Electroplating quality monitoring method, device and equipment for MLCC capacitor and medium

Abstract

The invention discloses an electroplating quality monitoring method, device and equipment for an MLCC capacitor and a medium, and relates to the technical field of MLCC capacitor electroplating. The electroplating quality monitoring method comprises the following steps: firstly, acquiring coating microstructure and ion migration data by adopting X-ray tomography and electrochemical impedance spectroscopy technologies, and performing defect classification in combination with appearance detection data; electroplating process parameters are obtained, and a multi-dimensional feature library is constructed through principal component analysis. Then based on a defect classification result and a feature library, training an improved random forest model to realize multi-defect prediction, screening key process parameters through sensitivity ranking, and establishing defect-parameter association; and then constructing an attribution model to analyze an influence mechanism of process parameters on defects. In actual production, real-time process parameters are input into the prediction model for online quality monitoring; and when defects appear, reversely tracing cause parameters through an attribution model. According to the method, a system of data acquisition, feature fusion, real-time monitoring and defect tracing is formed, and the electroplating quality can be effectively monitored.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ceramic capacitor electroplating, in particular to a method and device for monitoring the electroplating quality of MLCC capacitors, and a medium. BACKGROUND

[0002] In the process of MLCC capacitor electroplating, defects such as tin-lead layer thickness exceeding the upper limit, tin-lead layer thickness exceeding the lower limit, nickel layer thickness exceeding the upper limit, nickel layer thickness exceeding the lower limit, barrier layer delamination, barrier layer discontinuity, nickel-silver separation, plating solution intrusion, pitting, blistering, and plating layer peeling are common quality problems.

[0003] The occurrence of these defects is closely related to multiple key parameters in the electroplating process.

[0004] Firstly, uneven distribution of current density can cause local areas to have excessively thick or thin plating layers, resulting in pitting or plating layer peeling. Secondly, instability of electrolyte composition, such as fluctuations in additive concentration or excessively high impurity content, can directly affect the adhesion and uniformity of the plating layer, leading to blistering or insufficient plating layer bonding. In addition, improper control of the temperature and pH value of the electroplating solution can change the electrochemical reaction rate, affecting the density and crystalline quality of the plating layer.

[0005] In the electroplating process, the rotation speed of the drum is also an important influencing factor. Insufficient stirring can cause uneven distribution of MLCC capacitor metal ions in the electrolyte, resulting in differences in plating layer thickness. While too fast stirring can introduce bubbles, causing blistering or holes on the surface of the plating layer. At the same time, the length of the electroplating time also affects the density and thickness of the plating layer, and too short a time can result in incomplete coverage of the plating layer, while too long a time can cause increased stress in the plating layer, leading to peeling.

[0006] However, the above reasons and results are only speculations based on basic science. There has been a lack of mapping relationship between the type of capacitor electroplating defects and specific process parameters. Therefore, when capacitor electroplating defects occur, the skilled person in the art needs to adjust the process parameters multiple times to find process parameters that do not produce defects, greatly reducing production efficiency. SUMMARY

[0007] The present application provides a method and device for monitoring the electroplating quality of MLCC capacitors, and a medium to improve at least one of the above technical problems.

[0008] In a first aspect, the present application provides a method for monitoring the electroplating quality of MLCC capacitors, which comprises steps S1 to S8.

[0009] S1, acquire the MLCC capacitance and use X-ray tomography and electrochemical impedance spectroscopy testing technology for physical field coupling detection, collect the microstructure morphology of the plating layer and the interface ion migration data, and generate detection data.

[0010] S2, classify defects according to the appearance of the MLCC capacitance and the detection data, and obtain a classification result.

[0011] S3, acquire process parameter data of the electroplating process of the MLCC capacitance.

[0012] S4, data fusion of the process parameters is performed based on a principal component analysis method to generate a multi-dimensional feature library. The multi-dimensional feature library includes linear combinations of different process parameters.

[0013] S5, according to the classification result and the multi-dimensional feature library, an improved random forest model for multi-defect type classification task is trained to obtain a defect prediction model. The sensitivity ranking of the defect type and the process parameter is analyzed during the training, and according to the sensitivity ranking result, a threshold is set to screen out the key process parameters that have the greatest impact on the electroplating quality, so as to obtain the correlation between the defect type and the process parameter.

[0014] S6, according to the classification result and the multi-dimensional feature library, a defect attribution model between defect classification and process parameters is trained to determine the process parameters that cause defects.

[0015] S7, acquire the process parameters in the electroplating process of the MLCC capacitance, and input the defect prediction model to predict whether the electroplating of the capacitance has defects, so as to perform real-time monitoring on the electroplating process of the MLCC capacitance.

[0016] S8, when the electroplating of the MLCC capacitance has defects, or when it is predicted that the electroplating of the MLCC capacitance has defects, the production parameters and the defect type of the MLCC capacitance are input into the trained defect attribution model to obtain the process parameters that cause defects.

[0017] In an optional embodiment, step S4 specifically includes steps S41 to S46.

[0018] S41, pre-process the process parameter data to obtain pre-processed process parameters. The pre-processing includes data cleaning, missing value processing, and abnormal value detection and processing.

[0019] S42, standardize the pre-processed process parameter data to obtain standardized process parameters.

[0020] S43, multiply any two numerical parameters in the pre-processed process parameters to construct a plurality of new parameters. The plurality of new parameters are standardized to obtain standardized new parameters.

[0021] S44, using the principal component analysis method to reduce the dimension of the standardized process parameters and the new parameters, extracting the main components as the basic features, and saving the feature vector matrix of the principal component analysis model.

[0022] S45, inputting the standardized process parameters into a deep autoencoder fused with an attention mechanism for nonlinear feature extraction to obtain nonlinear features.

[0023] S46, fusing the basic features and the nonlinear features to obtain a multi-dimensional feature library.

[0024] Preferably, the deep autoencoder fused with the attention mechanism includes a first input layer, a first encoder, a first decoder, and a first output layer.

[0025] The number of neurons of the first input layer is set according to the number of features of the multi-dimensional feature library.

[0026] The first encoder includes a first hidden layer provided with a first fully connected layer and a first attention mechanism layer, a second hidden layer provided with a second fully connected layer and a second attention mechanism layer, a third fully connected layer, and a first encoding output layer. The number of neurons of the first fully connected layer is 128, and the activation function is ReLU. The first attention mechanism layer includes a first fully connected sublayer and a first Softmax function, the first fully connected sublayer maps a 128-dimensional vector to a 64-dimensional vector, and the first Softmax function converts it into attention weights, which are multiplied with corresponding elements of a 128-dimensional feature vector to make the model pay attention to important features. The number of neurons of the second fully connected layer is 64, and the activation function is ReLU. The second attention mechanism layer includes a second fully connected sublayer and a second Softmax function, which are used to calculate attention weights for a 64-dimensional feature vector and weight them. The number of neurons of the third fully connected layer is 32, and the activation function is ReLU. The number of neurons of the first encoding output layer is 16, and the activation function is ReLU, which is used to output the final low-dimensional feature encoding of the first encoder.

[0027] The first decoder includes a fourth fully connected layer, a fifth fully connected layer, and a third hidden layer provided with a sixth fully connected layer and a third attention mechanism layer. The number of neurons of the fourth fully connected layer is 32, and the activation function is ReLU. The number of neurons of the fifth fully connected layer is 64, and the activation function is ReLU. The number of neurons of the sixth fully connected layer is 128, and the activation function is ReLU. The third attention mechanism layer includes a third fully connected sublayer and a third Softmax function, which are used to calculate attention weights for a 128-dimensional feature vector to assist the model in accurately reconstructing important features.

[0028] The first output layer has the same number of neurons as the first input layer, and the activation function is a linear activation function, which is used to output the reconstructed continuous process parameter value.

[0029] In an optional embodiment, the improved random forest model comprises a second input layer, a first random forest layer, a second random forest layer, a SHAP framework layer, and a second output layer. The second input layer is adapted to expand the input features. The first random forest layer adopts a conventional random forest based on the Gini coefficient for preliminary feature screening to remove unimportant features. The second random forest layer adopts a conditional inference tree to eliminate the interference of highly correlated parameters on feature importance evaluation through permutation test, thereby improving the accuracy of feature importance evaluation. The SHAP framework layer is used to quantify the marginal contribution of each process parameter to each type of defect. The second output layer is provided with a Softmax function for normalization processing.

[0030] Preferably, in the training stage, step S5 specifically comprises steps S51 to S55.

[0031] S51, input the multi-dimensional feature library into the second input layer, and expand the time series features according to the process parameter data of the past preset time length by the second input layer. The expansion includes calculating the mean, variance and slope.

[0032] S52, input the multi-dimensional feature library and the new features obtained by expansion into the first random forest layer to calculate the feature importance by Gini impurity, preliminarily screen the key process parameters, generate the preliminary multi-defect classification probability, and obtain the preliminary screening data.

[0033] S53, the second random forest layer inputs the process parameters with the top 50% importance in the preliminary screening data, then eliminates the interference of multicollinearity on feature importance evaluation by permutation test, constructs an unbiased tree splitting rule, captures the conditional dependence relationship between parameters, obtains the parameter sensitivity ranking after eliminating the interference of correlation, and obtains the multi-defect type prediction probability distribution based on conditional probability, thereby obtaining the screening result.

[0034] S54, the SHAP framework layer calculates the marginal contribution value of each process parameter to each type of defect and quantifies the nonlinear interaction effect between parameters based on game theory according to the multi-dimensional feature library, the new features obtained by expansion, the preliminary screening data and the screening result, thereby obtaining the parameter contribution matrix of each process parameter to different defect types.

[0035] S55, the output layer is used to generate the final defect type probability output by Softmax normalization according to the screening result, and generate an interpretable process parameter sensitivity heat map based on the parameter contribution matrix.

[0036] Preferably, in the reasoning stage, the SHAP framework layer of the improved random forest model is deleted, and step S54 is removed. The multi-dimensional feature library obtained according to the real-time process parameter data is input into the trained improved random forest model, and only the final defect type probability output is output.

[0037] In an optional embodiment, step S7 specifically comprises steps S71 to S77.

[0038] S71, acquiring real-time process parameters in the MLCC capacitor electroplating process.

[0039] S72, based on the mean vector and the standard deviation vector saved in the training stage, performing standardization processing on the real-time process parameters to obtain standardized real-time process parameters. The standardization model is wherein z is the standardized real-time process parameter vector, x is the real-time process parameter vector, μ is the mean vector, and σ is the standard deviation vector.

[0040] S73, multiplying any two numerical parameters in the standardized real-time process parameters to construct a plurality of real-time new parameters. The plurality of real-time new parameters are standardized to obtain standardized real-time new parameters.

[0041] S74, based on the feature vector matrix saved in the training, performing principal component vector extraction on the standardized real-time process parameters and the standardized real-time new parameters to obtain real-time basic features. The principal component vector extraction model is y=Ez, wherein y is the principal component score vector, and E is the feature vector matrix.

[0042] S75, inputting the standardized real-time process parameters into the pre-trained deep feature encoding model to perform nonlinear feature extraction to obtain real-time nonlinear features.

[0043] S76, fusing the real-time basic features and the standardized real-time process parameters to obtain a real-time multi-dimensional feature library.

[0044] S77, inputting the real-time multi-dimensional feature library into the defect prediction model to predict the probability of defects existing in the capacitor electroplating.

[0045] In an optional embodiment, the defect attribution model comprises a third input layer, a sensitivity weighting module, a cross-modal interaction module, a root cause positioning module, and a third output layer. The third input layer comprises a text encoding sublayer and a numerical encoding sublayer for processing the text label of the classification result and the multi-dimensional feature library, respectively. The sensitivity weighting module injects the correlation as prior weights through Hadamard product for dynamic weighting of the multi-dimensional feature library. The cross-modal interaction module adopts a bidirectional cross-attention mechanism to establish the interactive association between defect semantics and process features. The root cause positioning module decodes high-order features to the original process parameter space through reverse PCA mapping to generate a root cause probability distribution. The third output layer outputs the root cause probability of the original process parameter based on a Sigmoid activation function.

[0046] Preferably, step S8 specifically comprises steps S81 to S86.

[0047] S81, input the text label of the classification result into the text encoding sublayer to generate a defect semantic vector through word embedding and attention pooling.

[0048] S82, input the multi-dimensional feature library into the numerical encoding sublayer to generate an initial process feature vector through standardization and full connection mapping.

[0049] S83, perform Hadamard product weighting on the initial process feature vector through the weight matrix generated according to the correlation, and optimize the weight distribution through a learnable full connection layer to generate a weighted process feature vector.

[0050] S84, in the cross-modal interaction module, calculate a defect→process attention vector with the defect semantic vector as Query and the weighted process feature vector as Key / Value, and calculate a process→defect attention vector with the weighted process feature vector as Query and the defect semantic vector as Key / Value, splice the bidirectional attention results, and generate a cross-modal joint feature vector through residual connection.

[0051] S85, extract high-order nonlinear features of the cross-modal joint feature vector through an eighth full connection layer, and perform reverse mapping based on a coefficient matrix of principal component analysis to generate a feature representation in the original process parameter space.

[0052] S86, calculate the root cause probability of each process parameter through a Sigmoid function according to the feature representation in the original process parameter space, and determine the defect root cause parameter according to a preset probability threshold to obtain the process parameter causing the defect.

[0053] In an optional embodiment, the defect attribution model uses the artificially labeled defect root cause parameters as a supervision signal in the training process, and jointly optimizes the parameters of the defect attribution model using Focal Loss and sensitivity ranking consistency loss. The defect root cause parameter represents a process parameter that causes a defect. The root cause probability is the probability that the process parameter belongs to the defect root cause parameter.

[0054] In an optional embodiment, the weight matrix generated according to the correlation relationship specifically includes step A4 in step A1.

[0055] A1, normalize the correlation relationship determined according to the sensitivity ranking to obtain a sensitivity value. The normalization model is:

[0056] A2, amplify the sensitivity value to obtain an initial weight. The calculation model of the initial weight is: initial weight = sensitivity 2 .

[0057] A3, L2-normalize the initial weight to obtain a final weight. The calculation model of the final weight is:

[0058] A4, map the final weight to the principal component space according to the coefficient matrix of the principal component analysis to obtain the weight matrix.

[0059] Preferably, the detection data includes plating layer thickness, porosity, and ion migration rate indicators.

[0060] Preferably, the defect classification includes tin-lead layer thickness exceeding the upper limit, tin-lead layer thickness exceeding the lower limit, nickel layer thickness exceeding the upper limit, nickel layer thickness exceeding the lower limit, barrier layer delamination, barrier layer discontinuity, nickel-silver separation, plating solution intrusion, pitting, blistering, plating layer peeling, and normal capacitance.

[0061] Preferably, the process parameter data includes operator, electroplating equipment, chemical concentration, current density, plating tank current, plating tank temperature, plating tank pH value, steel ball volume, steel ball diameter, plating barrel speed, electroplating time, and drying temperature.

[0062] In a second aspect, the application provides a quality inspection analysis device for an electroplating process of a ceramic capacitor, which includes a detection data acquisition module, a classification module, a process parameter acquisition module, a principal component analysis module, a defect prediction module, a defect attribution module, a monitoring module, and a cause analysis module.

[0063] The detection data acquisition module is used to acquire MLCC capacitors and perform physical field coupling detection using X-ray tomography and electrochemical impedance spectroscopy testing technology, collect plating layer microstructure morphology and interface ion migration data, and generate detection data.

[0064] The classification module is configured to classify defects according to the appearance of the MLCC capacitor and the detection data, and obtain a classification result.

[0065] The process parameter acquisition module is configured to acquire process parameter data of the electroplating process of the MLCC capacitor.

[0066] The principal component analysis module is configured to perform data fusion on the process parameters based on a principal component analysis method, and generate a multi-dimensional feature library. The multi-dimensional feature library includes linear combinations of different process parameters.

[0067] The defect prediction module is configured to train an improved random forest model for a multi-defect type classification task according to the classification result and the multi-dimensional feature library, and obtain a defect prediction model. The training simultaneously analyzes the sensitivity ranking of the defect type and the process parameters, and sets a threshold according to the sensitivity ranking result to filter out key process parameters that have the greatest impact on the electroplating quality, so as to obtain the correlation between the defect type and the process parameters.

[0068] The defect attribution module is configured to train a defect attribution model between the defect classification and the process parameters according to the classification result and the multi-dimensional feature library, so as to determine the process parameters that cause defects.

[0069] The monitoring module is configured to acquire process parameters in the electroplating process of the MLCC capacitor, and input the defect prediction model to predict whether the electroplating of the capacitor has defects, so as to perform real-time monitoring on the electroplating process of the MLCC capacitor.

[0070] The cause analysis module is configured to input the production parameters and the defect type of the MLCC capacitor into the trained defect attribution model when the electroplating of the MLCC capacitor has defects or when it is predicted that the electroplating of the MLCC capacitor has defects, so as to obtain the process parameters that cause defects.

[0071] In a third aspect, the present application provides a quality inspection and analysis device for a ceramic capacitor electroplating process, characterized in that it comprises a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement the MLCC capacitor electroplating quality monitoring method described in any one of the first aspect.

[0072] In a third aspect, the present application provides a computer readable storage medium, characterized in that the computer readable storage medium comprises a stored computer program, wherein the computer readable storage medium controls the device to execute the MLCC capacitor electroplating quality monitoring method described in any one of the first aspect when the computer program runs.

[0073] By adopting the technical scheme, the following technical effects can be achieved.

[0074] The plating quality monitoring method of the MLCC capacitor realizes plating quality monitoring of the MLCC capacitor through a multi-dimensional data fusion and machine learning cooperative mechanism. Based on principal component analysis-based feature library construction and improved random forest model training, the accuracy of multi-defect classification and the reliability of process parameter sensitivity identification are effectively improved. Through a bidirectional verification mechanism of a defect prediction model and an attribution model, the plating process anomalies are monitored in real time, and the process root of defect generation is traced in reverse, forming a preventive quality control closed loop of "detection-early warning-tracing". The scheme significantly improves the intelligent level of plating quality control of the MLCC capacitor. BRIEF DESCRIPTION OF DRAWINGS

[0075] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed to be used in the specific embodiments of the present application. It should be understood that the following drawings only show some specific embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0076] Fig. 1 is a flowchart of the plating quality monitoring method.

[0077] Fig. 2 is a network structure diagram of a deep autoencoder fused with an attention mechanism.

[0078] Fig. 3 is a network structure diagram of an improved random forest model.

[0079] Fig. 4 is a network structure diagram of a defect attribution model. DETAILED DESCRIPTION

[0080] The technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0081] Embodiment one, please refer to Figs. 1 to 4 The first embodiment of the present application provides a plating quality monitoring method of an MLCC capacitor, which comprises steps S1 to S8.

[0082] S1, obtain the MLCC capacitance and use X-ray tomography and electrochemical impedance spectroscopy testing technology for physical field coupling detection, collect the microstructure morphology of the plating layer and the interface ion migration data, and generate detection data. Preferably, the detection data includes plating layer thickness, porosity, and ion migration rate indicators. This multi-modal detection method can find micron-sized pores that are not visible to the naked eye, ensuring that the detection data covers both physical structure and electrochemical performance indicators.

[0083] S2, classify defects according to the appearance of the MLCC capacitance and the detection data, and obtain a classification result. Preferably, the defect classification includes tin-lead layer thickness exceeding the upper limit, tin-lead layer thickness exceeding the lower limit, nickel layer thickness exceeding the upper limit, nickel layer thickness exceeding the lower limit, barrier layer delamination, barrier layer discontinuity, nickel-silver separation, plating solution intrusion, pitting, blistering, plating layer peeling, and normal capacitance.

[0084] S3, obtain process parameter data of the electroplating process of the MLCC capacitance. Preferably, the process parameter data includes operator, electroplating equipment, chemical solution concentration, current density, plating tank current, plating tank temperature, plating tank pH value, steel ball volume, steel ball diameter, plating barrel rotation speed, electroplating time, and drying temperature. Among them, the current density refers to the distribution of the positive electrode.

[0085] S4, data fusion of the process parameters based on principal component analysis method, generating a multi-dimensional feature library. Wherein, the multi-dimensional feature library contains linear combinations of different process parameters. Preferably, step S4 specifically includes steps S41 to S46.

[0086] S41, pre-process the process parameter data to obtain pre-processed process parameters. Wherein, the pre-processing includes data cleaning, missing value processing, and outlier detection and processing.

[0087] S42, standardize the pre-processed process parameter data to obtain standardized process parameters.

[0088] S43, multiply any two numerical parameters in the pre-processed process parameters to construct a plurality of new parameters. And standardize the plurality of new parameters to obtain standardized new parameters.

[0089] S44, use principal component analysis method to reduce the dimension of the standardized process parameters and new parameters, extract the main components as the basic features, and save the feature vector matrix of the principal component analysis model.

[0090] S45, input the standardized process parameters into a deep autoencoder with an attention mechanism for nonlinear feature extraction to obtain nonlinear features.

[0091] Preferably, the deep autoencoder with attention mechanism comprises a first input layer, a first encoder, a first decoder, and a first output layer.

[0092] The number of neurons of the first input layer is set according to the number of features of the multi-dimensional feature library.

[0093] The first encoder comprises a first hidden layer provided with a first fully connected layer and a first attention mechanism layer, a second hidden layer provided with a second fully connected layer and a second attention mechanism layer, a third fully connected layer, and a first encoding output layer. The number of neurons of the first fully connected layer is 128, and the activation function is ReLU. The first attention mechanism layer comprises a first fully connected sub-layer and a first Softmax function, the first fully connected sub-layer maps a 128-dimensional vector to a 64-dimensional vector, and the first Softmax function converts it into attention weights, which are multiplied with corresponding elements of a 128-dimensional feature vector to make the model focus on important features. The number of neurons of the second fully connected layer is 64, and the activation function is ReLU. The second attention mechanism layer comprises a second fully connected sub-layer and a second Softmax function, which are used to calculate attention weights for a 64-dimensional feature vector and weight them. The number of neurons of the third fully connected layer is 32, and the activation function is ReLU. The number of neurons of the first encoding output layer is 16, and the activation function is ReLU, which is used to output the final low-dimensional feature encoding of the first encoder.

[0094] Specifically, the first fully connected layer is used to linearly transform the data of the first input layer and introduce a nonlinear factor, and map it to a 128-dimensional feature space. The second fully connected layer further compresses and transforms the features. The second attention mechanism layer operates in the same way as the first attention mechanism layer. The third fully connected layer is used to continue to reduce the dimension of the features.

[0095] The first decoder comprises a fourth fully connected layer, a fifth fully connected layer, and a third hidden layer provided with a sixth fully connected layer and a third attention mechanism layer. The number of neurons of the fourth fully connected layer is 32, and the activation function is ReLU. The number of neurons of the fifth fully connected layer is 64, and the activation function is ReLU. The number of neurons of the sixth fully connected layer is 128, and the activation function is ReLU. The third attention mechanism layer comprises a third fully connected sub-layer and a third Softmax function, which are used to calculate attention weights for a 128-dimensional feature vector to assist the model in accurately reconstructing important features.

[0096] Specifically, the fourth fully connected layer is used to perform inverse transformation on the 16-dimensional feature vector of the first encoder output layer. The fifth fully connected layer is used to expand the feature dimension.

[0097] The number of neurons of the first output layer is the same as that of the first input layer, and the activation function is a linear activation function, which is used to output the reconstructed continuous process parameter value.

[0098] S46, fuse the basic features and the nonlinear features to obtain a multi-dimensional feature library. Specifically, the fusion manner is splicing, or adding / multiplying the basic features and the nonlinear features extracted from the "standardized process parameters".

[0099] The combined feature engineering of principal component analysis and deep autoencoder has significant advantages: on the one hand, the nonlinear relationship between parameters is captured by constructing interaction terms (such as the product term of current density x plating bath temperature), and on the other hand, the attention mechanism makes the model focus on key features (such as assigning an attention weight of 0.35 to the PH value parameter).

[0100] In practical applications, the scheme first expands the original 12-dimensional process parameters into more dimensional features, and then reduces the feature dimension to a reasonable range through principal component analysis, thereby reducing the complexity of subsequent model training while ensuring information integrity and expansibility.

[0101] S5, according to the classification result and the multi-dimensional feature library, training an improved random forest model for multi-defect type classification task to obtain a defect prediction model. The training simultaneously analyzes the sensitivity ranking of defect type and process parameters, and according to the sensitivity ranking result, sets a threshold to screen out the key process parameters that have the greatest impact on electroplating quality, to obtain the correlation between defect type and process parameters.

[0102] The correlation is a mapping relationship. When training the random forest model, the feature importance of each process parameter can be calculated. The feature importance reflects the influence size of the parameter in the model decision process. When a product has a defect, the importance ranking of these process parameters can be viewed, and the process parameters with higher importance are more likely to be the cause of the defect.

[0103] Preferably, the improved random forest model includes a second input layer, a first random forest layer, a second random forest layer, a SHAP framework layer, and a second output layer. The second input layer is suitable for expanding the input features. The first random forest layer adopts a conventional random forest based on Gini coefficient for preliminary feature screening to remove unimportant features. The second random forest layer adopts a conditional inference tree to eliminate the interference of highly correlated parameters on feature importance evaluation through permutation test, thereby improving the accuracy of feature importance evaluation. The SHAP framework layer is used to quantify the marginal contribution of each process parameter to each type of defect. The second output layer is provided with a Softmax function for normalization processing.

[0104] Specifically, the improved random forest is designed to reduce the computation by 42% by using the conventional random forest to screen the features first, and then using conditional inference trees to eliminate the multicollinearity interference (such as the correlation coefficient between the plating bath temperature and the current density is reduced from 0.78 to 0.12). Finally, the SHAP framework is used to quantify the parameter contribution, making it more intuitive.

[0105] In the training phase of the improved random forest model described above, step S5 specifically includes steps S51 to S55.

[0106] S51, input the multi-dimensional feature library into the second input layer, and expand the time series features according to the process parameter data of the past preset time length by the second input layer. Wherein, the expansion includes calculating the mean, variance and slope.

[0107] S52, input the multi-dimensional feature library and the new features obtained by expansion into the first random forest layer, to calculate the feature importance by Gini impurity, preliminarily screen the key process parameters, and generate the preliminary multi-defect classification probability, to obtain the preliminary screening data. Wherein, the multi-defect classification probability is the original voting result without calibration.

[0108] S53, the second random forest layer inputs the process parameters with the top 50% importance in the preliminary screening data, then performs replacement test to eliminate the interference of multicollinearity on the evaluation of feature importance, and constructs unbiased tree splitting rules to capture the conditional dependence between parameters, to obtain the parameter sensitivity ranking after eliminating the correlation interference, and the multi-defect type prediction probability distribution based on conditional probability, thereby obtaining the screening result.

[0109] S54, the SHAP framework layer calculates the marginal contribution value of each process parameter to each type of defect and quantifies the nonlinear interaction effect between parameters based on game theory according to the multi-dimensional feature library, the new features obtained by expansion, the preliminary screening data and the screening result, thereby obtaining the parameter contribution matrix of each process parameter to different defect types.

[0110] S55, the output layer is used to generate the final defect type probability output by Softmax normalization according to the screening result, and generate the interpretable process parameter sensitivity heat map based on the parameter contribution matrix.

[0111] Preferably, in the inference phase of the improved random forest model described above, the SHAP framework layer of the improved random forest model is deleted, and step S54 is removed, and the multi-dimensional feature library obtained by processing the real-time acquired process parameter data is input into the trained improved random forest model, only the final defect type probability output is output.

[0112] The improved random forest model of the embodiment can continue to refine the principal component parameters step by step in three layers, from linear combination, to conditional dependence relationship, to nonlinear interaction contribution, and finally form an interpretable process-defect association network. Thus, whether the product has defects can be predicted in real time by monitoring the process parameters. The conditional inference tree output of the second layer will correct the feature screening threshold of the first layer in reverse (such as joint importance evaluation on highly correlated parameter groups). SHAP provides both global parameter importance ranking (for process optimization) and single-sample-level SHAP value analysis (for defect root cause tracing).

[0113] S6、According to the classification result and the multi-dimensional feature library, a defect attribution model between defect classification and process parameters is trained to determine the process parameters that cause defects. Preferably, the data processing process of the defect attribution model is shown in steps S81 to S86, and step S8 is the application of the defect attribution model in the inference stage.

[0114] Specifically, the defect attribution model includes a third input layer, a sensitivity weighting module, a cross-modal interaction module, a root cause positioning module, and a third output layer. The third input layer includes a text encoding sublayer and a numerical encoding sublayer for processing the text label of the classification result and the multi-dimensional feature library, respectively. The sensitivity weighting module injects the correlation relationship as a prior weight through Hadamard product for dynamic weighting of the multi-dimensional feature library. The cross-modal interaction module uses a bidirectional cross-attention mechanism to establish the interactive association between defect semantics and process features. The root cause positioning module decodes high-order features to the original process parameter space through reverse PCA mapping to generate a root cause probability distribution. The third output layer outputs the root cause probability of the original process parameter based on the Sigmoid activation function. The root cause probability is the probability that the process parameter belongs to the defect root cause parameter.

[0115] Preferably, the defect attribution model uses the artificially labeled defect root cause parameters as a supervision signal in the training process, and uses Focal Loss and sensitivity sorting consistency loss to jointly optimize the parameters of the defect attribution model. The defect root cause parameter represents the process parameter that causes defects. The defect root cause parameter is an artificially labeled process parameter that causes electroplating defects. This data is obtained by manually adjusting the machine to make the electroplating quality normal when electroplating defects occur in the actual production process.

[0116] The defect attribution model adopts a dual-channel architecture of text encoding and numerical encoding to process the semantic information of the defect classification label and the numerical information of the process parameter feature library respectively. The parameter sensitivity ranking obtained in the early stage is injected into the model as prior knowledge through a sensitivity weighting module. Then, the deep interaction between the defect semantic features and the process parameter features is realized by using a bidirectional cross-attention mechanism. Finally, the high-order features are decoded to the original process parameter space by reverse principal component mapping to output the root cause probability of each parameter, thereby realizing accurate tracing.

[0117] By introducing the sensitivity weighting mechanism, the prior knowledge of the influence of process parameters on defects is integrated into the model training, effectively improving the accuracy of root cause positioning. The cross-modal interaction module breaks through the limitations of traditional single-modal analysis, enabling the deep correlation between the text description of defect features and the numerical process parameters in the joint feature space, and capturing the synergistic effect of parameter combinations on complex defects. The reverse principal component mapping technology realizes the reversible conversion from the feature space to the original parameter space, so as to directly lock the problem process parameters according to the root cause probability matrix. The model has significant advantages in multiple collinearity interference elimination, nonlinear relationship modeling, and multi-defect type adaptability, and provides intelligent support for the rapid optimization of electroplating process.

[0118] S7, obtaining process parameters in the electroplating process of the MLCC capacitor and inputting the defect prediction model to predict whether the electroplating of the capacitor has defects, so as to monitor the electroplating process of the MLCC capacitor in real time. Preferably, step S7 specifically comprises steps S71 to S77.

[0119] S71, obtaining real-time process parameters in the electroplating process of the MLCC capacitor.

[0120] S72, based on the mean vector and the standard deviation vector saved in the training stage, standardizing the real-time process parameters to obtain standardized real-time process parameters. The standardization model is wherein z is the standardized real-time process parameter vector, x is the real-time process parameter vector, μ is the mean vector, and σ is the standard deviation vector.

[0121] S73, multiplying any two numerical parameters in the standardized real-time process parameters to construct a plurality of real-time new parameters. And standardizing the plurality of real-time new parameters to obtain standardized real-time new parameters.

[0122] S74, based on the feature vector matrix saved in the training, extracting principal component vectors from the standardized real-time process parameters and the standardized real-time new parameters to obtain real-time basic features. The principal component vector extraction model is y=Ez, wherein y is the principal component score vector and E is the feature vector matrix.

[0123] S75, input the standardized real-time process parameters into the pre-trained deep feature encoding model for nonlinear feature extraction to obtain real-time nonlinear features.

[0124] S76, fuse the real-time basic features and the standardized real-time process parameters to obtain a real-time multi-dimensional feature library.

[0125] S77, input the real-time multi-dimensional feature library into the defect prediction model to predict the probability of defects in the capacitive electroplating.

[0126] Specifically, principal component analysis is performed on a large amount of data, and only one value is obtained at each moment in the real-time monitoring process of the process parameters. Therefore, in this embodiment, the mean vector and the standard deviation vector in the training stage are used to perform principal component analysis on the real-time obtained process parameters. In other embodiments, the mean vector and the standard deviation can also be recalculated by monitoring the process parameter values in the preset market at the current moment, and then the real-time monitored process parameters are subjected to principal component analysis. The present application does not make specific limitations on this.

[0127] Step S7 uses the standardized model and the feature extraction model constructed in step S4 in the training stage to convert the real-time data into a real-time multi-dimensional feature library. Then the real-time multi-dimensional feature library is input into the defect prediction model constructed in step S5 in the training stage to predict the electroplating defect probability, thereby realizing the instant warning of electroplating abnormalities.

[0128] S8, when the electroplating of the MLCC capacitor has defects, or when it is predicted that the electroplating of the MLCC capacitor has defects, the production parameters and defect types of the MLCC capacitor are input into the trained defect attribution model to obtain the process parameters that cause the defects. Preferably, step S8 specifically includes steps S81 to S86.

[0129] S81, input the text label of the classification result into the text encoding sublayer to generate a defect semantic vector through word embedding and attention pooling.

[0130] S82, input the multi-dimensional feature library into the numerical encoding sublayer to generate an initial process feature vector through standardization and full connection mapping.

[0131] S83, perform Hadamard product weighting on the initial process feature vector through the weight matrix generated according to the correlation, and optimize the weight distribution through a learnable fully connected layer to generate a weighted process feature vector. Preferably, the weight matrix generated according to the correlation specifically includes step A4 in step A1.

[0132] A1, normalize the correlation determined according to the sensitivity ranking to obtain a sensitivity value. The normalization model is:

[0133] A2, amplify the sensitivity value to obtain an initial weight. Wherein the calculation model of the initial weight is: initial weight = sensitivity 2 .

[0134] A3, L2 normalize the initial weight to obtain the final weight. Wherein the calculation model of the final weight is:

[0135] A4, according to the coefficient matrix of principal component analysis, map the final weight to the principal component space to obtain the weight matrix.

[0136] S84, in the cross-modal interaction module, the defect semantic vector is taken as Query, and the weighted process feature vector is taken as Key / Value to calculate the defect→process attention vector. The weighted process feature vector is taken as Query, and the defect semantic vector is taken as Key / Value to calculate the process→defect attention vector. After splicing the bidirectional attention results, the cross-modal joint feature vector is generated through residual connection.

[0137] S85, extract high-order nonlinear features of the cross-modal joint feature vector through the eighth fully connected layer, and perform reverse mapping based on the coefficient matrix of principal component analysis to generate feature representation in the original process parameter space.

[0138] S86, according to the feature representation of the original process parameter space, calculate the root cause probability of each process parameter through the Sigmoid function, and determine the defect root cause parameter according to the preset probability threshold to obtain the process parameter causing the defect.

[0139] Specifically, first, the defect type text label and the multi-dimensional feature library are respectively encoded, the feature expression of the key process parameter is dynamically enhanced through the sensitivity weighting module, and then the deep association between the defect semantics and the process features is established by combining the bidirectional cross attention mechanism. Subsequently, the high-order features are decoded to the original parameter space by using reverse principal component mapping, and finally the root cause probability of each process parameter is output through the Sigmoid function, so as to accurately locate the specific process parameter causing the defect.

[0140] The core advantage of this step is to realize the multi-modal interaction analysis of defects and process parameters. By fusing the cross-modal attention mechanism of text semantics (i.e. defect type) and numerical features (i.e. process parameters), the implicit association between defect description and process parameters can be captured. Combined with the prior knowledge weighting of sensitivity sorting and reverse principal component mapping technology, the accuracy of root cause positioning is improved, and the interpretability of the result in the original parameter space is guaranteed.

[0141] The plating quality monitoring method of the MLCC capacitor of the present application realizes the plating quality monitoring of the MLCC capacitor through a multi-dimensional data fusion and machine learning collaborative mechanism. Based on the feature library construction and improved random forest model training of principal component analysis, the accuracy of multi-defect classification and the reliability of process parameter sensitivity identification are effectively improved. Through the bidirectional verification mechanism of the defect prediction model and the attribution model, the process root of the defect is traced back in reverse while the plating process anomalies are monitored in real time, forming a preventive quality control closed loop of "detection-early warning-tracing". The plating quality monitoring method of the present embodiment can significantly improve the intelligent level of the plating quality control of the MLCC capacitor.

[0142] In the second embodiment, the present application provides a quality inspection and analysis device for a plating process of a ceramic capacitor, which comprises a detection data acquisition module, a classification module, a process parameter acquisition module, a principal component analysis module, a defect prediction module, a defect attribution module, a monitoring module and a cause analysis module

[0143] The detection data acquisition module is used to acquire the MLCC capacitor and perform physical field coupling detection by adopting X-ray tomography and electrochemical impedance spectroscopy testing technology, collect plating layer microstructure morphology and interface ion migration data, and generate detection data.

[0144] The classification module is used to classify defects according to the appearance of the MLCC capacitor and the detection data, and acquire a classification result.

[0145] The process parameter acquisition module is used to acquire process parameter data of the plating process of the MLCC capacitor.

[0146] The principal component analysis module is used to perform data fusion on the process parameters based on the principal component analysis method, and generate a multi-dimensional feature library. The multi-dimensional feature library comprises linear combinations of different process parameters.

[0147] The defect prediction module is used to train an improved random forest model for a multi-defect type classification task according to the classification result and the multi-dimensional feature library, and acquire a defect prediction model. The defect prediction module simultaneously analyzes the sensitivity ranking of the defect type and the process parameters, and sets a threshold according to the sensitivity ranking result to screen out key process parameters that have the greatest impact on the plating quality, so as to acquire the correlation between the defect type and the process parameters.

[0148] The defect attribution module is used to train a defect attribution model between the defect classification and the process parameters according to the classification result and the multi-dimensional feature library, so as to determine the process parameters that cause defects.

[0149] The monitoring module is configured to acquire process parameters in the plating process of the MLCC capacitor, and input the process parameters into the defect prediction model to predict whether the plating of the capacitor is defective, so as to monitor the plating process of the MLCC capacitor in real time.

[0150] The cause analysis module is configured to input the production parameters and the defect type of the MLCC capacitor into the trained defect attribution model when the plating of the MLCC capacitor is defective or when it is predicted that the plating of the MLCC capacitor is defective, to acquire process parameters causing the defect.

[0151] In the third embodiment, the application provides a quality inspection and analysis device for a plating process of a ceramic capacitor, which comprises a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement the MLCC capacitor plating quality monitoring method according to any one of the first embodiment.

[0152] In the fourth embodiment, the application provides a computer readable storage medium, which comprises a stored computer program. When the computer program is running, the computer readable storage medium controls a device where the computer readable storage medium is located to execute the MLCC capacitor plating quality monitoring method according to any one of the first embodiment.

[0153] In the several embodiments of the application, it should be understood that the disclosed device and method can also be implemented in other ways. The device and method embodiments described above are only illustrative, and for example, the flowcharts and block diagrams in the drawings show the possible implementation architectures, functions and operations of the device, method and computer program product according to the embodiments of the application. In this regard, each block in the flowcharts or block diagrams can represent a module, a program segment or a part of code, which contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in different orders from those noted in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and sometimes they can be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and the combination of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system for performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0154] In addition, the functional modules in each of the embodiments of the application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0155] The functions described can be implemented in software, firmware, hardware, or any combination thereof. If implemented in software and as an independent application, the functions can be stored in one or more computer-readable storage media. In view of this understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, an electronic device, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes. It should be noted that in this paper, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the processes, methods, articles or devices that include a series of elements not only include those elements, but also include other elements that are not explicitly listed or inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "includes a" does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0156] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. The singular forms "a", "an" and "the" used in the embodiments of the present application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0157] It should be understood that the term "and / or" used herein is only to describe the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents that the front and rear associated objects are a "or" relationship.

[0158] Depending on the context, the word "if" as used herein can be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting". Similarly, depending on the context, the phrase "if it is determined" or "if (a stated condition or event) is detected" can be interpreted as "when it is determined" or "in response to determining" or "when (a stated condition or event) is detected" or "in response to detecting (a stated condition or event)".

[0159] The "first / second" mentioned in the embodiments are only to distinguish similar objects, and do not represent a specific order for the objects. Understandably, the "first / second" can be interchanged in a specific order or sequence as appropriate. It should be understood that the objects distinguished by "first / second" can be interchanged as appropriate, so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0160] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Various modifications and changes can be made to the present application by those skilled in the art. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method of monitoring plating quality of an MLCC capacitor, the method comprising: The method comprises steps S1 to S8. ​ S1, acquiring MLCC capacitance and performing physical field coupling detection by X-ray tomography and electrochemical impedance spectroscopy testing technology to collect data of plating layer microstructure morphology and interface ion migration, and generating detection data; S2, classifying defects according to the appearance of the MLCC capacitance and the detection data, and obtaining a classification result; S3, acquiring process parameter data of the plating process of the MLCC capacitance; S4, data fusion of the process parameters based on a principal component analysis method to generate a multi-dimensional feature library; wherein the multi-dimensional feature library comprises linear combinations of different process parameters; S5, training an improved random forest model for a multi-defect type classification task according to the classification result and the multi-dimensional feature library to obtain a defect prediction model; training simultaneously analyzes the sensitivity ranking of the defect type and the process parameter, and sets a threshold according to the sensitivity ranking result to screen out the key process parameters that have the greatest impact on the plating quality to obtain the correlation between the defect type and the process parameter; S6, training a defect attribution model between the defect classification and the process parameter according to the classification result and the multi-dimensional feature library to determine the process parameter that causes the defect; S7, acquiring process parameters in the plating process of the MLCC capacitance and inputting the defect prediction model to predict whether the plating of the capacitance has defects to monitor the plating process of the MLCC capacitance in real time; S8, when the plating of the MLCC capacitance has defects or it is predicted that the plating of the MLCC capacitance has defects, inputting the production parameters and the defect type of the MLCC capacitance into the trained defect attribution model to obtain the process parameter that causes the defect.

2. The method of claim 1, wherein the method is characterized by: Step S4 specifically comprises steps S41 to S46. S41, preprocessing the process parameter data to obtain preprocessed process parameters; wherein the preprocessing comprises data cleaning, missing value processing, and abnormal value detection and processing; S42, standardizing the preprocessed process parameter data to obtain standardized process parameters; S43, multiplying any two numerical parameters in the preprocessed process parameters to construct a plurality of new parameters; and standardizing the plurality of new parameters to obtain standardized new parameters; S44, using a principal component analysis method to reduce the dimension of the standardized process parameters and new parameters, extracting main components as basic features, and saving the feature vector matrix of the principal component analysis model; S45, inputting the standardized process parameters into a deep autoencoder with an attention mechanism for nonlinear feature extraction to obtain nonlinear features; S46, fusing the basic features and the nonlinear features to obtain a multi-dimensional feature library; The deep autoencoder with an attention mechanism comprises a first input layer, a first encoder, a first decoder, and a first output layer; The number of neurons of the first input layer is set according to the number of features of the multi-dimensional feature library; The first encoder comprises a first hidden layer provided with a first full connection layer and a first attention mechanism layer, a second hidden layer provided with a second full connection layer and a second attention mechanism layer, a third full connection layer, and a first encoding output layer; the number of neurons of the first full connection layer is 128, and the activation function is ReLU; the first attention mechanism layer comprises a first full connection sublayer and a first Softmax function, the first full connection sublayer maps a 128-dimensional vector to a 64-dimensional vector, and the first Softmax function converts it into attention weights, which are multiplied with corresponding elements of a 128-dimensional feature vector to make the model focus on important features; the number of neurons of the second full connection layer is 64, and the activation function is ReLU; the second attention mechanism layer comprises a second full connection sublayer and a second Softmax function, which are used to calculate attention weights of a 64-dimensional feature vector and weight them; the number of neurons of the third full connection layer is 32, and the activation function is ReLU; the number of neurons of the first encoding output layer is 16, and the activation function is ReLU, which are used to output the final low-dimensional feature encoding of the first encoder; The first decoder comprises a fourth full connection layer, a fifth full connection layer, and a third hidden layer provided with a sixth full connection layer and a third attention mechanism layer; the number of neurons of the fourth full connection layer is 32, and the activation function is ReLU; the number of neurons of the fifth full connection layer is 64, and the activation function is ReLU; the number of neurons of the sixth full connection layer is 128, and the activation function is ReLU; the third attention mechanism layer comprises a third full connection sublayer and a third Softmax function, which are used to calculate attention weights of a 128-dimensional feature vector and assist the model in accurately reconstructing important features; The number of neurons of the first output layer is the same as that of the first input layer, and the activation function is a linear activation function, which is used to output the reconstructed continuous process parameter value.

3. The method of claim 1, wherein the method is characterized by: The improved random forest model comprises a second input layer, a first random forest layer, a second random forest layer, a SHAP framework layer, and a second output layer; wherein the second input layer is adapted to expand the input features; the first random forest layer adopts a conventional random forest based on a Gini coefficient to preliminarily filter features and remove unimportant features; the second random forest layer adopts a conditional inference tree to eliminate the interference of highly correlated parameters on feature importance evaluation through a permutation test and improve the accuracy of feature importance evaluation; the SHAP framework layer is used to quantify the marginal contribution of each process parameter to each type of defect; and the second output layer is provided with a Softmax function for normalization processing; In the training stage, step S5 specifically comprises steps S51 to S55; S51, input the multi-dimensional feature library into the second input layer, and expand the time series features according to the process parameter data of the past preset time length by the second input layer; wherein the expansion comprises calculating the mean, variance and slope; S52, input the multi-dimensional feature library and the new features obtained by extension into a first random forest layer to calculate feature importance by Gini impurity to preliminarily screen key process parameters, generate a preliminary multi-defect classification probability, and obtain preliminary screening data; S53, input the process parameters with importance ranking in the top 50% in the preliminary screening data into a second random forest layer, then perform replacement inspection to eliminate the interference of multicollinearity on feature importance evaluation, construct unbiased tree splitting rules, capture the conditional dependence relationship between parameters, obtain parameter sensitivity ranking after eliminating the interference of correlation, and obtain multi-defect type prediction probability distribution based on conditional probability, so as to obtain screening results; S54, a SHAP framework layer calculates the marginal contribution value of each process parameter to each type of defect and quantifies the nonlinear interaction effect between parameters based on game theory according to the multi-dimensional feature library, the new features obtained by extension, the preliminary screening data and the screening results, so as to obtain a parameter contribution matrix of each process parameter to different defect types; S55, the output layer is used to generate final defect type probability output by Softmax normalization according to the screening results, and generate an interpretable process parameter sensitivity heat map based on the parameter contribution matrix; In the inference stage, the SHAP framework layer of the improved random forest model is deleted, and step S54 is deleted, the multi-dimensional feature library obtained by processing the real-time process parameter data is input into the trained improved random forest model, and only the final defect type probability output is output.

4. The method of claim 1, wherein the method is characterized by: Step S7 specifically includes steps S71 to S77; S71, obtain real-time process parameters in the MLCC capacitor plating process; S72, based on the mean vector and the standard deviation vector saved in the training stage, normalizing the real-time process parameters to obtain normalized real-time process parameters; the normalization model is wherein, is the normalized real-time process parameter vector, is the real-time process parameter vector, is the mean vector, is the standard deviation vector; S73, multiply any two numerical parameters in the standardized real-time process parameters to construct a plurality of real-time new parameters; and standardize the plurality of real-time new parameters to obtain standardized real-time new parameters; S74, based on the feature vector matrix saved during training, principal component vector extraction is performed on the standardized real-time process parameters and the standardized real-time new parameters to obtain real-time basic features; the principal component vector extraction model is where is the principal component score vector and is the feature vector matrix; S75, input the standardized real-time process parameters into a pre-trained deep feature encoding model to perform nonlinear feature extraction and obtain real-time nonlinear features; S76, fuse the real-time basic features and the standardized real-time process parameters to obtain a real-time multi-dimensional feature library; S77, input the real-time multi-dimensional feature library into the defect prediction model to predict the probability of defects in capacitor plating.

5. The method of claim 1 to 4, wherein the method is characterized by: The defect attribution model comprises a third input layer, a sensitivity weighting module, a cross-modal interaction module, a root cause positioning module, and a third output layer; wherein: the third input layer comprises a text encoding sublayer and a numerical encoding sublayer, which are respectively used for processing the text label of the classification result and the multi-dimensional feature library; the sensitivity weighting module injects the correlation as prior weights through Hadamard product for dynamic weighting of the multi-dimensional feature library; the cross-modal interaction module adopts a bidirectional cross-attention mechanism to establish the interactive association of defect semantics and process features; the root cause positioning module decodes high-order features to the original process parameter space through reverse PCA mapping to generate a root cause probability distribution; and the third output layer outputs the root cause probability of the original process parameter based on a Sigmoid activation function; Step S8 specifically comprises steps S81 to S86; S81, input the text label of the classification result into the text encoding sublayer to generate a defect semantic vector through word embedding and attention pooling; S82, input the multi-dimensional feature library into the numerical encoding sublayer to generate an initial process feature vector through standardization and full connection mapping; S83, perform Hadamard product weighting on the initial process feature vector through a weight matrix generated according to the correlation, and optimize the weight distribution through a learnable full connection layer to generate a weighted process feature vector; S84, in the cross-modal interaction module, calculate a defect→process attention vector with the defect semantic vector as Query and the weighted process feature vector as Key / Value, and calculate a process→defect attention vector with the weighted process feature vector as Query and the defect semantic vector as Key / Value, splice the bidirectional attention results, and generate a cross-modal joint feature vector through residual connection; S85, extract high-order nonlinear features of the cross-modal joint feature vector through an eighth full connection layer, and perform reverse mapping based on a coefficient matrix of principal component analysis to generate a feature representation in the original process parameter space; S86, calculate the root cause probability of each process parameter through a Sigmoid function according to the feature representation in the original process parameter space, and determine the defect root cause parameter according to a preset probability threshold to obtain the process parameter causing the defect.

6. The method of claim 5, wherein the method further comprises: In the training process, the defect attribution model uses artificially labeled defect root cause parameters as supervision signals, and jointly optimizes the parameters of the defect attribution model using Focal Loss and sensitivity ranking consistency loss; wherein the defect root cause parameter represents the process parameter causing the defect; and the root cause probability is the probability that the process parameter belongs to the defect root cause parameter; The weight matrix generated according to the correlation specifically comprises: The correlation determined according to the sensitivity ranking is normalized to obtain a sensitivity value; the normalization model is: ; Amplify the sensitivity value, get the initial weight; wherein the calculation model of the initial weight is: ​ The initial weight is subjected to L2 normalization to obtain the final weight; wherein a calculation model of the final weight is: ​ According to the coefficient matrix of principal component analysis, map the maximum weight to the principal component space to obtain the weight matrix.

7. The method of claim 1, wherein the method is characterized by: The detection data includes plating layer thickness, porosity, and ion migration rate indicators; The defect classification includes tin-lead layer thickness over the upper limit, tin-lead layer thickness over the lower limit, nickel layer thickness over the upper limit, nickel layer thickness over the lower limit, barrier layer delamination, barrier layer discontinuity, nickel-silver separation, plating solution intrusion, pitting, blistering, plating layer peeling, and normal capacitance; The process parameter data includes an operator, electroplating equipment, bath concentration, current density, plating bath current, plating bath temperature, plating bath pH, steel ball volume, steel ball diameter, plating barrel rotation speed, electroplating time, and drying temperature.

8. A quality inspection analysis device for a plating process of a ceramic capacitor, characterized by, Comprise: The detection data acquisition module is used for acquiring MLCC capacitors and performing physical field coupling detection by adopting X-ray tomography and electrochemical impedance spectrum testing technology, collecting plating layer microstructure morphology and interface ion migration data, and generating detection data; The classification module is used for defect classification according to the appearance of the MLCC capacitor and the detection data, and obtaining a classification result; The process parameter acquisition module is used for acquiring process parameter data of the electroplating process of the MLCC capacitor; The principal component analysis module is used for data fusion of the process parameters based on a principal component analysis method, to generate a multi-dimensional feature library; wherein the multi-dimensional feature library comprises a linear combination of different process parameters; The defect prediction module is used for training an improved random forest model for a multi-defect type classification task according to the classification result and the multi-dimensional feature library, to obtain a defect prediction model; the training simultaneously analyzes the sensitivity ranking of the defect type and the process parameter, and sets a threshold according to the sensitivity ranking result to screen out key process parameters that most affect the electroplating quality, to obtain the correlation between the defect type and the process parameter; The defect attribution module is used for training a defect attribution model between the defect classification and the process parameter according to the classification result and the multi-dimensional feature library, to determine the process parameter that causes the defect; The monitoring module is used for acquiring process parameters in the electroplating process of the MLCC capacitor, and inputting the defect prediction model to predict whether the electroplating of the capacitor has defects, to perform real-time monitoring on the electroplating process of the MLCC capacitor; The cause analysis module is used for inputting the production parameters and the defect type of the MLCC capacitor into the trained defect attribution model when the electroplating of the MLCC capacitor has defects, or when it is predicted that the electroplating of the MLCC capacitor has defects, to obtain the process parameter that causes the defect.

9. A quality inspection analysis device of a plating process of a ceramic capacitor, characterized by, The computer readable storage medium comprises a stored computer program, wherein the computer readable storage medium controls the device where the computer readable storage medium is located to execute the MLCC capacitor electroplating quality monitoring method according to any one of claims 1 to 7 when the computer program runs.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium comprises a stored computer program, wherein the computer readable storage medium controls the device where the computer readable storage medium is located to execute the MLCC capacitor electroplating quality monitoring method according to any one of claims 1 to 7 when the computer program runs.

Images