIPTION

[0026] This invention takes a typical high-pass filter as an example, and uses DBN, PSO and other algorithms to extract circuit fault features based on the fault data obtained by PSpice circuit simulation. The above technical solution is now further described in detail. The specific process is as follows Figure 2 As shown:

[0027] Step 1: High-pass filter modeling and fault injection based on PSpice:

[0028] First, a high-pass filter is digitally modeled according to the actual circuit diagram using PSpice circuit simulation software. The circuit model is as follows: Figure 3 As shown. The four-op-amp biquad high-pass filter circuit consists of 10 resistors, 2 capacitors and 4 operational amplifiers, where R1 = R2 = R3 = 6.2kΩ, R4 = 1.6kΩ, R5 = R6 = 5.1kΩ, R7 = R8 = R9 = R 10=10kΩ, C1=C2=5nF. Apply a single pulse signal with an amplitude of 5V and a duration of 10us. See Table 1 for specific settings.

[0029] Table 1

[0030]

[0031] Since the actual parameters of components cannot be completely equal to the nominal parameters, in order to make the simulation more consistent with the actual situation, the tolerance range of the resistor R and the capacitor C is set to ±10% of their nominal values. The PSpice software is used to collect data on the circuit simulation output voltage, which is the fault-free output data.

[0032] Generally speaking, when the parameter value of any component in the circuit deviates from ±50% of its nominal value, while the parameter values ​​of the remaining components are within the allowable tolerance range of their nominal values, the component is considered to have failed. Since circuit failures are often caused by key components, it is necessary to perform a sensitivity analysis on the circuit under test before collecting data to determine the components that have a greater response to the circuit output. Based on the sensitivity analysis results, the key components C1, C2, R1, R2, R3, and R4 are selected as faulty components. By changing the parameters of these key components (10% to 50% higher or lower than the nominal value), fault injection is performed on the circuit simulation model in the PSpice software to obtain multiple failure modes. The specific fault settings are shown in Table 2, which include 1 normal mode {F0} and 12 fault modes {F1, F2, F3, F4, F5, F6, F7, F8, F9, F10, F11, F12}. Among them, ↓fault and ↑fault indicate that the parameter value is lower than and higher than the nominal value [10% and 50%], respectively.

[0033] Table 2

[0034] Failure Mode Fault type Nominal value Tolerance Fault value F0 NF / / / F1 <![CDATA[C1↓]]> 5nF ±10％ [2.5nF,4.5nF] F2 <![CDATA[C1↑]]> 5nF ±10％ [5.5nF,7.5nF] F3 <![CDATA[C2↓]]> 5nF ±10％ <![CDATA[[2.5nF , 4.5nF]]> F4 <![CDATA[C2↑]]> 5nF ±10％ <![CDATA[[5.5nF , 7.5nF]]> F5 <![CDATA[R1↓]]> 6.2kΩ ±10％ <![CDATA[[3.1kΩ , 5.58kΩ]]> F6 <![CDATA[R1↑]]> 6.2kΩ ±10％ <![CDATA[[6.82kΩ , 9.3kΩ]]]> F7 <![CDATA[R2↓]]> 6.2kΩ ±10％ <![CDATA[[3.1kΩ , 5.58kΩ]]> F8 <![CDATA[R2↑]]> 6.2kΩ ±10％ <![CDATA[[6.82kΩ , 9.3kΩ]]]> F9 <![CDATA[R3↓]]> 6.2kΩ ±10％ [3.1kΩ,5.58kΩ] F10 <![CDATA[R3↑]]> 6.2kΩ ±10％ [6.82kΩ,9.3kΩ] F11 <![CDATA[R4↓]]> 1.6kΩ ±10％ [0.8kΩ,1.44kΩ] F12 <![CDATA[R4↑]]> 1.6kΩ ±10％ [1.76kΩ,2.4kΩ]

[0035] The output voltage data under different fault modes were collected using PSpice software. The results are as follows: Figure 4 As shown in the figure, the output response waveforms obtained by the circuit under different fault modes are basically similar, with only certain differences between the peaks and troughs. This requires further extraction of fault characteristics that can express the more essential and underlying variation patterns of the data. The output voltage with a sampling time between 0 and 250 ns is used as the raw signal, and a total of 200 sets of sample data are obtained through Monte Carlo simulation.

[0036] Step 2: Select the number of hidden layers of the DBN-based fault feature extraction model:

[0037] Through Monte Carlo simulation, 200 sets of sample data were obtained for each fault mode in step 1 through random sampling, and the data were normalized. This is because DBN is a probabilistic model, and the sigmoid function used in the RBM model has an activation probability distribution between [0, 1]. Linear normalization is used to map the collected raw signal to the interval [0, 1]. The calculation formula is as follows:

[0038]

[0039] Among them, x represents voltage data, x max and x min Represents the maximum and minimum values ​​of voltage respectively.

[0040] The normalized voltage data is input into the DBN model for fault feature extraction. The number of RBM stacks, i.e., the number of hidden layers, is set to 1, 2, 3, 4, and 5, respectively. In the process of setting the corresponding number of hidden neurons, on the one hand, in order to achieve the DBN model's deep feature extraction while minimizing the model's complexity and computational complexity, the number of neurons in the hidden layer is gradually reduced from the bottom to the top. On the other hand, in order to ensure that the final extracted feature vector has the same dimension, the number of neurons in the last hidden layer is uniformly set to 50. The training time required for the DBN model with different numbers of hidden layers is recorded, and the feature evaluation index based on the intra-class and inter-class distance is calculated. The calculation method is as follows:

[0041] Both intra-class distance and inter-class distance generally represent the Euclidean distance between samples. The intra-class distance refers to the mean squared distance between feature vectors within the same fault type, while the inter-class distance refers to the mean squared distance between feature vectors within different fault types. The smaller the intra-class distance within the same fault type and the larger the inter-class distance between different fault types, the better the separability of the fault feature classes and the higher the feature extraction quality of the DBN model.

[0042] Assuming there are M types of faults, after using DBN model to extract fault features, a total of W-dimensional feature vectors are obtained, with the intra-class distance d of each feature vector (1) and the inter-class distance d (2) Based on this, the concentration and dispersion of the feature vector are measured, and d (2) with d (1) The ratio of constructs the evaluation index of the feature vector:

[0043]

[0044] In the process of fault feature extraction, the ideal fault feature should satisfy the minimum distance within the class and the maximum distance between classes, providing a powerful condition for subsequent fault diagnosis. For the i-th feature vector, when the evaluation index λ i The larger the value is, the better the separability of the feature vector for the fault types in M ​​is, and the stronger its ability to characterize different fault modes is. Otherwise, the ability is weaker.

[0045] Calculate the average evaluation index of W eigenvectors and use it as the overall evaluation index of the eigenvector, which can be expressed as:

[0046]

[0047] where λ r represents the evaluation index of the rth eigenvector, The fault separation capabilities of all feature vectors are measured, which can generally reflect the extraction effect and quality of the feature extraction method.

[0048] The training time required for the DBN model with different numbers of hidden layers is recorded, and the overall evaluation index of the feature vector is calculated. The feature extraction effects of different numbers of hidden layers are shown in Table 3.

[0049] Table 3

[0050]

[0051]

[0052] The results in the table show that as the number of hidden layers increases, the DBN model's structure becomes increasingly complex, and the simulation time increases accordingly. The overall evaluation index of the feature vector shows a trend of first increasing and then decreasing, reaching its maximum value when the number of hidden layers is 3. This indicates that the DBN model has the strongest nonlinear mapping capability at this point, and the extracted feature vectors are most effective in distinguishing different fault modes. However, when too many RBMs are stacked, that is, when too many hidden layers are set, not only does it significantly increase training time, it also weakens the stability of the network structure and reduces the quality of feature extraction. Therefore, when using a DBN model for feature extraction, it is necessary to set a reasonable number of hidden layers (i.e., the number of RBMs) based on the actual situation. Therefore, in this example, the number of hidden layers is set to 3.

[0053] Step 3: Determine the number of neurons in the DBN hidden layer based on PSO:

[0054] When determining the structure of a DBN model, there's no clear standard for setting the number of neurons in the hidden layer. Manually determining the number and adjusting it based on the results is time-consuming and subjective, and feature extraction doesn't achieve optimal results. This paper introduces the particle swarm optimization algorithm (PSO) to determine the number of neurons in the hidden layer, optimizing the DBN model structure and further improving the model's feature extraction performance.

[0055] According to the principle of PSO algorithm, when using PSO algorithm to search for the optimal parameters of DBN model, it is necessary to determine the fitness function used in PSO algorithm in advance. In order to ensure the quality of fault feature extraction, the root mean square error RMSE and the overall evaluation index of the feature vector before and after data isomorphism are comprehensively considered in the study. The fitness of the weighted combination of the two is used to evaluate the fitness of the PSO algorithm. The expression formula is as follows:

[0056]

[0057] Among them, σ1 and σ2 represent RMSE and The weight coefficient is used to describe the degree of influence on the fitness function. In this example, σ1 = 0.1 and σ2 = 0.9.

[0058] During the iteration process, when the value of the fitness function is smaller, it means that the searched hidden layer neuron number combination is optimal, and the corresponding DBN model feature extraction performance is also better. The algorithm flow of using the PSO algorithm to optimize the number of hidden layer neurons in the DBN model is as follows: Figure 2 As shown, specifically for this example, the steps are as follows:

[0059] a) In the PSO algorithm, the total number of particles is set to 20, the number of iterations is set to 100, the range of the inertia weight ω is set to [0.4, 0.9], and the range of the learning factors c1 and c2 is set to [0.5, 2.5]

[0060] b) Using the fitness function in the PSO algorithm, draw the curve of the fitness function, such as Figure 5 As shown in the figure.

[0061] It is found that when the convergence times k = 57, the fitness function in the PSO algorithm converges and remains unchanged for a long time. c) The optimal number of hidden layer neurons is 139, 171 and 35 respectively. At this time, the PSO-DBN model extracts

[0062] The overall evaluation index of the eigenvector After all the training is completed, the deep features of the original voltage signal are extracted layer by layer through the stacked RBM, and the final output result is the extracted features.

[0063] d) To further verify the effectiveness of the proposed method, it is compared with the common signal analysis method Wavelet Packet Decomposition (WPD), and the As the evaluation index, the results are shown in Table 4.

[0064] Table 4

[0065]

[0066] Through the comparative experiments of the two methods, it can be seen that the feature evaluation index obtained by the PSO-DBN method is higher than the corresponding index obtained by the wavelet packet transform method, indicating that the feature extraction effect of the PSO-DBN method is better, and it can adaptively extract the deep features of the data and describe the essential laws of the original input signal.

Claims

1. A high-pass filter fault feature extraction method based on PSO-DBN, characterized by: Here are the steps: Step 1: High-pass filter modeling and fault injection based on PSpice: Use PSpice software to build a circuit model for a high-pass filter. By changing the values ​​of key components, different fault modes can be injected. The software is then used to collect data on the output under different fault modes. Step 2: Select the number of hidden layers of the fault feature extraction model based on the deep belief network DBN: A certain amount of sample data under each fault mode was obtained through Monte Carlo simulation and then normalized. The number of RBM stacks in the DBN model was set to 1, 2, 3, 4, and 5, and feature extraction was performed respectively. The feature evaluation index and running time were compared, and the most appropriate number of RBM stacks, i.e., the number of hidden layers, was selected. The feature evaluation index was the average ratio of the intra-class distance to the inter-class distance. Step 3: Determine the number of neurons in the DBN hidden layer based on PSO: After determining the number of hidden layers in step 2, the particle swarm optimization algorithm (PSO) is used to calculate and obtain the change curve of the fitness function, determine the number of iterations, and then determine the number of hidden neurons in each layer, and finally obtain a complete high-pass filter fault feature extraction model; In step 1, a high-pass filter is first digitally modeled on the PSpice circuit simulation software; the four-op-amp biquad high-pass filter circuit consists of 10 resistors, 2 capacitors, and 4 operational amplifiers, where R1 = R2 = R3 = 6.2kΩ, R4 = 1.6kΩ, R5 = R6 = 5.1kΩ, R7 = R8 = R9 = R 10 =10kΩ, C1=C2=5nF; apply a single pulse signal with an amplitude of 5V and a duration of 10us; In step 2, there are M types of faults. After extracting fault features using the DBN model, a total of W-dimensional feature vectors are obtained. The intra-class distance d of each feature vector is (1) and the inter-class distance d (2) Based on this, the concentration and dispersion of the feature vector are measured, and d (2) with d (1) The ratio of constructs the evaluation index of the feature vector: In the process of fault feature extraction, the ideal fault feature should satisfy the minimum distance within the class and the maximum distance between classes. For the i-th feature vector, when the evaluation index λ i The larger the value, the better the separability of the feature vector for the fault types in M, and the stronger its ability to characterize different fault modes, and vice versa. Calculate the average evaluation index of W eigenvectors and use it as the overall evaluation index of the eigenvector, expressed as: Among them, λ r represents the evaluation index of the rth eigenvector, The fault separation capability of all feature vectors is measured, which can generally reflect the extraction effect and quality of the feature extraction method; In step three, the particle swarm optimization algorithm PSO is introduced to determine the number of hidden layer neurons; when using the PSO algorithm to search for the optimal parameters of the DBN model, it is necessary to determine the fitness function used in the PSO algorithm in advance, considering the root mean square error RMSE before and after data isomorphism and the overall evaluation index of the eigenvector The fitness of the weighted combination of the two is used to evaluate the fitness of the PSO algorithm. The expression formula is as follows: Among them, σ1 and σ2 represent RMSE and The weight coefficient is used to describe the degree of influence on the fitness function.

2. The method for extracting high-pass filter fault features based on PSO-DBN according to claim 1, characterized in that: In step 1, the tolerance ranges of the resistor R and the capacitor C are set to ±10% of their nominal values, and the circuit simulation output voltage data is collected using PSpice software. This is the output data without faults.

3. The method for extracting fault features of a high-pass filter based on PSO-DBN according to claim 1 or 2, characterized in that: In step 1, when the parameter value of any component in the circuit deviates from ±50% of its nominal value, while the parameter values ​​of the remaining components are within the allowable tolerance range of their nominal values, the component is considered to have failed; based on the sensitivity analysis results, key components C1, C2, R1, R2, R3, and R4 are selected as faulty components. By changing the parameters of these key components, fault injection is performed on the circuit simulation model in the PSpice software to obtain multiple fault modes; the fault modes include one normal mode F0 and 12 fault modes F1, F2, F3, F4, F5, F6, F7, F8, F9, F10, F11, and F12.

4. The method for extracting high-pass filter fault features based on PSO-DBN according to claim 1, characterized in that: In step 2, 200 sets of sample data are obtained under each fault mode in step 1 through Monte Carlo simulation random sampling, and the data are normalized. The sigmoid function used in the RBM model has an activation probability distribution between [0, 1]. Linear normalization is used to map the collected original signal to the interval [0, 1]. The calculation formula is as follows: Among them, x represents voltage data, x max and x min Represents the maximum and minimum values ​​of voltage respectively.

5. The method for extracting fault features of a high-pass filter based on PSO-DBN according to claim 1, characterized in that: In the process of setting the number of hidden neurons, the number of neurons in the hidden layer is gradually reduced from the bottom to the top, and the number of neurons in the last hidden layer is uniformly set to 50; the training time required for the DBN model with different numbers of hidden layers is recorded, and the feature evaluation indicators based on intra-class and inter-class distances are calculated.

6. The method for extracting fault features of a high-pass filter based on PSO-DBN according to claim 1, characterized in that: In step 3, during the iteration process, when the value of the fitness function is smaller, it means that the combination of the number of hidden layer neurons found is optimal, and the corresponding DBN model feature extraction performance is also better. The steps of using the PSO algorithm to optimize the number of hidden layer neurons in the DBN model are as follows: a) In the PSO algorithm, the total number of particles is set to 20, the number of iterations is set to 100, the range of the inertia weight ω is set to [0.4, 0.9], and the range of the learning factors c1 and c2 is set to [0.5, 2.5]; b) Using the fitness function in the PSO algorithm, draw the change curve of the fitness function; when the convergence number k = 57, the fitness function in the PSO algorithm converges and remains unchanged for a long time; c) The optimal number of hidden layer neurons is 139, 171 and 35 respectively. At this time, the overall evaluation index of the feature vector extracted by the PSO-DBN model is After all training is completed, the deep features of the original voltage signal are extracted layer by layer through the stacked RBM, and the final output is the extracted features; d) Compare the extracted features with the signal analysis method wavelet packet decomposition (WPD), and use as an evaluation indicator.

Images