MLCC residual life prediction method based on machine learning

Abstract

The invention belongs to the technical field of MLCC reliability evaluation, and particularly relates to an MLCC residual life prediction method based on a machine learning model, and the method comprises the steps: S1, carrying out the high-acceleration life test of an MLCC, and collecting an aging current response curve of the MLCC; s2, extracting features from the current response curve; s3, the extracted features are optimized and screened, and a feature combination is obtained; and S4, constructing a machine learning model by using the obtained feature combination so as to predict the life of the MLCC. According to the method, the service life of the individual multilayer ceramic capacitor can be accurately predicted in real time, the equipment maintenance strategy is optimized, and the fault occurrence rate is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of MLCC reliability evaluation, and particularly relates to a method for predicting the remaining life of MLCC based on a machine learning model. Background Art

[0002] Multi-Layer Ceramic Capacitors (MLCCs) are widely used in fields such as consumer electronics, automotive, and medical applications. Their reliability issues have become increasingly prominent. Especially under extreme conditions such as high voltage and high temperature, the long-term stability of MLCCs is affected by factors such as oxygen vacancy migration, resulting in a gradual decline in the performance of the capacitors and even failure. Traditional life prediction methods mainly rely on predicting the mean time to failure of a group of capacitors. However, these methods are difficult to accurately evaluate the remaining life of an individual capacitor, especially in the case of early failures and large individual differences.

[0003] Existing research mainly focuses on predicting the average life of capacitors through physical degradation models or empirical models based on historical data. However, these methods often ignore the unique behavior of an individual capacitor under specific operating conditions and have limited accuracy in real-time monitoring and early fault prediction.

[0004] Therefore, how to accurately and real-time predict the service life of individual multi-layer ceramic capacitors has become a key issue in improving the reliability of capacitors and optimizing maintenance strategies. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention provides a method for predicting the remaining life of MLCC based on machine learning, which can accurately and real-time predict the service life of individual multi-layer ceramic capacitors, optimize equipment maintenance strategies, and reduce the failure rate.

[0006] A method for predicting the life of MLCC based on machine learning includes: Step S1: Conduct a high-accelerated life test on the MLCC and collect its aging current response curve; Step S2: Extract features from the current response curve; Step S3: Optimize and screen the extracted features to obtain a feature combination; Step S4: Use the obtained feature combination to construct a machine learning model to predict the life of the MLCC.

[0007] As an optional implementation, the method for predicting the life of MLCC based on machine learning further includes: Step S5: Optimize the constructed machine learning model to ensure prediction accuracy, and use cross-validation and resampling methods for model evaluation to improve the generalization ability of the model and avoid overfitting.

[0008] As an alternative implementation, the high-accelerated life test on the MLCC includes: aging the MLCC under preset high temperature and preset high voltage conditions, recording its current response curve until the MLCC fails or malfunctions.

[0009] As an alternative implementation, the features extracted from the current response curve at least include: The starting time t of the current observation period (defined as the current curve interval for feature extraction) 1 and the corresponding current value I 1 , the ending time t of the current observation period 2 and the corresponding current value I 2 , and the slope k of the current change.

[0010] As an alternative implementation, the starting time t of the current observation period 1 and the ending time t of the current observation period 2 satisfy: t 2 = nt 1 , where n is the multiple of the current observation period.

[0011] As an alternative implementation, optimizing and screening the extracted features to obtain a feature combination includes: Taking uniformly spaced values of n from the m-th preset number to the o-th preset number, performing machine learning modeling, obtaining the trend of model performance varying with the n value, and determining the optimal n value; After determining the optimal n value, performing SHAP analysis on the features corresponding to the optimal n value to analyze the feature contribution degree and the dependence between features; And performing permutation and combination on the features, respectively modeling to obtain performance, and selecting a group of feature combinations with the best performance as the optimal feature combination.

[0012] As an alternative implementation, using the obtained feature combination to construct a machine learning model to predict the life of the MLCC includes: Based on the prediction accuracy and training speed, comparing the predictions of preset algorithms, and taking the random forest model as the optimal model of the machine learning model. The preset algorithms at least include random forest, XGBoost, gradient boosting decision tree, support vector regression, artificial neural network (these were actually added just to make up the numbers) model.

[0013] As an alternative implementation, the MLCC life prediction method based on machine learning further includes: enhancing the robustness of the model through 10-fold cross-validation and optimizing the hyperparameters through grid search.

[0014] In a second aspect of the present invention, there is provided a readable storage medium storing a computer program, which when executed by a processor performs the steps of the machine learning-based MLCC life prediction method as described in the first aspect of the present invention.

[0015] The beneficial effects of the present invention are as follows. By combining high-accelerated life tests and machine learning models, the remaining life of MLCCs can be accurately and quickly predicted. Especially in fields with high reliability requirements, such as automotive and medical equipment, the reliability and safety of the system can be significantly improved. An efficient and accurate prediction tool is provided, which helps to optimize equipment maintenance strategies and reduce the failure rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic flowchart of a method for predicting the remaining life of MLCCs based on machine learning provided by an embodiment of the present invention;

[0017] Figure 2 It is a current curve graph obtained by an MLCC in a highly accelerated aging experiment provided by an embodiment of the present invention;

[0018] Figure 3 It is a schematic diagram of a feature extraction method provided by an embodiment of the present invention;

[0019] Figure 4 It is a schematic flowchart of another method for predicting the remaining life of MLCCs based on machine learning provided by an embodiment of the present invention;

[0020] Figure 5 It is a schematic diagram of the basic process of machine learning modeling provided by an embodiment of the present invention;

[0021] Figure 6 It is a schematic diagram of the prediction result of the best model provided by an embodiment of the present invention.

[0022] Figure 7 It is a comparison graph of the effects of different machine learning models of an actual MLCC life prediction method provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] To further illustrate the content, features and actual effects of the present invention, the present invention will be described in detail below with reference to embodiments. It should be noted that the modified methods designed in the present invention are not limited to these specific embodiments. Without departing from the spirit and connotation of the design of the present invention, equivalent replacements and modifications made by those skilled in the art on the basis of reading the content of the present invention are also within the scope of protection required by the present invention.

[0024] In the field of capacitor life prediction, ordinary capacitors (such as electrolytic capacitors) usually rely on indicators such as capacitance decay, equivalent series resistance (ESR) change, and the number of charge and discharge cycles to evaluate their life. This is because the decline of electrolytic capacitors mainly stems from the evaporation of the electrolyte, corrosion, and decomposition of the electrolyte, resulting in an increase in their internal impedance and a decrease in capacitance. Therefore, by measuring the change in capacitance or ESR, the aging state of the capacitor can be effectively reflected.

[0025] However, for MLCCs, due to the differences in their working principles and material structures, the aging mechanism is more complex. The dielectric of MLCCs is usually composed of ceramic materials, and their aging process often manifests as cracks, bubbles, or damage caused by stress in the ceramic materials. These changes are not easily captured accurately by capacitance decay or ESR change. Therefore, leakage current, as an important performance parameter of MLCCs, can more sensitively reflect their internal damage and aging process, especially in high-voltage and high-temperature environments.

[0026] As the usage time of MLCCs increases, the leakage current usually shows a gradually increasing trend, and this change is often closely related to the deterioration of the ceramic dielectric and crack propagation.

[0027] Therefore, leakage current becomes a more intuitive and sensitive indicator for evaluating the life of MLCCs. Compared with indicators such as capacitance decay and equivalent series resistance (ESR) change, the change in leakage current is more sensitive and can reflect the aging process of MLCCs under harsh conditions such as high voltage and high temperature at an early stage, providing a more accurate life prediction.

[0028] In view of this, the first aspect of the present invention provides a method for predicting the remaining life of MLCCs based on machine learning, mainly aiming at scenarios with high reliability requirements for MLCCs, and performing life prediction through a method combining high-accelerated life testing and machine learning models.

[0029] As Figure 1 shown, the present invention provides a method for predicting the remaining life of MLCCs based on machine learning, mainly including the following steps.

[0030] Before performing the steps of the method, it is necessary to configure the MLCC model for modeling, and correspondingly select a part of the samples of the same MLCC model to participate in life prediction, and apply different stress conditions to cover different application scenarios and performance requirements. These samples should be representative and diverse, and can reflect various situations of MLCCs in actual use.

[0031] Step S1: High-accelerated life test stage. Specifically, it means performing a high-accelerated life test on the MLCC and collecting its aging current response curve.

[0032] Specifically, the present invention performs a highly accelerated life test on MLCCs to obtain the current response curve during the capacitor aging process; by setting different temperature and voltage conditions, and in order to have a wide and uniform life range, an accelerated aging test is carried out by applying a constant voltage load.

[0033] In one embodiment of the present invention, the highly accelerated life test on MLCCs includes: aging the MLCCs under preset high temperature and preset high voltage conditions, recording their current response curves until the MLCCs fail.

[0034] Since the highly accelerated life test stage requires a capacitor aging board that matches the sample size, multiple capacitors can be placed in the experiment at one time.

[0035] Specifically, let the capacitor operate under preset high temperature such as 120 °C to 150 °C and preset high voltage such as 400 V to 600 V, record the current response curve until the capacitor fails or malfunctions, and the criterion for judging failure or malfunction is that the leakage current of the capacitor increases sharply.

[0036] Such as Figure 2 and Figure 3 shown, step S2: Current curve feature extraction, specifically refers to extracting features from the current response curve, which can effectively characterize the aging process of the capacitor and provide necessary input data for subsequent machine learning modeling.

[0037] Specifically, key features are extracted from the current response curve collected during the highly accelerated life test. These features include but are not limited to: the starting time t 1 of the current observation period and the corresponding current value I 1 , the ending time t 2 of the current observation period and the corresponding current value I 2 , the slope k of the current change, as well as the ambient temperature and voltage, etc. Among them, the current observation period is a self-defined observation range, and the relaxation end time (the stationary point where the current changes from decreasing to increasing) can be set as the starting point; the ambient temperature and voltage are specifically the environmental stresses applied during the aging process, specifically corresponding to high voltage and high temperature, and remain unchanged during the test process.

[0038] Please continue to refer to Figure 3 , Figure 3 , the straight line at the top represents the change of current with time during the actual use of the device until the end of the device life (ttf, Time To Failure). In one embodiment of the present invention, the starting time t 1 of the current observation period and the ending time t 2 of the current observation period satisfy: t 2 = nt 1, where n is the multiple of the current observation period.

[0039] Step S3: Feature optimization and screening, specifically referring to optimizing and screening the extracted features to obtain a feature combination.

[0040] Through experimental analysis and training of a machine learning model, the present invention optimizes the feature extraction method and removes redundant features to improve the accuracy and efficiency of the model.

[0041] Specifically, as Figure 3 and Figure 4 shown, in step S3, different multiples n of the current observation period are used to screen out the best feature extraction method. The optimization and screening of the extracted features to obtain a feature combination includes: Step S31: Take uniform interval values of n from the m-th preset number to the o-th preset number, perform machine learning modeling, obtain the trend of model performance changing with the n value, and determine the optimal n value; the uniform interval value can be an integer value or a fractional value, and can be selected according to the implementation requirements. For example, m is 2, o is 10, the uniform interval value is an integer value, and the optimal n value obtained is 7; After determining the optimal n value, perform SHAP analysis on the features corresponding to the optimal n value to analyze the feature contribution degree and the dependence between features; Perform permutation and combination on the features, model respectively to obtain the performance, and select a set of feature combinations with the best performance as the optimal feature combination.

[0042] Specifically, n determines the length of the multiple of the current observation period (the time range for taking features), and also determines the effect of the feature extraction range. The present invention determines the feature extraction range with the best effect by obtaining different n values. In the present invention, for each n value taken, the same modeling evaluation and comparison are performed, and the best n value is obtained after comparison. After the n value is determined, the best feature extraction method is correspondingly screened out.

[0043] SHAP (SHapley Additive exPlanations) analysis is a method for explaining the predictions of machine learning models. It is derived from the concept of Shapley values in game theory and assigns importance values to each feature of the model, thereby explaining the prediction process of the model. This analysis method can be applied to any machine learning model, such as linear regression, decision trees, random forests, gradient boosting models, and neural networks, etc. SHAP values can accurately reflect the contribution of each feature to a single prediction, and when the actual impact of a feature increases, its SHAP value will not decrease. The sum of the SHAP values of all features is equal to the difference between the model prediction value and the average prediction value. Through SHAP values, it is possible to intuitively see which features have the greatest impact on the model prediction results, explain the prediction results of individual data points, help understand why the model makes a certain prediction, and discover abnormal data points and potential problem features.

[0044] Furthermore, perform SHAP analysis on the features with the optimal n value to analyze the feature contribution degree and the dependence between features, so as to screen the feature combinations. For each selected feature combination, build a model to evaluate the results and compare them to obtain the optimal feature combination. Exemplarily, through SHAP analysis, it is possible to obtain the influence degree of each feature on the model output. For example, the slope k of the current change has the greatest impact, and it is also possible to obtain the most relevant feature combination. For example, it is the slope k of the current change, the start time t of the current observation period 1 and the corresponding current value I at the end time of the current observation period 2 The composed feature combination.

[0045] Step S4: Machine learning model construction, specifically referring to using the obtained feature combination to construct a machine learning model to predict the lifespan of MLCC.

[0046] Step S5: Model optimization and evaluation, specifically referring to optimizing the constructed machine learning model to ensure the prediction accuracy, and using cross-validation and resampling methods for model evaluation to improve the generalization ability of the model and avoid overfitting.

[0047] In the present invention, the machine learning models used include but are not limited to RF (Random Forest), XGBoost (eXtreme Gradient Boosting), GBDT (Gradient Boosting Decision Tree), Support Vector Regression (SVR), ANN (Artificial Neural Network), Gaussian process, principal component analysis, generalized linear model, etc. These models are trained and learned through the input feature data.

[0048] Exemplarily, in the present invention, when using the random forest regression model for MLCC life prediction, the input features include the start time t of the current observation period 1 , the initial current value I 1 , the end time t 2 , the end current value I 2 , the slope k of the current change, the ambient temperature, and the voltage, and the output is the life value of the MLCC, that is, the predicted failure occurrence time

[0049] Since feature selection has been combined and screened in feature engineering, and the optimal feature combination has been obtained, the present invention further normalizes these features to eliminate the dimensional differences between features and ensure the effect during model training

[0050] As Figure 5 and Figure 6 shown Figure 5 shows a schematic diagram of the basic process of a machine learning model in an embodiment of the present invention Figure 6 shows a schematic diagram of the prediction result of the best model in an embodiment of the present invention

[0051] Specifically, in the present invention, 250 HALT experiment (Highly Accelerated Life Test) leakage current curves (a type of current response curve) are collected, and 31 feature combinations and 30 observation periods are screened through the feature engineering provided in step S3 to obtain the final dataset for machine learning modeling. The dataset is divided into a training set accounting for 70% of the total samples and a test set accounting for 30% of the total samples. The training set is used for model fitting, and the test set is used to evaluate the performance of the model

[0052] The model of the training set accounting for 70% of the total samples is trained, and the model of the test set accounting for 30% of the total samples is evaluated using 2 test metrics. At the same time, 10-fold cross-validation is used to ensure robustness. Based on this, the same process is repeated for 5 models to obtain the optimal model of each model. The five optimal models are compared to obtain the finally selected final model. The final model obtained in this example is the random forest model. Inputting the prediction data into the optimal model can obtain the MLCC life prediction value

[0053] Exemplarily, the present invention uses RandomForestRegressor in the sklearn library of the Python language to establish a random forest regression model and sets relevant parameters. Common parameters include: n_estimators (the number of trees), criterion (the criterion for measuring the splitting quality, such as mse or mae), max_depth (the maximum depth of the tree), min_samples_split (the minimum number of samples required to split an internal node), min_samples_leaf (the minimum number of samples required for a leaf node), and max_features (the maximum number of features to consider at each split). The selection of these parameters will directly affect the performance and complexity of the model.

[0054] Grid Search (GridSearchCV) or Random Search (RandomizedSearchCV) is used to automatically tune the above parameters to ensure the best performance of the model.

[0055] During the model training process, RandomForestRegressor constructs and optimizes each tree, and captures the potential patterns in the data by integrating multiple decision trees.

[0056] After training is completed, the test set is used to evaluate the model. The prediction accuracy of the model is mainly measured by calculating one or more of the following metrics: mean squared error (MSE), mean absolute error (MAE), root mean squared error (RMSE), or R 2 value. If the error of the model meets the predetermined error threshold Error, the training is completed and the trained model is saved; otherwise, the model can be tuned according to the error feedback and retrained.

[0057] The purpose of optimizing the constructed machine learning model in the present invention is to ensure its high prediction accuracy. The specific method is to optimize the hyperparameters of the machine learning model through the grid search method to ensure its consistency and robustness on different data sets.

[0058] Finally, the trained model can be saved and loaded for prediction in the actual production environment. In actual life prediction, those skilled in the art can input new current observation period data, and the trained model can output the corresponding life prediction value. This process makes full use of the powerful ability of the random forest regression model in dealing with complex data relationships.

[0059] The evaluation method is as follows: The model is evaluated using cross-validation and resampling methods to improve the generalization ability of the model and avoid overfitting. The evaluation metrics used are R 2 and RMSE.

[0060] To improve the generalization ability of the model and avoid overfitting, the present invention adopts 10-fold cross-validation and resampling methods for model evaluation. 10-fold cross-validation can effectively utilize each part of the dataset. Through multiple trainings and tests, it ensures the stable performance of the model on different data and reduces the accidental errors caused by data partitioning.

[0061] In specific operations, first divide the training set into 10 subsets. Each time, use one of the subsets as the validation set and the other parts as the training set. In this way, each data point will appear in the test set once, and the model will be trained and evaluated on different data subsets.

[0062] Finally, the evaluation result of the model is the average of the evaluation results of all folds, which can more accurately reflect the actual performance of the model and avoid the bias caused by data splitting.

[0063] In addition, the resampling method can also be used to further improve the stability of the model. By randomly sampling samples from the training data with replacement for training, the resampling method allows each sample to participate in training multiple times, thus better fitting the training data and avoiding the overfitting problem that may be caused by a single dataset partitioning.

[0064] Through the above optimization steps, the optimal parameters of each model can be obtained and the best model can be compared.

[0065] After the modeling is completed, the life prediction of other samples of this model can be carried out.

[0066] As Figure 7 shown, in an application scenario of the present invention, taking an actual MLCC (such as X7R, EIA0603, 99230pF, 50V voltage grade) as an example, when performing highly accelerated life tests, the test temperature range is from 120°C to 150°C, and the voltage range is from 400V to 600V.

[0067] Through this test method, we can obtain the current curve data of multiple capacitors under different conditions and extract key characteristic quantities from them, such as the starting point, ending point of the current observation period, and the slope of the current change, etc.

[0068] In the feature optimization stage, different observation period multiples n (such as n = 2 - 10) are used for screening. By comparing various machine learning algorithms, the random forest is finally selected for modeling.

[0069] After optimization, the corresponding R values of RF, XGB, GBDT, ANN, SVR are respectively 2Are: XGB: 0.8234; RF: 0.8971; GBDT: 0.8871; ANN: 0.80305; SVR: 0.7557; It can be seen that RF has the highest R 2 , in addition, Figure 6 It also shows that RF also has the lowest RMSE. Therefore, it is proved that RF has high prediction accuracy in practical applications.

[0070] In the present invention, by measuring the initial current curve of the MLCC in use, the features required for modeling are obtained and input into the model to obtain the lifetime. In this way, the present invention can effectively improve the accuracy of MLCC lifetime prediction and can adapt to capacitors of different types and under different conditions, making it a prediction method that can be widely applied in practical engineering.

[0071] It should be noted that after those skilled in the art understand the method provided by the present invention, they can also optimize or test through other machine learning algorithms to obtain the best model and the corresponding highest prediction accuracy, which still falls within the scope of the rights protected by the present invention.

[0072] The second aspect of the present invention provides a readable storage medium storing a computer program, and the computer program is executed by a processor to perform the steps of the method described in any one of the above embodiments.

[0073] The computer-readable storage medium may include: any entity or device capable of carrying a computer program, a recording medium, a USB flash drive, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), and a software distribution medium, etc. The computer program includes computer program code. The computer program code may be in the form of source code, object code, an executable file, or some intermediate form, etc. The computer-readable storage medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), and a software distribution medium, etc.

[0074] Any process or method description represented in a flowchart or otherwise described herein can be understood to represent a module, segment, or portion of code including one or more executable instructions for implementing a specific logical function or process, and the scope of the preferred embodiments of the present invention includes additional implementations where functions may be executed not in the order shown or discussed, including in a substantially simultaneous manner according to the functions involved or in a reverse order, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

[0075] The logic and / or steps represented in a flowchart or otherwise described herein, for example, can be considered a sequenced list of executable instructions for implementing a logical function and can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, a system including a processing module, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device.

[0076] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments or equivalently replace some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting MLCC life based on machine learning, characterized in that: include: Step S1: Perform a highly accelerated life test on the MLCC and collect its aging current response curve; Step S2: extracting features from the current response curve; Step S3: Optimizing and screening the extracted features to obtain feature combinations; Step S4: Use the obtained feature combination to build a machine learning model to predict the life of MLCC.

2. The MLCC life prediction method based on machine learning according to claim 1, characterized in that: It also includes: Step S5: optimizing the constructed machine learning model to ensure prediction accuracy, and using cross-validation and resampling methods to evaluate the model to improve the generalization ability of the model and avoid overfitting.

3. The MLCC life prediction method based on machine learning according to claim 1, characterized in that: The highly accelerated life test of the MLCC includes: aging the MLCC under conditions of a preset high temperature and a preset high voltage, and recording its current response curve until the MLCC fails or malfunctions.

4. The MLCC life prediction method based on machine learning according to claim 1 is characterized in that: The features extracted from the current response curve include at least: Start time of current observation period t 1 and the corresponding current value I 1. End time of current observation period t 2 and the corresponding current value I 2, and the slope of the current change k .

5. The MLCC life prediction method based on machine learning according to claim 4 is characterized in that: The start time of the current observation period t 1 and the end time of the current observation period t 2 Satisfaction: t 2= nt 1, where n is the current observation period multiple.

6. The MLCC life prediction method based on machine learning according to claim 5 is characterized in that: The extracted features are optimized and screened to obtain a feature combination, including: right n Take from m Preset number to o Preset the uniform interval value between the numbers, perform machine learning modeling, and obtain the model performance n The trend of value changes determines the best n value; Determine the best n After the value, the best n The features corresponding to the values ​​are subjected to SHAP analysis to analyze the feature contribution and the dependency between features; And the features are arranged and combined, the performance is modeled respectively, and a set of feature combinations with the best performance are selected as the optimal feature combination.

7. The MLCC life prediction method based on machine learning according to claim 1 is characterized in that: The obtained feature combination is used to construct a machine learning model to predict the life of MLCC, including: The preset algorithms are compared based on prediction accuracy and training speed, and the random forest model is used as the optimal model of the machine learning model. The preset algorithms include at least random forest, XGBoost, gradient boosting decision tree, support vector machine, and artificial neural network.

8. The MLCC life prediction method based on machine learning according to claim 2 is characterized in that: Also includes: The robustness of the model was enhanced through 10-fold cross validation, and the hyperparameters were optimized through grid search.

9. A readable storage medium storing a computer program, characterized in that: The computer program is executed by a processor to perform the steps of the MLCC life prediction method based on machine learning as described in any one of claims 1-8.

Images