A method for identifying the state of a switching power supply based on the integration of fault data

Abstract

The present invention discloses a method for identifying the state of a switching power supply based on fault data integration, which relates to the technical field of data processing and identification, and includes: comprehensively collecting various parameter data of the switching power supply during operation; cleaning the collected original data to remove noise, outliers and missing values, and performing standardization processing on the data; extracting time-frequency domain features and wavelet transform features reflecting the working state; extracting the timing features of the switching power supply by analyzing the information accumulation and transformation of the input data at different time steps; extracting the spatial features of the switching power supply through a convolutional neural network; screening the features with a high degree of influence on the identification of the switching power supply state. By converting one-dimensional timing data into two-dimensional images by using Gram angle and field coding, extracting the spatial features of the switching power supply through a convolutional neural network, and adaptively fusing various key features, the real-time monitoring and accurate identification of the working state of the switching power supply are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing and recognition, and specifically relates to a method for identifying the state of a switching power supply based on integrated fault data. Background Art

[0002] In modern power systems, as a key power conversion device, the stable operation of a switching power supply is crucial for the reliability and safety of the entire system. However, due to factors such as complex and variable working environments, equipment aging, and load fluctuations, various faults will inevitably occur during the operation of the switching power supply. These faults not only affect the performance of the power supply itself but may also cause a chain reaction in the entire power system, resulting in greater economic losses and safety risks.

[0003] Traditional methods for monitoring the state of switching power supplies mainly rely on manual inspections and regular maintenance. This method is not only inefficient but also difficult to capture the moment of fault occurrence and the dynamic change process. With the rapid development of sensor technology, Internet of Things technology, big data technology, and artificial intelligence technology, data-driven fault identification and state monitoring methods have gradually become a research hotspot. Summary of the Invention

[0004] To solve the above technical problems, a method for identifying the state of a switching power supply based on integrated fault data is provided, and the present technical solution solves the problems raised in the above background art.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] A method for identifying the state of a switching power supply based on integrated fault data, comprising:

[0007] Using at least one sensor and monitoring device to comprehensively collect various parameter data of the switching power supply during operation, obtain the changes in the operating data of the switching power supply before and after a fault occurs, and construct a fault data set, where the parameter data includes voltage, current, temperature, and frequency;

[0008] Clean the collected original data, remove noise, outliers, and missing values, and perform standardization processing on the data;

[0009] Based on the working principle and fault characteristics of the switching power supply, extract time-frequency domain features and wavelet transform features reflecting the working state;

[0010] Using a long short-term memory network, by analyzing the information accumulation and transformation of the input data at at least one time step, extract the temporal characteristics of the switching power supply;

[0011] By converting the time series into the polar coordinate system, calculating the Gram matrix using trigonometric functions, mapping one-dimensional data into a two-dimensional image, and generating the corresponding two-dimensional image based on the result of field coding, the spatial features of the switching power supply are extracted through a convolutional neural network;

[0012] Feature selection methods based on mutual information and tree models are adopted to screen the features that have a high degree of influence on the identification of the switching power supply state, and output them as key features. The key features are adaptively fused to construct a multi-dimensional feature fusion state recognition model;

[0013] The trained model is deployed into the actual system, the monitoring data of the switching power supply is received in real time, and the model is used to predict the input data. The prediction result is output in the form of a classification label to judge the current working state of the switching power supply;

[0014] The new fault data is fed back into the system to update and optimize the model and continuously learn new fault patterns and data features.

[0015] Preferably, the process of converting the time series into the polar coordinate system, calculating the Gram matrix using trigonometric functions, mapping one-dimensional data into a two-dimensional image, and generating the corresponding two-dimensional image based on the result of field coding, and extracting the spatial features of the switching power supply through a convolutional neural network specifically includes:

[0016] The normalized time series data is converted into the polar coordinate system. In this conversion process, the time series data is mapped onto a unit circle, and each data point corresponds to a point on the unit circle. Its angle is calculated through trigonometric functions;

[0017] The cosine function is used to calculate the sum and difference of the angles of the data points between each time stamp;

[0018] For any two data points in the time series, their cosine similarity is calculated. By calculating the cosine similarity between each data point, the Gram matrix is constructed to reflect the correlation between each time stamp in the time series;

[0019] Each element of the calculated Gram matrix is mapped to the range of gray values according to the value range of 0 - 255, and used as the pixel value of the image;

[0020] The structure of the convolutional neural network is designed. The input layer receives the preprocessed data. The convolutional layer contains at least one convolutional kernel to extract the local features of the data. The number, size and stride parameters of the convolutional kernel are adjusted based on the task and data. The pooling layer is after the convolutional layer to reduce the dimension of the data and reduce the amount of calculation. The fully connected layer maps the features extracted by the convolutional layer to the output space and is used for classification or regression tasks. The output layer selects the output layer function based on the task requirements;

[0021] In the convolutional layer, the convolutional kernel slides over the input data, performs element-wise multiplication and summation operations, and extracts local features;

[0022] Each convolutional kernel generates a new feature map, reflecting specific patterns or features in the input data;

[0023] For a switching power supply, the spatial features include the morphological features of waveform data and the spatial distribution features in image data. The morphological features of the waveform data include peak values, valley values, and waveform widths, and the spatial distribution features in the image data include hot spots and abnormal regions;

[0024] Input the two-dimensional image converted from the time series into a convolutional neural network. Through convolutional operations, the convolutional neural network automatically learns and extracts the spatial features in the image.

[0025] Preferably, the feature selection method based on mutual information and tree model is adopted to screen the features with high influence on the state recognition of the switching power supply and output them as key features. The specific steps of adaptively fusing the key features to construct a multi-dimensional feature fusion state recognition model include:

[0026] Use the mutual information formula to calculate the mutual information value between each feature and the target variable. The mutual information value is positively correlated with the contribution of the feature to the target variable;

[0027] Sort the features based on the mutual information value, and select the features with high mutual information values as key features. The key features are used for the subsequent multi-dimensional feature fusion state recognition model;

[0028] Construct a tree structure to split nodes, and evaluate the importance of features based on the loss function value reduced when the features split the nodes;

[0029] Use the decision tree model to model the switching power supply state data. During the training process of the tree model, record the loss reduction amount of each feature when the nodes split, and accumulate to obtain the importance score of the feature;

[0030] Sort the features based on the feature importance score, and select the features with high importance scores as key features;

[0031] After screening out the key features, adaptively fuse the key features to construct a multi-dimensional feature fusion state recognition model, and weight each feature based on its importance or contribution degree;

[0032] Normalize the screened key features to ensure that they have the same scale during the fusion process;

[0033] Based on the importance score or mutual information value of the features, a weight is assigned to each feature. The magnitude of the weight reflects the contribution degree of the feature to the target variable. During the fusion process, each feature is multiplied by its corresponding weight to achieve adaptive weighting;

[0034] The weighted features are concatenated to form a new multi-dimensional feature vector, which is used as the input of the multi-dimensional feature fusion state recognition model.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] The key features are adaptively fused to construct a multi-dimensional feature fusion state recognition model. This method can significantly improve the recognition accuracy and robustness of the model. By real-time monitoring the working state of the switching power supply, potential faults can be detected and processed in a timely manner, which can significantly improve the reliability and stability of the equipment, reduce the fault downtime and maintenance costs. At the same time, this method can also provide a scientific basis for the preventive maintenance and health management of the equipment, extend the service life of the equipment, and reduce the overall operation and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flowchart of the switching power supply state recognition method based on fault data integration of the present invention;

[0038] Figure 2 It is a flowchart of the method for cleaning the collected original data of the present invention;

[0039] Figure 3 It is a flowchart of the method for extracting time-frequency domain features and wavelet transform features reflecting the working state of the present invention;

[0040] Figure 4 It is a flowchart of the method for analyzing the information accumulation and transformation of the input data at at least one time step and extracting the timing features of the switching power supply of the present invention;

[0041] Figure 5 It is a flowchart of the method for extracting the spatial features of the switching power supply by a convolutional neural network of the present invention;

[0042] Figure 6 It is a flowchart of the method for adaptively fusing the key features and constructing a multi-dimensional feature fusion state recognition model of the present invention;

[0043] Figure 7 It is a flowchart of the method for judging the current working state of the switching power supply of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0044] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments in the following description are only examples, and other obvious variations can be conceived by those skilled in the art.

[0045] Referring to Figure 1 As shown, a method for identifying the state of a switching power supply based on integrated fault data includes:

[0046] Using at least one sensor and monitoring device, comprehensively collect various parameter data of the switching power supply during operation, obtain the changes in the operating data of the switching power supply before and after a fault occurs, and construct a fault data set. The parameter data includes voltage, current, temperature, and frequency;

[0047] Clean the collected raw data, remove noise, outliers, and missing values, and perform standardization processing on the data;

[0048] Based on the working principle and fault characteristics of the switching power supply, extract time-frequency domain features and wavelet transform features reflecting the working state;

[0049] Using a long short-term memory network, by analyzing the information accumulation and transformation of the input data at at least one time step, extract the timing features of the switching power supply;

[0050] By converting the time series into a polar coordinate system, calculating the Gram matrix using trigonometric functions, mapping one-dimensional data into a two-dimensional image, and based on the result of field coding, generating the corresponding two-dimensional image, and extracting the spatial features of the switching power supply through a convolutional neural network;

[0051] Adopt feature selection methods based on mutual information and tree models to screen features with a high degree of influence on the state identification of the switching power supply, and output them as key features. Perform adaptive fusion on each key feature to construct a multi-dimensional feature fusion state identification model;

[0052] Deploy the trained model to the actual system, receive the monitoring data of the switching power supply in real time, and use the model to predict the input data. The prediction result is output in the form of a classification label to judge the current working state of the switching power supply;

[0053] Feed the new fault data back into the system to update and optimize the model and continuously learn new fault patterns and data features.

[0054] Referring to Figure 2 As shown, cleaning the collected raw data, removing noise, outliers, and missing values, and performing standardization processing on the data specifically includes:

[0055] Based on the reasonable value range and mutual relationship of variables, check the normativity and logic of the data, and use the sample mean of each variable to replace invalid values and missing values;

[0056] Perform relevance verification on multi-source data and select characteristic attributes by verifying the relevance between data;

[0057] Adopt a filtering method to remove noise;

[0058] For random noise, use mean filtering to calculate the neighborhood average for smoothing;

[0059] For salt-and-pepper noise, use median filtering to select the median of the window data as the output;

[0060] For Gaussian noise, use Gaussian filtering to weight and smooth the signal;

[0061] For high-frequency noise, use wavelet denoising to process the noise part through signal decomposition;

[0062] Perform standardization processing on the data. Take the base-10 logarithm of each observation value of each variable and divide it by the base-10 logarithm of the maximum value of that variable.

[0063] For time series data, it may be necessary to consider the mean or other statistics within the time window to replace missing values to maintain the temporal characteristics of the data, analyze the relevance between data from different sources, such as comparing the same physical quantity measured by different sensors, and evaluate the relevance between data through correlation analysis such as Pearson correlation coefficient and Spearman rank correlation coefficient. Select highly correlated and useful characteristic attributes for state recognition for subsequent analysis.

[0064] Refer to Figure 3 As shown, based on the working principle and fault characteristics of the switching power supply, the time-frequency domain characteristics and wavelet transform characteristics reflecting the working state are extracted, specifically including:

[0065] The switching power supply controls the energy transfer of the power supply through the rapid on-off operation of electronic switching elements. The rapid on-off process generates changes in voltage and current in at least one energy storage element to achieve voltage conversion and stabilization;

[0066] Observe the voltage and current waveforms of the switching power supply, measure its zero-crossing rate and the number of slope changes to reflect the working stability of the power supply and the load change situation;

[0067] Calculate the statistical characteristics of the mean, maximum value, minimum value, peak value, valley value, and variance of voltage and current to evaluate the output performance and stability of the power supply;

[0068] Use Fourier transform to convert the voltage and current signals from the time domain to the frequency domain, analyze the amplitude and phase of at least one frequency component, and identify the harmonic components and noise interference in the power supply;

[0069] Analyze the power spectral density of the power supply signal, display the distribution of the signal power as a function of frequency, and identify abnormal frequency components in the power supply;

[0070] Based on the characteristics of the switched-mode power supply signal, select the Haar wavelet function as the wavelet basis function for wavelet transform, perform continuous wavelet transform on the voltage and current signals of the switched-mode power supply, and obtain wavelet coefficients at at least one scale.

[0071] The Haar wavelet function is: ,

[0072] where, is the input data, is the output data.

[0073] Referring to Figure 4 shown, using the long short-term memory network, by analyzing the information accumulation and transformation of the input data at at least one time step, the specific timing features of the switched-mode power supply are extracted, including:

[0074] Obtain the operation data of the switched-mode power supply at at least one time step. The data exists in the form of a time series, and each time step has a corresponding data point;

[0075] Based on the characteristics of the data and the task requirements, design the structure of the long short-term memory network, determine the number of long short-term memory network layers, the number of neurons in each layer, and the dimension of the input data;

[0076] Preprocess the operation data, convert the data into three-dimensional data of the number of samples, time step length, and number of features, and input the preprocessed data into the long short-term memory network;

[0077] Perform hidden layer design. In the long short-term memory network, the hidden layer extracts features in the time series, and through at least one long short-term memory network hidden layer, deeper features are gradually extracted;

[0078] Based on the task requirements, set the structure and activation function of the output layer. In the task of extracting timing features, the output layer is a fully connected layer that maps the features extracted by the long short-term memory network to the target space;

[0079] Select the mean squared error function as the loss function for constructing the long short-term memory network;

[0080] Input the preprocessed data into the long short-term memory network, perform forward propagation and backward propagation, and update the weights of the model in real time through the optimizer;

[0081] Judge whether the preset number of training epochs is reached or whether the loss function converges. If so, end the training. If not, do not make an output;

[0082] After training is completed, extract the features of the time series from the hidden layer of the long short-term memory network. The features of the time series include the information accumulation and transformation of the switching power supply at at least one time step.

[0083] The mean squared error function is: ,

[0084] In the formula, is the mean squared error function, is the total number of samples, is the true value of the i-th sample, is the predicted value of the i-th sample.

[0085] Referring to Figure 5 As shown, by converting the time series into a polar coordinate system, calculating the Gram matrix using trigonometric functions, mapping one-dimensional data into a two-dimensional image, and generating a corresponding two-dimensional image based on the result of field coding, extracting the spatial features of the switching power supply through a convolutional neural network specifically includes:

[0086] Convert the normalized time series data into a polar coordinate system. In this conversion process, map the time series data onto a unit circle, and each data point corresponds to a point on the unit circle. Calculate its angle through trigonometric functions;

[0087] Use the cosine function to calculate the sum and difference of the angles between data points at each time stamp;

[0088] For any two data points in the time series, calculate the cosine similarity between them. By calculating the cosine similarity between each data point, construct a Gram matrix to reflect the correlation between each time stamp in the time series;

[0089] Map each element of the calculated Gram matrix to the range of gray values according to the value range of 0-255, and use it as the pixel value of the image;

[0090] Design the structure of the convolutional neural network. The input layer receives the preprocessed data. The convolutional layer contains at least one convolutional kernel to extract the local features of the data. Adjust the number, size, and stride parameters of the convolutional kernel based on the task and data. The pooling layer is after the convolutional layer to reduce the dimension of the data and reduce the amount of calculation. The fully connected layer maps the features extracted by the convolutional layer to the output space and is used for classification or regression tasks. The output layer selects the output layer function based on the task requirements;

[0091] In the convolutional layer, the convolutional kernel slides on the input data to perform element-wise multiplication and summation operations and extract local features;

[0092] Each convolutional kernel generates a new feature map to reflect specific patterns or features in the input data;

[0093] For a switching power supply, the spatial features include the morphological features of waveform data and the spatial distribution features in image data. The morphological features of the waveform data include peak value, valley value, and waveform width. The spatial distribution features in the image data include hot spots and abnormal regions;

[0094] Input the two-dimensional image converted from the time series into a convolutional neural network. Through convolutional operations, the convolutional neural network automatically learns and extracts the spatial features in the image.

[0095] Map the normalized time series data onto a unit circle, where each data point in the time series corresponds to a point on the unit circle. Use trigonometric functions such as the arctangent function to calculate the angle corresponding to each data point, so that the time series data is converted into an angle series.

[0096] Refer to Figure 6 As shown, adopt feature selection methods based on mutual information and tree models to screen out features with a high degree of influence on the identification of the switching power supply state, and output them as key features. Adaptive fusion of each key feature is carried out to construct a multi-dimensional feature fusion state recognition model, which specifically includes:

[0097] Use the mutual information formula to calculate the mutual information value between each feature and the target variable. The mutual information value is positively correlated with the contribution of this feature to the target variable;

[0098] Rank the features based on the mutual information value, and select the features with high mutual information values as key features. The key features are used for the subsequent multi-dimensional feature fusion state recognition model;

[0099] Construct a tree structure to split nodes, and evaluate the importance of features based on the loss function value reduced when the features split the nodes;

[0100] Use a decision tree model to model the switching power supply state data. During the training process of the tree model, record the loss reduction amount of each feature when the nodes split, and accumulate to obtain the importance score of the feature;

[0101] Rank the features based on the feature importance score, and select the features with high importance scores as key features;

[0102] After screening out the key features, adaptively fuse the key features to construct a multi-dimensional feature fusion state recognition model, and each feature is weighted based on its importance or contribution degree;

[0103] Perform standardization processing on the screened key features to ensure that they have the same scale during the fusion process;

[0104] Based on the importance scores or mutual information values of the features, a weight is assigned to each feature. The magnitude of the weight reflects the contribution degree of the feature to the target variable. During the fusion process, each feature is multiplied by its corresponding weight to achieve adaptive weighting;

[0105] The weighted features are concatenated to form a new multi-dimensional feature vector, which is used as the input of the multi-dimensional feature fusion state recognition model.

[0106] The mutual information formula is as follows: ,

[0107] In the formula, is the mutual information value between each feature and the target variable, is the feature, is the target variable, is the joint probability distribution of X and Y, are the marginal probability distributions of X and Y respectively.

[0108] Referring to Figure 7 as shown, the trained model is deployed into the actual system to receive the monitoring data of the switching power supply in real time, and the model is used to predict the input data. The prediction results are output in the form of classification labels. Judging the current working state of the switching power supply specifically includes:

[0109] Classify the working state based on the working mode. When in the continuous conduction mode, the inductor current is continuous, suitable for medium or heavy load conditions. When in the discontinuous conduction mode, the inductor current is discontinuous, suitable for light load conditions. When in the hybrid mode, the continuous conduction mode and the discontinuous conduction mode are automatically switched based on the load conditions to achieve efficiency optimization;

[0110] Classify the working state based on the output state. When in the normal working state, the output voltage and current are stable, with fluctuations within the allowable range, high efficiency, and low loss. The abnormal working states include overvoltage protection state, overcurrent protection state, short-circuit protection state, and overheat protection state. When in the standby state, the output voltage and current are low, maintaining a fixed standby power consumption;

[0111] Classify the working state based on the fault type. Hardware faults include component damage and circuit board faults. Software faults include control program anomalies and parameter setting errors;

[0112] Classify the working state based on the load change. The light load state means the load is light and the output power is low. The full load state means the load reaches the rated value and the output power is high. The heavy load state means the load exceeds the rated value and is within the tolerable range of the power supply. The overload state means the load exceeds the rated value and will cause damage to the power supply.

[0113] Based on the output classification tags and combined with the preset state judgment logic, determine the current specific working state of the switching power supply. The state judgment logic can be implemented based on methods such as threshold judgment and rule matching. According to the judgment result, corresponding response measures are taken, such as triggering alarms, cutting off the power supply and other protection measures in abnormal working states; reducing power consumption in standby states; adjusting loads or upgrading power supplies in heavy load or overload states, etc.

[0114] Furthermore, this solution also proposes a computer-readable storage medium, on which a computer-readable program is stored. When the computer-readable program is called, it executes the above-mentioned switching power supply state recognition method based on fault data integration.

[0115] It can be understood that the storage medium can be a magnetic medium, such as a floppy disk, a hard disk, a magnetic tape; an optical medium, such as a DVD; or a semiconductor medium, such as a solid-state disk (SSD), etc.

[0116] In summary, the advantages of the present invention are as follows: Adaptive integration of various key features to construct a multi-dimensional feature fusion state recognition model. This method can significantly improve the recognition accuracy and robustness of the model. By real-time monitoring the working state of the switching power supply, potential faults can be discovered and processed in a timely manner, which can significantly improve the reliability and stability of the device, reduce the fault downtime and maintenance costs. At the same time, this method can also provide a scientific basis for the preventive maintenance and health management of the device, extend the service life of the device, and reduce the overall operation and maintenance costs.

[0117] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification is only the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying the state of a switching power supply based on fault data integration, characterized in that: include: Using at least one sensor and monitoring equipment, comprehensively collect various parameter data of the switching power supply during operation, obtain the change of the switching power supply operation data before and after the fault occurs, and construct a fault data set, wherein the parameter data includes voltage, current, temperature and frequency; Clean the collected raw data, remove noise, outliers and missing values, and standardize the data; Based on the working principle and fault characteristics of the switching power supply, the time-frequency domain features and wavelet transform features reflecting the working status are extracted; By using the long short-term memory network, the timing characteristics of the switching power supply are extracted by analyzing the information accumulation and transformation of the input data in at least one time step; By converting the time series into a polar coordinate system and using trigonometric functions to calculate the Gram matrix, the one-dimensional data is mapped into a two-dimensional image. Based on the results of field coding, the corresponding two-dimensional image is generated, and the spatial features of the switching power supply are extracted through a convolutional neural network. The feature selection method based on mutual information and tree model is used to screen the features with high influence on the state recognition of the switch power supply, and output them as key features. The key features are adaptively fused to build a multi-dimensional feature fusion state recognition model. Deploy the trained model to the actual system, receive the monitoring data of the switching power supply in real time, and use the model to predict the input data. The prediction results are output in the form of classification labels to determine the current working status of the switching power supply. Feed new fault data back into the system, update the optimization model and continuously learn new fault modes and data features; Based on the working principle and fault characteristics of the switching power supply, the time-frequency domain features and wavelet transform features reflecting the working status are extracted, including: The switching power supply controls the energy transfer of the power supply through the rapid on-off operation of the electronic switching element. The rapid on-off process produces changes in voltage and current in at least one energy storage element, achieving voltage conversion and stabilization; Observe the voltage and current waveforms of the switching power supply, measure its zero-crossing rate and slope change times, and reflect the working stability and load changes of the power supply; Calculate the statistical characteristics of the mean, maximum, minimum, peak, valley and variance of voltage and current to evaluate the output performance and stability of the power supply; Convert voltage and current signals from the time domain to the frequency domain using Fourier transform, analyze the amplitude and phase of at least one frequency component, and identify harmonic components and noise interference in the power supply; Analyze the power spectrum density of the power supply signal, display the distribution of signal power as a function of frequency, and identify abnormal frequency components in the power supply; Based on the characteristics of the switching power supply signal, the Haar wavelet function is selected as the wavelet basis function for wavelet transform, and the voltage and current signals of the switching power supply are subjected to continuous wavelet transform to obtain wavelet coefficients under at least one scale.

2. A method for identifying a switching power supply state based on fault data integration according to claim 1, characterized in that: The cleaning of the collected raw data, removal of noise, outliers and missing values, and standardization of the data specifically include: Based on the reasonable value range and mutual relationship of variables, check the standardization and logic of data, and use the sample mean of each variable to replace invalid values ​​and missing values; Verify the correlation between data from multiple sources and select feature attributes by verifying the correlation between data; Use filtering methods to remove noise; For random noise, the mean filter is used to calculate the neighborhood average for smoothing; For salt and pepper noise, use median filtering to select the median of the window data as output; For Gaussian noise, use Gaussian filtering to weight and smooth the signal; For high-frequency noise, wavelet denoising is used to process the noise part through signal decomposition; The data were standardized by taking the base 10 logarithm of each observed value of each variable and dividing it by the base 10 logarithm of the maximum value of the variable.

3. A method for identifying a switching power supply state based on fault data integration according to claim 2, characterized in that: The method of extracting the timing characteristics of the switching power supply by using the long short-term memory network and analyzing the information accumulation and transformation of the input data in at least one time step specifically includes: Obtaining the operating data of the switching power supply in at least one time step, the data exists in the form of a time series, and each time step has a corresponding data point; Based on the characteristics of the data and task requirements, design the structure of the LSTM network, determine the number of LSTM network layers, the number of neurons in each layer, and the dimension of the input data; Preprocess the running data, convert the data into three-dimensional data of sample number, time step and feature number, and input the preprocessed data into the long short-term memory network; Design the hidden layer. In the long short-term memory network, the hidden layer extracts the features in the time series. Through at least one long short-term memory network hidden layer, deeper features are gradually extracted. Based on the task requirements, the structure and activation function of the output layer are set. In the task of extracting time series features, the output layer is a fully connected layer that maps the features extracted by the long short-term memory network to the target space. Select the mean square error function as the loss function to construct the long short-term memory network; The preprocessed data is input into the long short-term memory network for forward and backward propagation, and the weights of the model are updated in real time through the optimizer; Determine whether the preset number of training rounds has been reached or whether the loss function has converged. If so, end the training; if not, do not output. After the training is completed, the characteristics of the time series are extracted from the hidden layer of the long short-term memory network, and the characteristics of the time series include information accumulation and transformation of the switching power supply in at least one time step.

4. A method for identifying a switching power supply state based on fault data integration according to claim 3, characterized in that: The method converts the time series into a polar coordinate system, calculates the Gram matrix using trigonometric functions, maps the one-dimensional data into a two-dimensional image, generates a corresponding two-dimensional image based on the result of field coding, and extracts the spatial features of the switching power supply through a convolutional neural network, specifically including: The normalized time series data is converted into a polar coordinate system. In this conversion process, the time series data is mapped onto a unit circle. Each data point corresponds to a point on the unit circle, and its angle is calculated by trigonometric functions. Use the cosine function to calculate the angle sum and angle difference of the data points between each time stamp; For any two data points in the time series, the cosine similarity between them is calculated. By calculating the cosine similarity between each data point, a Gram matrix is ​​constructed to reflect the correlation between each time stamp in the time series; Map each element of the calculated Gram matrix to a grayscale value range according to a value range of 0-255, and use it as the pixel value of the image; Design the convolutional neural network structure. The input layer receives the preprocessed data. The convolution layer contains at least one convolution kernel to extract the local features of the data. The number, size and step size of the convolution kernel are adjusted based on the task and data. The pooling layer is after the convolution layer to reduce the dimension of the data and the amount of calculation. The fully connected layer maps the features extracted by the convolution layer to the output space and is used for classification or regression tasks. The output layer selects the output layer function based on the task requirements. In the convolution layer, the convolution kernel slides over the input data, performs element-wise multiplication and summation operations, and extracts local features; Each convolution kernel generates a new feature map that reflects a specific pattern or feature in the input data; For the switching power supply, the spatial features include morphological features of the waveform data and spatial distribution features in the image data, wherein the morphological features of the waveform data include peak values, valley values, and waveform width, and the spatial distribution features in the image data include hot spots and abnormal areas; The two-dimensional image obtained by time series conversion is input into the convolutional neural network. Through the convolution operation, the convolutional neural network automatically learns and extracts the spatial features in the image.

5. A method for identifying a switching power supply state based on fault data integration according to claim 4, characterized in that: The feature selection method based on mutual information and tree model is used to screen the features with high influence on the state recognition of the switch power supply, and output them as key features. The key features are adaptively fused to construct a multi-dimensional feature fusion state recognition model, which specifically includes: The mutual information formula is used to calculate the mutual information value between each feature and the target variable. The mutual information value is positively correlated with the contribution of the feature to the target variable. Sort the features based on the mutual information value, and select the features with high mutual information value as key features, which are used in the subsequent multi-dimensional feature fusion state recognition model; Splitting nodes by building a tree structure and evaluating the importance of features based on the loss function value that is reduced when the feature is split at the node; Use the decision tree model to model the switching power supply state data. During the tree model training process, record the loss reduction of each feature when the node is split, and accumulate the importance scores of the features. Sort the features based on their importance scores and select features with high importance scores as key features; After the key features are screened out, the key features are adaptively fused to construct a multi-dimensional feature fusion state recognition model, and each feature is weighted based on its importance or contribution; Standardize the selected key features to ensure that they have the same scale during the fusion process; Based on the importance score or mutual information value of the feature, a weight is assigned to each feature. The size of the weight reflects the contribution of the feature to the target variable. During the fusion process, each feature is multiplied by its corresponding weight to achieve adaptive weighting; The weighted features are concatenated to form a new multi-dimensional feature vector, which is used as the input of the multi-dimensional feature fusion state recognition model.

6. A method for identifying a switching power supply state based on fault data integration according to claim 5, characterized in that: The trained model is deployed to the actual system, the monitoring data of the switching power supply is received in real time, and the model is used to predict the input data. The prediction result is output in the form of a classification label. The current working state of the switching power supply is judged specifically including: The working state is classified based on the working mode. In continuous conduction mode, the inductor current is continuous, which is suitable for medium or heavy load conditions. In discontinuous conduction mode, the inductor current is discontinuous, which is suitable for light load conditions. In mixed mode, the continuous conduction mode and discontinuous conduction mode are automatically switched based on the load conditions to achieve efficiency optimization. The working state is classified based on the output state. In the normal working state, the output voltage and current are stable, the fluctuation is within the allowable range, the efficiency is high, and the loss is low. The abnormal working state includes overvoltage protection state, overcurrent protection state, short circuit protection state and overheating protection state. In the standby state, the output voltage and current are low, and the standby power consumption is kept fixed. Classify the working status based on the fault type. Hardware faults include component damage and circuit board failure. Software faults include control program abnormalities and parameter setting errors. The working status is classified based on load changes. The light load state is light load and low output power. The full load state is that the load reaches the rated value and the output power is high. The heavy load state is that the load exceeds the rated value and is within the tolerable range of the power supply. The overload state is that the load exceeds the rated value and will cause damage to the power supply.

Images