High-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction

Abstract

The invention discloses a high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction. Acquiring a voltage signal of the high-voltage power switch by using a sensor, and performing synchronous sampling; a wavelet threshold value correction noise reduction method is adopted to carry out noise reduction processing on the collected signals, wavelet detail coefficients are calculated through multi-scale decomposition, a threshold value is adaptively corrected based on the peak sum ratio, and the noise removal effect is optimized; thirdly, performing normalization processing on the denoised signal, mapping the signal to a polar coordinate system, constructing a two-dimensional Gramer angle field containing an included angle cosine value and amplitude information, and realizing time sequence-space conversion of the signal; and finally, generating two-dimensional image data of the voltage signal of the high-voltage power switch, and providing feature input for subsequent state evaluation and fault detection. According to the method, wavelet transform and two-dimensional feature mapping are combined, noise interference can be effectively reduced, the signal distinguishability and the information retention capacity are improved, and the method is suitable for state monitoring and intelligent diagnosis of a power system.

Description

Technical Field

[0001] The present invention belongs to the technical field of high-voltage power switch fault diagnosis, and in particular relates to a high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction. Background Art

[0002] High-voltage power switches are critical control and protection equipment in power systems. Common power switches include circuit breakers, disconnectors, and load switches. Failures can lead to serious losses in power system stability and reliability. With the development of intelligent power systems, the safety and reliability requirements for high-voltage power switches are becoming increasingly stringent. High-voltage power switch failures can be diverse, with various mechanical failures being common. Therefore, fault diagnosis of high-voltage power switches, providing a basis for condition-based maintenance, is crucial for ensuring the safe operation of power systems.

[0003] The current in the opening and closing coils, motor current, and travel curves of high-voltage power switches are important sources of fault diagnosis information. To facilitate data processing and analysis on the acquisition platform, angular displacement sensors and current Hall effect sensors are used to collect this information and convert it into a voltage signal, which serves as the processing signal for fault diagnosis. This signal can accurately reflect the fault type and various damage characteristics of the power switch. This voltage signal can also reflect various faults, such as normal operation of the power switch, spring weakness, spring sticking, improper opening and closing, and undervoltage operation. In recent years, wavelet threshold denoising has been a common method for noise reduction of acquired signals. Traditional wavelet threshold denoising methods set a constant threshold for each layer of wavelet coefficients. However, noise varies across wavelet coefficients, so using a fixed threshold results in poor adaptability and less than ideal noise reduction. Furthermore, some researchers have directly extracted feature values from the acquired signals to preserve data integrity. While this method preserves data integrity, it increases the data processing workload, and noise interference can also affect the accuracy of fault diagnosis. In these data processing methods, traditional wavelet filtering uses a fixed threshold, which can result in data loss and incomplete noise removal. Directly collecting eigenvalues will make subsequent neural network training more difficult and the fault diagnosis of power switches will not be accurate enough.

[0004] In the prior art, Chinese patent application CN114118159A discloses a high-voltage circuit breaker fault diagnosis method based on image recognition, which relates to the field of online monitoring and fault diagnosis of electrical equipment. The method comprises the following steps: collecting historical current data of the high-voltage circuit breaker operating coil and preprocessing it; using a mapping function to convert the preprocessed one-dimensional time series of the current historical data into a two-dimensional tensor, and drawing a current image with uniform pixel size; processing the current image to obtain a current grayscale image, which is combined with historical data labels to form high-voltage circuit breaker fault samples, thereby obtaining a high-voltage circuit breaker fault sample library; training a pre-established two-dimensional convolutional neural network model based on the high-voltage circuit breaker fault samples to obtain a fault diagnosis model that meets accuracy requirements; online monitoring of the high-voltage circuit breaker operating coil current, converting it into an online current grayscale image after data preprocessing, and inputting the online current grayscale image into the trained fault diagnosis model to obtain a fault diagnosis result. However, the patent uses the Sqtwolog criterion to calculate the noise reduction threshold. This method may lead to over-smoothing of the signal when the noise is strong, resulting in loss of fault characteristics and affecting diagnostic accuracy. In addition, this patent uses a mapping function to convert a one-dimensional current signal into a two-dimensional tensor. Its conversion method is relatively simple. Due to the strong time series characteristics of the current signal, the simple mapping method may lead to the loss of feature information, making the fault mode insufficiently expressive in the image, thereby affecting the feature extraction and fault identification capabilities of the convolutional neural network (CNN). Summary of the Invention

[0005] In response to the above problems, the present invention proposes a high-voltage power switch fault diagnosis method based on wavelet threshold correction noise reduction and GAF-CNN. On the noise reduction problem, in order to make the threshold value more consistent with the noise change law, a wavelet threshold correction noise reduction algorithm is proposed. A correction factor is added on the basis of the original threshold to make the noise reduction effect better. In the traditional direct eigenvalue extraction, the present invention adopts a two-dimensional image conversion algorithm. This method fully considers the multimodal characteristics of the power switch and uses the Gram angular field (GAF) image encoding method to change the one-dimensional high-voltage vacuum circuit breaker vibration signal into two dimensions. This method amplifies the time domain change information of the signal and strengthens the time characteristics of the signal, which can not only retain the complete information of the signal, but also maintain the temporal nature of the signal. Finally, the designed CNN network realizes the eigenvalue fusion diagnosis and outputs the diagnosis result.

[0006] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] In one aspect, the present invention provides a high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction, comprising the following steps:

[0009] Collect voltage signals from high-voltage power switches;

[0010] The voltage signal is denoised using the wavelet threshold correction noise reduction method.

[0011] The GAF method is used to perform two-dimensional conversion on the voltage signal after noise reduction to obtain two-dimensional image data of the voltage signal of the high-voltage power switch;

[0012] A pre-trained convolutional neural network is used to extract features and classify faults from two-dimensional image data, identify the operating status of high-voltage power switches, and determine whether a fault exists and its specific fault type.

[0013] Furthermore, the collecting of the voltage signal of the high-voltage power switch specifically includes:

[0014] A displacement sensor is installed on the transmission rod of the high-voltage power switch to detect the stroke change of the transmission rod and convert the stroke signal into a voltage signal;

[0015] Current Hall sensors are installed on the opening coil, closing coil and motor current inlet to collect current signals and convert them into corresponding voltage signals;

[0016] Place voltage sensors at appropriate locations in the control loop to monitor voltage changes during switching operations in real time;

[0017] Set the sampling time T and sampling frequency f s ,The synchronous sampling technology is used to sample the high voltage power switch through ,each sensor to obtain the voltage signal of the high voltage power switch.

[0018] Furthermore, the voltage signal of the high-voltage power switch includes voltage signals collected by a displacement sensor, a current Hall sensor, and a voltage sensor.

[0019] Furthermore, the voltage signal is subjected to noise reduction processing using a wavelet threshold correction noise reduction method, specifically including:

[0020] Perform wavelet transform on the voltage signal and use discrete wavelet transform to decompose the signal into multiple scales to obtain the wavelet detail coefficients d at each frequency level in the voltage signal. j,k and its approximate coefficients, where d j,k Indicates the kth wavelet detail coefficient of the voltage signal at this scale in the jth layer decomposition;

[0021] According to the wavelet detail coefficient d of each frequency level in the voltage signal j,k Calculate the peak-sum ratio S of the wavelet detail coefficients of each layer j , where S jis the peak-sum ratio of the j-th layer decomposition wavelet detail coefficients;

[0022] Based on the peak-sum ratio S of the wavelet detail coefficients of each layer j Calculate the correction factor F j , and combined with ln(j+1) to modify the traditional wavelet threshold, we get the improved threshold λ j ;

[0023] For the wavelet detail coefficient d j,k Applying improved threshold λ j , and processed using an improved threshold function;

[0024] Using the processed wavelet detail coefficients The inverse wavelet transform is performed on the denoised voltage signal and its approximate coefficients to reconstruct the denoised voltage signal.

[0025] Furthermore, the peak-sum ratio S of the wavelet detail coefficients of each layer is calculated. j , the calculation formula is:

[0026]

[0027] Among them, max(|d j |) is the maximum absolute value of the j-th layer wavelet coefficient, is the total absolute value and total wavelet coefficients of the jth layer, N j Represents the total number of wavelet detail coefficients in the jth layer, that is, the number of samples of the coefficients in this layer;

[0028] The peak-sum ratio S based on the wavelet detail coefficients of each layer j Calculate the correction factor F j , the calculation formula is:

[0029]

[0030] Among them L j is the length of the j-layer wavelet detail coefficient;

[0031] The improved threshold λ j The calculation formula is:

[0032]

[0033] Among them, λ j is the improved threshold of the j-th layer wavelet detail coefficient, σ=M mid / 0.675,M mid is the median of the lowest level wavelet coefficients, F j is the correction factor of the j-th layer wavelet detail coefficient;

[0034] The improved threshold function formula is:

[0035]

[0036] in, is the wavelet detail coefficient of the kth wavelet detail coefficient of the jth layer after applying the improved threshold, sign(d j,k ) is a sign function that represents the sign of the wavelet detail coefficient. It outputs 1 for positive values and -1 for negative values.

[0037] Furthermore, the GAF method is used to perform two-dimensional conversion on the voltage signal after noise reduction to obtain two-dimensional image data of the voltage signal of the high-voltage power switch, specifically including:

[0038] Normalizing the noise-reduced voltage signal to obtain a normalized voltage signal, wherein the normalized voltage signal has a value range of [-1, 1];

[0039] Each element in the normalized voltage signal sequence is taken as the angle cosine value α, and is transformed into a polar coordinate system to obtain its corresponding angle cosine value α and radius;

[0040] Using the angle cosine value α and radius obtained by the polar coordinate transformation, a two-dimensional Gram angle field is constructed;

[0041] The two-dimensional Gram angle field is used as two-dimensional image data of the voltage signal after noise reduction.

[0042] Furthermore, the voltage signal after noise reduction is X={x1, x2,··x i ··, x1}, where x i represents the i-th sampling point of the voltage signal;

[0043] The voltage signal after noise reduction is normalized, and the calculation formula is:

[0044]

[0045] in, is the i-th element in the normalized voltage signal, max(X) is the maximum value in the voltage signal sequence X after noise reduction, and min(X) is the minimum value in the voltage signal sequence X after noise reduction.

[0046] Furthermore, the method of taking each element in the normalized voltage signal sequence as the angle cosine value α and transforming it into a polar coordinate system to obtain the corresponding angle cosine value α and radius specifically includes:

[0047] For each element in the normalized voltage signal sequence Calculate the corresponding angle cosine value α, the formula is:

[0048]

[0049] Among them, α i is the normalized voltage signal The corresponding angle cosine value is in the range [0,π], is the i-th element in the normalized voltage signal;

[0050] According to the sequence value t of each element in the normalized voltage signal sequence i , calculate the corresponding radius value r in the polar coordinate system i , the calculation formula is:

[0051]

[0052] Among them, r i is the voltage signal t in the polar coordinate system i The radius of the sampling point, t i is the sequence value of the i-th element in the voltage signal sequence, indicating the position of the element in the sequence, and N is the total number of sampling points in the voltage signal sequence, that is, the length of the signal.

[0053] Furthermore, the angle cosine value α and radius obtained by the polar coordinate transformation are used to construct a two-dimensional Gram angle field, and the formula is:

[0054]

[0055] Among them, G ADF is the two-dimensional Gram angle field, which represents the geometric structure of the voltage signal in the polar coordinate system.

[0056] Furthermore, the convolutional neural network training process includes:

[0057] For different types of high-voltage power switches, including circuit breakers, disconnectors, and load switches, the voltage signals of the high-voltage power switches in normal states and various fault states are collected. The fault states include spring weakness, spring jamming, inadequate opening and closing, and undervoltage operation. For each fault state, the corresponding voltage signal is collected;

[0058] Label the collected voltage signals to clarify the specific status corresponding to each signal;

[0059] Classify all marked voltage signals according to fault type and normal state, and establish a signal database of fault and normal state;

[0060] Dividing the signal data in the database into a training set and a validation set;

[0061] The voltage signals in the training set and the validation set are converted into two-dimensional images using the GAF method to obtain their corresponding two-dimensional image data;

[0062] Construct a convolutional neural network structure, including multiple convolutional layers, pooling layers, and fully connected layers. Add a maximum pooling layer after each convolutional layer, and set a batch normalization layer and ReLU activation function after the convolutional layer.

[0063] The two-dimensional image data and corresponding labels in the training set are input into the convolutional neural network for training. Specifically, the two-dimensional image data is used as the input of the convolutional neural network, supervised learning is performed according to the corresponding labels, the network weights are optimized by calculating the loss function and performing backpropagation, and in each training iteration, the model is evaluated using the validation set and the network hyperparameters are adjusted to finally form a convolutional neural network model with diagnostic capabilities.

[0064] Compared with the prior art, the present invention has the following advantages:

[0065] (1) To address the signal loss or residual noise caused by the fixed threshold of traditional wavelet thresholding methods, the present invention achieves adaptive threshold denoising by calculating the peak-to-sum ratio of the wavelet detail coefficients of each layer, dynamically adjusting the correction factor, and combining it with an improved threshold function. Compared with fixed-threshold wavelet denoising methods, the present invention can more accurately remove noise while maximally retaining the effective characteristic information of the signal, enhancing the expressiveness of fault characteristics and improving the accuracy of subsequent feature extraction and classification.

[0066] (2) To address the problem of feature loss caused by simple mapping methods in the prior art, the present invention constructs a two-dimensional Gram angle field based on the cosine value of the angle and the radius, and converts the time series signal into a two-dimensional tensor through polar coordinate transformation, making its feature expression in the spatial domain more intuitive. The present invention overcomes the feature information loss caused by directly using one-dimensional time series signals for mapping conversion in the prior art, improves the accuracy of signal conversion, makes the fault features clearer in the two-dimensional image, and enhances the feature extraction capability and fault identification accuracy of the convolutional neural network (CNN).

[0067] (3) To address the problems of gradient vanishing and overfitting that may exist in the existing CNN structure, the present invention introduces batch normalization and ReLU activation function into the CNN network, and adds a maximum pooling layer after each convolutional layer to improve the convergence speed and generalization ability of the model.

[0068] (4) The present invention installs sensors at key locations on the high-voltage power switch and uses synchronous sampling technology to accurately collect voltage signals. Displacement sensors monitor changes in the transmission rod stroke, current Hall sensors monitor current changes, and voltage sensors monitor voltage fluctuations in real time, ensuring that the operating status data of the power switch is complete and accurate. These technical measures improve the accuracy and real-time performance of fault diagnosis and ensure the stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a flow chart of wavelet threshold denoising according to an embodiment of the present invention;

[0070] Figure 2 This is a fault diagnosis flowchart of an embodiment of the present invention;

[0071] Figure 3 Schematic diagram of a convolutional neural network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0072] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0073] Example 1:

[0074] This embodiment provides a high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction, including the following steps:

[0075] Collect voltage signals from high-voltage power switches;

[0076] The voltage signal is denoised using the wavelet threshold correction noise reduction method.

[0077] The GAF method is used to perform two-dimensional conversion on the voltage signal after noise reduction to obtain two-dimensional image data of the voltage signal of the high-voltage power switch;

[0078] A pre-trained convolutional neural network is used to extract features and classify faults from two-dimensional image data, identify the operating status of high-voltage power switches, and determine whether a fault exists and its specific fault type.

[0079] Furthermore, the collecting of the voltage signal of the high-voltage power switch specifically includes:

[0080] A displacement sensor is installed on the transmission rod of the high-voltage power switch to detect the stroke change of the transmission rod and convert the stroke signal into a voltage signal;

[0081] Current Hall sensors are installed on the opening coil, closing coil and motor current inlet to collect current signals and convert them into corresponding voltage signals;

[0082] Place voltage sensors at appropriate locations in the control loop to monitor voltage changes during switching operations in real time;

[0083] Set the sampling time T and sampling frequency f s ,The synchronous sampling technology is used to sample the high voltage power switch through ,each sensor to obtain the voltage signal of the high voltage power switch.

[0084] Furthermore, the voltage signal of the high-voltage power switch includes voltage signals collected by a displacement sensor, a current Hall sensor, and a voltage sensor.

[0085] The present invention collects several signals, including displacement, current, and voltage. By simultaneously monitoring these signals, the operating status of high-voltage power switches can be comprehensively understood from multiple dimensions, improving the accuracy and reliability of fault diagnosis.

[0086] First, the displacement signal from the transmission rod is collected primarily to monitor the mechanical operation of the switch. The displacement of the transmission rod reflects the change in the switch's travel and can determine whether the switch is operating within the normal range. Mechanical anomalies (such as jamming and spring fatigue) can directly affect the switch's operation. Therefore, acquiring this signal can promptly detect mechanical faults. Compared with existing diagnostic methods, traditional diagnostic methods mostly focus on monitoring current and voltage signals, neglecting the precise monitoring of mechanical operation. However, the present invention supplements this information with displacement sensors, providing a more comprehensive reflection of the device's operating status and addressing shortcomings in the existing technology. Second, the current signal is collected, specifically by installing current Hall effect sensors on the opening coil, closing coil, and motor current input terminals to monitor changes in switch current. Current signals are key to diagnosing power switch faults and can reflect the health of the electrical components. For example, problems such as overload, short circuit, or current imbalance can be promptly detected using current signals. Compared to existing fault diagnosis methods that rely solely on current signals, the present invention uses current Hall effect sensors to monitor key components, more accurately capturing abnormal electrical conditions and improving the accuracy of fault diagnosis. Furthermore, the purpose of collecting voltage signals is to monitor voltage fluctuations during the switching process. Voltage fluctuations are usually closely related to electrical faults. By deploying voltage sensors, voltage anomalies such as voltage drops and overvoltages can be detected in real time. Compared with the existing technology, traditional methods usually rely more on local voltage monitoring, while the present invention uses voltage signals as part of real-time monitoring, which can better capture abnormal conditions of the electrical system, especially changes in voltage during the operation of the power switch, and timely discover potential electrical problems. Through synchronous sampling technology, the present invention can achieve synchronous acquisition of various signals, ensure the time consistency of all signals, and enable data from different sources to be analyzed at the same time point. This is different from the problems of incomplete synchronization of signal acquisition or inconsistent sampling time in the existing technology, and avoids diagnostic errors caused by time errors. Synchronous sampling can accurately capture various changes during the operation of the switch and make fault judgments more accurate.

[0087] Furthermore, the voltage signal is subjected to noise reduction processing using a wavelet threshold correction noise reduction method, specifically including:

[0088] Perform wavelet transform on the voltage signal and use discrete wavelet transform to decompose the signal into multiple scales to obtain the wavelet detail coefficients d at each frequency level in the voltage signal. j,k and its approximate coefficients, where d j,k Indicates the kth wavelet detail coefficient of the voltage signal at this scale in the jth layer decomposition;

[0089] According to the wavelet detail coefficient d of each frequency level in the voltage signal j,k Calculate the peak-sum ratio S of the wavelet detail coefficients of each layerj , where S j is the peak-sum ratio of the j-th layer decomposition wavelet detail coefficients;

[0090] Based on the peak-sum ratio S of the wavelet detail coefficients of each layer j Calculate the correction factor F j , and combined with ln(j+1) to modify the traditional wavelet threshold, we get the improved threshold λ j ;

[0091] For the wavelet detail coefficient d j,k Applying improved threshold λ j , and processed using an improved threshold function;

[0092] Using the processed wavelet detail coefficients The inverse wavelet transform is performed on the denoised voltage signal and its approximate coefficients to reconstruct the denoised voltage signal.

[0093] The voltage signal is subjected to wavelet transform, and the signal is decomposed into multiple scales using discrete wavelet transform to obtain the wavelet detail coefficients d of each frequency level in the voltage signal. j,k and its approximate coefficients, including:

[0094] Selecting a wavelet basis function, wherein the wavelet basis function is the Daubechies wavelet (DbN), where N is an even number between 4 and 8 to achieve a balance between time and frequency resolution;

[0095] Determine the number of wavelet decomposition layers, which is based on the sampling frequency f of the signal s And the main spectral components, take

[0096]

[0097] Among them, f c is the cutoff frequency of the main frequency component in the signal, represents the floor operation, J is the number of decomposition levels;

[0098] Perform discrete wavelet transform on the voltage signal, use the wavelet basis function selected above to perform multi-scale decomposition on the voltage signal, and obtain the wavelet detail coefficients d at each frequency level in the voltage signal. j,k and its approximate coefficients, where d j,k It represents the kth wavelet detail coefficient of the voltage signal at this scale in the jth layer decomposition.

[0099] Furthermore, the peak-sum ratio S of the wavelet detail coefficients of each layer is calculated. j , the calculation formula is:

[0100]

[0101] Among them, max(|d j |) is the maximum absolute value of the j-th layer wavelet coefficient, is the total absolute value and total wavelet coefficients of the jth layer, N j Represents the total number of wavelet detail coefficients in the jth layer, that is, the number of samples of the coefficients in this layer;

[0102] The peak-sum ratio S based on the wavelet detail coefficients of each layer j Calculate the correction factor F j , the calculation formula is:

[0103]

[0104] Among them L j is the length of the j-layer wavelet detail coefficient;

[0105] The improved threshold λ j The calculation formula is:

[0106]

[0107] Among them, λ j is the improved threshold of the j-th layer wavelet detail coefficient, σ=M mid / 0.675,M mid is the median of the lowest level wavelet coefficients, F j is the correction factor of the j-th layer wavelet detail coefficient;

[0108] The improved threshold function formula is:

[0109]

[0110] in, is the wavelet detail coefficient of the kth wavelet detail coefficient of the jth layer after applying the improved threshold, sign(d j,k ) is a sign function that represents the sign of the wavelet detail coefficient. It outputs 1 for positive values and -1 for negative values.

[0111] The purpose of calculating the peak-to-sum ratio is to quantify the ratio of signal to noise in the wavelet detail coefficients of each layer, which in turn helps us adaptively adjust the wavelet threshold of each layer, thereby achieving more effective noise reduction. Specifically, the main purpose of the calculation is to determine the relative ratio of signal information and noise information in the wavelet detail coefficients of each layer by evaluating the energy distribution of the signal at each scale. In the wavelet transform, the signal is decomposed into wavelet detail coefficients at different scales, and the detail coefficients of each layer represent the high-frequency components at that scale, that is, the rapidly changing part of the signal. Usually, the useful information in the signal is concentrated in the larger coefficients, while the noise is usually concentrated in the smaller coefficients. By calculating the peak-to-sum ratio, the energy ratio of noise and signal at each layer can be quantified, helping us determine which scales mainly contain useful signals and which scales mainly contain noise.

[0112] Based on the peak-sum ratio of the wavelet detail coefficients of each layer, the correction factor is further calculated, and the traditional wavelet threshold is corrected in combination with ln(j+1), thereby obtaining the improved threshold λ j This correction process allows the threshold to adapt not only to the characteristics of the signal itself but also to the distribution of noise. Compared with the traditional fixed threshold method, this modified threshold method has greater flexibility and accuracy, and can better meet the noise reduction needs under different signal and noise conditions.

[0113] The improved threshold function of this invention is designed to improve signal denoising while preserving the signal's effective information by dynamically adjusting the wavelet detail coefficients. This threshold function uses different processing strategies for different wavelet coefficient sizes, aiming to enhance adaptability and avoid some of the problems of traditional denoising methods.

[0114] When the wavelet detail coefficients are greater than or equal to the threshold, a soft-thresholding-like processing method is used. In this case, the coefficient amplitude is smoothed, reducing the effect of noise while preserving the coefficient sign. In this way, coefficients greater than the threshold are "corrected" without losing critical information, effectively reducing noise interference and avoiding over-smoothing. This processing helps preserve important characteristic components of the signal.

[0115] When the absolute value of a wavelet detail coefficient is less than a threshold, the improved threshold function employs a different approach. This avoids the problem of traditional hard threshold denoising methods, which simply set coefficients below the threshold to zero. Instead, it smoothes smaller coefficients, mitigating the effects of noise without completely losing the signal's low-frequency information. Thus, smaller coefficients are still processed but not directly ignored, preserving signal continuity and detail.

[0116] Traditional wavelet threshold methods usually use a fixed threshold, which is not adaptive to different signal levels and noise distributions. The present invention, by correcting the threshold, adaptively adjusts the degree of denoising according to the noise distribution characteristics at different scales, thereby significantly improving the accuracy and adaptability of denoising. The improved threshold function of the present invention does not use overly radical zeroing for smaller detail coefficients, but adjusts the amplitude of the coefficient in a smoothing manner, effectively avoiding the loss of signal details. In particular, the signal in the low-frequency part can be better retained, ensuring the continuity and integrity of the signal at all scales. By smoothing the wavelet coefficients greater than the threshold, the noise suppression effect is stronger. In the denoising process, the high-frequency noise components of the signal are effectively weakened, the characteristics of the low-frequency signal are retained, and excessive signal distortion is avoided.

[0117] Furthermore, the GAF method is used to perform two-dimensional conversion on the voltage signal after noise reduction to obtain two-dimensional image data of the voltage signal of the high-voltage power switch, specifically including:

[0118] Normalizing the noise-reduced voltage signal to obtain a normalized voltage signal, wherein the normalized voltage signal has a value range of [-1, 1];

[0119] Each element in the normalized voltage signal sequence is taken as the angle cosine value α, and is transformed into a polar coordinate system to obtain its corresponding angle cosine value α and radius;

[0120] Using the angle cosine value α and radius obtained by the polar coordinate transformation, a two-dimensional Gram angle field is constructed;

[0121] The two-dimensional Gram angle field is used as two-dimensional image data of the voltage signal after noise reduction.

[0122] Furthermore, the voltage signal after noise reduction is X={x1, x2,··x i ··, x1}, where x i represents the i-th sampling point of the voltage signal;

[0123] The voltage signal after noise reduction is normalized, and the calculation formula is:

[0124]

[0125] in, is the i-th element in the normalized voltage signal, max(X) is the maximum value in the voltage signal sequence X after noise reduction, and min(X) is the minimum value in the voltage signal sequence X after noise reduction.

[0126] Furthermore, the method of taking each element in the normalized voltage signal sequence as the angle cosine value α and transforming it into a polar coordinate system to obtain the corresponding angle cosine value α and radius specifically includes:

[0127] For each element in the normalized voltage signal sequence Calculate the corresponding angle cosine value α, the formula is:

[0128]

[0129] Among them, α i is the normalized voltage signal The corresponding angle cosine value is in the range [0,π], is the i-th element in the normalized voltage signal;

[0130] According to the sequence value t of each element in the normalized voltage signal sequence i , calculate the corresponding radius value r in the polar coordinate system i , the calculation formula is:

[0131]

[0132] Among them, r i is the voltage signal t in the polar coordinate system i The radius of the sampling point, t i is the sequence value of the i-th element in the voltage signal sequence, indicating the position of the element in the sequence, and N is the total number of sampling points in the voltage signal sequence, that is, the length of the signal.

[0133] Each element in the normalized voltage signal sequence is taken as the angle cosine value α, and transformed into the polar coordinate system to obtain the corresponding angle cosine value α and radius. This conversion step is to convert the voltage signal from a one-dimensional sequence to a two-dimensional data structure with spatial geometric meaning. In the polar coordinate system, each data point of the signal corresponds to a position (defined by the angle cosine value and the radius), so that the one-dimensional signal data can be converted into a two-dimensional signal image. The advantage of this polar coordinate transformation is that it can capture the time domain and frequency domain characteristics of the signal. Especially when the signal exhibits periodic or complex fluctuations, the polar coordinate system can better display these changes, which is convenient for subsequent visualization and pattern recognition.

[0134] Furthermore, the angle cosine value α and radius obtained by the polar coordinate transformation are used to construct a two-dimensional Gram angle field, and the formula is:

[0135]

[0136] Among them, G ADFis the two-dimensional Gram angle field, which represents the geometric structure of the voltage signal in the polar coordinate system.

[0137] The strength or amplitude (radius) of a voltage signal is closely related to the signal's energy. By multiplying the sinusoidal difference between the angles by the signal's amplitude, we can more accurately capture the actual changes in the signal at various time points. The amplitude of a signal reflects its energy. In some cases, a signal with a larger amplitude may be more significant, while a signal with a smaller amplitude may be merely noise or unimportant. Therefore, by introducing the radius as a multiplier, we can emphasize signal components with larger amplitudes and higher energy, allowing the signal's strength to be fully considered when constructing the two-dimensional Gram angle field, thereby improving the accuracy of signal processing and analysis.

[0138] Secondly, multiplying by the radius enhances the geometric representation of the signal in space. In polar coordinates, the radius determines the spatial extent of the signal, further influencing its geometry. By combining the radius with the angle difference, a two-dimensional image can be generated that reflects the signal's variations in both time and frequency domains. This allows the signal's variations to be captured not only in terms of relative angles but also in terms of amplitude, providing more clues and features for subsequent pattern recognition and fault detection.

[0139] Furthermore, multiplying by the radius can enhance signal distinguishability. Signals from complex devices like power switches often have varying amplitudes. Incorporating this amplitude information into the two-dimensional image data can effectively improve signal resolution, particularly in the presence of noise or unstable signals, enhancing the ability to identify valid signals. By factoring in amplitude, the two-dimensional Gram angle field not only reveals the timing and frequency characteristics of the signal, but also reflects the intensity distribution of the signal at different time points, thereby improving analysis accuracy.

[0140] Furthermore, the convolutional neural network training process includes:

[0141] For different types of high-voltage power switches, including circuit breakers, disconnectors, and load switches, the voltage signals of the high-voltage power switches in normal states and various fault states are collected. The fault states include spring weakness, spring jamming, inadequate opening and closing, and undervoltage operation. For each fault state, the corresponding voltage signal is collected;

[0142] Label the collected voltage signals to clarify the specific status corresponding to each signal;

[0143] Classify all marked voltage signals according to fault type and normal state, and establish a signal database of fault and normal state;

[0144] Dividing the signal data in the database into a training set and a validation set;

[0145] The voltage signals in the training set and the validation set are converted into two-dimensional images using the GAF method to obtain their corresponding two-dimensional image data;

[0146] Construct a convolutional neural network structure, including multiple convolutional layers, pooling layers, and fully connected layers. Add a maximum pooling layer after each convolutional layer, and set a batch normalization layer and ReLU activation function after the convolutional layer.

[0147] The two-dimensional image data and corresponding labels in the training set are input into the convolutional neural network for training. Specifically, the two-dimensional image data is used as the input of the convolutional neural network, supervised learning is performed according to the corresponding labels, the network weights are optimized by calculating the loss function and performing backpropagation, and in each training iteration, the model is evaluated using the validation set and the network hyperparameters are adjusted to finally form a convolutional neural network model with diagnostic capabilities.

[0148] Example 2:

[0149] The parts not mentioned in this embodiment are the same as those in embodiment 1.

[0150] This embodiment provides a high-voltage power switch fault diagnosis method based on wavelet threshold correction noise reduction and GAF-CNN, including the following steps:

[0151] Step 1: Common high-voltage power switches include circuit breakers, disconnectors, and load switches. By collecting voltage signals from high-voltage power switches in normal states and various fault states (normal operation, spring weakness, spring jamming, inadequate opening and closing, and undervoltage operation), a signal database is established based on a large number of experimental tests. The database is divided into a training set and a validation set.

[0152] Step 1.1: For high-voltage power switches (circuit breakers, disconnectors, and load switches), displacement sensors are installed on the drive rods to convert the travel changes of the drive rods into voltage travel curves. Current Hall effect sensors are installed on the incoming terminals of the closing and opening coils and the motor current to convert the current signals into voltage signals. When the switch is actuated, the acquisition system begins to operate and obtains the output voltage signal. The acquisition process is performed at each completed actuation, with an acquisition time of t, a sampling rate of fs, and a total number of signal sampling points of n. Where t is 3s, the sampling rate and the number of signal acquisition points vary depending on the switch configuration.

[0153] Step 2: Apply wavelet threshold correction to the voltage signal collected in step 1.1 to reduce noise. This removes environmental noise from the output voltage waveform, making the output signal closer to the ideal collected signal.

[0154] The wavelet threshold correction noise reduction mentioned in step 2 is a signal noise reduction method that applies wavelet theory. It mainly reduces the interference of noise while retaining important information in the signal. The noise of the output voltage signal mostly exists in the form of Gaussian white noise. The wavelet transform is used to perform multi-resolution decomposition of the signal. Since the wavelet transform has the characteristic of removing data correlation, the energy of the useful signal and the noise can be separated. The effective information in the signal is mainly concentrated in the larger wavelet coefficients, while the noise is mostly distributed in the smaller coefficients. Therefore, by setting a threshold, the coefficients below the threshold can be treated as noise and removed to achieve the purpose of filtering. The wavelet threshold noise reduction process is as follows: Figure 1 As shown:

[0155] The specific steps of wavelet threshold denoising are as follows:

[0156] (1) Multi-scale decomposition: According to the characteristics of the noisy signal, the appropriate wavelet basis and decomposition layer number are selected, and the wavelet detail coefficients d of each layer are obtained through discrete wavelet transform. j,k ;

[0157] (2) Threshold denoising: By determining the threshold λ and the threshold function, the wavelet detail coefficient d j,k Processing is performed to obtain the processed wavelet detail coefficients of each layer

[0158] (3) Wavelet reconstruction: Based on the obtained wavelet detail coefficients The power quality disturbance signal is reconstructed by the sum of the approximate coefficients to obtain the denoised disturbance signal x′(t).

[0159] Among these steps, the selection of threshold and threshold function is the key to wavelet threshold denoising, directly affecting the quality of the reconstructed signal. If the threshold is too large, the useful signal will be filtered out as noise; if the threshold is too small, the noise will not be filtered out thoroughly. The traditional hard and soft threshold functions are as follows:

[0160] Definition of hard threshold function:

[0161]

[0162] Definition of soft threshold function:

[0163]

[0164] Where sign(d j,k ) is the sign function, N is the signal length;

[0165] However, the traditional wavelet threshold denoising method sets the wavelet coefficient threshold of each layer to be constant, but the noise is not the same in each layer of wavelet coefficients. Therefore, the fixed threshold has poor adaptability and the denoising effect is not ideal. In order to solve the above problems, an improved wavelet threshold denoising algorithm is proposed, which can adaptively modify the threshold according to the distribution of noise, and its threshold function can realize a variety of soft and hard features through variable parameters, making it more suitable for voltage signal processing. And the distribution of noise is random, so it is necessary to take into account the wavelet detail coefficient of the noise, where S j is the peak-to-peak ratio of the wavelet detail coefficients of the jth layer. In the multi-scale decomposition of the wavelet, the useful information of the signal is mainly concentrated on the larger wavelet detail coefficients, while the noise components are scattered in the wavelet detail coefficients of each layer. j When the value is large, it means that there is a large coefficient at this scale, indicating that the layer contains more useful information; j When the value is small, it means that there are smaller coefficients at this scale, indicating that the layer contains more noise. The formula for the peak-to-sum ratio of the wavelet detail coefficients of the jth layer is as follows:

[0166]

[0167] Among them, max(|d j |) is the maximum absolute value of the j-th layer wavelet coefficient, is the total absolute value and total wavelet coefficients of the jth layer, N j Represents the total number of wavelet detail coefficients in the jth layer, that is, the number of samples of the coefficients in this layer;

[0168] The peak-sum ratio S based on the wavelet detail coefficients of each layer j Calculate the correction factor F j , the calculation formula is:

[0169]

[0170] Among them L j is the length of the j-layer wavelet detail coefficient;

[0171] The improved threshold λ j The calculation formula is:

[0172]

[0173] Among them, λ j is the improved threshold of the j-th layer wavelet detail coefficient, σ=M mid / 0.675,M mid is the median of the lowest level wavelet coefficients, F j is the correction factor of the j-th layer wavelet detail coefficient;

[0174] The improved threshold function formula is:

[0175]

[0176] in, is the wavelet detail coefficient of the kth wavelet detail coefficient of the jth layer after applying the improved threshold, sign(d j,k ) is a sign function that represents the sign of the wavelet detail coefficient. It outputs 1 for positive values and -1 for negative values.

[0177] Step 3: Perform a GAF two-dimensional transformation on the voltage signal after noise reduction. The voltage signal output when a high-voltage power switch fails is a one-dimensional time series signal. Although it contains the original characteristics of the signal, its fault characteristics are not very prominent. The decomposed signal contains time domain and frequency domain characteristics, but processing all components at the same time will increase the amount of calculation. Many deep learning models, including the model used in this article, are pre-trained on large-scale image data sets. By converting one-dimensional data into two-dimensional images, these pre-trained models can be well used for training. Training with pre-trained models can also make up for the disadvantage of insufficient data volume. Therefore, the present invention introduces the GAF method as a conversion method.

[0178] Step 3.1: The specific change is to divide the voltage signal obtained by the acquisition system into X = {x1, x2, ··x i ··, x1}, normalized to the interval [-1,1]. The sequence consists of n sampling points and the value x of the vibration signal corresponding to each sampling point. The calculation is as follows:

[0179]

[0180] x i is the i-th element in the one-dimensional time series X; max(X) means selecting the element with the largest value in X; min(X) operation means selecting the element with the smallest value in X, n is the total number of samples in the one-dimensional time series X, is the normalized one-dimensional time series element.

[0181] The values in the normalized data sequence are used as the cosine value of the angle α, the one-dimensional time series is transformed into the polar coordinate system, the voltage signal value and the corresponding sampling point sequence value are associated with the angle and radius, and the inverse proportional function arccos is used for calculation.

[0182]

[0183] Where t i is the order value of the sampling point, r i for The corresponding polar coordinate radius, N is the constant factor of the generated space in the polar coordinate system, which is numerically equal to the total number of sampling points.

[0184] This method preserves the time information and obtains the Gram angle field G through trigonometric transformation. ADF :

[0185]

[0186] Among them, G ADF is the two-dimensional Gram angle field, which represents the geometric structure of the voltage signal in the polar coordinate system.

[0187] Step 4: After the data processing of steps 2 and 3, a two-dimensional tensor with significant eigenvalues is obtained, which is then brought into the convolutional neural network for eigenvalue extraction. In order to extract key feature information and reduce the influence of noise, a maximum pooling layer is added after each convolution layer; secondly, in order to reduce overfitting, accelerate model training, and prevent gradient explosion and gradient disappearance problems during back propagation, a batch normalization layer (BN) and ReLU activation function are set after the convolution layer. The convolutional neural network is trained to form a diagnostic model. The convolutional neural network of this embodiment is as follows: Figure 3 As shown in the figure, it includes convolutional layer 1, batch normalization processing, ReLU activation function, convolutional layer 2, batch normalization processing, ReLU activation function, maximum pooling layer, fully connected layer, and output layer softmax.

[0188] Step 5: For the power switch in an unknown operating state, collect its voltage signal and use the above method to obtain its two-dimensional time domain resolution diagram. Compare it with the normal state of the power switch and the conditions under different faults to determine whether the power switch has a fault and what kind of fault has occurred.

[0189] The beneficial effects of the present invention are:

[0190] Wavelet threshold correction denoising: This algorithm adaptively adjusts the threshold based on the noise distribution. Its threshold function, with variable parameters, can achieve a variety of soft and hard characteristics, making it more suitable for voltage signal processing. Compared to traditional wavelet denoising, this algorithm calculates the peak-to-sum ratio of the wavelet detail coefficients at each scale, adaptively determines the number of wavelet decomposition layers, and incorporates a correction factor to improve the universal threshold. This allows the voltage vibration signal to approach the desired ideal signal.

[0191] The characteristic signal is more significant: In order to fully utilize the advantages of CNN in two-dimensional field learning, the present invention performs Gram Angular Field (GAF) transformation on the denoised signal to obtain a two-dimensional data set, which significantly improves the time domain feature resolution of the voltage vibration signal and overcomes the limitations of traditional algorithms in time domain resolution.

[0192] Improve fault detection accuracy: In data processing, wavelet threshold optimization noise reduction and Gram angle field change are used to convert the one-dimensional signal into a two-dimensional tensor form while retaining the original signal, and G ADF The data is divided into training set and validation set, and input into CNN for training to obtain a training model. Based on the better data processing method, the final fault detection accuracy is higher. The flow chart of the high-voltage power switch fault diagnosis method based on wavelet threshold correction noise reduction and GAF-CNN is as follows: Figure 2 shown.

[0193] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0194] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction, characterized in that: The following steps are involved: Collect voltage signals from high-voltage power switches; The voltage signal is denoised using the wavelet threshold correction noise reduction method; The GAF method is used to perform two-dimensional conversion on the voltage signal after noise reduction to obtain two-dimensional image data of the voltage signal of the high-voltage power switch; A pre-trained convolutional neural network is used to extract features and classify faults from two-dimensional image data, identify the operating status of high-voltage power switches, and determine whether a fault exists and its specific fault type.

2. A high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction according to claim 1, characterized in that: The collecting of the voltage signal of the high-voltage power switch specifically includes: A displacement sensor is installed on the transmission rod of the high-voltage power switch to detect the stroke change of the transmission rod and convert the stroke signal into a voltage signal; Current Hall sensors are installed on the opening coil, closing coil and motor current inlet to collect current signals and convert them into corresponding voltage signals; Place voltage sensors at appropriate locations in the control loop to monitor voltage changes during switching operations in real time; Set the sampling time T and sampling frequency f s ,The synchronous sampling technology is used to sample the high voltage power switch through ,each sensor to obtain the voltage signal of the high voltage power switch.

3. The high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction according to claim 2 is characterized in that: The voltage signal of the high-voltage power switch includes voltage signals collected by a displacement sensor, a current Hall sensor, and a voltage sensor.

4. The high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction according to claim 1 is characterized in that: The noise reduction process of the voltage signal using the wavelet threshold correction noise reduction method specifically includes: Perform wavelet transform on the voltage signal and use discrete wavelet transform to decompose the signal into multiple scales to obtain the wavelet detail coefficients d at each frequency level in the voltage signal. j,k and its approximate coefficients, where d j,k Indicates the kth wavelet detail coefficient of the voltage signal at this scale in the jth layer decomposition; According to the wavelet detail coefficient d of each frequency level in the voltage signal j,k Calculate the peak-sum ratio S of the wavelet detail coefficients of each layer j , where S j is the peak-sum ratio of the j-th layer decomposition wavelet detail coefficients; Based on the peak-sum ratio S of the wavelet detail coefficients of each layer j Calculate the correction factor F j , and combined with ln(j+1) to modify the traditional wavelet threshold, we get the improved threshold λ j ; For the wavelet detail coefficient d j,k Applying improved threshold λ j , and processed using an improved threshold function; Using the processed wavelet detail coefficients The inverse wavelet transform is performed on the denoised voltage signal and its approximate coefficients to reconstruct the denoised voltage signal.

5. The high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction according to claim 4 is characterized in that: The peak sum ratio S of the wavelet detail coefficients of each layer is calculated j , the calculation formula is: Among them, max(|d j |) is the maximum absolute value of the j-th layer wavelet coefficient, is the total absolute value and total wavelet coefficients of the jth layer, N j Represents the total number of wavelet detail coefficients in the jth layer, that is, the number of samples of the coefficients in this layer; The peak-sum ratio S based on the wavelet detail coefficients of each layer j Calculate the correction factor F j , the calculation formula is: Among them L j is the length of the j-layer wavelet detail coefficient; The improved threshold λ j The calculation formula is: Among them, λ j is the improved threshold of the j-th layer wavelet detail coefficient, σ=M mid / 0.675,M mid is the median of the lowest level wavelet coefficients, F j is the correction factor of the j-th layer wavelet detail coefficient; The improved threshold function formula is: in, is the wavelet detail coefficient of the kth wavelet detail coefficient of the jth layer after applying the improved threshold, sign(d j,k ) is a sign function that represents the sign of the wavelet detail coefficient, outputting 1 for positive values and -1 for negative values.

6. The high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction according to claim 1 is characterized in that: The GAF method is used to perform two-dimensional conversion on the voltage signal after noise reduction to obtain two-dimensional image data of the voltage signal of the high-voltage power switch, specifically including: Normalizing the noise-reduced voltage signal to obtain a normalized voltage signal, wherein the normalized voltage signal has a value range of [-1, 1]; Each element in the normalized voltage signal sequence is taken as the angle cosine value α, and is transformed into a polar coordinate system to obtain its corresponding angle cosine value α and radius; Using the angle cosine value α and radius obtained by the polar coordinate transformation, a two-dimensional Gram angle field is constructed; The two-dimensional Gram angle field is used as two-dimensional image data of the voltage signal after noise reduction.

7. The high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction according to claim 6 is characterized in that: The voltage signal after noise reduction is X={x1, x2, ··x i ··, x1}, where x i represents the i-th sampling point of the voltage signal; The voltage signal after noise reduction is normalized, and the calculation formula is: in, is the i-th element in the normalized voltage signal, max(X) is the maximum value in the voltage signal sequence X after noise reduction, and min(X) is the minimum value in the voltage signal sequence X after noise reduction.

8. The high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction according to claim 6 is characterized in that: The method of taking each element in the normalized voltage signal sequence as the angle cosine value α and transforming it into a polar coordinate system to obtain the corresponding angle cosine value α and radius specifically includes: For each element in the normalized voltage signal sequence Calculate the corresponding angle cosine value α, the formula is: Among them, α i is the normalized voltage signal The corresponding angle cosine value is in the range [0,π], is the i-th element in the normalized voltage signal; According to the sequence value t of each element in the normalized voltage signal sequence i , calculate the corresponding radius value r in the polar coordinate system i , the calculation formula is: Among them, r i is the voltage signal t in the polar coordinate system i The radius of the sampling point, t i is the sequence value of the i-th element in the voltage signal sequence, indicating the position of the element in the sequence, and N is the total number of sampling points in the voltage signal sequence, that is, the length of the signal.

9. The high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction according to claim 6 is characterized in that: The angle cosine value α and radius obtained by the polar coordinate transformation are used to construct a two-dimensional Gram angle field, and the formula is: Among them, G ADF is the two-dimensional Gram angle field, which represents the geometric structure of the voltage signal in the polar coordinate system.

10. The high-voltage power switch fault diagnosis method based on wavelet threshold correction and noise reduction according to claim 1, characterized in that: The convolutional neural network training process includes: For different types of high-voltage power switches, including circuit breakers, disconnectors, and load switches, the voltage signals of the high-voltage power switches in normal states and various fault states are collected. The fault states include spring weakness, spring jamming, inadequate opening and closing, and undervoltage operation. For each fault state, the corresponding voltage signal is collected; Label the collected voltage signals to clarify the specific status corresponding to each signal; Classify all marked voltage signals according to fault type and normal state, and establish a signal database of fault and normal state; Dividing the signal data in the database into a training set and a validation set; The voltage signals in the training set and the validation set are converted into two-dimensional images using the GAF method to obtain their corresponding two-dimensional image data; Construct a convolutional neural network structure, including multiple convolutional layers, pooling layers, and fully connected layers. Add a maximum pooling layer after each convolutional layer, and set a batch normalization layer and ReLU activation function after the convolutional layer. The two-dimensional image data and corresponding labels in the training set are input into the convolutional neural network for training. Specifically, the two-dimensional image data is used as the input of the convolutional neural network, supervised learning is performed according to the corresponding labels, the network weights are optimized by calculating the loss function and performing backpropagation, and in each training iteration, the model is evaluated using the validation set and the network hyperparameters are adjusted to finally form a convolutional neural network model with diagnostic capabilities.

Images