Layered detection method and system for intelligent power module

Abstract

The invention provides a hierarchical detection method and system for an intelligent power module, and the method comprises the steps: collecting the data of the power module through a sensor, and carrying out the preprocessing of the data through an advanced noise filtering and abnormal value processing technology; key features are extracted from the preprocessed data, and then an automatic feature selection technology is applied to optimize the key features so as to improve the fault recognition capability of the model; performing preliminary fault detection on the power module by using a deep learning model, and then performing fault cause analysis and classification by using traditional algorithms such as a decision tree and the like; an online learning mechanism is adopted, and deep learning model parameters are continuously adjusted and optimized according to newly collected data of the power module; and according to the trained model, the operation state of the power module is evaluated in real time, and the fault development trend is predicted. According to the method, the precision and efficiency of fault detection are improved, and the user friendliness and self-adaptability of the system are greatly improved by introducing interpretability and a user feedback mechanism.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent detection, and in particular relates to a hierarchical detection method and system for an intelligent power module. Background Art

[0002] In modern power systems, safe and efficient operation and maintenance of transmission lines is the key to maintaining stable operation of the power grid. As the scale of the power grid continues to expand, traditional line monitoring technology faces various challenges, especially in terms of data processing efficiency and monitoring accuracy. At present, most transmission line monitoring systems rely on basic image acquisition and manual inspection, which is not only inefficient, but also easily affected by human factors and environmental conditions. Factors such as lighting and weather often affect image quality, thereby affecting the accurate interpretation of data.

[0003] In existing technologies, although image recognition technology is used to assist in the monitoring of transmission lines, most of these systems still rely on preset rules for image processing and analysis, such as simple threshold segmentation, edge detection, etc. These methods are often unable to handle complex or blurred images. In addition, existing monitoring systems often generate a large amount of invalid data during the data screening process, which increases the burden of data processing, and fail to effectively use these data for prediction and risk assessment, thus failing to prevent potential hidden dangers in advance.

[0004] In addition, the data classification in existing technologies is usually rough and lacks the capture of details, which directly affects the accuracy and efficiency of operation and maintenance decisions. In terms of efficiency monitoring, existing systems mostly rely on post-event data statistics and analysis, lack real-time monitoring and response mechanisms, and cannot send key information to staff in a timely manner, thus delaying the best time for maintenance response. These technical deficiencies illustrate the limitations of existing systems in terms of intelligent and automated processing capabilities, and there is an urgent need for a new solution that can comprehensively improve image processing quality and data analysis efficiency. Summary of the invention

[0005] The purpose of the present invention is to design a hierarchical detection method and system for intelligent power modules to achieve efficient and accurate hierarchical detection of intelligent power modules. Each step is innovated and optimized for the limitations of the prior art mentioned in the background technology. The entire system forms a closed-loop intelligent detection process from data acquisition to fault warning, which not only improves the accuracy and timeliness of fault detection, but also enhances the user interaction and self-optimization capabilities of the system.

[0006] In order to achieve the above object, a first aspect of the present invention provides a delamination detection method for an intelligent power module, the method comprising:

[0007] S1. Collect data from the power module using sensors and pre-process the data using advanced noise filtering and outlier processing techniques.

[0008] S2. Extract key features from the preprocessed data, and then optimize the key features using automated feature selection technology to improve the fault identification capability of the model; wherein the extraction of basic statistical features from the preprocessed data includes time domain feature extraction, frequency domain feature extraction, and time-frequency domain feature extraction;

[0009] The time domain feature extraction extracts basic statistical features from the standardized voltage and current signals. The basic statistical features include mean μ, standard deviation σ, skewness γ and kurtosis k, which are expressed as follows:

[0010]

[0011] Among them, x i represents a single data point, N represents the number of samples, and i represents data point i;

[0012] The frequency domain feature extraction uses fast Fourier transform to convert the signal from the time domain to the frequency domain, and extracts the energy and frequency values ​​of the main frequency components, which are expressed as follows:

[0013]

[0014] Among them, X k represents the Fourier transform result at frequency k, E k Represents the corresponding energy, and n represents the order of the filter;

[0015] The time-frequency domain feature extraction uses short-time Fourier transform to analyze the signal, obtain the changes in frequency components at different times, and extract dynamic features, which are expressed as follows:

[0016] S(t,f)=∫x(t+τ)w(τ)e -j2πfτ dτ

[0017] Where S(t,f) represents the amplitude at time t and frequency f in the time-frequency representation;

[0018] S3. Use deep learning models to perform preliminary fault detection on power modules, and then use traditional algorithms such as decision trees to analyze and classify fault causes.

[0019] S4, using online learning mechanism to continuously adjust and optimize deep learning model parameters based on newly collected power module data;

[0020] S5. Based on the trained model, the operating status of the power module is evaluated in real time and the fault development trend is predicted.

[0021] In some embodiments, the method further comprises:

[0022] S6, using explainable algorithms to show users the logic and reasons behind each fault detection and prediction;

[0023] S7. Establish a user feedback mechanism and optimize the power module fault model based on user feedback;

[0024] S8. Integrate and deploy the hierarchical detection method to the cloud platform or local server to achieve continuous stream processing and real-time update of power module data.

[0025] In some embodiments, step S1 specifically includes:

[0026] S101, sampling the output signal of the power module to obtain sampling data, the sampling frequency f s Set to twice the maximum operating frequency of the module, f s =2×f max , where f max is the maximum operating frequency of the power module;

[0027] S102, filtering the sampled data using a digital low-pass filter to remove high-frequency noise;

[0028] S103, performing outlier detection on the filtered sampled data, and using a statistical method to identify and eliminate possible erroneous data;

[0029] S104. Standardize the processed sampled data to make the data numerically stable during machine learning training.

[0030] In some embodiments, in step S2, the extracted key features are fused and optimized as follows:

[0031] All the extracted key features are integrated into a feature matrix X, where each row represents a sample and each column represents a feature, and then standardized, as shown below:

[0032]

[0033] Among them, X std represents the standardized data matrix, μ and σ represent the mean and standard deviation of the original data matrix X respectively; the covariance matrix is ​​calculated to show the linear relationship between data features, which is expressed as follows:

[0034]

[0035] Where n represents the number of samples, Σ represents the covariance matrix;

[0036] The eigenvalue λ and eigenvector v of the covariance matrix Σ are calculated according to the covariance matrix, which are expressed as follows:

[0037] Σv=λv;

[0038] According to the size of the eigenvalue λ, the eigenvectors corresponding to the largest k eigenvalues ​​are selected to form a transformation matrix P, and then the selected eigenvectors are used to transform the original data into the new subspace, which is expressed as follows:

[0039] Y=X std P

[0040] Where Y is the transformed data matrix.

[0041] In some embodiments, the structure of the deep learning model in step S3 includes a bottom layer model, a middle layer model and a top layer model;

[0042] The underlying model is used to find a classification hyperplane that maximizes the margin between different classes;

[0043] The mid-level model is used to improve the classification results through random forest;

[0044] The top-level model is used to integrate the outputs of the bottom-level and middle-level models for fault prediction and health status assessment through a gradient boosting machine.

[0045] In some embodiments, step S4 specifically includes:

[0046] S401, continuously monitor the data flow of the power module, online learning and update the model, as shown below:

[0047] D memory =D old ∪D new

[0048] Among them, D memory Indicates the fusion of old data D old and new data D new Datasets;

[0049] S402. Use regularization technology to design the loss function to avoid destroying old task knowledge, as shown below:

[0050]

[0051] Among them, L new (θ) represents the loss function on new data, F i represents the diagonal elements in the fisherman information matrix from the old task, measuring the parameter θ i The importance of θ i,old represents the value of the parameter in the previous task, and λ represents the regularization strength;

[0052] S403: Continue to monitor the power module and update the model.

[0053] In some embodiments, step S5 specifically includes:

[0054] S501. Use the seasonal autoregressive integrated moving average model to predict the time series data of equipment performance, as shown below:

[0055]

[0056] Among them, Y t represents the equipment performance index at time t, B represents the back-shift operator, s represents the seasonal cycle, and denote non-seasonal and seasonal differences, ε t is the error term;

[0057] S502, by calculating the mean μ of each window w and standard deviation σ w , dynamically set the threshold for real-time anomaly detection, as shown below:

[0058] Threshold=μ w +kσ w

[0059] S503, using machine learning technology to model and predict the monitoring trend of the power module, and the health score of the power module is dynamically adjusted according to the prediction result, as shown below:

[0060] H(t)=β1f1(X t )+β2f2(X t )+…+β n f n (X t )

[0061] Among them, X t represents the feature vector at time t, f i represents the i-th decision tree in the random forest, β i Represents the weight of the corresponding tree;

[0062] S504, regularly evaluating the accuracy of the prediction model based on the latest power module data, and automatically adjusting model parameters or retraining the model to adapt to new trends and data patterns.

[0063] In some embodiments, in step S6, based on the explainability algorithm, additional contextual data is integrated and features are extracted from the additional contextual data, and added to the prediction calculation of each power module to dynamically adjust the weights. At the same time, a visualization tool is constructed to adjust and optimize the model based on user feedback.

[0064] In some embodiments, in step S7, a feedback database is designed for user feedback, and then a performance key indicator formula of the power module is designed, which is expressed as follows:

[0065]

[0066] Among them, P eff represents the prediction efficiency, R time represents the average response time;

[0067] Design optimization strategies based on the feedback database and key indicator results and provide feedback to users.

[0068] In a second aspect of the present invention, a delamination detection system for an intelligent power module is provided, the system comprising:

[0069] A data collection module, used to collect data from the power module using sensors and pre-process the data using advanced noise filtering and outlier processing techniques;

[0070] A key feature extraction module is used to extract key features from the preprocessed data, and then optimize the key features by applying automated feature selection technology to improve the fault identification capability of the model; wherein the basic statistical features extracted from the preprocessed data include time domain feature extraction, frequency domain feature extraction, and time-frequency domain feature extraction;

[0071] The time domain feature extraction extracts basic statistical features from the standardized voltage and current signals. The basic statistical features include mean μ, standard deviation σ, skewness γ and kurtosis k, which are expressed as follows:

[0072]

[0073] Among them, x i represents a single data point, N represents the number of samples, and i represents data point i;

[0074] The frequency domain feature extraction uses fast Fourier transform to convert the signal from the time domain to the frequency domain, and extracts the energy and frequency values ​​of the main frequency components, which are expressed as follows:

[0075]

[0076] Among them, X krepresents the Fourier transform result at frequency k, E k Represents the corresponding energy, and n represents the order of the filter;

[0077] The time-frequency domain feature extraction uses short-time Fourier transform to analyze the signal, obtain the changes in frequency components at different times, and extract dynamic features, which are expressed as follows:

[0078] S(t,f)=∫x(t+τ)w(τ)e -j2πfτ dτ

[0079] Where S(t,f) represents the amplitude at time t and frequency f in the time-frequency representation;

[0080] A preliminary fault detection module, which uses a deep learning model to perform preliminary fault detection on the power module, and then uses traditional algorithms such as decision trees to analyze and classify the fault causes;

[0081] An online learning module, which is used to continuously adjust and optimize the deep learning model parameters based on newly collected power module data using an online learning mechanism;

[0082] The real-time evaluation module evaluates the operating status of the power module in real time and predicts the fault development trend based on the trained model;

[0083] The system also includes at least 0 of the following modules:

[0084] The fault display module is used to show users the logic and reasons behind each fault detection and prediction through an explainable algorithm;

[0085] Fault optimization module: establishes a user feedback mechanism and optimizes the fault model of the power module based on user feedback;

[0086] The system integration module integrates and deploys the hierarchical detection method to the cloud platform or local server to achieve continuous processing and real-time update of power module data.

[0087] The beneficial technical effects of the present invention are at least as follows:

[0088] (1) This system uses an advanced image acquisition module and combines it with adaptive image enhancement technology to automatically adjust image parameters such as contrast and brightness according to actual lighting and weather conditions to ensure that the acquired images always maintain high quality, effectively overcoming the application limitations of traditional technologies in harsh environments.

[0089] (2) By introducing a deep learning algorithm, the present invention not only enables the system to identify key information in images, but also intelligently screen data with high fuzziness. Only data exceeding a preset fuzziness threshold is identified as valid line data, which greatly reduces the amount of invalid data to be processed and improves the overall data processing efficiency and accuracy.

[0090] (3) The present invention uses advanced image recognition and natural language processing technology to establish a mapping relationship between effective line data and preset keywords, automatically construct a data-keyword database, provide accurate data support for operation and maintenance decisions, and significantly improve the accuracy and practicality of data classification.

[0091] (4) The present invention is provided with an efficiency monitoring module, which can monitor and analyze in real time the efficiency of staff in processing line records, operation and maintenance information and hidden danger information, and timely generate and send reminder information to the staff's work terminal to ensure that problems can be responded to and handled quickly, thereby effectively improving the timeliness and accuracy of maintenance work.

[0092] (5) The present invention realizes efficient and accurate hierarchical detection of intelligent power modules through a series of organically connected steps, each of which is innovative and optimized with respect to the limitations of the prior art mentioned in the background technology. The entire system, from data collection to fault warning, forms a closed-loop intelligent detection process, which not only improves the accuracy and timeliness of fault detection, but also enhances the user interaction and self-optimization capabilities of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0094] Figure 1 The present invention is a flow chart of a layered detection method for an intelligent power module.

[0095] Figure 2 This is a framework diagram of a hierarchical detection system for intelligent power modules according to the present invention. DETAILED DESCRIPTION

[0096] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0097] In one or more embodiments, Figure 1As shown, the present invention discloses a delamination detection method for an intelligent power module, the method comprising the following steps 1-5:

[0098] S1. Collect data from the power module using sensors and pre-process the data using advanced noise filtering and outlier processing techniques, including:

[0099] S101, sampling the output signal of the power module to obtain sampling data, the sampling frequency f s Set to twice the highest operating frequency of the module (according to the Nyquist theorem) to ensure complete capture of the signal. Set f s =2×f max , where f max is the maximum operating frequency of the power module;

[0100] S102, using a digital low-pass filter to filter the sampled data to remove high-frequency noise; wherein the filter is designed as a Butterworth filter, and the specific expression is:

[0101]

[0102] Where f represents the signal frequency, f c is the cutoff frequency, n is the order of the filter, which is usually set to 4 to ensure sufficient smoothing effect without introducing phase delay.

[0103] S103, performing outlier detection on the filtered sampled data, using a statistical method to identify and eliminate possible erroneous data; wherein the Z score of each data point is calculated:

[0104]

[0105] Where x is a single data point, μ and σ are the mean and standard deviation of the data respectively. If |Z|>3, the data point is considered an outlier and is removed;

[0106] S104, standardize the processed sampled data to make the data numerically stable during machine learning training, wherein the standardization formula is:

[0107]

[0108] Among them, x' is the standardized data, min(x) and max(x) are the minimum and maximum values ​​of the original data respectively.

[0109] S2. Extract key features from the preprocessed data, and then optimize the key features using automated feature selection technology to improve the fault identification capability of the model; wherein the extraction of basic statistical features from the preprocessed data includes time domain feature extraction, frequency domain feature extraction, and time-frequency domain feature extraction;

[0110] The time domain feature extraction extracts basic statistical features from the standardized voltage and current signals. The basic statistical features include mean μ, standard deviation σ, skewness γ and kurtosis k, which are expressed as follows:

[0111]

[0112] Among them, x i represents a single data point, N represents the number of samples, and i represents data point i;

[0113] The frequency domain feature extraction uses fast Fourier transform to convert the signal from the time domain to the frequency domain, extracting the energy and frequency values ​​of the main frequency components, including the main frequency (f main ) and the corresponding energy (E main ), which is expressed as follows:

[0114]

[0115] Among them, X k represents the Fourier transform result at frequency k, E k Represents the corresponding energy, and n represents the order of the filter;

[0116] The time-frequency domain feature extraction uses short-time Fourier transform to analyze the signal, obtain the changes of frequency components at different times, and extract dynamic features such as the rate of change (r) and energy peak (E peak ), which is expressed as follows:

[0117] S(t,f)=∫x(t+τ)w(τ)e -j2πfτ dτ (6)

[0118] Where S(t,f) represents the amplitude at time t and frequency f in the time-frequency representation;

[0119] To further explain, in the feature engineering of intelligent power modules, in order to reduce the computational load of subsequent model training and improve the generalization ability of the model, it is necessary to effectively reduce the dimensionality of the extracted multi-dimensional features. Principal component analysis (PCA) is a commonly used statistical method that converts multiple variables that may be related into linearly independent variables (principal components) through orthogonal transformation. These principal components are arranged in descending order of variance to select the most representative features. Therefore, the above features are fused and optimized as follows:

[0120] All extracted features (time domain, frequency domain, and time-frequency domain features) are integrated into a feature matrix X, where each row represents a sample and each column represents a feature, and then the data is standardized:

[0121]

[0122] Where X std is the standardized data matrix, μ and σ are the mean and standard deviation of the original data matrix X (calculated for each column respectively);

[0123] Compute the covariance matrix:

[0124]

[0125] Where n is the number of samples and Σ is the covariance matrix;

[0126] Solve the eigenvalues ​​and eigenvectors. The eigenvalues ​​of the covariance matrix and the corresponding eigenvectors determine the principal components of the data. Calculate the eigenvalues ​​λ and eigenvectors v of Σ:

[0127] Σv=λv (9)

[0128] The variance of the data in the direction of the eigenvector with a larger eigenvalue is larger, so the amount of information is also greater;

[0129] According to the size of the eigenvalue, the eigenvectors corresponding to the largest k eigenvalues ​​are selected to form a transformation matrix P. The k usually selected is the minimum k value that makes the cumulative contribution rate exceed a certain threshold (such as 95\%);

[0130] Transform the original data into the new subspace using the selected eigenvectors:

[0131] Y=X std P (10)

[0132] Y is the transformed data matrix, which contains only the most important k principal components, greatly reducing the number of features while retaining as much information as possible.

[0133] S3. Use deep learning models to perform preliminary fault detection on power modules, and then use traditional algorithms such as decision trees to analyze and classify the causes of faults.

[0134] The specific construction steps are as follows:

[0135] Bottom-layer model construction: In this layer, support vector machine (SVM) is used for preliminary fault classification. For a given feature vector x, the goal of the SVM model is to find a classification hyperplane that maximizes the interval between different categories. The classification function can be expressed as:

[0136] f(x)=w T x+b (11)

[0137] Where w is the weight vector and b is the bias term. The kernel trick is used to map the data into a high-dimensional space to handle nonlinear classification problems.

[0138] Middle-level model design: Based on the output of the underlying SVM, a random forest is used to perform more detailed fault analysis and important feature identification. Random forest improves the classification results by building multiple decision trees and voting. The decision function of each tree is:

[0139] g i (x) = decisionpath(x; Θ i ) (12)

[0140] Among them, Θ i Represents the parameter set of the i-th tree, and decisionpath is a function that selects a path in the decision tree based on the input features.

[0141] Top-level model integration: Use the gradient boosting machine (GBM) to integrate the outputs of the bottom-level and middle-level models for final fault prediction and health status assessment. The gradient boosting machine optimizes the model by iteratively reducing the prediction error, which can be expressed as:

[0142]

[0143] Among them, h k (x) is the model prediction at the kth iteration, α k is the corresponding model weight, and K is the number of iterations.

[0144] S4, using online learning mechanism to continuously adjust and optimize deep learning model parameters based on newly collected power module data, including:

[0145] S401, continuously monitor the data flow of the power module, learn and update the model online, the model continuously collects data from the sensors of the power module, ensures that the data including new operating conditions and possible new failure modes are captured and analyzed in real time, and uses the experience replay mechanism to not only update the model with new data, but also regularly review old data samples to prevent the knowledge learned in the past from being overwritten by new data, as shown below:

[0146] D memory =D old ∪D new (14)

[0147] Among them, D memory Indicates the fusion of old data D old and new data D new Datasets;

[0148] S402. Use regularization techniques to design loss functions to avoid destroying old task knowledge. Specifically, use regularization techniques such as elastic weight sharing (EWC) to take into account the importance of old tasks when updating the model to avoid destroying old task knowledge when learning new tasks, as shown below:

[0149]

[0150] Among them, L new (θ) represents the loss function on new data, F i represents the diagonal elements in the fisherman information matrix from the old task, measuring the parameter θ i The importance of θ i,old represents the value of the parameter in the previous task, and λ represents the regularization strength;

[0151] S403, continuously monitoring the power module and updating the model;

[0152] To further illustrate, we continuously monitor the performance of the model on both new and old data to ensure that new learning does not harm previous performance levels. We apply a dual validation strategy, where we evaluate the model on both new and old datasets. During the model update process, we introduce incremental validation, where we quickly validate the model after each significant update to ensure that balanced performance on new and old tasks is maintained.

[0153] Through these enhanced continuous learning technologies, we can more effectively cope with the various challenges encountered by intelligent power modules in complex and ever-changing practical applications, including the identification of new fault modes and rapid adaptation to environmental changes, while protecting existing knowledge from being forgotten. This approach significantly improves the practicality and reliability of the system, and is particularly suitable for industrial applications that operate for a long time and require continuous optimization.

[0154] S5. Based on the trained model, the operating status of the power module is evaluated in real time, and the fault development trend is predicted, including:

[0155] S501. Use the seasonal autoregressive integrated moving average (SARIMA) model to predict the time series data of equipment performance, as shown below:

[0156]

[0157] Among them, Y t represents the equipment performance index at time t, B represents the back-shift operator, s represents the seasonal cycle, and denote non-seasonal and seasonal differences, ε t is the error term;

[0158] S502: Perform real-time anomaly detection in combination with statistical analysis of moving windows to capture immediate changes that are inconsistent with historical data patterns. By calculating the mean μ of each window w and standard deviation σ w , dynamically set the threshold, expressed as follows:

[0159] Threshold=μ w +kσ w (17)

[0160] When the device performance indicator exceeds this threshold, an alarm is triggered;

[0161] S503. Use machine learning techniques such as random forests to model and predict the long-term health trend of the equipment. The prediction model will learn from historical data and predict the future health status of the equipment. The equipment health score H(t) is dynamically adjusted according to the prediction results, as shown below:

[0162] H(t)=β1f1(X t )+β2f2(X t )+…+β n f n (X t ) (18)

[0163] Among them, X t represents the feature vector at time t, f i represents the i-th decision tree in the random forest, β i Represents the weight of the corresponding tree;

[0164] S504, regularly evaluating the accuracy of the prediction model based on the latest power module data, and automatically adjusting model parameters or retraining the model to adapt to new trends and data patterns;

[0165] To further illustrate, the accuracy of the prediction model is regularly evaluated based on the latest data, and the model parameters are automatically adjusted or the model is retrained to adapt to new trends and data patterns. The model weights and parameters are regularly updated using online learning techniques to ensure that the model reflects the latest equipment status and environmental conditions.

[0166] In some embodiments, steps S6-S8 are also included:

[0167] S6. The explainability (SHAP) algorithm is used to show users the logic and reasons behind each fault detection and prediction.

[0168] To further illustrate, on the basis of the conventional SHAP method, additional contextual data is integrated, such as ambient temperature, humidity, equipment operation time, etc. These factors may affect module performance and failure occurrence. Then the input feature set is modified to add these contextual information as additional features.

[0169] For each prediction of the intelligent power module, the SHAP value is calculated not only considering the traditional input characteristics, but also the environmental and operating status factors mentioned above:

[0170]

[0171] Among them, C represents the context feature set, f x is a prediction model that takes context into account, and other symbols are the same as standard SHAP;

[0172] Then, based on historical data, we analyze the statistical significance of the impact of contextual features on prediction results, dynamically adjust the weights of these contextual features in the SHAP value calculation, and implement an automatic learning mechanism to continuously optimize the weights of contextual features based on equipment operation data and fault records to reflect their true impact under different conditions.

[0173] Finally, an interactive visualization tool is provided to show not only the SHAP values ​​of traditional input features but also the contribution of contextual features. A user feedback mechanism is designed to collect users’ satisfaction with the explanations and improvement suggestions for further adjustment and optimization of the model.

[0174] S7. Establish a user feedback mechanism and optimize the power module fault model based on user feedback.

[0175] To further explain, the system regularly collects user feedback, including the accuracy of fault diagnosis, ease of use of the system, and clarity of explanations. The feedback is classified and coded into several main categories: functional issues, performance issues, user interface, and explanation issues. A feedback database is designed to store and query data using the following model:

[0176] F(u,t,c,s)={f1,f2,…,f n} (20)

[0177] Among them, F represents the feedback set, u is the user ID, t is the submission time, c is the classification label, s is the satisfaction score, and f i It is the specific feedback content.

[0178] In the embodiment of the present invention, the performance indicators of each part of the system, such as prediction accuracy, response time, etc., can be regularly evaluated, and statistical analysis and machine learning methods can be used to identify system performance bottlenecks and improvement points. Key indicators may include:

[0179]

[0180] Among them, P eff represents the prediction efficiency, R time Indicates the average response time.

[0181] Further explanation: Based on the feedback and performance analysis results, develop specific optimization strategies. This may include algorithm adjustments, feature enhancements, user interface redesign, etc. Implement an iterative development plan to periodically update the system and make adjustments based on user feedback and performance test results.

[0182] Feedback the optimization results and upcoming improvement measures directly to users through system update notifications or emails, and set up a user panel where users can view in real time how their feedback is adopted by the system and the progress of improvements.

[0183] S8. Integrate and deploy the hierarchical detection method to the cloud platform or local server to achieve continuous stream processing and real-time update of power module data.

[0184] To further explain, a modular system architecture is defined to ensure that each part such as data acquisition, processing, model operation and user interface can be updated and maintained independently. The dependency relationship between modules is described using the system integration diagram G(V,E), where V represents the components of the system and E represents the data flow and control flow between components;

[0185] Develop corresponding deployment strategies according to different operating environments (such as cloud servers, local servers, or edge computing devices), and use containerization technology (such as Docker) to package system components to ensure consistency and portability in different environments;

[0186] Define container configuration parameters C(p,v,m), where p is the port number, v is the volume mount configuration, and m is the memory and CPU quota;

[0187] Establish an automated Cl / CD process to automate and seamlessly update and deploy the system, and build an automated test script T(s,r), where s represents the test script and r represents the expected result, to ensure that the system functions are stable and the performance meets the standards after each update;

[0188] Integrated system monitoring tools monitor the system's operating status in real time, including resource usage, component operating status, etc. Use the monitoring indicator function M(x,y,z) for resource monitoring, where x is CPU usage, y is memory usage, and z is network traffic. Perform regular system maintenance and optimization, and adjust system configuration and performance based on monitoring data and user feedback.

[0189] In one or more embodiments, Figure 2 As shown, the present invention discloses a layered detection system for an intelligent power module, the system comprising:

[0190] A data collection module 101 is used to collect data of the power module using sensors, and pre-process the data using advanced noise filtering and outlier processing techniques;

[0191] The key feature extraction module 102 is used to extract key features from the preprocessed data, and then optimize the key features by applying the automated feature selection technology to improve the fault identification capability of the model; wherein the basic statistical features extracted from the preprocessed data include time domain feature extraction, frequency domain feature extraction and time-frequency domain feature extraction;

[0192] The time domain feature extraction extracts basic statistical features from the standardized voltage and current signals. The basic statistical features include mean μ, standard deviation σ, skewness γ and kurtosis k, which are expressed as follows:

[0193]

[0194] Among them, x i represents a single data point, N represents the number of samples, and i represents data point i;

[0195] The frequency domain feature extraction uses fast Fourier transform to convert the signal from the time domain to the frequency domain, and extracts the energy and frequency values ​​of the main frequency components, which are expressed as follows:

[0196]

[0197] Among them, X k represents the Fourier transform result at frequency k, E k Represents the corresponding energy, and n represents the order of the filter;

[0198] The time-frequency domain feature extraction uses short-time Fourier transform to analyze the signal, obtain the changes in frequency components at different times, and extract dynamic features, which are expressed as follows:

[0199] S(t,f)=∫x(t+τ)w(τ)e -j2πfτ dτ

[0200] Where S(t,f) represents the amplitude at time t and frequency f in the time-frequency representation;

[0201] A preliminary fault detection module 103 is used to perform preliminary fault detection on the power module using a deep learning model, and then use traditional algorithms such as decision trees to analyze and classify the causes of the faults;

[0202] An online learning module 104 is used to continuously adjust and optimize deep learning model parameters according to newly collected power module data using an online learning mechanism;

[0203] A real-time evaluation module 105 evaluates the operating status of the power module in real time and predicts the fault development trend according to the trained model;

[0204] The system also includes at least 0 of the following modules:

[0205] A fault display module 106 is used to display the logic and reasons behind each fault detection and prediction to the user through an explainable algorithm;

[0206] Fault optimization module 107, establishing a user feedback mechanism, and optimizing the fault model of the power module according to user feedback;

[0207] The system integration module 108 integrates and deploys the layered detection method to a cloud platform or a local server to achieve continuous stream processing and real-time update of power module data.

[0208] In summary, the present invention ensures data quality through advanced noise filtering and outlier processing techniques, and provides accurate and consistent input for subsequent feature engineering and model training. This step solves the problem of insufficient data processing in the prior art and lays the foundation for efficient fault detection. Then, the input features are optimized through automated feature selection, which not only improves the fault identification ability of the model, but also makes the model more focused on the variables that have the greatest impact on fault detection. This step effectively solves the problem of manual and inefficient feature processing in traditional technology. The multi-layer model strategy is adopted, combining the advantages of deep learning and traditional machine learning to improve the accuracy and interpretability of fault diagnosis. This method overcomes the limitations of single model application and provides a more comprehensive fault detection solution. Real-time update and adaptive learning of the model ensure that the system is continuously optimized over time and adapts to new fault modes, thereby improving the accuracy of prediction. This step solves the problem that the system in the prior art is difficult to adapt to new fault types, and enhances the adaptability and future applicability of the system. Fault prediction and health assessment combine real-time data with trained models to conduct continuous health status assessment and future fault prediction, which not only warns of possible problems in advance, but also helps the maintenance team optimize maintenance plans. This step effectively addresses the limitation of lack of prediction capabilities in the prior art. Interpretable result output By integrating the interpretable algorithm into the system, maintenance personnel can understand the judgment logic of the model, which increases the transparency of the system and the trust of users. This solves the problem of "black box" operation of intelligent systems in the existing technology and improves user acceptance and satisfaction. User feedback and system optimization Through direct feedback from users, the system can be adjusted and optimized more accurately to make the model more in line with the actual operating environment and needs. This step provides a mechanism for continuous improvement and enhances the dynamic adjustment ability of the system. Complete system integration and deployment integrates all steps into a unified platform, ensuring that every link from data collection to fault warning can run efficiently and seamlessly. This step ensures the integrity and stability of the system, enabling it to achieve efficient fault detection in practical applications.

[0209] The above disclosures are only some preferred embodiments of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiments and equivalent changes made according to the claims of the present invention are still within the scope of the invention.

Claims

1. A delamination detection method for an intelligent power module, characterized in that: The method comprises the following steps: S1. Collect data from the power module using sensors and pre-process the data using advanced noise filtering and outlier processing techniques. S2. Extract key features from the preprocessed data, and then optimize the key features using automated feature selection technology to improve the fault identification capability of the model; wherein the extraction of basic statistical features from the preprocessed data includes time domain feature extraction, frequency domain feature extraction, and time-frequency domain feature extraction; The time domain feature extraction extracts basic statistical features from the standardized voltage and current signals. The basic statistical features include mean μ, standard deviation σ, skewness γ and kurtosis k, which are expressed as follows: Among them, x i represents a single data point, N represents the number of samples, and i represents data point i; The frequency domain feature extraction uses fast Fourier transform to convert the signal from the time domain to the frequency domain, and extracts the energy and frequency values ​​of the main frequency components, which are expressed as follows: Among them, X k represents the Fourier transform result at frequency k, E k represents the corresponding energy, n represents the order of the filter, and j represents the imaginary unit; The time-frequency domain feature extraction uses short-time Fourier transform to analyze the signal, obtain the changes in frequency components at different times, and extract dynamic features, which are expressed as follows: S(t,f)=∫x(t+τ)w(τ)e- j2πfτ dτ Where S(t,f) represents the amplitude at time t and frequency f in the time-frequency representation; S3. Use deep learning models to perform preliminary fault detection on power modules, and then use traditional algorithms such as decision trees to analyze and classify fault causes. S4, using online learning mechanism to continuously adjust and optimize deep learning model parameters based on newly collected power module data; S5. Based on the trained model, the operating status of the power module is evaluated in real time and the fault development trend is predicted.

2. A delamination detection method for an intelligent power module according to claim 1, characterized in that: The method further comprises: S6, using explainable algorithms to show users the logic and reasons behind each fault detection and prediction; S7. Establish a user feedback mechanism and optimize the power module fault model based on user feedback; S8. Integrate and deploy the hierarchical detection method to the cloud platform or local server to achieve continuous stream processing and real-time update of power module data.

3. The delamination detection method for an intelligent power module according to claim 1, characterized in that: The step S1 specifically includes: S101, sampling the output signal of the power module to obtain sampling data, the sampling frequency f s Set to twice the maximum operating frequency of the module, f s =2×f max , where f max is the maximum operating frequency of the power module; S102, filtering the sampled data using a digital low-pass filter to remove high-frequency noise; S103, performing outlier detection on the filtered sampled data, and using a statistical method to identify and eliminate possible erroneous data; S104. Standardize the processed sampled data to make the data numerically stable during machine learning training.

4. The delamination detection method for an intelligent power module according to claim 1, characterized in that: In step S2, the extracted key features are fused and optimized as follows: All the extracted key features are integrated into a feature matrix X, where each row represents a sample and each column represents a feature, and then standardized, as shown below: Among them, X std represents the standardized data matrix, μ and σ represent the mean and standard deviation of the original data matrix X respectively; The covariance matrix is ​​calculated to show the linear relationship between data features, which is expressed as follows: Where n represents the number of samples, Σ represents the covariance matrix; The eigenvalue λ and eigenvector v of the covariance matrix Σ are calculated according to the covariance matrix, which are expressed as follows: 2v=λv; According to the size of the eigenvalue λ, the eigenvectors corresponding to the largest k eigenvalues ​​are selected to form a transformation matrix P, and then the selected eigenvectors are used to transform the original data into the new subspace, which is expressed as follows: Y=X std P Where Y is the transformed data matrix.

5. The delamination detection method for an intelligent power module according to claim 1, characterized in that: The structure of the deep learning model in step S3 includes a bottom layer model, a middle layer model and a top layer model; The underlying model is used to find a classification hyperplane that maximizes the margin between different classes; The mid-level model is used to improve the classification results through random forest; The top-level model is used to integrate the outputs of the bottom-level and middle-level models for fault prediction and health status assessment through a gradient boosting machine.

6. The delamination detection method for an intelligent power module according to claim 1, characterized in that: The step S4 specifically includes: S401, continuously monitor the data flow of the power module, online learning and update the model, as shown below: D memory =D old ∪D new Among them, D memory Indicates the fusion of old data D old and new data D new Datasets; S402. Use regularization technology to design the loss function to avoid destroying old task knowledge, as shown below: Among them, L new (θ) represents the loss function on new data, F i represents the diagonal elements in the fisherman information matrix from the old task, measuring the parameter θ i The importance of θ i,old represents the value of the parameter in the previous task, and λ represents the regularization strength; S403: Continue to monitor the power module and update the model.

7. The delamination detection method for an intelligent power module according to claim 1, characterized in that: The step S5 specifically includes: S501. Use the seasonal autoregressive integrated moving average model to predict the time series data of equipment performance, as shown below: Among them, Y t represents the equipment performance index at time t, B represents the back-shift operator, s represents the seasonal cycle, and denote non-seasonal and seasonal differences, ε t is the error term; S502, by calculating the mean μ of each window w and standard deviation σ w , dynamically set the threshold for real-time anomaly detection, as shown below: Threshold=μ w +kσ w S503, using machine learning technology to model and predict the monitoring trend of the power module, and the health score of the power module is dynamically adjusted according to the prediction result, as shown below: H(t)=β1f1(X t )+β2f2(X t )+…+β n f n (X t ) Among them, X t represents the feature vector at time t, f i represents the i-th decision tree in the random forest, β i Represents the weight of the corresponding tree; S504, regularly evaluating the accuracy of the prediction model based on the latest power module data, and automatically adjusting model parameters or retraining the model to adapt to new trends and data patterns.

8. The delamination detection method for an intelligent power module according to claim 2, characterized in that: In step S6, based on the explainability algorithm, additional contextual data is integrated and features are extracted from the additional contextual data, and added to the prediction calculation of each power module to dynamically adjust the weights. At the same time, a visualization tool is built to adjust and optimize the model based on user feedback.

9. The delamination detection method for an intelligent power module according to claim 2, characterized in that: In step S7, a feedback database is designed for user feedback, and then a key performance indicator formula of the power module is designed, which is expressed as follows: Among them, P eff represents the prediction efficiency, R time represents the average response time; Design optimization strategies based on the feedback database and key indicator results and provide feedback to users.

10. A layered detection system for an intelligent power module, characterized in that: The system comprises: A data collection module, used to collect data from the power module using sensors and pre-process the data using advanced noise filtering and outlier processing techniques; A key feature extraction module is used to extract key features from the preprocessed data, and then optimize the key features by applying automated feature selection technology to improve the fault identification capability of the model; wherein the basic statistical features extracted from the preprocessed data include time domain feature extraction, frequency domain feature extraction, and time-frequency domain feature extraction; The time domain feature extraction extracts basic statistical features from the standardized voltage and current signals. The basic statistical features include mean μ, standard deviation σ, skewness γ and kurtosis k, which are expressed as follows: Among them, x i represents a single data point, N represents the number of samples, and i represents data point i; The frequency domain feature extraction uses fast Fourier transform to convert the signal from the time domain to the frequency domain, and extracts the energy and frequency values ​​of the main frequency components, which are expressed as follows: Among them, X k represents the Fourier transform result at frequency k, E k represents the corresponding energy, and n represents the order of the filter; The time-frequency domain feature extraction uses short-time Fourier transform to analyze the signal, obtain the changes in frequency components at different times, and extract dynamic features, which are expressed as follows: S(t,f)=∫x(t+τ)w(τ)e -j2πfτ dτ Where S(t,f) represents the amplitude at time t and frequency f in the time-frequency representation; A preliminary fault detection module, which uses a deep learning model to perform preliminary fault detection on the power module, and then uses traditional algorithms such as decision trees to analyze and classify the fault causes; An online learning module, which is used to continuously adjust and optimize the deep learning model parameters based on newly collected power module data using an online learning mechanism; The real-time evaluation module evaluates the operating status of the power module in real time and predicts the fault development trend based on the trained model; The system also includes at least 0 of the following modules: The fault display module is used to show users the logic and reasons behind each fault detection and prediction through an explainable algorithm; Fault optimization module: establishes a user feedback mechanism and optimizes the fault model of the power module based on user feedback; The system integration module integrates and deploys the hierarchical detection method to the cloud platform or local server to achieve continuous stream processing and real-time update of power module data.

Images