Switch cabinet monitoring device and signal processing method thereof

Abstract

The invention discloses a switch cabinet monitoring device and a signal processing method thereof, and the device comprises an extension probe, a socket connector, a mechanical filter, a first probe, a vibration-charge converter, a charge power amplification module, a signal preprocessing module and a signal processing module. The mechanical filter is connected with the first probe through the socket connector, the first probe is connected with the vibration-charge converter, signal acquisition, filtering and classification are finally completed through the signal processing module, signal preprocessing adopts a band-pass filter, fast Fourier transform and normalized denoising filtering, signal processing adopts short-time Fourier transform, and the vibration-charge converter is connected with the vibration-charge converter. And the convolutional neural network extracts depth features, and classification and anomaly detection are carried out through a full connection layer and a Softmax classifier. According to the invention, the signal capture sensitivity and the anti-interference capability are improved through the spring ejector pin type probe and the multilayer composite mechanical filter structure, meanwhile, a cross entropy loss optimization model is supported, and the classification precision and the anomaly detection reliability are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of power equipment monitoring, and in particular relates to a switch cabinet monitoring device and a signal processing method thereof. Background Art

[0002] Once a fault occurs during the operation of the switch cabinet, abnormal noise will be generated inside. Abnormal noise monitoring has historically adopted a variety of technologies, such as ultrasound, or transient voltage to ground combined with electrical signal analysis. This method relies too much on the signal after analog-to-digital conversion and loses the original sound detection solution. The main reason is still due to the interference of environmental noise on site. If the environmental noise interference can be solved to a certain extent, then sound measurement is still an intuitive, practical, economical and popular solution. In addition, in some on-site monitoring processes, it is difficult to ensure that the monitored equipment has no leakage. The existing technical sensors are too close to the operator, and it is easy to connect high voltages, introducing the risk of electric shock. It is necessary to solve practical problems such as sensitivity and personal safety under leakage conditions. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a switch cabinet monitoring device and a signal processing method thereof which are safe and reliable and improve signal capture sensitivity and anti-interference ability by using a spring pin probe and a multi-layer composite mechanical filter structure, while improving classification accuracy and anomaly detection reliability.

[0004] The technical solution of the present invention is: a switch cabinet monitoring device, including: an extended probe, a socket, a mechanical filter, a first probe, a vibration-charge converter, a charge power amplifier module, a signal preprocessing module and a signal processing module, the extended probe is connected to the mechanical filter through the socket, the mechanical filter is connected to the first probe through the socket, the socket is sleeved on the mechanical filter, the first probe is connected to the vibration-charge converter, the vibration-charge converter is connected to the charge power amplifier module, the charge power amplifier module is connected to the signal preprocessing module, and the signal preprocessing module is connected to the signal processing module.

[0005] Furthermore, the extension probe is a hollow tubular structure, the diameter of the extension probe is less than 2.5 mm, and the length is not less than 15 cm.

[0006] Furthermore, the first probe is a spring-type ejector structure.

[0007] Furthermore, the socket is a metal ring.

[0008] Furthermore, the mechanical filter is a multi-layer composite structure, including: a metal ring and a plastic ring, the metal ring is sleeved on the plastic ring, and the plastic ring is sleeved with the extension probe and the first probe respectively.

[0009] Furthermore, the vibration-charge converter contains piezoelectric material.

[0010] Furthermore, the charge power amplifier module includes a high input impedance circuit.

[0011] A signal preprocessing method for a switch cabinet monitoring device comprises the following steps:

[0012] Step 1: Sampling signal: The signal x(t) captured by the vibration-charge converter is sampled and quantized into a discrete signal x[n]. The sampling formula is:

[0013] x[n]=x(nT s ),

[0014] Where x(t) represents a continuous-time signal, x[n] represents a discrete-time signal, n represents a sampling point index, and T s represents the sampling period, i.e., the time interval for discretizing the continuous-time signal x(t) in the time domain;

[0015] Step 2, design a bandpass filter: Design a discrete bandpass filter to retain the signal in the target frequency band and suppress other frequency bands. The transfer function expression of the discrete bandpass filter is:

[0016]

[0017] Among them, f L represents the low cutoff frequency of the filter, f H represents the high cutoff frequency of the filter, and H(f) represents the frequency domain transfer function of the filter;

[0018] Step 3, filter implementation: Use a digital filter to implement the filtering process, the digital filter includes a finite impulse response filter and an infinite impulse response filter, and the expression of the finite impulse response filter is:

[0019]

[0020] Where y[n] represents the output signal, h[k] represents the impulse response of the filter, M represents the filter order, x[n] represents the input signal, x[nk] represents the value of the input signal delayed by k sampling points, and k represents the accumulation index, that is, the offset when the filter delays the input signal;

[0021] The expression of the infinite impulse response filter is:

[0022]

[0023] Wherein, y[n] represents the output signal, x[n] represents the input signal, x[nk] represents the value of the input signal delayed by k sampling points, b[k] represents the filter coefficient of the input signal, a[k] represents the feedback coefficient of the output signal, N represents the order of the input signal, M represents the order of the output signal, and y[nk] represents the value of the output signal delayed by k sampling points;

[0024] Step 4: Filtered signal output: Output the filtered signal y[n] for subsequent analysis.

[0025] Furthermore, after step 3 is completed, frequency domain filtering is also performed, and the frequency domain filtering includes the following steps:

[0026] S1. Perform fast Fourier transform on the signal x[n]:

[0027]

[0028] Where X[k] represents the frequency domain representation of the signal, x[n] represents the input signal, and N represents the number of signal points;

[0029] S2. Apply a frequency domain window function to X[k] to retain the target frequency band:

[0030] Y[k]=H[k]·X[k],

[0031] Wherein, X[k] represents the frequency domain representation of the signal, H[k] represents the frequency domain transfer function of the frequency domain filter, and Y[k] represents the frequency domain signal after filtering;

[0032] S3. Apply inverse fast Fourier transform to Y[k] to obtain the time domain signal:

[0033]

[0034] Where y[n] represents the discrete value of the time domain signal, N represents the number of signal points, Y[k] represents the frequency domain representation of the input signal, k represents the frequency index, and n represents the time domain index.

[0035] A signal processing method for a switch cabinet monitoring device comprises the following steps:

[0036] Step 1: Input signal preprocessing, including the following steps:

[0037] S1, sampling signal: the filtered signal x[n] output by the acquisition signal preprocessing module;

[0038] S2. Normalization, the normalized expression is:

[0039]

[0040] Where x′[n] represents the value of the normalized signal at the nth sampling point, x[n] represents the value of the input signal at the nth sampling point, and μ x represents the mean of the signal x[n], σ x Represents the standard deviation of the signal x[n];

[0041] S3, Window segmentation: Segment the signal and use a sliding window to extract signal fragments:

[0042] x i =[x′[i],x′[i+1],...,x′[i+L-1]],

[0043] Among them, x i represents the i-th signal segment, x′[i] represents the standardized discrete signal, and L represents the window length;

[0044] Step 2, feature extraction, includes the following steps:

[0045] S4. Time-frequency analysis: for signal segment x i Perform short-time Fourier transform and extract the time-frequency feature matrix:

[0046]

[0047] Among them, X i (t,f) represents the time-frequency characteristics of the signal, w(τ-t) represents the window function, using Hanning window or Gaussian window, t represents time, and f represents frequency;

[0048] S5. Generate time-frequency graph: transform the time-frequency feature matrix X i (t,f) is converted into a two-dimensional image representation and used as the input of the model;

[0049] Step 3: Deep feature extraction, including the following steps:

[0050] S6. Convolutional neural network: Design a convolutional neural network to extract deep features from the time-frequency graph. The operation formula of the convolutional neural network is:

[0051]

[0052] in, represents the activation value output by the convolution kernel of the lth layer, represents the weight of the convolution kernel of the lth layer, represents the value of the input feature map at the position (i+m,j+n) in the l-1th layer, b (l) represents the bias term of the lth layer, σ represents the activation function;

[0053] S7, pooling layer: reduce the dimension of convolution features and extract key information:

[0054] y ij =max(x ij ),

[0055] Among them, y ij represents the feature value after pooling, x ij Represents the input feature value, and the pooling method is maximum pooling or average pooling;

[0056] Step 4: Classification and anomaly detection, including the following steps:

[0057] S8, fully connected layer: Input the deep features into the fully connected layer and calculate the score of each category:

[0058]

[0059] Among them, z k represents the score of category k, w kj represents the weight of the fully connected layer, b k represents the bias term of the fully connected layer, y j represents the input feature value, k represents the index of the output unit, and j represents the index of the input unit;

[0060] S9, Softmax classifier: Calculate the probability of each category:

[0061]

[0062] Among them, P k represents the probability of belonging to category k, N represents the total number of categories, z k represents the score of category k, and j represents the cumulative index;

[0063] S10, abnormal identification: if max(P k ) is less than the set threshold θ, it is judged as an unknown abnormal mode:

[0064]

[0065] Among them, max(P k ) represents the probability P of all categories k The largest probability value among them;

[0066] Step 5: Model training, including the following steps:

[0067] S11. Loss function: Use the cross entropy loss function to optimize the model:

[0068]

[0069] Among them, L represents the total loss value, N represents the total number of categories in the classification problem, and yk represents the distribution of the true category, P k Represents the model's predicted probability for the kth class;

[0070] S12. Optimization algorithm: Use the optimization algorithm Adam or SGD to update the model parameters.

[0071] Beneficial effects of the present invention:

[0072] 1. High signal capture sensitivity: The extended probe adopts a spring-loaded pin design to ensure sensitive signal transmission and stable contact, and can accurately capture mechanical vibration signals even in complex environments;

[0073] 2. Strong anti-interference ability: The mechanical filter adopts a multi-layer composite structure to effectively filter out vibration noise in non-target frequency bands and improve signal quality;

[0074] 3. Efficient signal processing: The signal preprocessing module combines bandpass filters and fast Fourier transforms to retain the signal in the target frequency band and remove environmental noise, providing high-quality input for subsequent processing;

[0075] 4. Strong feature extraction capability: The signal processing module uses short-time Fourier transform to generate time-frequency features, combined with convolutional neural network to extract deep features, which improves the classification of equipment operation status and the accuracy of anomaly detection; 5. High reliability of anomaly detection: The Softmax classifier combined with the maximum probability threshold judgment method can effectively detect unknown abnormal patterns and is suitable for complex and changeable operating environments;

[0076] 6. Strong structural stability of the equipment: through the combination of metal sockets and elastic elements, the stability of the device in long-term operation is ensured to avoid the influence of mechanical jitter on signal capture;

[0077] 7. Improve the safety of the power system: The present invention can timely detect the abnormal operation and potential faults of the switch cabinet, prevent the spread of faults, and provide reliable operation guarantee for the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 The present invention is a schematic diagram of the connection relationship between the components of a switch cabinet monitoring device.

[0079] Figure 2 It is a structural schematic diagram of a first probe of a switch cabinet monitoring device of the present invention.

[0080] In the figure: 1-extended probe, 2-set of machines, 3-first probe. DETAILED DESCRIPTION

[0081] like Figure 1-2As shown, a switch cabinet monitoring device includes: an extended probe 1, a socket 2, a mechanical filter, a first probe 3, a vibration-charge converter, a charge power amplifier module, a signal preprocessing module and a signal processing module, the extended probe 1 is connected to the mechanical filter through the socket 2, the mechanical filter is connected to the first probe 3 through the socket 2, the socket 2 is sleeved on the mechanical filter, the first probe 3 is connected to the vibration-charge converter, the vibration-charge converter is connected to the charge power amplifier module, the charge power amplifier module is connected to the signal preprocessing module, and the signal preprocessing module is connected to the signal processing module.

[0082] Preferably, the extension probe 1 is a hollow tubular structure, the diameter of the extension probe 1 is less than 2.5 mm, and the length is not less than 15 cm.

[0083] Preferably, the first probe 3 is a spring-type ejector structure.

[0084] Preferably, the socket connector 2 is a metal ring.

[0085] Preferably, the mechanical filter is a multi-layer composite structure, including: a metal ring and a plastic ring, the metal ring is sleeved on the plastic ring, and the plastic ring is sleeved with the extension probe 1 and the first probe 3 respectively.

[0086] Preferably, the vibration-charge converter contains piezoelectric material.

[0087] Preferably, the charge power amplifier module comprises a high input impedance circuit.

[0088] A signal preprocessing method for a switch cabinet monitoring device comprises the following steps:

[0089] Step 1: Sampling signal: The signal x(t) captured by the vibration-charge converter is sampled and quantized into a discrete signal x[n]. The sampling formula is:

[0090] x(n)=x(nT s ),

[0091] Where x(t) represents a continuous-time signal, x[n] represents a discrete-time signal, n represents a sampling point index, and T s represents the sampling period, i.e., the time interval for discretizing the continuous-time signal x(t) in the time domain;

[0092] Step 2, design a bandpass filter: Design a discrete bandpass filter to retain the signal in the target frequency band and suppress other frequency bands. The transfer function expression of the discrete bandpass filter is:

[0093]

[0094] Among them, f L represents the low cutoff frequency of the filter, f H represents the high cutoff frequency of the filter, and H(f) represents the frequency domain transfer function of the filter;

[0095] Step 3, filter implementation: Use a digital filter to implement the filtering process, the digital filter includes a finite impulse response filter and an infinite impulse response filter, and the expression of the finite impulse response filter is:

[0096]

[0097] Where y[n] represents the output signal, h[k] represents the impulse response of the filter, M represents the filter order, x[n] represents the input signal, x[nk] represents the value of the input signal delayed by k sampling points, and k represents the accumulation index, that is, the offset when the filter delays the input signal;

[0098] The expression of the infinite impulse response filter is:

[0099]

[0100] Wherein, y[n] represents the output signal, x[n] represents the input signal, x[nk] represents the value of the input signal delayed by k sampling points, b[k] represents the filter coefficient of the input signal, a[k] represents the feedback coefficient of the output signal, N represents the order of the input signal, M represents the order of the output signal, and y[nk] represents the value of the output signal delayed by k sampling points;

[0101] Step 4: Filtered signal output: Output the filtered signal y[n] for subsequent analysis.

[0102] Preferably, after step 3 is completed, frequency domain filtering is further performed, and the frequency domain filtering includes the following steps:

[0103] S1. Perform fast Fourier transform on the signal x[n]:

[0104]

[0105] Where X[k] represents the frequency domain representation of the signal, x[n] represents the input signal, and N represents the number of signal points;

[0106] S2. Apply a frequency domain window function to X[k] to retain the target frequency band:

[0107] Y[k]=H[k]·X[k],

[0108] Wherein, X[k] represents the frequency domain representation of the signal, H[k] represents the frequency domain transfer function of the frequency domain filter, and Y[k] represents the frequency domain signal after filtering;

[0109] S3. Apply inverse fast Fourier transform to Y[k] to obtain the time domain signal:

[0110]

[0111] Where y[n] represents the discrete value of the time domain signal, N represents the number of signal points, Y[k] represents the frequency domain representation of the input signal, k represents the frequency index, and n represents the time domain index.

[0112] A signal processing method for a switch cabinet monitoring device comprises the following steps:

[0113] Step 1: Input signal preprocessing, including the following steps:

[0114] S1, sampling signal: the filtered signal x[n] output by the acquisition signal preprocessing module;

[0115] S2. Normalization, the normalized expression is:

[0116]

[0117] Where x′[n] represents the value of the normalized signal at the nth sampling point, x[n] represents the value of the input signal at the nth sampling point, and μ x represents the mean of the signal x[n], σ x Represents the standard deviation of the signal x[n];

[0118] S3, Window segmentation: Segment the signal and use a sliding window to extract signal fragments:

[0119] x i =[x′[i],x′[i+1],...,x′[i+1L-1]],

[0120] Among them, x i represents the i-th signal segment, x′[i] represents the standardized discrete signal, and L represents the window length;

[0121] Step 2, feature extraction, includes the following steps:

[0122] S4. Time-frequency analysis: for signal segment x i Perform short-time Fourier transform and extract the time-frequency feature matrix:

[0123]

[0124] Among them, X i (t,f) represents the time-frequency characteristics of the signal, w(τ-t) represents the window function, using Hanning window or Gaussian window, t represents time, and f represents frequency;

[0125] S5. Generate time-frequency graph: transform the time-frequency feature matrix X i (t,f) is converted into a two-dimensional image representation and used as the input of the model;

[0126] Step 3: Deep feature extraction, including the following steps:

[0127] S6. Convolutional neural network: Design a convolutional neural network to extract deep features from the time-frequency graph. The operation formula of the convolutional neural network is:

[0128]

[0129] in, represents the activation value output by the convolution kernel of the lth layer, represents the weight of the convolution kernel of the lth layer, represents the value of the input feature map at the position (i+m,j+n) in the l-1th layer, b (l) represents the bias term of the lth layer, σ represents the activation function;

[0130] S7, pooling layer: reduce the dimension of convolution features and extract key information:

[0131] y ij =max(x ij ),

[0132] Among them, y ij represents the feature value after pooling, x ij Represents the input feature value, and the pooling method is maximum pooling or average pooling;

[0133] Step 4: Classification and anomaly detection, including the following steps:

[0134] S8, fully connected layer: Input the deep features into the fully connected layer and calculate the score of each category:

[0135]

[0136] Among them, z k represents the score of category k, w kj represents the weight of the fully connected layer, b k represents the bias term of the fully connected layer, y j represents the input feature value, k represents the index of the output unit, and j represents the index of the input unit;

[0137] S9, Softmax classifier: Calculate the probability of each category:

[0138]

[0139] Among them, P krepresents the probability of belonging to category k, N represents the total number of categories, z k represents the score of category k, and j represents the cumulative index;

[0140] S10, abnormal identification: if max(P k ) is less than the set threshold θ, it is judged as an unknown abnormal mode:

[0141]

[0142] Among them, max(P k ) represents the probability P of all categories k The largest probability value among them;

[0143] Step 5: Model training, including the following steps:

[0144] S11. Loss function: Use the cross entropy loss function to optimize the model:

[0145]

[0146] Among them, L represents the total loss value, N represents the total number of categories in the classification problem, and y k represents the distribution of the true category, P k Represents the model's predicted probability for the kth class;

[0147] S12. Optimization algorithm: Use the optimization algorithm Adam or SGD to update the model parameters.

[0148] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments and various changes can be made thereto within the knowledge scope of those skilled in the art.

Claims

1. A switch cabinet monitoring device, characterized in that: include: An extended probe (1), a socket (2), a mechanical filter, a first probe (3), a vibration-charge converter, a charge power amplifier module, a signal preprocessing module and a signal processing module. The extended probe (1) is connected to the mechanical filter via the socket (2), the mechanical filter is connected to the first probe (3) via the socket (2), the socket (2) is sleeved on the mechanical filter, the first probe (3) is connected to the vibration-charge converter, the vibration-charge converter is connected to the charge power amplifier module, the charge power amplifier module is connected to the signal preprocessing module, and the signal preprocessing module is connected to the signal processing module.

2. A switch cabinet monitoring device according to claim 1, characterized in that: The extension probe (1) is a hollow tubular structure, the diameter of the extension probe (1) is less than 2.5 mm, and the length is not less than 15 cm.

3. A switch cabinet monitoring device according to claim 1, characterized in that: The first probe (3) is a spring-type ejector structure.

4. A switch cabinet monitoring device according to claim 1, characterized in that: The socket connector (2) is a metal ring.

5. A switch cabinet monitoring device according to claim 1, characterized in that: The mechanical filter is a multi-layer composite structure, comprising: a metal ring and a plastic sleeve ring, wherein the metal ring is sleeved on the plastic sleeve ring, and the plastic sleeve ring is sleeved with the extension probe (1) and the first probe (3) respectively.

6. A switch cabinet monitoring device according to claim 1, characterized in that: The vibration-charge transducer contains piezoelectric material.

7. A switch cabinet monitoring device according to claim 1, characterized in that: The charge power amplifier module includes a high input impedance circuit.

8. A signal preprocessing method for a switch cabinet monitoring device according to claim 1, characterized in that: The following steps are involved: Step 1: Sampling signal: The signal x(t) captured by the vibration-charge converter is sampled and quantized into a discrete signal x[n]. The sampling formula is: x[n]=x(nT s ), Where x(t) represents a continuous-time signal, x[n] represents a discrete-time signal, n represents a sampling point index, and T s represents the sampling period, i.e., the time interval for discretizing the continuous-time signal x(t) in the time domain; Step 2, design a bandpass filter: Design a discrete bandpass filter to retain the signal in the target frequency band and suppress other frequency bands. The transfer function expression of the discrete bandpass filter is: Among them, f L represents the low cutoff frequency of the filter, f H represents the high cutoff frequency of the filter, and H(f) represents the frequency domain transfer function of the filter; Step 3, filter implementation: Use a digital filter to implement the filtering process, the digital filter includes a finite impulse response filter and an infinite impulse response filter, and the expression of the finite impulse response filter is: Where y[n] represents the output signal, h[k] represents the impulse response of the filter, M represents the filter order, x[n] represents the input signal, x[nk] represents the value of the input signal delayed by k sampling points, and k represents the accumulation index, that is, the offset when the filter delays the input signal; The expression of the infinite impulse response filter is: Wherein, y[n] represents the output signal, x[n] represents the input signal, x[nk] represents the value of the input signal delayed by k sampling points, b[k] represents the filter coefficient of the input signal, a[k] represents the feedback coefficient of the output signal, N represents the order of the input signal, M represents the order of the output signal, and y[nk] represents the value of the output signal delayed by k sampling points; Step 4: Filtered signal output: Output the filtered signal y[n] for subsequent analysis.

9. The signal preprocessing method of the switch cabinet monitoring device according to claim 8, characterized in that: After step 3 is completed, frequency domain filtering is further performed, and the frequency domain filtering includes the following steps: S1. Perform fast Fourier transform on the signal x[n]: Where X[k] represents the frequency domain representation of the signal, x[n] represents the input signal, and N represents the number of signal points; S2. Apply a frequency domain window function to X[k] to retain the target frequency band: Y[k]=H[k]·X[k], Wherein, X[k] represents the frequency domain representation of the signal, H[k] represents the frequency domain transfer function of the frequency domain filter, and Y[k] represents the frequency domain signal after filtering; S3. Apply inverse fast Fourier transform to Y[k] to obtain the time domain signal: Where y[n] represents the discrete value of the time domain signal, N represents the number of signal points, Y[k] represents the frequency domain representation of the input signal, k represents the frequency index, and n represents the time domain index.

10. A signal processing method for a switch cabinet monitoring device according to claim 1, characterized in that: The following steps are involved: Step 1: Input signal preprocessing, including the following steps: S1, sampling signal: the filtered signal x[n] output by the acquisition signal preprocessing module; S2. Normalization, the normalized expression is: Where x′[n] represents the value of the normalized signal at the nth sampling point, x[n] represents the value of the input signal at the nth sampling point, and μ x represents the mean of the signal x[n], σ x Represents the standard deviation of the signal x[n]; S3, Window segmentation: Segment the signal and use a sliding window to extract signal fragments: x i =[x′[i],x′[i+1],...,x′[i+L-1]], Among them, x i represents the i-th signal segment, x′[i] represents the standardized discrete signal, and L represents the window length; Step 2, feature extraction, includes the following steps: S4. Time-frequency analysis: for signal segment x i Perform short-time Fourier transform and extract the time-frequency feature matrix: Among them, X i (t, f) represents the time-frequency characteristics of the signal, w(τ-t) represents the window function, using Hanning window or Gaussian window, t represents time, and f represents frequency; S5. Generate time-frequency graph: transform the time-frequency feature matrix X i (t, f) is converted into a two-dimensional image representation and used as the input of the model; Step 3: Deep feature extraction, including the following steps: S6. Convolutional neural network: Design a convolutional neural network to extract deep features from the time-frequency graph. The operation formula of the convolutional neural network is: in, represents the activation value output by the convolution kernel of the lth layer, represents the weight of the convolution kernel of the lth layer, represents the value of the input feature map at the position (i+m, j+n) in the l-1th layer, b (l) represents the bias term of the lth layer, σ represents the activation function; S7, pooling layer: reduce the dimension of convolution features and extract key information: and ij =max(x ij ), Among them, y ij represents the feature value after pooling, x ij Represents the input feature value, and the pooling method is maximum pooling or average pooling; Step 4: Classification and anomaly detection, including the following steps: S8, fully connected layer: Input the deep features into the fully connected layer and calculate the score of each category: Among them, z k represents the score of category k, w kj represents the weight of the fully connected layer, b k represents the bias term of the fully connected layer, y j represents the input feature value, k represents the index of the output unit, and j represents the index of the input unit; S9, Softmax classifier: Calculate the probability of each category: Among them, P k represents the probability of belonging to category k, N represents the total number of categories, z k represents the score of category k, and j represents the cumulative index; S10, abnormal identification: If max(P k ) is less than the set threshold θ, it is judged as an unknown abnormal mode: Among them, max(P k ) represents the probability P of all categories k The largest probability value among them; Step 5: Model training, including the following steps: S11. Loss function: Use the cross entropy loss function to optimize the model: Among them, L represents the total loss value, N represents the total number of categories in the classification problem, and y k represents the distribution of the true category, P k Represents the model's predicted probability for the kth class; S12. Optimization algorithm: Use the optimization algorithm Adam or SGD to update the model parameters.

Images