Power module online evaluation and early warning system based on multi-parameter fusion

Abstract

The application discloses a power module online evaluation and early warning system based on multi-parameter fusion, relates to the technical field of power module evaluation, and comprises: a feature fusion module, which is used for monitoring multi-dimensional parameters of a power module in real time and generating a fusion feature set; a model construction and evaluation module, which is used for building a collaborative architecture of a time sequence trend evaluation model and an abnormal isolation evaluation model, the time sequence trend evaluation model is based on an improved LSTM network output state fitting degree, and the abnormal isolation evaluation model is based on an improved isolated forest output abnormal isolation degree; a fusion evaluation score is generated; a warning and tracing module, which is used for dividing a warning level according to the fusion score, calculating parameter contribution degrees and analyzing fault roots, and generating a structured warning log; and a model optimization module, which is used for collecting operation and maintenance feedback samples, fine-tuning model parameters through incremental training, and regularly evaluating model structure adaptability, so that the accuracy and timeliness of fault warning are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of power module evaluation technology, specifically to an online evaluation and early warning system for power modules based on multi-parameter fusion. Background Technology

[0002] As a core component of electronic devices, the reliability of power modules directly affects the stable operation of the entire system. Currently, power module monitoring technology mainly relies on single parameter threshold judgment or simple multi-parameter combination analysis. For example, by setting fixed thresholds for key parameters such as voltage and current, an alarm is triggered when the parameters exceed the preset range. Some existing technologies use traditional machine learning methods, such as support vector machines and random forests, to evaluate the state of power modules and train models using historical data to achieve anomaly detection.

[0003] However, existing technologies have obvious limitations. First, the single-parameter threshold method cannot effectively capture the complex correlations and temporal change patterns between parameters, which can easily lead to false alarms or missed alarms. For example, the parameters of power modules fluctuate greatly under dynamic loads, and fixed thresholds are difficult to adapt to different operating conditions, which may misjudge normal fluctuations as abnormalities. Second, traditional machine learning methods have limited ability to model temporal features and are difficult to accurately identify progressive or compound faults. For example, parameter drift caused by capacitor aging often has the characteristic of slow change, and traditional methods may fail to provide timely warnings due to insufficient sensitivity.

[0004] Furthermore, existing technologies lack dynamic optimization and adaptive adjustment mechanisms. Factors such as the aging of power modules and environmental changes can cause their performance to gradually degrade. However, existing methods typically use static models, which cannot dynamically adjust the evaluation strategy based on real-time operating data. For example, model parameters and warning thresholds are often fixed and cannot adapt to the state changes of the power module throughout its entire life cycle, resulting in a decrease in evaluation accuracy over time. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides an online evaluation and early warning system for power modules based on multi-parameter fusion. By real-time monitoring of the multi-dimensional parameters of the power module, fusion of time-series and correlation feature analysis, dynamic adjustment of model weights and early warning thresholds, it solves the problems of high false alarm and false negative rates of single-parameter threshold methods, insufficient time-series modeling capabilities of traditional machine learning models, and the inability of static models to adapt to the state changes of the power module throughout its entire life cycle, thus significantly improving the accuracy and timeliness of fault early warning.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention is realized through the following technical solutions: An online evaluation and early warning system for power modules based on multi-parameter fusion, including:

[0009] A feature fusion module, which is used to monitor the multi-dimensional parameters of the power module in real time, divide the parameter priorities through the fault impact weight method, and generate a fusion feature set by using normalization processing and time series correlation analysis;

[0010] A model construction and evaluation module, which is used to build a collaborative architecture of a time series trend evaluation model and an anomaly isolation evaluation model. The time series trend evaluation model outputs the state fitting degree based on an improved LSTM network, and the anomaly isolation evaluation model outputs the anomaly isolation degree based on an improved isolation forest; Dynamically weight and fuse the results of the two models according to the working conditions to generate a fusion evaluation score;

[0011] An early warning and traceability module, which is used to divide the early warning levels according to the fusion score, determine the anomaly type by combining the output results of the double models, calculate the parameter contribution degree and analyze the root cause of the fault, and generate a structured early warning log;

[0012] A model optimization module, which is used to collect operation and maintenance feedback samples, fine-tune the model parameters through incremental training, dynamically update the sample library, and regularly evaluate the adaptability of the model structure.

[0013] Furthermore, the multi-dimensional parameters of the power module include the output voltage accuracy of the power module, the output current fluctuation value, the case temperature rise rate, the capacitance equivalent series resistance, the ripple peak value, the fan speed, the input voltage stability, and the load regulation rate; The correlation degree of each parameter is calculated by the fault impact weight method, and the parameters are divided into three levels: the correlation degree > G2 is the first-level parameter; the correlation degree of the second-level parameter is G1~G2; the correlation degree < G1 is the third-level parameter.

[0014] Furthermore, for the collected single parameter value, calculate its normal fluctuation range. If the parameter value exceeds this range, verify the parameter change trend in the adjacent 3 acquisition cycles: calculate the trend slope through linear fitting. If the trend slope is opposite to the change direction of the abnormal value, it is determined as an invalid abnormal value and repaired by the cubic spline interpolation method; If the trend slope is consistent with the change direction of the abnormal value, it is determined as a potential fault precursor, and the data is retained and marked with a key monitoring label.

[0015] Furthermore, the parameter values are processed by working condition adaptive normalization, and at the same time, time series features and correlation features are extracted. The time series features include the maximum value, minimum value, mean value, variance, and change rate within m minutes, and the correlation features include the Pearson correlation coefficient between parameters, forming a fusion feature set containing multiple dimensions.

[0016] Furthermore, the time-series trend evaluation model employs an improved LSTM network, whose structure includes an input layer, an LSTM layer, an attention mechanism layer, a fully connected layer, and an output layer. The output layer outputs the original matching score in the 0-1 range through a Sigmoid activation function, and multiplies the original matching score by 100 to map it to a state fit score of 0-100. The anomaly isolation evaluation model employs an improved isolation forest, which contains N decision trees. When randomly splitting parameters, parameter priority weights are introduced, and the anomaly isolation degree is output by calculating the path length of the fused feature set in the normal sample distribution.

[0017] Furthermore, the power module's operating condition is monitored in real time. By collecting load rate and temperature data for three consecutive collection cycles, it is determined whether the current operating condition is steady-state or dynamic. Based on the operating condition, the weights of the time-series trend evaluation model and the anomaly isolation evaluation model are set respectively. The fusion evaluation score is calculated by using the linear weighted summation method based on the state fit degree and the anomaly isolation degree.

[0018] Furthermore, the fusion score is divided into four warning levels: normal level (85-100 points), attention level (70-84 points), general warning level (50-69 points), and emergency warning level (<50 points). The threshold is dynamically calibrated based on the power module's runtime and aging degree: for every year of runtime increase, the warning level threshold is lowered by 1 point, with a maximum reduction of 10 points; the aging degree is corrected by ESR aging rate, with an additional reduction of 2 points when the ESR aging rate is >50%.

[0019] If the state fit score is <70 and the anomaly isolation score is <60, it is judged as a progressive anomaly; if the state fit score is ≥70 and the anomaly isolation score is ≥80, it is judged as a sudden anomaly; if the state fit score is <70 and the anomaly isolation score is ≥80, it is judged as a complex anomaly.

[0020] Furthermore, the top 3 parameters with the highest contribution are extracted; the root cause is analyzed by combining correlation features: if the top 3 parameters include temperature and fan speed, and the correlation features between the two show a negative correlation, it is determined that the abnormal temperature is caused by fan failure; if the top 3 parameters include ESR and ripple peak value, and the correlation features between the two show a positive correlation, it is determined that the capacitor aging failure is caused; the timestamp, real-time operating conditions, fusion score, anomaly type and top 3 parameters of each warning are automatically recorded to form a structured warning log.

[0021] Furthermore, cases confirmed as faults within 24 hours after the warning signal is triggered are marked as valid samples; cases that are found to be fault-free after the warning signal is triggered and operate normally for 72 consecutive hours are marked as calibration samples; and cases that do not trigger a warning but experience a fault within 72 hours are marked as incremental samples.

[0022] For valid samples, the attention weight of the corresponding fault precursor parameters in the time-series trend assessment model is increased, with a single adjustment increment of 0.1, and the adjusted weight ≤ 0.5; for corrected samples, the priority weight of misjudged parameters in the anomaly isolation assessment model is reduced, with a single adjustment increment of 0.05, and the adjusted weight ≥ 0.05.

[0023] Furthermore, a two-tiered structure of a basic sample library and an incremental sample library is constructed. The basic sample library stores typical samples throughout the entire lifecycle of the power module, including normal samples and various abnormal samples. The incremental sample library stores collected valid samples and supplementary samples. For every M incremental samples accumulated, samples with a feature similarity ≥ the similarity threshold with the basic sample library are deleted. The similarity is calculated using cosine similarity. Valid samples and supplementary samples are retained. Incremental samples are integrated into the basic sample library at a fixed ratio, while the oldest sample of the same number in the basic sample library is deleted.

[0024] Periodically perform an adaptive evaluation of the model structure, calculate the ESR aging rate of the power module, and if the aging rate is greater than the aging rate threshold, increase the number of neurons in the LSTM layer of the time-series trend evaluation model by a fixed proportion; count the number of newly added anomaly types, and if there are ≥3 new anomaly types, increase the number of decision trees in the anomaly isolation evaluation model by a fixed proportion; and perform offline validation again after structural optimization.

[0025] (III) Beneficial Effects

[0026] This invention provides an online evaluation and early warning system for power modules based on multi-parameter fusion, which has the following beneficial effects:

[0027] (1) By monitoring the multi-dimensional parameters of the power module in real time, the priority is divided by the fault impact weight method. Combined with normalization processing and time series correlation analysis, a fusion feature set is generated to improve the comprehensiveness and accuracy of parameter monitoring, effectively distinguish between invalid anomalies and potential fault precursors, and enhance the adaptability to complex working conditions through dynamic weight allocation and time series feature extraction. This provides a high-quality data foundation for subsequent model evaluation, significantly reduces false alarms and false alarms, and thus improves the timeliness and reliability of fault warning.

[0028] (2) Through the collaborative architecture of the improved LSTM network and the isolated forest algorithm, the accurate evaluation of the power module status is realized. The improved LSTM network, combined with the attention mechanism, effectively captures the temporal characteristics and improves the sensitivity to progressive faults. The isolated forest introduces parameter weights to enhance the pertinence of anomaly detection. The dynamic weighted fusion strategy adapts to different working conditions and improves the reliability of the evaluation results. Through the comprehensive scoring of state fitting degree and anomaly isolation degree, the false alarm and false alarm rates are significantly reduced, providing a high-precision basis for fault early warning.

[0029] (3) By dynamically classifying the early warning level through fusion score, and accurately identifying the progressive, sudden and compound anomaly types by combining the state fitting degree and the anomaly isolation degree, the contribution of the top 3 parameters is calculated to analyze the root cause of the fault, realize a multi-level early warning response mechanism, improve the accuracy of fault classification, quickly locate the fault source through parameter contribution, shorten the investigation time, record complete diagnostic information in the structured early warning log, support retrospective analysis and decision optimization, dynamically calibrate the threshold to adapt to equipment aging, reduce the false alarm and missed alarm rate, and significantly improve the system reliability and operation and maintenance efficiency.

[0030] (4) Through incremental training and dynamic sample library updates, the model is continuously optimized and adaptively adjusted. The model parameters are fine-tuned based on operation and maintenance feedback samples to improve the sensitivity to specific faults and the ability to suppress false alarms. The model structure is dynamically adjusted through a two-layer sample library structure and periodic evaluation to adapt to the aging of power modules and the addition of new abnormal types. The optimization effect is verified by combining performance indicators to ensure that the model maintains high accuracy throughout its entire life cycle and significantly improves the long-term reliability and adaptability of the system. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the online evaluation and early warning system for power modules based on multi-parameter fusion according to the present invention;

[0032] Figure 2 This is a schematic diagram of the online evaluation and early warning process for power modules based on multi-parameter fusion, as described in this invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Please see Figures 1-2 This invention provides an online evaluation and early warning system for power modules based on multi-parameter fusion, comprising: a feature fusion module, a model construction and evaluation module, an early warning and source tracing module, and a model optimization module, wherein:

[0035] The feature fusion module is used to monitor the multi-dimensional parameters of the power module in real time. It prioritizes parameters by using the fault impact weighting method and generates a fused feature set by normalization processing and time series correlation analysis.

[0036] Specifically, select the output voltage accuracy, output current fluctuation value, case temperature rise rate, capacitance equivalent series resistance (ESR), ripple peak value, fan speed, input voltage stability, and load regulation rate of the power supply module as core parameters; use the fault impact weight method to calculate the correlation degree of each parameter, correlation degree = (number of faults caused by the abnormality of this parameter / total number of faults) × 100%, combined with historical fault data, at least including 500 fault records, and divide the parameters into three levels: first-level parameters, such as temperature, voltage accuracy, current fluctuation, correlation degree > G2, such as 70%; second-level parameters, such as ripple peak value, ESR, correlation degree G1~G2; third-level parameters, such as fan speed, input voltage stability, load regulation rate, correlation degree < G1, such as 40%, G2 > G1, and the value ranges of G1 and G2 are 0~100%. At the same time, set the priority weights of each level of parameters, and the value range of the priority weights is 0~1. By default, the weight of the first-level parameter is 0.5, the weight of the second-level parameter is 0.3, and the weight of the third-level parameter is 0.2;

[0037] Real-time monitor the load rate of the power supply module. When the load rate < 30% is the light load condition, the acquisition frequency is set to 1 time / minute. When 30% ≤ load rate ≤ 80% is the rated load condition, it is set to 1 time / 10 seconds. When the load rate > 80% is the heavy load condition, it is set to 1 time / second. When the load rate change rate > 10% / second is the impact load condition, it is set to 1 time / second, and automatically switch the acquisition frequency;

[0038] For the single parameter value collected, calculate its normal fluctuation range, based on the ±3σ interval of the normal operation data in the past 3 months, where σ is the standard deviation; if the parameter value exceeds this range, verify the parameter change trend in the adjacent 3 acquisition cycles: calculate the trend slope through linear fitting. If the trend slope is opposite to the change direction of the abnormal value, such as the abnormal value suddenly increases but the trend slope is negative, it is determined as an invalid abnormal value and repaired by the cubic spline interpolation method. With time as the horizontal axis and the parameter value as the vertical axis, interpolate the 5 effective data points before and after; if the trend slope is the same as the change direction of the abnormal value, it is determined as a potential fault precursor, retain the data and mark the key monitoring label;

[0039] Adopt the working condition adaptive normalization formula: normalization value = (actual parameter value - parameter minimum value) / (parameter maximum value - parameter minimum value), where the parameter maximum value and parameter minimum value are taken from the factory rated range of this power supply module; the normalization weight coefficient is set according to the priority, and the value range of the weight coefficient is from 0 to 1. By default, it is 0.4 for the first-level parameter, 0.3 for the second-level parameter, and 0.3 for the third-level parameter; at the same time, extract the time series features and correlation features. The time series features include the maximum value, minimum value, mean value, variance, and change rate within m minutes (such as 5 minutes, 10 minutes), and the correlation features include the Pearson correlation coefficient between parameters, forming a fusion feature set including multiple dimensions;

[0040] Calculate the data integrity and consistency of the fused feature set. Data integrity: effective data percentage = effective data volume / total collected data volume. Consistency: feature change rate between adjacent periods = |current feature value - previous period feature value| / previous period feature value. If integrity < 95%, trigger the acquisition module self-test, and check the sensor power supply, signal line impedance, and AD sampling accuracy in sequence. If consistency > 20%, return to re-execute the outlier repair. If both meet the criteria, output the fused feature set.

[0041] By monitoring the multidimensional parameters of the power module in real time, prioritizing them using the fault impact weighting method, and generating a fusion feature set by combining normalization processing and time-series correlation analysis, the comprehensiveness and accuracy of parameter monitoring are improved. This effectively distinguishes invalid anomalies from potential fault precursors. Through dynamic weight allocation and time-series feature extraction, the adaptability to complex operating conditions is enhanced. This provides a high-quality data foundation for subsequent model evaluation, significantly reduces false alarms and false negatives, and thus improves the timeliness and reliability of fault early warning.

[0042] The model building and evaluation module is used to build a collaborative architecture between the time-series trend evaluation model and the anomaly isolation evaluation model. The time-series trend evaluation model outputs the state fit based on an improved LSTM network, and the anomaly isolation evaluation model outputs the anomaly isolation degree based on an improved isolated forest. The results of the two models are dynamically weighted and fused according to the working conditions to generate a fused evaluation score.

[0043] Specifically, a collaborative architecture is constructed for the time-series trend assessment model and the anomaly isolation assessment model. The time-series trend assessment model adopts an improved LSTM network. The network structure includes an input layer, an LSTM layer, an attention mechanism layer, a fully connected layer, and an output layer. The input layer takes in multi-dimensional fusion features. The LSTM layer has 128 neurons with the tanh activation function and a forgetting gate threshold of 0.5. The attention mechanism layer assigns attention to the 128-dimensional features output by the LSTM according to parameter priority weights, generating a weighted feature vector. The fully connected layer has 64 neurons with the ReLU activation function. The output layer outputs the state fit. The state fit generation logic is as follows: the output layer uses a Sigmoid function... The live function outputs the original matching score in the range of 0 to 1. This score reflects the similarity between the current fused features and the normal temporal features learned by the model. Multiplying the original matching score by 100 maps it to a state fit score of 0 to 100. The higher the score, the better the current parameter temporal changes match the normal operating conditions, and the more stable the operating state. The anomaly isolation evaluation model uses an improved isolated forest, which contains N decision trees, such as 100 or 150 trees. When randomly splitting parameters, parameter priority weights are introduced. The splitting probability = parameter priority weight / sum of all parameter priority weights.

[0044] Data from the entire lifecycle of the power module was collected and divided into training, validation, and test sets in a 7:2:1 ratio. An offline time-series trend evaluation model was trained using the fused feature set of the training set as input and normal time-series patterns as labels. Normal time-series patterns are defined as time-series features without anomalies over the past three months. Mean squared error (MSE) was used as the loss function, and training was continued until the validation set loss value was <0.05. An anomaly isolation evaluation model was trained using the fused feature set of normally operating data from the training set as input, and training was continued until the anomaly identification accuracy was >90%. During model deployment, pre-trained parameters were initialized and loaded online, and the factory parameters and historical operating data of the current power module over the past three months were read to complete the model's adaptation to the target power module.

[0045] The current fused feature set is simultaneously input into two models for parallel inference. The time-series trend evaluation model outputs the state fit score (0-100 points) based on the learned normal time-series pattern, with a higher score indicating a better fit between the current state and the normal time-series. The anomaly isolation evaluation model outputs the anomaly isolation score (0-100 points) based on the path length of the feature set in the normal sample distribution, with a higher score indicating a greater deviation from the normal distribution.

[0046] The system monitors the power module's operating condition in real time. By collecting load rate and temperature data over three consecutive collection cycles, it determines whether the operating condition is steady-state or dynamic. Steady-state conditions are defined as load fluctuation ≤ 5% and temperature fluctuation ≤ 2℃, while dynamic conditions are defined as load fluctuation > 5% or temperature fluctuation > 2℃. Under steady-state conditions, the system sets the weights for the time-series trend evaluation model and the anomaly isolation evaluation model, with a total range of 0 to 1. The default weight for the time-series trend evaluation model is set to 0.7, and the weight for the anomaly isolation evaluation model is set to 0.3. Under dynamic conditions, the weights are adjusted to 0.5:0.5. The fusion evaluation score is calculated using the formula: Fusion Score = State Fit Degree × Weight 1 + (100 - Anomaly Isolation Degree) × Weight 2, with a score range of 0 to 100.

[0047] Calculate the Pearson correlation coefficient r between the evaluation results of the two models. If r ≥ 0.6, the results are considered consistent, and the fusion score is output. If r < 0.6, return to extract the fusion feature set again and infer again. If r < 0.6 for 3 consecutive times, trigger the subsequent online fine-tuning process of the model.

[0048] By employing a collaborative architecture of an improved LSTM network and an isolated forest algorithm, accurate assessment of the power module status is achieved. The improved LSTM network, combined with an attention mechanism, effectively captures temporal features and enhances sensitivity to progressive faults. The isolated forest algorithm introduces parameter weights to enhance the targeting of anomaly detection. The dynamic weighted fusion strategy adapts to different operating conditions and improves the reliability of the assessment results. Through a comprehensive score of state fit and anomaly isolation, the false alarm and false negative rates are significantly reduced, providing a high-precision basis for fault early warning.

[0049] The early warning and source tracing module is used to classify early warning levels based on the fusion score, identify the anomaly type by combining the output results of the dual models, calculate the parameter contribution and analyze the root cause of the fault, and generate structured early warning logs.

[0050] Specifically, the fusion score is divided into four warning levels: normal level (85-100 points), attention level (70-84 points), general warning level (50-69 points), and emergency warning level (<50 points). The threshold is dynamically calibrated based on the power module's runtime and aging degree: for every year the runtime increases, the warning level threshold is lowered by 1 point, with a maximum reduction of 10 points. The aging degree is corrected by the ESR aging rate, which is calculated as (current ESR - initial ESR) / initial ESR × 100%. When the ESR aging rate is >50%, an additional 2 points are deducted.

[0051] The anomaly type is determined by combining the output results of the two models. If the state fit score is <70 and the anomaly isolation score is <60, it is determined to be a gradual anomaly, such as slow parameter drift caused by capacitor aging. If the state fit score is ≥70 and the anomaly isolation score is ≥80, it is determined to be a sudden anomaly, such as current surge caused by load change. If the state fit score is <70 and the anomaly isolation score is ≥80, it is determined to be a compound anomaly, such as aging combined with fan failure.

[0052] Based on the warning level and anomaly type, corresponding signals are generated. For the normal level, only the operation log is recorded to the local database; for the attention level, a yellow LED indicator is output to alert, and "[Parameter Name] slight deviation, it is recommended to strengthen monitoring" is displayed on the monitoring interface; for the warning level, an audible and visual alarm is triggered, and an operation and maintenance notification is pushed to the management personnel terminal; for the emergency warning level, in addition to the audible and visual alarm and notification, load reduction protection is activated, and the load controller is controlled through the PWM signal to reduce the load rate to below 50%, and a snapshot of the current operating parameters is saved.

[0053] Calculate the contribution of each parameter to the fusion score: Contribution = Parameter weight × Parameter anomaly deviation, where Parameter anomaly deviation = |Parameter normalized value - 0.5| / 0.5; Extract the top 3 parameters with the highest contribution; Analyze the root cause based on correlation features: If the top 3 parameters include temperature and fan speed, and the correlation features between the two show a negative correlation, it is determined that the abnormal temperature is caused by a fan failure; If the top 3 parameters include ESR and ripple peak value, and the correlation features between the two show a positive correlation, it is determined that the capacitor aging failure is the cause; Automatically record the timestamp, real-time operating conditions, fusion score, anomaly type, and top 3 parameters for each warning to form a structured warning log;

[0054] By dynamically classifying early warning levels through score fusion, and accurately identifying progressive, sudden, and compound anomaly types by combining state fit and anomaly isolation, and calculating the contribution of the top 3 parameters to analyze the root cause of the fault, a multi-level early warning response mechanism is realized, improving the accuracy of fault classification. The fault source can be quickly located through parameter contribution, shortening the troubleshooting time. The structured early warning log records complete diagnostic information, supports retrospective analysis and decision optimization, and dynamically calibrates thresholds to adapt to equipment aging, reducing false alarm and missed alarm rates, and significantly improving system reliability and operation and maintenance efficiency.

[0055] The model optimization module is used to collect operation and maintenance feedback samples, fine-tune model parameters through incremental training, dynamically update the sample library, and periodically evaluate the adaptability of the model structure.

[0056] Specifically, collect operation and maintenance information feedback: cases confirmed as faults within 24 hours after the warning signal is triggered are marked as valid samples; cases that are found to be fault-free after the warning signal is triggered and operate normally for 72 consecutive hours are marked as correction samples; cases that do not trigger a warning but fail within 72 hours are marked as incremental samples; each sample needs to be associated with the warning log and the fault investigation report, including the fault location and maintenance measures;

[0057] The dual-model parameters are fine-tuned using incremental training. The online fine-tuning process includes: for valid samples, increasing the attention weight of the corresponding fault precursor parameters in the time-series trend assessment model, such as for capacitor aging samples, increasing the attention weight of the ESR parameter from the initial 0.2 to 0.3, with a single adjustment increment of 0.1, and the adjusted weight ≤ 0.5; for correction samples, reducing the priority weight of misjudged parameters in the anomaly isolation assessment model, such as for false alarms caused by fan speed fluctuations, decreasing the priority weight of fan speed from 0.2 to 0.15, with a single adjustment increment of 0.05, and the adjusted weight ≥ 0.05.

[0058] A two-tiered structure of a basic sample library and an incremental sample library is constructed. The basic sample library stores typical samples throughout the entire lifecycle of the power module, including normal samples and various abnormal samples. The incremental sample library stores collected valid samples and supplementary samples. Every M new samples (e.g., 100 or 150) are added, samples with a feature similarity ≥95% with the basic sample library are deleted (similarity is calculated using cosine similarity). Valid samples and supplementary samples are retained. 10% of the incremental samples are integrated into the basic sample library, while the same number of the oldest samples in the basic sample library are deleted.

[0059] Periodically evaluate the model structure for adaptation. For example, every three months, calculate the ESR aging rate of the power module. If the aging rate is >50%, increase the number of LSTM layer neurons in the time-series trend evaluation model by a fixed percentage (e.g., 10% or 15%). Count the number of newly added anomaly types. If there are ≥3 new anomaly types, increase the number of decision trees in the anomaly isolation evaluation model by a fixed percentage. After structural optimization, re-perform offline validation to ensure that the model performance is improved by ≥5%.

[0060] The optimization effect is verified through three core indicators: early warning accuracy, false alarm rate, and false negative rate. Early warning accuracy = number of true early warnings / total number of early warnings; false alarm rate = number of false early warnings / total number of early warnings ≤ 5%; false negative rate = number of faults without early warning / total number of faults ≤ 3%. To meet the target, the early warning accuracy must be ≥ 90%, the false alarm rate ≤ 5%, and the false negative rate ≤ 3%. If the target is met, the optimized model parameters and sample library are saved. If the target is not met, the parameters are readjusted, i.e., the online fine-tuning process is implemented. If the target is not met after 3 consecutive optimizations, the model reconstruction process is triggered, and the model building and pre-training in the model building and evaluation module are re-executed.

[0061] Through incremental training and dynamic sample library updates, continuous optimization and adaptive adjustment of the model are achieved. Model parameters are fine-tuned based on operation and maintenance feedback samples to improve sensitivity to specific faults and the ability to suppress false alarms. Through a two-layer sample library structure and periodic evaluation, the model structure is dynamically adjusted to adapt to the aging of power modules and the addition of new anomaly types. The optimization effect is verified by combining performance indicators to ensure that the model maintains high accuracy throughout its entire life cycle, significantly improving the long-term reliability and adaptability of the system.

[0062] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The coefficients in the formulas are set by those skilled in the art according to the actual situation.

[0063] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, and combinations thereof. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0064] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0065] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A power module online evaluation and early warning system based on multi-parameter fusion, characterized in that: Comprise: The feature fusion module is used for real-time monitoring of the multi-dimensional parameters of the power module, the multi-dimensional parameters of the power module including the output voltage accuracy, output current fluctuation value, shell temperature rise rate, equivalent series resistance of the capacitor, ripple peak value, fan speed, input voltage stability and load regulation of the power module; the fault influence weight method is used to calculate the correlation degree of each parameter, and the parameters are divided into three levels: the correlation degree > G2 is the first level parameter; the correlation degree of the second level parameter is G1~G2; the correlation degree < G1 is the third level parameter; the normalized processing and time sequence correlation analysis are used to generate a fusion feature set; The model construction and evaluation module is used for building a collaborative architecture of a time sequence trend evaluation model and an abnormal isolation evaluation model, the time sequence trend evaluation model adopts an improved LSTM network, the network structure including an input layer, an LSTM layer, an attention mechanism layer, a full connection layer and an output layer; the output layer outputs the original matching score in the interval of 0~1 through the Sigmoid activation function, and the original matching score is multiplied by 100 to map the state fitting degree to 0~100 points; the abnormal isolation evaluation model adopts an improved isolation forest, containing N decision trees, introducing a parameter priority weight when randomly dividing parameters, the division probability = parameter priority weight / sum of all parameter priority weights, and outputting the abnormal isolation degree by calculating the path length of the fusion feature set in the normal sample distribution; the results of the two models are dynamically weighted and fused according to the working conditions to generate a fusion evaluation score; The early warning and tracing module is used for dividing the early warning level according to the fusion score, combining the output results of the double models to determine the abnormal type, if the state fitting degree < 70 points and the abnormal isolation degree < 60 points, it is determined as a gradual abnormality; if the state fitting degree ≥ 70 points and the abnormal isolation degree ≥ 80 points, it is determined as a sudden abnormality; if the state fitting degree < 70 points and the abnormal isolation degree ≥ 80 points, it is determined as a compound abnormality; the parameter contribution degree is calculated, the contribution degree = parameter weight × parameter abnormal deviation degree, and the Top3 parameters with the highest contribution degree are extracted; the root cause is analyzed combined with the correlation characteristics: if the Top3 parameters contain temperature and fan speed, and the correlation characteristics of the two show a negative correlation, it is determined that the fan failure causes the temperature abnormality; if the Top3 parameters contain ESR and ripple peak value, and the correlation characteristics of the two show a positive correlation, it is determined that the capacitor aging failure; the timestamp, real-time working condition, fusion score, abnormal type and Top3 parameters of each early warning are automatically recorded to form a structured early warning log; The model optimization module is used for collecting operation and maintenance feedback samples, fine-tuning the model parameters through incremental training, dynamically updating the sample library, and regularly evaluating the model structure adaptability.

2. The multi-parameter fusion based power module online evaluation and early warning system according to claim 1, characterized in that: For the collected single parameter value, the normal fluctuation range is calculated, if the parameter value exceeds the range, the parameter trend of the adjacent 3 collection periods is verified: the trend slope is calculated through linear fitting, if the trend slope is opposite to the change direction of the abnormal value, it is determined as an invalid abnormal value, which is repaired through the cubic spline interpolation method; if the trend slope is consistent with the change direction of the abnormal value, it is determined as a potential failure precursor, the single parameter value is retained and marked with a key monitoring label.

3. The multi-parameter fusion based power module online evaluation and early warning system according to claim 2, characterized in that: The parameter values are normalized by self-adaptive processing, and the time sequence features and correlation features are extracted. The time sequence features include the maximum, minimum, mean, variance and change rate within m minutes, and the correlation features include the Pearson correlation coefficient between parameters, forming a fusion feature set containing multiple dimensions.

4. The multi-parameter fusion based power module online evaluation and early warning system according to claim 1, characterized in that: The working condition of the power module is monitored in real time. The load rate and temperature data of the last three acquisition cycles are used to determine whether the current working condition is steady or dynamic. The weights of the time trend evaluation model and the abnormal isolation evaluation model are set according to the working condition. The fusion evaluation score is calculated by linear weighted summation method based on the state fitting degree and abnormal isolation degree.

5. The multi-parameter fusion based power module online evaluation and early warning system according to claim 1, characterized in that: The fusion score is divided into four warning levels: normal level 85-100 points, attention level 70-84 points, general warning level 50-69 points, and emergency warning level <50 points. The threshold value is dynamically calibrated based on the running time and aging degree of the power module. For every 1 year of running time, the warning level threshold is lowered by 1 point, with a maximum reduction of 10 points. The aging degree is corrected by ESR aging rate. When the ESR aging rate is greater than 50%, the threshold is lowered by an additional 2 points.

6. The multi-parameter fusion based power module online evaluation and early warning system according to claim 1, characterized in that: Cases confirmed as failures within 24 hours after the warning signal is triggered are marked as valid samples. Cases that are not faulty and run normally for 72 hours after the warning signal is triggered are marked as correction samples. Cases that do not trigger a warning but fail within 72 hours are marked as incremental samples. For valid samples, the attention weight of the corresponding failure precursor parameter in the time trend evaluation model is enhanced, with a single adjustment range of 0.

1. The adjusted weight is ≤0.

5. For correction samples, the priority weight of the misjudgment parameter in the abnormal isolation evaluation model is reduced, with a single adjustment range of 0.

05. The adjusted weight is ≥0.

05.

7. The multi-parameter fusion based power module online evaluation and early warning system according to claim 6, characterized in that: A dual-layer structure of basic sample library and incremental sample library is constructed. The basic sample library stores typical samples of the power module throughout its life cycle, including normal samples and various abnormal samples. The incremental sample library stores valid samples and supplementary samples. Every M incremental samples are deleted if their similarity to the features in the basic sample library is greater than or equal to the similarity threshold. The similarity is calculated by cosine similarity. The valid samples and supplementary samples are retained. The incremental samples are integrated into the basic sample library at a fixed ratio, and the oldest samples in the basic sample library are deleted at the same ratio. The model structure is periodically evaluated for adaptability. The ESR aging rate of the power module is calculated. If the aging rate is greater than the aging rate threshold, the number of LSTM layer neurons in the time trend evaluation model is increased by a fixed ratio. The number of new abnormal types is counted. If the number of new abnormal types is greater than or equal to 3, the number of decision trees in the abnormal isolation evaluation model is increased by a fixed ratio. The structure is optimized and re-verified offline.

Images