Method and system for automatically adjusting production parameters of gold bonding wire

Abstract

The invention provides an automatic adjustment method for gold bonding wire production parameters, and relates to the technical field of information, and the method comprises the steps: collecting a new smelting parameter data sample in real time, inputting the data sample into a parameter self-inspection neural network model, obtaining a predicted quality index value, comparing the predicted value with an actually measured quality index, and determining the production parameters of the gold bonding wire. Calculating a prediction error; and an incremental learning algorithm is adopted, online updating and fine tuning are performed on the parameter self-checking neural network model by using the newly collected data sample, and the connection weight of the parameter self-checking neural network model is adjusted according to the prediction error. Real-time self-inspection of parameters in the smelting process and dynamic prediction of quality indexes are achieved, abnormity can be found in time, production parameters can be optimized, and the quality stability and production efficiency of smelted products are effectively improved. Meanwhile, a self-adaptive optimization mechanism of the model ensures continuous improvement of prediction precision, and powerful support is provided for intelligent control of the smelting process.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method and system for automatically adjusting production parameters of bonding gold wires. Background Art

[0002] In the smelting process, how to build an intelligent model system that can accurately predict quality indicators and adaptively optimize? This system needs to collect and process a large amount of smelting parameter data in real time, and dynamically adjust the model structure and parameters based on these data. However, the smelting process is complex and changeable, and there are complex nonlinear relationships between parameters. It is difficult for traditional static models to accurately capture the mapping relationship between process parameters and quality indicators. At the same time, the smelting process and raw material properties may change over time, and the prediction accuracy of the model will gradually decrease. How to design a model system with self-learning and self-optimization capabilities so that it can continuously adapt to process changes and maintain high prediction accuracy? In addition, the automatic optimization of the model structure involves a large number of hyperparameter combinations. How to quickly search for the optimal model structure under limited computing resources? While ensuring the performance of the model, how to balance the complexity and computational efficiency of the model so that it can meet the needs of real-time control of the smelting process? The solution to these problems is of great significance to improving the intelligence level of the smelting process and realizing the accurate prediction and control of quality indicators. Summary of the invention

[0003] The present invention provides a method for automatically adjusting production parameters of a gold bonding wire, which mainly includes:

[0004] Constructing the parameter self-checking neural network model, the parameter self-checking neural network model comprises an input layer, a hidden layer and an output layer, the number of nodes in the input layer is the same as the number of smelting process parameters, and the number of nodes in the output layer is the same as the number of quality indicators;

[0005] Collecting new smelting parameter data samples in real time, inputting the data samples into the parameter self-checking neural network model, obtaining predicted quality index values, comparing the predicted values ​​with the actually measured quality indexes, and calculating the prediction errors;

[0006] Adopting an incremental learning algorithm, using the newly collected data samples to perform online updating and fine-tuning on the parameter self-checking neural network model, and adjusting the connection weights of the parameter self-checking neural network model according to the prediction error;

[0007] Regularly evaluate the performance of the parameter self-checking neural network model, calculate the average prediction error within a specified time period, and trigger an automatic optimization process of the model structure if the error exceeds a preset threshold;

[0008] A genetic algorithm is used to optimize and search the structure of the parameter self-checking neural network model, the hyperparameters of the parameter self-checking neural network model are encoded as genes, and candidate models are randomly generated as the initial population; each candidate model in the population is evaluated, and the fitness function value is calculated, and a new generation of population is generated through genetic operators such as selection, crossover, and mutation. It is iteratively optimized until the termination condition is met, and the model with the highest fitness in the final population is used as the optimized parameter self-checking neural network model.

[0009] As a preferred solution of the method for automatically adjusting the production parameters of the bonding wire of the present invention, the parameter self-checking neural network model is constructed, and the parameter self-checking neural network model comprises an input layer, a hidden layer and an output layer, the number of nodes in the input layer is the same as the number of smelting process parameters, and the number of nodes in the output layer is the same as the number of quality indicators, including:

[0010] Obtaining the number of smelting process parameters, determining the number of nodes in the input layer of the parameter self-checking neural network model to make them the same;

[0011] Obtaining the number of quality indicators, determining the number of nodes in the output layer of the parameter self-checking neural network model to make them the same;

[0012] Using correlation analysis and principal component analysis, select smelting process parameters whose correlation with the quality index is higher than a preset threshold as input features of the parameter self-checking neural network model;

[0013] The parameter self-checking neural network model is trained by a back propagation algorithm, and weights and biases are updated by a gradient descent method to minimize the mean square error between the predicted output and the actual quality index.

[0014] As a preferred solution of the method for automatically adjusting the production parameters of the gold bonding wire of the present invention, the real-time acquisition of new smelting parameter data samples, inputting the data samples into the parameter self-checking neural network model, obtaining predicted quality index values, comparing the predicted values ​​with the actually measured quality indexes, and calculating the prediction errors include:

[0015] Acquire smelting parameter data samples collected in real time, pre-process the data samples, and input them into a pre-built parameter self-checking neural network model to obtain predicted smelting product quality index values;

[0016] Obtaining the quality index value of the actual measurement, and calculating the error index between the predicted value and the measured value;

[0017] If the error index exceeds a preset threshold, an incremental learning algorithm is used to retrain the parameter self-check neural network model using newly collected data samples, update the model parameters, and obtain a parameter self-check neural network model with better performance, which is applied to subsequent real-time smelting parameter data.

[0018] As a preferred solution of the method for automatically adjusting the production parameters of the bonding wire of the present invention, the incremental learning algorithm is used to update and fine-tune the parameter self-checking neural network model online using the newly collected data samples, and the connection weight of the parameter self-checking neural network model is adjusted according to the prediction error, including:

[0019] Obtaining data samples, preprocessing the data samples, removing missing values ​​and outliers, extracting key features, and constructing feature vectors;

[0020] According to the error loss function, the back propagation algorithm is used to calculate the model parameter gradient, and the stochastic gradient descent algorithm is used to update the model weights to minimize the loss function;

[0021] The trained model is used to predict and detect anomalies on new data. If the prediction result exceeds the preset threshold, it is determined that an anomaly exists and a detection report is generated.

[0022] As a preferred solution of the method for automatically adjusting the production parameters of the bonding wire of the present invention, the performance of the parameter self-checking neural network model is regularly evaluated, and the average prediction error within a specified time period is calculated. If the error exceeds a preset threshold, the automatic optimization process of the model structure is triggered, including:

[0023] Get the prediction results and true values ​​of the self-checking model within a specified time period, and calculate the average prediction error within the time period;

[0024] Determine whether the average prediction error exceeds the preset error threshold. If it exceeds the threshold, the following automatic optimization process of the model structure is triggered: for the hyperparameters of the self-checking model, such as the number of hidden layers, the number of neurons in each layer, and the learning rate, a grid search algorithm is used to traverse different parameter combinations within the preset value range and train multiple candidate models;

[0025] Based on the comprehensive performance indicators of the candidate model on the validation set, such as the weighted average of accuracy, recall, and F1 score, the model hyperparameter combination with the best performance is selected;

[0026] Using the optimal hyperparameter combination, retrain the self-checking model on the training set to obtain the optimized model;

[0027] Apply the optimized self-check model to online prediction to improve the prediction accuracy of the system;

[0028] Continuously monitor the online performance of the optimized self-checking model. When the prediction error exceeds the threshold again, repeat the above automatic optimization process to achieve continuous iterative update of the model and keep the model's prediction performance stable at a high level.

[0029] As a preferred solution of the method for automatically adjusting the production parameters of the bonding wire of the present invention, the method uses a genetic algorithm to optimize the structure of the parameter self-checking neural network model, encodes the hyperparameters of the parameter self-checking neural network model into genes, and randomly generates candidate models as the initial population, including:

[0030] Obtaining a preset number of neural network models as an initial population, wherein the hyperparameters of the parameter self-checking neural network model are encoded as binary gene sequences;

[0031] For each of the parameter self-checking neural network models, training and testing are performed on a validation data set, and the accuracy of the parameter self-checking neural network model on the validation set is used as its fitness value;

[0032] Determine whether the average fitness value of the current population has not been significantly improved in multiple consecutive iterations, or whether the number of iterations has reached a preset upper limit;

[0033] If so, the parameter self-checking neural network model with the highest fitness value in the current population is determined as the optimal model.

[0034] As a preferred solution of the method for automatically adjusting the production parameters of the bonding wire of the present invention, each candidate model in the population is evaluated, the fitness function value is calculated, a new generation of population is generated by genetic operators such as selection, crossover, and mutation, and iterative optimization is performed until the termination condition is met, and the model with the highest fitness in the final population is used as the optimized parameter self-checking neural network model, including:

[0035] Acquire characteristic parameters of the candidate model, construct a fitness function, and use the fitness function value as an indicator to measure the quality of the candidate model;

[0036] A roulette wheel selection operator is used to select at least one high-quality model with a fitness higher than a preset threshold from the current population according to the fitness values ​​of the candidate models;

[0037] For the high-quality models selected by the roulette wheel selection operator, an arithmetic crossover operator is used to randomly pair them in pairs and cross-recombine their characteristic parameters to obtain a recombined candidate model.

[0038] An automatic adjustment system for gold bonding wire production parameters, comprising:

[0039] Data collection module, used to collect sample data and perform preprocessing;

[0040] Parameter self-checking module, used to compare the collected sample parameters and calculate the prediction error;

[0041] The incremental update module is used to collect new data samples and update the data of the parameter self-check module; the evaluation module is used to evaluate the performance of the candidate model

[0042] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0043] The present invention discloses a method for self-checking and quality prediction of smelting process parameters. The method constructs a neural network model, collects smelting parameter data in real time and inputs it into the model, predicts quality indicators and compares them with actual values. Through an incremental learning algorithm, the model is updated and fine-tuned online using new data. The present invention regularly evaluates model performance, and when the prediction error exceeds a threshold, triggers automatic optimization of the model structure. The optimization process uses a genetic algorithm to encode the model hyperparameters as genes, and obtains the optimal model through iterative optimization. This method realizes real-time self-checking of smelting process parameters and dynamic prediction of quality indicators, can detect anomalies in time and optimize production parameters, and effectively improves the quality stability and production efficiency of smelting products. At the same time, the adaptive optimization mechanism of the model ensures the continuous improvement of prediction accuracy, providing strong support for the intelligent control of the smelting process. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 The present invention is a flow chart of a method for automatically adjusting production parameters of bonding gold wires.

[0045] Figure 2 It is a schematic diagram of a method for automatically adjusting production parameters of bonding gold wire according to the present invention.

[0046] Figure 3 It is another schematic diagram of a method for automatically adjusting production parameters of a bonding gold wire according to the present invention. DETAILED DESCRIPTION

[0047] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. The present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the relevant inventions, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0048] like Figure 1-3 In this embodiment, a method for automatically adjusting production parameters of a gold bonding wire may specifically include:

[0049] Step S101, constructing the parameter self-checking neural network model, the parameter self-checking neural network model comprises an input layer, a hidden layer and an output layer, the number of nodes in the input layer is the same as the number of smelting process parameters, and the number of nodes in the output layer is the same as the number of quality indicators.

[0050] The number of smelting process parameters is obtained, and the number of nodes in the input layer of the parameter self-checking neural network model is determined to be the same; the number of quality indicators is obtained, and the number of nodes in the output layer of the parameter self-checking neural network model is determined to be the same. Correlation analysis and principal component analysis are used to select smelting process parameters with correlation with the quality indicators higher than a preset threshold as input features of the parameter self-checking neural network model. The parameter self-checking neural network model is trained using a back propagation algorithm, and weights and biases are updated using a gradient descent method to minimize the mean square error between the predicted output and the actual quality indicator.

[0051] Exemplarily, according to the number of smelting process parameters, the number of nodes in the input layer of the neural network model is determined to be the same. According to the number of quality indicators, the number of nodes in the output layer of the neural network model is determined to be the same. The original data is preprocessed, including data cleaning, missing value processing, outlier processing, etc. The smelting process parameters with high correlation with the quality indicators are selected as the input features of the model by using methods such as correlation analysis and principal component analysis. A three-layer feedforward neural network model is constructed, the input layer receives the selected smelting process parameters, the hidden layer uses the Sigmoid activation function, and the output layer outputs the predicted quality indicators. The hyperparameters such as the number of layers and the number of nodes of the hidden layer are optimized by methods such as grid search. The neural network model is trained by the back propagation algorithm, and the weights and biases are continuously updated by the gradient descent method to minimize the mean square error between the predicted output and the actual quality indicator. Set appropriate training parameters such as learning rate, regularization coefficient and number of iterations. Evaluate the performance of the trained neural network model on the test set, and calculate evaluation indicators such as mean square error and determination coefficient. If the model performance meets the requirements, save the current model parameters; otherwise, adjust the hyperparameters, and retrain and test. Use the trained neural network model to input new smelting process parameters and predict the corresponding quality indicators. Compare the predicted values ​​with the actual values ​​and analyze the error size and distribution. If the predicted quality indicators meet the metallurgical production requirements, the current smelting process parameters are used as optimized process parameters; otherwise, perform a local search near the current parameters, fine-tune the parameter values, and re-enter the neural network model for prediction until the optimal parameter combination that meets the requirements is found.

[0052] Step S102, collect new smelting parameter data samples in real time, input the data samples into the parameter self-checking neural network model, obtain predicted quality index values, compare the predicted values ​​with the actually measured quality indexes, and calculate the prediction errors.

[0053] A sample of smelting parameter data collected in real time is obtained, and after preprocessing the sample, the sample is input into a pre-built parameter self-check neural network model to obtain a predicted smelting product quality index value; the actual measured quality index value is obtained, and the error index between the predicted value and the measured value is calculated; if the error index exceeds a preset threshold, an incremental learning algorithm is used to retrain the parameter self-check neural network model using the newly collected data sample, and the model parameters are updated to obtain a parameter self-check neural network model with better performance, which is applied to subsequent real-time smelting parameter data.

[0054] Exemplarily, a sample of smelting parameter data collected in real time is obtained, and the data is preprocessed, including data cleaning, feature selection, etc., and the preprocessed data is input into a pre-built parameter self-checking neural network model. The parameter self-checking neural network model can select a commonly used machine learning model such as a support vector machine or a random forest. The input data sample is processed according to the parameter self-checking neural network model to obtain a predicted smelting product quality index value, such as component content, hardness, etc. The actual measured quality index value corresponding to the predicted quality index value is obtained, and the commonly used error indicators such as the mean square error between the predicted value and the measured value are calculated. It is determined whether the error index exceeds a preset threshold, such as the mean square error exceeds 1. If it exceeds the threshold, it is determined that the prediction performance of the parameter self-checking neural network model does not meet the requirements, triggering the retraining of the model. Through incremental learning algorithms such as gradient descent, the parameter self-checking neural network model is retrained using the newly collected smelting parameter data samples, and the parameters of the model are updated to obtain a parameter self-checking neural network model with better performance. The updated parameter self-checking neural network model is applied to subsequent real-time smelting parameter data to predict the quality indicators of smelting products and improve the real-time and accuracy of smelting process quality control. By continuously iterating the above steps, the online learning and self-optimization of the parameter self-checking neural network model can be realized to adapt to the dynamic changes of the smelting process.

[0055] Step S103, adopting an incremental learning algorithm, using the newly collected data samples to perform online updating and fine-tuning on the parameter self-checking neural network model, and adjusting the connection weights of the parameter self-checking neural network model according to the prediction error.

[0056] Obtain data samples, preprocess the data samples, remove missing values ​​and outliers, extract key features, and construct feature vectors; use the back propagation algorithm to calculate the gradient of model parameters according to the error loss function, and use the stochastic gradient descent algorithm to update the model weights to minimize the loss function; use the trained model to predict and detect anomalies on new data. If the prediction result exceeds the preset threshold, it is determined that an anomaly exists and a detection report is generated.

[0057] Exemplarily, a newly collected data sample is obtained, and the data is preprocessed, including data cleaning operations such as removing missing values ​​and outliers, and key features are extracted to construct a normalized feature vector. The preprocessed feature vector is input into the established parameter self-checking neural network model, and the predicted output of the model is obtained through the forward propagation calculation process. The difference between the predicted output and the true label is used to construct a mean square error loss function to measure the size of the model prediction error. Based on the error loss function, the back propagation algorithm is used to calculate the gradient of the parameters of each layer of the model, and the stochastic gradient descent optimization algorithm is used to update the connection weights of the model to minimize the loss function. Repeat steps 2-4, and use new data samples to perform multiple rounds of iterative training on the model so that the model gradually adapts to the new data distribution. During the training process, the exponential decay strategy is used to dynamically adjust the learning rate, and the L2 regularization method is used to control the model complexity to avoid overfitting. After each round of iterative training, the data is divided into a training set and a validation set using the holdout method, the prediction accuracy of the model on the validation set is evaluated, the generalization performance of the model is monitored, and the training is terminated in time according to the verification results. Use the trained model to predict and detect anomalies in newly collected business data, output the prediction results, determine whether there are anomalies based on the set threshold, generate a detection report, and provide a reference for business operations.

[0058] Step S104, regularly evaluate the performance of the parameter self-checking neural network model, calculate the average prediction error within a specified time period, and if the error exceeds a preset threshold, trigger the automatic optimization process of the model structure.

[0059] Get the prediction results and true values ​​of the self-check model within the specified time period, and calculate the average prediction error within the time period. Determine whether the average prediction error exceeds the preset error threshold. If it exceeds the threshold, the following automatic optimization process of the model structure is triggered: For the hyperparameters of the self-check model, such as the number of hidden layers, the number of neurons in each layer, and the learning rate, a grid search algorithm is used to traverse different parameter combinations within the preset value range to train multiple candidate models. According to the comprehensive performance indicators of the candidate model on the validation set, such as the weighted average of the accuracy, recall rate, and F1 score, the model hyperparameter combination with the best performance is screened out. Using the optimal hyperparameter combination, retrain the self-check model on the training set to obtain the optimized model. Apply the optimized self-check model to online prediction to improve the prediction accuracy of the system. Continuously monitor the performance of the optimized self-check model online. When the prediction error exceeds the threshold again, repeat the above automatic optimization process to achieve continuous iterative update of the model and keep the prediction performance of the model stable at a high level.

[0060] Exemplarily, the prediction results and true values ​​of the self-check model within a specified time period are obtained, and the average prediction error within the time period is calculated. It is determined whether the average prediction error exceeds the preset error threshold. If it exceeds the threshold, the following automatic optimization process of the model structure is triggered: for the hyperparameters of the self-check model, such as the number of hidden layers, the number of neurons in each layer, the learning rate, etc., a grid search algorithm is used to traverse different parameter combinations within the preset value range to train multiple candidate models. According to the comprehensive performance indicators of the candidate model on the validation set, such as the weighted average of the accuracy, recall rate, and F1 score, the model hyperparameter combination with the best performance is screened out. Using the optimal hyperparameter combination, the self-check model is retrained on the training set to obtain the optimized model. The optimized self-check model is applied to online prediction to improve the prediction accuracy of the system. The online performance of the optimized self-check model is continuously monitored. When the prediction error exceeds the threshold again, the above automatic optimization process is repeated to achieve continuous iterative updating of the model and keep the prediction performance of the model stable at a high level.

[0061] Step S105, using a genetic algorithm to optimize the structure of the parameter self-checking neural network model, encoding the hyperparameters of the parameter self-checking neural network model into genes, and randomly generating candidate models as an initial population.

[0062] A preset number of neural network models are obtained as an initial population, wherein the hyperparameters of the parameter self-checking neural network model are encoded as binary gene sequences; for each of the parameter self-checking neural network models, training and testing are performed on a validation data set, and the accuracy of the parameter self-checking neural network model on the validation set is used as its fitness value; it is determined whether the average fitness value of the current population has not been significantly improved in multiple consecutive iterations, or whether the number of iterations has reached a preset upper limit; if so, the parameter self-checking neural network model with the highest fitness value in the current population is determined as the optimal model.

[0063] Exemplarily, a preset number of neural network models are randomly generated as the initial population, and the hyperparameters of each model are encoded as binary gene sequences. Hyperparameters include the number of hidden layers, the number of neurons in each layer, the type of activation function, the learning rate, etc. For each neural network model, training and testing are performed on the validation data set, and the accuracy of the model on the validation set is used as its fitness value. According to the fitness value of the neural network model, a roulette selection algorithm is used to select several models from the current population as parent individuals. In the roulette selection algorithm, the higher the fitness value, the greater the probability of the individual being selected. For the selected parent individuals, two individuals are randomly selected, and their gene sequences are recombined by single-point crossover, that is, a position is randomly selected and the gene fragments after the position are exchanged, thereby generating two new offspring neural network models. At the same time, a bit reversal mutation is performed on each bit in the gene sequence with a certain probability to introduce random disturbances. The generated offspring neural network model is added to the current population, and the fitness values ​​of all models in the population are calculated at the same time, and a part of the models with the lowest fitness value are eliminated to keep the population size unchanged. Determine whether the average fitness value of the current population has not been significantly improved for multiple consecutive iterations, or whether the number of iterations has reached the preset upper limit. If so, the neural network model with the highest fitness value in the current population is taken as the optimal model, and its hyperparameter combination is taken as the optimal solution; if not, return to step 2 to continue the next round of iterative optimization. Based on the obtained optimal hyperparameter combination, build and train the final neural network model for subsequent application deployment. Retrain the model on the new training data set and evaluate its performance on the test set to further verify the effect of hyperparameter optimization.

[0064] Step S106, evaluate each candidate model in the population, calculate the fitness function value, generate a new generation of population through genetic operators such as selection, crossover, and mutation, iterate and optimize until the termination condition is met, and use the model with the highest fitness in the final population as the optimized parameter self-checking neural network model.

[0065] Acquire characteristic parameters of candidate models, construct a fitness function, and use the fitness function value as an indicator to measure the quality of the candidate models; adopt a roulette wheel selection operator to select at least one high-quality model with a fitness higher than a preset threshold from the current population according to the fitness value of the candidate models; for the high-quality models screened out by the roulette wheel selection operator, adopt an arithmetic crossover operator to randomly pair them in pairs and cross-recombine their characteristic parameters to obtain the recombined candidate models.

[0066] Exemplarily, for the characteristic parameters of the candidate model, a fitness function for evaluating the model performance is constructed, and the fitness function value is used as an indicator to measure the quality of the candidate model. The candidate model adopts a floating point encoding method, and each candidate model contains n characteristic parameters, which respectively represent the various attributes of the model. M candidate models are randomly generated to form an initial population. For each candidate model in the population, the corresponding characteristic parameters are substituted, and its fitness function value is calculated to obtain the fitness value of all models in the current population. A roulette selection operator is used to select a part of high-quality models with higher fitness from the current population according to the fitness value of the candidate model. The basic principle of roulette selection is that the higher the fitness of the model, the greater the probability of being selected. For the high-quality models selected by the roulette selection operator, an arithmetic crossover operator is used to randomly pair them and cross-recombine their characteristic parameters. Arithmetic crossover refers to the weighted average of the same characteristic parameters of two models to generate new characteristic parameter values. For the new candidate models generated by arithmetic crossover recombination, a uniform mutation operator is used to randomly perturb their characteristic parameters with a certain probability to introduce new diversity. Uniform mutation refers to making small random perturbations to each feature parameter with equal probability within the range of parameter values. New candidate models generated by genetic operators such as roulette wheel selection, arithmetic crossover, and uniform mutation are added to the population. An elite retention strategy is adopted to directly retain the top k% models with the highest fitness in the current population to the next generation, and the remaining models are replaced by newly generated models to form a new generation of populations, and the model parameters are iteratively optimized. The average fitness and optimal fitness of the current population are calculated to determine whether one of the following termination conditions is met: (1) The number of iterations reaches the preset maximum number of iterations; (2) The average fitness increase of several consecutive generations of populations is less than the preset convergence threshold; (3) The optimal fitness reaches the preset target threshold. If the termination condition is met, the iteration is stopped and the model with the highest fitness in the population is output as the optimal model; otherwise, return to step 2 and continue iterative optimization.

[0067] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for automatically adjusting production parameters of bonding wire, characterized in that: The method comprises: Constructing the parameter self-checking neural network model, the parameter self-checking neural network model comprises an input layer, a hidden layer and an output layer, the number of nodes in the input layer is the same as the number of smelting process parameters, and the number of nodes in the output layer is the same as the number of quality indicators; Collecting new smelting parameter data samples in real time, inputting the data samples into the parameter self-checking neural network model, obtaining predicted quality index values, comparing the predicted values ​​with the actually measured quality indexes, and calculating the prediction errors; Adopting an incremental learning algorithm, using the newly collected data samples to perform online updating and fine-tuning on the parameter self-checking neural network model, and adjusting the connection weights of the parameter self-checking neural network model according to the prediction error; Regularly evaluate the performance of the parameter self-checking neural network model, calculate the average prediction error within a specified time period, and trigger an automatic optimization process of the model structure if the error exceeds a preset threshold; A genetic algorithm is used to optimize and search the structure of the parameter self-checking neural network model, hyperparameters of the parameter self-checking neural network model are encoded as genes, and candidate models are randomly generated as an initial population; Each candidate model in the population is evaluated, the fitness function value is calculated, a new generation of population is generated by selecting, crossing over, and mutating genetic operators, and iterative optimization is performed until the termination condition is met, and the model with the highest fitness in the final population is used as the optimized parameter self-checking neural network model.

2. The method according to claim 1, characterized in that The parameter self-checking neural network model is constructed, and the parameter self-checking neural network model comprises an input layer, a hidden layer and an output layer, the number of nodes in the input layer is the same as the number of smelting process parameters, and the number of nodes in the output layer is the same as the number of quality indicators, including: Obtaining the number of smelting process parameters, determining the number of nodes in the input layer of the parameter self-checking neural network model to make them the same; Obtaining the number of quality indicators, determining the number of nodes in the output layer of the parameter self-checking neural network model to make them the same; Using correlation analysis and principal component analysis, select smelting process parameters whose correlation with the quality index is higher than a preset threshold as input features of the parameter self-checking neural network model; The parameter self-checking neural network model is trained by a back propagation algorithm, and weights and biases are updated by a gradient descent method to minimize the mean square error between the predicted output and the actual quality index.

3. The method according to claim 2, characterized in that The real-time collection of new smelting parameter data samples, inputting the data samples into the parameter self-checking neural network model, obtaining predicted quality index values, comparing the predicted values ​​with the actually measured quality indexes, and calculating the prediction errors include: Acquire smelting parameter data samples collected in real time, pre-process the data samples, and input them into a pre-built parameter self-checking neural network model to obtain predicted smelting product quality index values; Obtaining the quality index value of the actual measurement, and calculating the error index between the predicted value and the measured value; If the error index exceeds a preset threshold, an incremental learning algorithm is used to retrain the parameter self-check neural network model using newly collected data samples, update the model parameters, and obtain a parameter self-check neural network model with better performance, which is applied to subsequent real-time smelting parameter data.

4. The method according to claim 3, characterized in that The incremental learning algorithm is used to update and fine-tune the parameter self-checking neural network model online using the newly collected data samples, and the connection weights of the parameter self-checking neural network model are adjusted according to the prediction error, including: Obtaining data samples, preprocessing the data samples, removing missing values ​​and outliers, extracting key features, and constructing feature vectors; According to the error loss function, the back propagation algorithm is used to calculate the model parameter gradient, and the stochastic gradient descent algorithm is used to update the model weights to minimize the loss function; The trained model is used to predict and detect anomalies on new data. If the prediction result exceeds the preset threshold, it is determined that an anomaly exists and a detection report is generated.

5. The method according to claim 4, characterized in that The periodic evaluation of the performance of the parameter self-checking neural network model, calculation of the average prediction error within a specified time period, and triggering the automatic optimization process of the model structure if the error exceeds a preset threshold, include: Get the prediction results and true values ​​of the self-checking model within a specified time period, and calculate the average prediction error within the time period; Determine whether the average prediction error exceeds the preset error threshold. If it exceeds the threshold, the following automatic optimization process of the model structure is triggered: for the hyperparameters of the self-checking model, such as the number of hidden layers, the number of neurons in each layer, and the learning rate, a grid search algorithm is used to traverse different parameter combinations within the preset value range to train multiple candidate models; Based on the comprehensive performance indicators of the candidate model on the validation set, such as the weighted average of accuracy, recall, and F1 score, the model hyperparameter combination with the best performance is selected; Using the optimal hyperparameter combination, retrain the self-checking model on the training set to obtain the optimized model; Apply the optimized self-check model to online prediction to improve the prediction accuracy of the system; Continuously monitor the online performance of the optimized self-checking model. When the prediction error exceeds the threshold again, repeat the above automatic optimization process to achieve continuous iterative update of the model and keep the model's prediction performance stable at a high level.

6. The method according to claim 5, characterized in that The method of using a genetic algorithm to optimize the structure of the parameter self-checking neural network model, encoding the hyperparameters of the parameter self-checking neural network model into genes, and randomly generating candidate models as an initial population includes: Obtaining a preset number of neural network models as an initial population, wherein the hyperparameters of the parameter self-checking neural network model are encoded as binary gene sequences; For each of the parameter self-checking neural network models, training and testing are performed on a validation data set, and the accuracy of the parameter self-checking neural network model on the validation set is used as its fitness value; Determine whether the average fitness value of the current population has not been significantly improved in multiple consecutive iterations, or whether the number of iterations has reached a preset upper limit; If so, the parameter self-checking neural network model with the highest fitness value in the current population is determined as the optimal model.

7. The method according to claim 6, characterized in that The method of evaluating each candidate model in the population, calculating the fitness function value, generating a new generation of population by selecting, crossing, and mutating genetic operators, iteratively optimizing until the termination condition is met, and taking the model with the highest fitness in the final population as the optimized parameter self-checking neural network model, comprises: Acquire characteristic parameters of the candidate model, construct a fitness function, and use the fitness function value as an indicator to measure the quality of the candidate model; A roulette wheel selection operator is used to select at least one high-quality model with a fitness higher than a preset threshold from the current population according to the fitness values ​​of the candidate models; For the high-quality models selected by the roulette wheel selection operator, an arithmetic crossover operator is used to randomly pair them in pairs and cross-recombine their characteristic parameters to obtain a recombined candidate model.

8. An automatic adjustment system for production parameters of bonding wires implementing any of the methods 1 to 7, characterized in that: include: Data collection module, used to collect sample data and perform preprocessing; Parameter self-checking module, used to compare the collected sample parameters and calculate the prediction error; The incremental update module is used to collect new data samples and update the data of the parameter self-check module; Evaluation module, which is used to evaluate the performance of candidate models.

Images