Intelligent temperature-controlled lead-free tin bar production method and device

Abstract

The invention relates to the technical field of intelligent temperature control, and discloses an intelligent temperature-controlled lead-free tin bar production method and device, wherein the method comprises: collecting and preprocessing data of multiple sensors of a production line to obtain a time series data set; performing attention-guided position-sensitive multiple interpolation to obtain a target data set; constructing and training an abnormality recognition model based on the target data set; analyzing normal production data screened by the abnormality recognition model and an implicit subspace recognition recurrent neural network based on orthogonal basis decomposition and reconstruction to obtain a production process parameter model; based on the production process parameter model, performing parameter optimization using a multi-task feature fusion framework to obtain target production process parameters. The method can effectively handle the problem of missing data in the production process, accurately identify abnormal situations, and can perform adaptive process parameter optimization according to the characteristics of different production lines or batches, thereby improving the production quality and efficiency of lead-free tin bars.

Description

Technical Field

[0001] The present invention relates to the field of intelligent temperature control technology, and in particular to an intelligent temperature control lead-free tin bar production method and device. Background Art

[0002] The traditional lead-free solder bar production method has problems such as unstable temperature control, inaccurate raw material ratio, and difficulty in timely detection of abnormal conditions during the production process, which seriously affects product quality and production efficiency. At the same time, due to the complex interaction of multiple parameters involved in the production process, the optimization of process parameters has become a major problem.

[0003] Although existing intelligent manufacturing technologies have improved the control and management of the production process to a certain extent, they are still insufficient when dealing with high-dimensional and nonlinear production data. In particular, when faced with problems such as data loss, anomaly detection, and multi-task optimization in the production process, traditional methods are difficult to meet actual needs. In addition, the differences between different production lines or batches also bring challenges to the optimization of process parameters. Summary of the invention

[0004] The present invention provides a method and device for producing an intelligent temperature-controlled lead-free solder bar. The present invention can effectively handle the problem of missing data in the production process, accurately identify abnormal situations, and adaptively optimize process parameters according to the characteristics of different production lines or batches, thereby improving the production quality and efficiency of the lead-free solder bar.

[0005] In a first aspect, the present invention provides a method for producing an intelligent temperature-controlled lead-free solder bar, the method comprising:

[0006] Collect and preprocess data from multiple sensors on the production line to obtain a time series data set containing temperature, pressure, flow and composition;

[0007] Performing attention-guided position-sensitive multiple interpolation on missing values ​​in the time series dataset to obtain a target dataset;

[0008] Based on the target data set, a fuzzy neural network with a self-organizing loop structure is constructed and trained to obtain an anomaly recognition model;

[0009] According to the normal production data screened by the abnormality recognition model, the implicit subspace recognition regression neural network based on orthogonal basis decomposition and reconstruction is analyzed to obtain a production process parameter model;

[0010] Based on the production process parameter model, a multi-task feature fusion framework is used to perform parameter optimization to obtain target production process parameters.

[0011] In a second aspect, the present invention provides an intelligent temperature-controlled lead-free solder bar production device, the intelligent temperature-controlled lead-free solder bar production device comprising:

[0012] The acquisition module is used to collect and preprocess the data of multiple sensors on the production line to obtain a time series data set containing temperature, pressure, flow and composition;

[0013] An interpolation module, configured to perform attention-guided position-sensitive multiple interpolation on missing values ​​in the time series dataset to obtain a target dataset;

[0014] A construction module is used to construct and train a fuzzy neural network with a self-organizing loop structure based on the target data set to obtain an anomaly recognition model;

[0015] An analysis module is used to analyze the normal production data screened by the abnormality recognition model and obtain a production process parameter model based on an implicit subspace recognition regression neural network decomposed and reconstructed by an orthogonal basis;

[0016] The optimization module is used to optimize the parameters based on the production process parameter model using a multi-task feature fusion framework to obtain target production process parameters.

[0017] The third aspect of the present invention provides an intelligent temperature-controlled lead-free tin bar production device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the intelligent temperature-controlled lead-free tin bar production device executes the above-mentioned intelligent temperature-controlled lead-free tin bar production method.

[0018] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned intelligent temperature-controlled lead-free solder bar production method.

[0019] In the technical solution provided by the present invention, the position-sensitive multiple interpolation method guided by attention effectively solves the problem of missing data in the production process and improves the data quality and reliability. The fuzzy neural network with a self-organizing cyclic structure is used to construct an abnormality recognition model, which can accurately detect abnormal conditions in the production process and reduce production risks. The implicit subspace recognition regression neural network based on orthogonal basis decomposition and reconstruction realizes the accurate modeling of production process parameters and provides a reliable basis for parameter optimization. The multi-task feature fusion framework is introduced to consider the differences between different production lines or batches, and more accurate and personalized process parameter optimization is achieved. The ResNet structure and variational autoencoder enhanced by the attention mechanism are used to improve the effect of feature extraction and representation and enhance the generalization ability of the model. Through the multi-objective optimization method, multiple objectives such as product quality, production efficiency and energy consumption are considered at the same time, and the comprehensive optimization of the production process is achieved. From data acquisition, preprocessing, abnormality detection to parameter optimization, the present invention comprehensively improves the intelligent level of lead-free tin bar production.

[0020] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0021] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic diagram of an embodiment of a method for producing a lead-free tin bar with intelligent temperature control in an embodiment of the present invention;

[0023] Figure 2 A schematic diagram of an embodiment of an intelligent temperature-controlled lead-free solder bar production device in an embodiment of the present invention;

[0024] Figure 3 It is a schematic diagram of an embodiment of an intelligent temperature-controlled lead-free solder bar production equipment in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0027] To facilitate understanding of this embodiment, firstly, a method for producing an intelligent temperature-controlled lead-free solder bar disclosed in an embodiment of the present invention is described in detail. Figure 1 As shown, the method comprises the following steps:

[0028] 101. Collect and preprocess data from multiple sensors on the production line to obtain a time series data set including temperature, pressure, flow rate and composition;

[0029] It is understandable that the execution subject of the present invention may be an intelligent temperature-controlled lead-free solder bar production device, or a terminal or a server, which is not specifically limited here. The embodiment of the present invention is described by taking a server as an execution subject as an example.

[0030] Specifically, data is collected from the temperature sensors, pressure sensors, flow sensors and component analyzers on the production line to obtain the original data set. In order to ensure the timeliness and accuracy of the data, the collected original data set is sampled at a frequency of once per second to obtain the sampled data set. In order to improve the quality and availability of the data, the data points in the sampled data set that exceed the preset physical range are eliminated to eliminate the outliers caused by equipment failure or external interference to form a preliminary screening data set. A sliding window of 60 seconds is applied to the preliminary screening data set. In each window, the average value of the data is calculated to smooth the data to eliminate the impact of short-term fluctuations and obtain a smoothed data set. The smoothed data set is normalized to unify the numerical range of the data to between 0 and 1, eliminate the impact of different sensor data units and dimensional differences, and obtain a normalized data set. Based on the normalized data set, the correlation coefficient matrix between the parameters is calculated to obtain the correlation data and reveal the relationship between different parameters. The normalized data set and the correlation data are combined according to the timestamp to form a complete time series data set, which contains the time series information of multiple parameters such as temperature, pressure, flow and composition.

[0031] 102. Perform attention-guided position-sensitive multiple interpolation on missing values ​​in a time series dataset to obtain a target dataset;

[0032] Specifically, the time series data set is scanned to identify the location of missing values ​​and their corresponding parameter types, including temperature, pressure, flow rate and composition parameters, to obtain missing value information and clarify their location and properties in the time series. Based on the missing value information, the time interval and parameter correlation coefficient between each missing value and the 10 non-missing values ​​before and after are calculated to construct a position weight matrix. This matrix reveals the relative position of the missing value in the time series and provides correlation information between adjacent data points, so that the subsequent interpolation process can be based on the local structure of the data rather than isolated processing. In this way, the interpolation results are ensured to be consistent with the surrounding known data points, and the credibility of the data is enhanced. According to the position weight matrix, different parameter types are assigned corresponding weights. The temperature parameter is assigned a weight of 1.5, the pressure parameter a weight of 1.3, the flow parameter a weight of 1.2, and the composition parameter a weight of 1.1. The weighting mechanism is designed according to the importance of each parameter in the production process, which can make the model pay more attention to key data when interpolating and improve the accuracy of the final interpolation value. Through weight assignment, the parameter attention distribution vector is obtained. Based on the parameter attention distribution vector, the 10 non-missing values ​​before and after each missing value are weighted averaged. The weight is the product of the attention score and the inverse of the time interval, which emphasizes the importance of non-missing values ​​close in time to the interpolation result. The weighted average process can generate an initial estimate. In order to enhance the randomness and diversity of the interpolation, a random perturbation following the normal distribution N(0,0.1) is added on the basis of the initial estimate. A certain degree of randomness is introduced so that the interpolation result is not too concentrated. At the same time, 50 candidate interpolation values ​​are generated to form a candidate value set. The candidate value set is input into the pre-trained masked attention network for processing. The network contains three fully connected layers, the size of the hidden layer is set to 64, and the ReLU activation function is used to introduce nonlinear features. Through this network, the attention score of each candidate value is calculated to reflect the importance of each candidate value in the overall interpolation process. The candidate value weight vector is used to perform a weighted average on the candidate value set, and the calculation of the weight depends on the attention score after the softmax function. In this way, the most representative candidate value is extracted to obtain the final interpolation value. The interpolation values ​​are filled back into the missing positions of the original time series data set to ensure the continuity and integrity of the data. In order to make the interpolated data smoother, linear interpolation is performed on the filled data to eliminate the noise introduced during the interpolation process, improve the smoothness and reliability of the data, and obtain the target data set.

[0033] Among them, the candidate value set is normalized. By mapping each candidate value to the range of [-1,1], the dimension difference between different candidate values ​​is eliminated, making the subsequent calculation more reasonable and effective, and obtaining a normalized candidate value set. The normalized candidate value set is concatenated with the corresponding timestamp information to form an input vector. Taking the temporal characteristics of the candidate values ​​into consideration, the timestamp can provide additional contextual information for the network, making the model more accurate in judging the validity of the candidate values. The formed input vector is input into the first fully connected layer of the masked attention network. In this layer, the network will perform a linear transformation and generate the first layer output. The ReLU activation function is applied to the output to introduce a nonlinear transformation and obtain the activation value of the first layer. The use of the ReLU activation function effectively solves the gradient vanishing problem, making the network more expressive when learning complex patterns. The first layer activation value is input into the second fully connected layer of the masked attention network, a linear transformation is performed, the second layer result is output, and the ReLU activation function is applied to the output to generate the activation value of the second layer. By passing information layer by layer, the network gradually learns the deep relationship between the candidate values ​​and their temporal characteristics. After passing through the second layer, the activation values ​​of the second layer are input into the third fully connected layer of the masked attention network to generate the original attention score. The original attention score is the network's preliminary evaluation of the importance of each candidate value. The Softmax function is applied to the original attention score to normalize the original score to the range of [0,1] to ensure that the sum of the scores of all candidate values ​​is 1, forming a set of attention weights with relative importance. Based on the attention weights, the masked attention score corresponding to each candidate value is calculated to reflect the importance of each candidate value in the interpolation process, and the candidate value attention score vector is obtained. The candidate value attention score vector is normalized to ensure its effectiveness and consistency in subsequent use, and the candidate value weight vector is obtained.

[0034] 103. Construct and train a fuzzy neural network with a self-organizing loop structure based on the target data set to obtain an anomaly recognition model;

[0035] Specifically, the target data set is segmented into time windows, the window size is set to 100 time steps, and the step length is 10, which effectively extracts the dynamic features in the time series and obtains the input sequence set. This process enables the network to capture abnormal trends that may occur in the production process in different time periods. A fuzzy neural network with a five-layer structure is constructed, including an input layer, a fuzzification layer, a rule layer, a defuzzification layer, and an output layer. The number of nodes in the input layer is the same as the number of monitoring parameters, while the output layer is designed as a single node to indicate the degree of abnormality. Through this structural design, the network can effectively integrate multi-dimensional monitoring data and improve the accuracy of anomaly detection. The input sequence set is fuzzified. Each input value is mapped to the interval [0,1] using a Gaussian membership function to obtain a fuzzy input. This mapping not only helps to simplify data processing, but also enhances the robustness of the network to the uncertainty of input data. The initial number of fuzzy rules is set to 10 in the rule layer, and each rule is represented by an IF-THEN statement, where the IF part corresponds to the fuzzy set of the input variable and the THEN part corresponds to the fuzzy set of the output variable. In order to effectively utilize historical information and balance the importance of information at the current moment, the network introduces a recursive parameter α, ranging from 0 to 1. By splicing the fuzzified input with the α-fold result of the previous moment output, an extended input is formed to ensure that the model not only considers the current input but also absorbs the influence of historical data. The extended input is passed to the rule layer, the activation strength of each rule is calculated, and the algebraic product is used as the T-norm operator to obtain the activation of the rule. The calculation of the activation directly affects the final anomaly recognition effect. In the defuzzification layer, the center average method is used to weighted average the rule activation and the center value of the THEN part of each rule to obtain a clear output, converting fuzzy information into a clear output result, so that the network can output the specific value of the degree of anomaly. In the training and optimization process of the network, the network structure adjustment indicators are defined, including coverage, discrimination and importance. Initially, the threshold of coverage is set to 0.8, and the thresholds of discrimination and importance are set to 0.2 and 0.1 respectively. These indicators are used to dynamically adjust the number of fuzzy rules in the network to ensure the adaptability of the network under different data conditions. For example, when the coverage is lower than the set threshold, the network will add new fuzzy rules to ensure that the model fully covers the data; when the discrimination or importance is lower than the corresponding threshold, the redundant rules will be deleted to optimize the network structure. In order to ensure the efficient learning of network parameters, the improved stochastic gradient descent method is used for optimization. The initial learning rate is set to 0.01, and the learning rate decays to 0.9 times the original after every 1000 iterations. This strategy can not only effectively prevent overfitting, but also improve the convergence speed of the model. At the same time, the improved recursive least squares method is applied to process the output weights of the network to improve the accuracy of the model.In the process of iterative training, the model convergence and stability are ensured until the loss function value on the validation set changes less than 0.001 for 10 consecutive iterations or the maximum number of iterations reaches 10,000. An efficient anomaly recognition model is obtained. This model not only has the ability to detect anomalies in the production process in real time, but also can adapt to different production environments and data characteristics through the design of self-organizing structures.

[0036] 104. The normal production data is screened by the abnormal recognition model, and the implicit subspace recognition regression neural network based on orthogonal basis decomposition and reconstruction is analyzed to obtain the production process parameter model;

[0037] Specifically, the target data set is input into the anomaly recognition model to obtain the anomaly score at each time point. The anomaly score is thresholded, and the threshold is set to 0.7. Data points with anomaly scores below the threshold are marked as normal data. The marking process can effectively filter out stable and reliable normal production data from the original data. The normal production data is segmented by sliding windows. The window size is set to 1000 time steps and the step length is 100, which can generate a data window set. Each window data represents the production status within a period of time. This segmentation method is used to analyze the dynamic changes of production in different time periods. Fast Fourier transform is performed on each window data in the data window set to obtain its frequency domain representation. The frequency domain representation can effectively reveal the frequency characteristics of the signal and provide a basis for subsequent feature selection. The first K frequency components with an energy share of 90% are selected from the frequency domain features as the frequency domain features after dimensionality reduction. The dimensionality reduction process simplifies the dimension of the data, retains the main information of the signal, and avoids the interference of redundant data on model training. Based on the frequency domain features after dimensionality reduction, an implicit subspace recognition regression neural network with an encoder-decoder structure is constructed. The encoder and decoder of the network both contain three fully connected layers, and the number of nodes in the hidden layer is 128, 64 and 32 respectively. This structural design enables the network to effectively capture the potential characteristics of the data during the learning process and realize complex nonlinear mapping. The frequency domain features after dimensionality reduction are input into the encoder to obtain the latent space representation. Based on the latent space representation, a nonlinear autoregressive moving average model is constructed, and the order of the model is determined by the Akaike information criterion to ensure the effectiveness and accuracy of the model when fitting the data. The output of the system dynamic model is reconstructed by the decoder to obtain time domain reconstructed data to reflect the actual state of the production process. The reconstruction error is calculated based on the time domain reconstructed data and the original normal production data. The performance and accuracy of the model are evaluated by comparing the difference between the reconstructed data and the real data. The reconstruction error is combined with the L1 regularization term to form a new loss function. The model is trained using the Adam optimizer, and the initial learning rate is set to 0.001, which decays by 10% after every 50 epochs. The learning rate decay strategy is designed to improve the convergence speed of the model and prevent overfitting in the later stages of training. The goal of model training is to make the loss on the validation set no longer decrease for 10 consecutive epochs, or to reach the maximum number of iterations of 5000. Through the settings, the model training process is effectively monitored to ensure that its performance on the validation set is continuously optimized, and finally an efficient production process parameter model is obtained.

[0038] Furthermore, the frequency domain features after dimensionality reduction are input into the first fully connected layer of the encoder, and the first weight matrix and the first bias vector are used for linear transformation to generate the output of the first layer. The ReLU activation function is applied to the output for nonlinear transformation to obtain the activation value of the first layer. Through this process, the model can enhance its learning ability of complex patterns while maintaining linear features. The activation value of the first layer is passed to the second fully connected layer of the encoder. The second layer uses the corresponding second weight matrix and the second bias vector for linear transformation, and the ReLU activation function is applied to obtain the activation value of the second layer. After processing, the information gradually extracts more abstract features in the process of layer-by-layer transmission. The activation value of the second layer is input into the third fully connected layer of the encoder, and the third weight matrix and the third bias vector are also used for linear transformation to obtain the latent space representation. Based on the latent space representation, a nonlinear autoregressive moving average model is constructed. The order p of the model increases gradually from 1 to 10, and the Akaike Information Criterion (AIC) value is calculated for each order to select the optimal order p*. The calculation process of the AIC value effectively balances the complexity and fit of the model, and selects the order that minimizes the AIC value as the order of the final model. At the same time, in order to improve the effectiveness of the model, the order q of the moving average term is set to p* / 2, and a system dynamic model expressed as NARMA(p*,q) is obtained. The system dynamic model is used to perform time series prediction on the latent space representation to generate prediction results. The prediction results are input into the first fully connected layer of the decoder. The first layer of the decoder uses the fourth weight matrix and the fourth bias vector for linear transformation, and applies the ReLU activation function to obtain the activation value of the first layer of the decoder. After the first layer is processed, the information is converted into a format suitable for decoding. The activation value of the first layer of the decoder is passed to the second fully connected layer of the decoder, linearly transformed, and the fifth weight matrix and the fifth bias vector are used, and the ReLU activation function is applied to obtain the activation value of the second layer of the decoder. This series of transformations ensures that information is effectively transmitted and transformed during the decoding process, making the final result more accurate. The activation value of the second layer of the decoder is input to the third fully connected layer of the decoder, and a linear transformation is performed using the sixth weight matrix and the sixth bias vector, followed by an inverse fast Fourier transform to convert the frequency domain data back to the time domain to obtain the time domain reconstructed data. The features transformed in the latent space are restored to the time domain signals required in the production process to achieve comprehensive analysis and monitoring of process parameters.

[0039] 105. Based on the production process parameter model, the multi-task feature fusion framework is used to optimize the parameters and obtain the target production process parameters.

[0040] Specifically, the output data of the production process parameter model are grouped according to different production lines or batches to obtain multiple subtask data sets. A subpopulation is constructed for each subtask data set, and the initial population size is set to 100. Each individual represents a specific set of production process parameters, forming an initial multi-task population, which provides a variety of candidate solutions for subsequent parameter optimization to ensure that the best parameter combination can be found in different production scenarios. Based on the initial multi-task population, a ResNet structure feature extractor with enhanced attention mechanism is constructed. The feature extractor contains 9 residual blocks, each of which consists of two convolutional layers and a short-circuit connection. In order to improve the effectiveness of feature extraction, a self-attention layer is added between the residual blocks so that the network can focus on important information areas when extracting features. This structure can not only effectively capture complex patterns in the data, but also share knowledge between different tasks and enhance the generalization ability of the model. The output of the feature extraction network is input into the variational autoencoder. Both the encoder and decoder use a three-layer fully connected network, with the hidden layer sizes set to 64 and 32 respectively, and the dimension of the latent space is set to 16. Through this process, the model can map the input data to the latent space and obtain the task feature representation. Based on the task feature representation, the similarity matrix between different tasks is calculated, and the task correlation matrix is ​​obtained using cosine similarity as the metric. Using this matrix, the features of each subtask are weighted and fused, and the weight is determined by the softmax value of the similarity. Through the weighted fusion process, the feature information of multiple subtasks can be effectively integrated to generate a fused feature vector. The fused feature vector is input into the production process parameter model to obtain the objective function value of each subtask, including important parameters such as product quality indicators, production efficiency and energy consumption. After these objectives are normalized, they are weighted and summed to obtain a comprehensive evaluation index that reflects the overall performance of the production process parameters. Based on the comprehensive evaluation index, optimization operations are performed on the multi-task population, including tournament selection, simulated binary crossover and polynomial mutation. These operations help to explore a better combination of production process parameters. Through these evolution strategies, the optimal solution can be effectively found in the multidimensional parameter space to obtain the target production process parameters.

[0041] The task feature representation is L2 normalized to ensure that the modulus of each feature vector is 1, and the feature representation is converted into a unit vector form to eliminate the dimensional difference between features. Based on the feature representation in unit vector form, the dot product between any two task feature vectors is calculated to generate the original similarity matrix for each pair of tasks, which can reflect the degree of correlation between tasks. The original similarity matrix is ​​symmetric to eliminate the asymmetry in the matrix. By calculating the average value of the original similarity matrix S and its transposed matrix ST, that is, taking (S + ST) / 2, a symmetric similarity matrix is ​​obtained to ensure that the similarity relationship can be reflected bidirectionally between tasks. The softmax function is applied to each row of the symmetric similarity matrix to convert the similarity value into a probability distribution to obtain the task correlation matrix. Through the softmax conversion, the similarity value is normalized to a probability value between 0 and 1, which can more intuitively reflect the degree of correlation between tasks. Based on the task correlation matrix, the feature vector of each subtask is weighted and summed, and the weight is the element value of the corresponding row in the correlation matrix, which effectively integrates the feature information of different tasks to generate the initial fusion feature. The initial fusion feature is layer-normalized. By calculating the mean and variance of the features and performing standardization, the distribution of each feature can be made more balanced, improving the training effect of the subsequent model. After standardization, the standardized fusion features obtained are better adapted to the learning needs of the model. The standardized fusion features are residually connected with the original task features, and the two are added to obtain the fused feature representation. The introduction of residual connections helps alleviate the gradient vanishing problem in deep networks, allowing information to propagate more smoothly in the network. The ReLU activation function is applied to the fused feature representation for nonlinear transformation to enhance the expressive power of the features, ensure that the model can learn more complex patterns, and improve the prediction and optimization effects in the production process. Obtain a fused feature vector.

[0042] In the embodiment of the present invention, the position-sensitive multiple interpolation method guided by attention effectively solves the problem of missing data in the production process and improves the data quality and reliability. The fuzzy neural network with a self-organizing loop structure is used to construct an abnormality recognition model, which can accurately detect abnormal conditions in the production process and reduce production risks. The implicit subspace recognition regression neural network based on orthogonal basis decomposition and reconstruction realizes the accurate modeling of production process parameters and provides a reliable basis for parameter optimization. The multi-task feature fusion framework is introduced to consider the differences between different production lines or batches, and more accurate and personalized process parameter optimization is achieved. The ResNet structure and variational autoencoder enhanced by the attention mechanism are used to improve the effect of feature extraction and representation, and enhance the generalization ability of the model. Through the multi-objective optimization method, multiple objectives such as product quality, production efficiency and energy consumption are considered at the same time, and the comprehensive optimization of the production process is achieved. From data acquisition, preprocessing, abnormality detection to parameter optimization, the present invention comprehensively improves the intelligent level of lead-free tin bar production.

[0043] In a specific embodiment, the process of executing step 101 may specifically include the following steps:

[0044] Collect data from the temperature sensor, pressure sensor, flow sensor and component analyzer on the production line to obtain an original data set, and sample the original data set at a frequency of once per second to obtain a sampled data set;

[0045] Data points in the sampled data set that are beyond the preset physical range are removed to obtain a preliminary screening data set, and a 60-second sliding window is applied to the preliminary screening data set to calculate the average value of the data in the window to obtain a smoothed data set;

[0046] Normalize the smoothed data set to unify the value range to between 0 and 1 to obtain a normalized data set, and calculate the correlation coefficient matrix between the parameters based on the normalized data set to obtain correlation data;

[0047] The normalized data set and the correlation data are combined according to the timestamp to obtain a time series data set.

[0048] Specifically, data is collected from the temperature sensors, pressure sensors, flow sensors, and component analyzers on the production line, including the readings of each sensor within a specific time period, including temperature values ​​(T), pressure values ​​(P), flow values ​​(F), and component concentrations (C). The original data set is sampled at a frequency of once per second to obtain a uniformly distributed sampling data set. Through sampling, the dynamic changes in the production process can be effectively captured. Data points in the sampled data set that exceed the preset physical range are removed to ensure data quality. In order to reduce the impact of data fluctuations on the analysis results, a 60-second sliding window is applied to the preliminary screening data set to calculate the average value of the data in the window. For any time t, the average value of the data in the window is expressed by the following formula:

[0049] ;

[0050] in, represents the average value at time t, N is the size of the window (60), and is the data at each time point in the window. By sliding the window, the smoothed data set can effectively remove short-term fluctuations, making the data more stable and reflecting the trend of the production process more clearly. The smoothed data set is normalized to unify all numerical ranges between 0 and 1. The purpose of normalization is to eliminate the influence of different dimensions and magnitudes so that each parameter can be compared and analyzed on the same scale. The normalization formula is usually:

[0051] ;

[0052] in, is the normalized value, is the original data, and Represent the minimum and maximum values ​​in the data set, respectively. After normalization, the normalized data set obtained maps all sensor readings to the range of 0 to 1. Based on the normalized data set, the correlation coefficient matrix between each parameter is calculated to understand the relationship between different sensor readings. The correlation coefficient matrix is ​​obtained by calculating the Pearson correlation coefficient between each pair of variables. The formula of the Pearson correlation coefficient is:

[0053] ;

[0054] in, represents the correlation coefficient between variables X and Y, is the covariance of X and Y, and and are the standard deviations of X and Y, respectively. The calculated correlation data can reveal the intrinsic connection between the sensors. For example, if the correlation coefficient between temperature and pressure is high, it means that the two are closely related in the production process. The normalized data set and the correlation data are combined according to the timestamp to obtain a time series data set. The time series data set combines the readings of each sensor at the same time point and the correlation information between them to form a rich source of information. For example, when performing fault detection and predictive maintenance, the information in the time series data set can more quickly identify potential problems and take appropriate measures to ensure the smooth operation of the production line.

[0055] In a specific embodiment, the process of executing step 102 may specifically include the following steps:

[0056] Scan the time series data set to identify the location of missing values ​​and the corresponding parameter types, including temperature, pressure, flow rate, and composition parameters, and obtain missing value information;

[0057] Based on the missing value information, the time interval and parameter correlation coefficient between each missing value and the 10 non-missing values ​​before and after it are calculated to obtain the position weight matrix;

[0058] According to the position weight matrix, the temperature parameter is assigned a weight of 1.5, the pressure parameter is assigned a weight of 1.3, the flow parameter is assigned a weight of 1.2, and the composition parameter is assigned a weight of 1.1, and the parameter attention distribution vector is obtained;

[0059] Based on the parameter attention distribution vector, the weighted average of the 10 non-missing values ​​before and after each missing value is taken, and the weight is the product of the attention score and the inverse of the time interval to obtain the initial estimate. A random perturbation following the normal distribution N(0,0.1) is added to the initial estimate to generate 50 candidate interpolation values ​​and obtain a set of candidate values.

[0060] The candidate value set is input into the pre-trained masked attention network, which contains 3 fully connected layers and a hidden layer size of 64. The ReLU activation function is used to calculate the attention score of each candidate value to obtain the candidate value weight vector;

[0061] Based on the candidate value weight vector, the candidate value set is weighted averaged, and the weight is the attention score after the softmax function processing to obtain the final interpolation value. The final interpolation value is filled into the missing position of the original time series data set, and the filled data is linearly interpolated and smoothed to obtain the target data set.

[0062] Specifically, the time series data set is scanned to identify the location of missing values ​​and the corresponding parameter types, including temperature, pressure, flow rate, and composition parameters. Through this process, a missing value information set is obtained, which records the timestamp of each missing value and its corresponding parameter type. Based on the missing value information, the time interval and parameter correlation coefficient between each missing value and the 10 non-missing values ​​before and after are calculated to obtain the position weight matrix. For each missing value, the time interval is expressed as ,in is the timestamp of the missing value, is the corresponding non-missing value timestamp. The parameter correlation coefficient is calculated using the Pearson correlation coefficient formula, expressed as:

[0063] ;

[0064] in, is the covariance of X and Y, and and are the standard deviations of X and Y, respectively. The calculation results help to better understand the relationship between missing values ​​and their adjacent non-missing values, and provide a basis for subsequent interpolation. Different weights are assigned to different parameters according to the position weight matrix. The weight of the temperature parameter is 1.5, the pressure parameter is 1.3, the flow parameter is 1.2, and the composition parameter is 1.1. The parameter attention distribution vector formed is used for subsequent weighted averaging. The setting of weights is based on the evaluation of the importance of each parameter in the production process. For example, temperature is more critical than flow in the production of lead-free solder bars, and its weight is higher. These weights are used to perform a weighted average of the 10 non-missing values ​​before and after each missing value. The calculated weight will be the product of the attention score and the inverse of the time interval. Non-missing values ​​closer to the timestamp of the missing value will receive higher weights, so that they can better reflect the true situation of the missing value during interpolation. The final initial estimate is expressed as:

[0065] ;

[0066] in, is the initial estimate of the missing values, is the corresponding weight, and is the non-missing value at time t+i. To increase the diversity of interpolation, add a normal distribution to the initial estimate 50 candidate interpolation values ​​are generated by random perturbation. The candidate value set is input into a pre-trained masked attention network. The network contains three fully connected layers, each with a hidden layer size of 64 and a ReLU activation function. Through this network, the attention score of each candidate value is calculated to obtain the candidate value weight vector. The relative importance of the candidate values ​​is learned through the network, making the final interpolation more accurate. The candidate value set is weighted averaged based on the candidate value weight vector, and the weight is the attention score after processing by the softmax function. This step ensures that the candidate values ​​with higher weights have a greater impact on the result during the final interpolation, and the final interpolation value is obtained. The final interpolation value is filled into the missing position of the original time series data set, and the filled data is linearly interpolated and smoothed to remove the mutation caused by the interpolation to obtain the target data set.

[0067] In a specific embodiment, the step of executing the candidate value set is input into a pre-trained masked attention network, the masked attention network includes 3 fully connected layers, the hidden layer size is 64, and the ReLU activation function is used to calculate the attention score of each candidate value. The process of obtaining the candidate value weight vector can specifically include the following steps:

[0068] Normalize the candidate value set, map each candidate value to the range of [-1,1], and obtain a normalized candidate value set;

[0069] The normalized candidate value set is concatenated with the corresponding timestamp information to form an input vector, which is input into the first fully connected layer of the masked attention network to obtain the first layer output, and the ReLU activation function is applied to the first layer output for nonlinear transformation to obtain the first layer activation value;

[0070] The first layer activation value is input into the second fully connected layer of the masked attention network to obtain the second layer output, and the ReLU activation function is applied again to obtain the second layer activation value;

[0071] The second layer activation value is input into the third fully connected layer of the masked attention network to obtain the original attention score, and the Softmax function is applied to the original attention score to normalize the score to the range of [0,1] to obtain the attention weight;

[0072] Based on the attention weight, the masked attention score corresponding to each candidate value is calculated to obtain the candidate value attention score vector, and the candidate value attention score vector is normalized to obtain the candidate value weight vector.

[0073] Specifically, the candidate value set is normalized, each candidate value is mapped to the range of [-1,1], the scale differences between different candidate values ​​are eliminated, the training efficiency and stability of the model are improved, and a normalized candidate value set is obtained. The normalized candidate value set is concatenated with the corresponding timestamp information to form an input vector. The input vector obtained by combining each normalized candidate value with its corresponding timestamp is used for subsequent network calculations. After the input vector is sent to the first fully connected layer of the masked attention network, the network is linearly transformed and the first layer output is generated. The linear transformation is expressed by the formula:

[0074] ;

[0075] in, is the weight matrix, is the input vector, is the bias vector, and is the output of the first layer. The ReLU activation function is applied to the output of the first layer for nonlinear transformation. The definition of the ReLU function is:

[0076] ;

[0077] Through this operation, the activation values ​​of the first layer are restricted to the non-negative interval, which helps to improve the expressiveness of the model. For example, if the output of the first layer is -0.2, 0.3, 0.5, the activation values ​​will become 0, 0.3, 0.5 after ReLU processing. The first layer activation values ​​are input to the second fully connected layer of the masked attention network, and a linear transformation is performed to obtain the second layer output. The ReLU activation function is applied to generate the activation values ​​of the second layer. The second layer activation values ​​are input to the third fully connected layer of the network, which generates the original attention score. The attention score is calculated using a linear transformation form, namely:

[0078] ;

[0079] in, is the weight matrix of the third layer, is the activation value of the second layer, and is the original attention score. After generating the attention score, the Softmax function is applied to normalize it to the range of [0,1]. The Softmax function is defined as:

[0080] ;

[0081] in, Indicates attention scores, and the denominator is the exponential sum of all scores. This process ensures that the sum of all attention scores is 1, so that it can be regarded as a probability distribution. Through this process, attention weights are obtained to measure the importance of each candidate value in the final interpolation process. Based on the calculated attention weights, the masked attention score corresponding to each candidate value is calculated to obtain a candidate value attention score vector, which reflects the relative importance of different candidate values ​​in the interpolation process. For each candidate value, its attention score is multiplied by the attention weight to obtain a weight vector of the candidate value.

[0082] In a specific embodiment, the process of executing step 103 may specifically include the following steps:

[0083] The target data set is split into time windows, with the window size set to 100 time steps and the step length set to 10, to obtain the input sequence set;

[0084] A fuzzy neural network with a five-layer structure is constructed, including input layer, fuzzification layer, rule layer, defuzzification layer and output layer. The number of nodes in the input layer is the same as the number of monitoring parameters, and the output layer has a single node to represent the degree of abnormality.

[0085] Perform fuzzification on the input sequence set, use Gaussian membership function to map each input value to the interval [0,1] to obtain fuzzy input;

[0086] In the rule layer, the number of initial fuzzy rules is set to 10. Each rule is represented by an IF-THEN statement. The IF part is the fuzzy set of the input variable, and the THEN part is the fuzzy set of the output variable.

[0087] The recursive parameter α is introduced to balance the importance of the current moment and historical information. The fuzzy input is concatenated with the α-fold result of the previous moment output to obtain the extended input, where 0<α<1.

[0088] Pass the extended input into the rule layer, calculate the activation strength of each rule, and use the algebraic product as the T-norm operator to get the rule activation degree;

[0089] In the defuzzification layer, the center average method is used to perform a weighted average of the rule activation and the center value of the THEN part of each rule to obtain a clear output;

[0090] Define network structure adjustment indicators, including coverage, discrimination and importance. The coverage threshold is initially set to 0.8, the discrimination threshold is 0.2, and the importance threshold is 0.1.

[0091] Based on the structural adjustment index, the number of network rules is dynamically adjusted: when the coverage is lower than the threshold, new fuzzy rules are added; when the discrimination or importance is lower than the threshold, redundant rules are deleted; after processing every 1000 samples, the coverage threshold is reduced by 0.01, and the discrimination and importance thresholds are increased by 0.01;

[0092] The improved stochastic gradient descent method is used to optimize the network parameters. The learning rate is initially set to 0.01, and the learning rate decays to 0.9 times the original after every 1000 iterations. At the same time, the improved recursive least squares method is used to process the network output weights. The iterative training is carried out until the loss function value on the validation set changes less than 0.001 for 10 consecutive iterations or reaches the maximum number of iterations of 10000 times, and the anomaly recognition model is obtained.

[0093] Specifically, the target data set is segmented into time windows and the time series data is converted into a form suitable for neural network input. The window size is set to 100 time steps, the step length is 10, and the sliding window moves forward 10 time steps each time to generate multiple input sequences. For example, if the original data sequence is , the generated input sequence set contains , Multiple windows such as . A fuzzy neural network with a five-layer structure is constructed. The network includes an input layer, a fuzzification layer, a rule layer, a defuzzification layer and an output layer. The number of nodes in the input layer is the same as the number of monitoring parameters, while the output layer is a single node, which represents the degree of abnormality. In the fuzzification layer, the Gaussian membership function is used to fuzzify the input sequence, and each input value is mapped to the interval [0, 1] to obtain the fuzzy input, thereby improving the performance of the network in an uncertain environment. In the rule layer, the number of fuzzy rules is initially set to 10, and each rule is represented by an IF-THEN statement. The IF part consists of a fuzzy set of input variables, while the THEN part is a fuzzy set of output variables. The recursive parameter α is introduced to balance the importance of the current moment and historical information. The fuzzified input is spliced ​​with the α-fold result of the output of the previous moment to form an extended input. The range of parameter α is 0 to 1. The closer the value is to 1, the greater the influence of historical information. The extended input is passed to the rule layer to calculate the activation strength of each rule. In this process, the algebraic product is used as the T-norm operator to obtain the rule activation degree. The calculation method is expressed by the following formula:

[0094] ;

[0095] in, represents the activation degree of the i-th rule, is the fuzzy set of input variables, is the number of input variables. This calculation ensures that the activation of the fuzzy rule can reasonably reflect the characteristics of the current input. In the defuzzification layer, the center average method is used to weight the rule activation and the center value of each rule THEN part to obtain a clear output. The fuzzy output is converted into a specific value that can be used for further analysis. For example, if the activation of a rule is 0.7 and the corresponding center value is 0.8, then the contribution of this rule to the final output is . In order to ensure the effectiveness of the fuzzy neural network, the network structure adjustment indicators are defined, including coverage, discrimination and importance. Among them, the coverage threshold is initially set to 0.8, the discrimination threshold is 0.2, and the importance threshold is 0.1. Based on the structure adjustment indicators, the number of rules of the network is dynamically adjusted. When the coverage is lower than the threshold, new fuzzy rules are added, and when the discrimination or importance is lower than the threshold, redundant rules are deleted. After processing every 1000 samples, the coverage threshold is reduced by 0.01, and the discrimination and importance thresholds are increased by 0.01. This strategy helps the network better adapt to the changing data characteristics. The improved stochastic gradient descent method is used to optimize the network parameters. The initial learning rate is set to 0.01, and the learning rate decays to 0.9 times the original after every 1000 iterations. At the same time, the improved recursive least squares method is applied to process the network output weights. The iterative training process will continue until the loss function value on the validation set changes less than 0.001 for 10 consecutive iterations or reaches the maximum number of iterations of 10,000 times, and the anomaly recognition model is obtained.

[0096] In a specific embodiment, the process of executing step 104 may specifically include the following steps:

[0097] Input the target data set into the anomaly recognition model to obtain the anomaly degree score at each time point, and perform threshold processing on the anomaly degree score. The threshold is set to 0.7, and the data points with anomaly degree scores below the threshold are marked as normal data to obtain normal production data;

[0098] The normal production data is segmented into sliding windows, the window size is set to 1000 time steps, the step length is 100, and a data window set is obtained. Each window data in the data window set is subjected to a fast Fourier transform to obtain a frequency domain representation;

[0099] Based on the frequency domain representation, the first K frequency components with energy accounting for 90% are selected to obtain the frequency domain features after dimensionality reduction;

[0100] An implicit subspace recognition recurrent neural network with an encoder-decoder structure is constructed. Both the encoder and decoder contain three fully connected layers, and the number of hidden layer nodes is 128, 64, and 32 respectively.

[0101] The frequency domain features after dimensionality reduction are input into the encoder to obtain the latent space representation, and based on the latent space representation, a nonlinear autoregressive moving average model is constructed. The model order is determined by the Akaike information criterion to obtain the system dynamic model. The output of the system dynamic model is reconstructed through the decoder to obtain time domain reconstructed data;

[0102] The reconstruction error is calculated based on the time-domain reconstructed data and the original normal production data, and the reconstruction error is combined with the L1 regularization term as the loss function. The Adam optimizer is used for model training. The learning rate is initially set to 0.001 and decays by 10% every 50 epochs. The training is carried out until the loss on the validation set does not decrease for 10 consecutive epochs or reaches the maximum number of iterations of 5000, and the production process parameter model is obtained.

[0103] Specifically, the target data set is input into the anomaly recognition model to obtain the anomaly score at each time point. In order to effectively analyze the score, the threshold is set to 0.7, which means that only when the anomaly score is lower than this value can the data point be marked as normal data, and normal production data can be extracted from the original data. The normal production data is segmented into sliding windows, and the window size is set to 1000 time steps and the step length is 100. On the data set, 100 time steps are moved each time to form multiple data windows. For example, if the original normal production data sequence is , multiple windows are obtained by sliding window segmentation such as , Etc. For each data window, apply fast Fourier transform to get its frequency domain representation. Converting the time domain signal into the frequency domain signal helps to analyze the frequency components of the signal. Based on the frequency domain representation, select the first K frequency components with an energy share of 90%. By calculating the energy of each frequency component in the frequency domain and sorting them, the cumulative energy share is obtained, thereby determining the value of K. For example, if the energy calculation formula of the frequency domain signal is:

[0104] ;

[0105] in, represents the energy of the kth frequency component, and is the corresponding frequency domain data. By selecting high-energy frequency components, simplified frequency domain features can be obtained after dimensionality reduction, retaining the main information and removing noise. Construct an implicit subspace recognition recurrent neural network with an encoder-decoder structure. The encoder and decoder of the network both contain three fully connected layers, and the number of hidden layer nodes is 128, 64 and 32, respectively. This design enables the network to effectively extract potential features. In this process, the frequency domain features after dimensionality reduction are input into the encoder to obtain the latent space representation. The representation of the latent space enables the network to capture the implicit structure of the data. Based on the latent space representation, a nonlinear autoregressive moving average model is constructed, and its model order is determined by the Akaike Information Criterion (AIC) in order to select the appropriate model complexity. The calculation formula of AIC is:

[0106] ;

[0107] in, is the number of parameters of the model, is the likelihood function of the model on the data. By calculating the AIC values ​​of models of different orders, the model with the minimum AIC value is selected to obtain the optimal order of the system dynamic model. The output of the system dynamic model is reconstructed through the decoder to obtain time domain reconstructed data, so that the system can recover the time domain characteristics of the original data from the representation of the latent space. Based on the time domain reconstructed data and the original normal production data, the reconstruction error is calculated, and the reconstruction error is combined with the L1 regularization term to form a loss function. The calculation of the reconstruction error is expressed as:

[0108] ;

[0109] in, is the reconstruction error, is the original normal data, is the reconstructed data, is the number of data points. The L1 regularization term is:

[0110] ;

[0111] in, is the regularization loss, are model parameters, is the regularization coefficient. By adding the reconstruction error to the L1 regularization term, the model can be effectively prevented from overfitting and its generalization ability on new data can be improved. The model is trained using the Adam optimizer, with the initial learning rate set to 0.001 and decayed by 10% every 50 epochs. The training process will continue until the loss on the validation set does not decrease for 10 consecutive epochs or the maximum number of iterations is 5000. Through optimization and training, the final production process parameter model will have strong prediction and anomaly recognition capabilities, thereby providing effective support for the optimization of the production process.

[0112] In a specific embodiment, the execution step inputs the frequency domain features after dimensionality reduction into the encoder to obtain a latent space representation, and based on the latent space representation, constructs a nonlinear autoregressive moving average model, the model order is determined by the Akaike information criterion, and the system dynamic model is obtained. The output of the system dynamic model is reconstructed through a decoder to obtain the time domain reconstructed data. The process can specifically include the following steps:

[0113] The frequency domain features after dimension reduction are input into the first fully connected layer of the encoder, and a first weight matrix and a first bias vector are used for linear transformation to obtain the first layer output, and a ReLU activation function is applied to the first layer output for nonlinear transformation to obtain the first layer activation value;

[0114] Input the first layer activation value into the second fully connected layer of the encoder, perform linear transformation using the second weight matrix and the second bias vector, and apply the ReLU activation function to obtain the second layer activation value;

[0115] Input the second layer activation value into the third fully connected layer of the encoder, and use the third weight matrix and the third bias vector to perform linear transformation to obtain the latent space representation;

[0116] Based on the latent space representation, a nonlinear autoregressive moving average model is constructed. The model order p is gradually increased from 1 to 10. The Akaike Information Criterion AIC value is calculated for each order, and the order with the smallest AIC value is selected as the optimal order p*. The system dynamic model expressed as NARMA(p*,q) is obtained, where q is the order of the moving average term, which is set to p* / 2;

[0117] Use the system dynamics model to perform time series prediction on the latent space representation to obtain a prediction result, and input the prediction result into the first fully connected layer of the decoder, use the fourth weight matrix and the fourth bias vector to perform linear transformation, and apply the ReLU activation function to obtain the first layer activation value of the decoder;

[0118] Input the activation value of the first layer of the decoder into the second fully connected layer of the decoder, perform linear transformation using the fifth weight matrix and the fifth bias vector, and apply the ReLU activation function to obtain the activation value of the second layer of the decoder;

[0119] The activation value of the second layer of the decoder is input into the third fully connected layer of the decoder, linearly transformed using the sixth weight matrix and the sixth bias vector, and an inverse fast Fourier transform is applied to obtain time domain reconstructed data.

[0120] Specifically, the frequency domain features after dimension reduction are input into the first fully connected layer of the encoder, and the first weight matrix is ​​used and the first bias vector Transform and calculate the output of the first layer. It is expressed by the following formula:

[0121] ;

[0122] in, represents the output of the first layer, is the frequency domain feature vector after dimensionality reduction. Apply the ReLU activation function to the output to introduce nonlinear characteristics and obtain the activation value of the first layer. The ReLU function is defined as:

[0123] ;

[0124] The first layer activation value It is expressed as:

[0125] ;

[0126] The first layer activation value Input to the second fully connected layer of the encoder, perform linear transformation, and use the second weight matrix and the second bias vector The output of this layer is The calculation method is:

[0127] ;

[0128] Likewise, Apply the ReLU activation function to get the activation value of the second layer

[0129] ;

[0130] The second layer activation value Input to the third fully connected layer of the encoder, using the third weight matrix and the third bias vector Perform linear transformation to obtain the latent space representation

[0131] ;

[0132] Latent Space Representation Captures important features of the input data. Based on latent space representation , construct a nonlinear autoregressive moving average model (NARMA), the model order Increase gradually from 1 to 10. Calculate the Akaike Information Criterion (AIC) value for each order to determine the optimal model complexity. The calculation formula for AIC is:

[0133] ;

[0134] in, is the number of model parameters, is the likelihood function of the model on the given data. By gradually calculating the AIC values ​​of different orders, the order that minimizes the AIC value is selected as the optimal order. According to this order, the expression of the NARMA model is ,in is the order of the moving average term, set it to . Use the constructed system dynamics model to perform time series prediction on the latent space representation and obtain the prediction results. Input the prediction results into the first fully connected layer of the decoder. Use the fourth weight matrix and the fourth bias vector Perform a linear transformation and calculate the first layer activation value of the decoder:

[0135] ;

[0136] Apply the ReLU activation function to the output to get the first layer activation value of the decoder

[0137] ;

[0138] The first layer activation value of the decoder Input to the second fully connected layer of the decoder, using the fifth weight matrix and the fifth bias vector Perform a linear transformation to obtain the second layer activation value

[0139] ;

[0140] Similarly, the ReLU activation function is applied to it to obtain the second layer activation value of the decoder:

[0141] ;

[0142] The second layer activation value of the decoder Input to the third fully connected layer of the decoder, using the sixth weight matrix and the sixth bias vector Perform linear transformation and finally obtain time domain reconstructed data

[0143] ;

[0144] In order to recover the time domain signal, the final linear transformation output is applied with an inverse fast Fourier transform to convert the frequency domain signal back to the time domain. The specific operation is as follows:

[0145] ;

[0146] in, is the time domain signal, is the signal obtained in the frequency domain, is the total number of sample points, is an imaginary unit. Through this transformation, the data in the frequency domain is converted back to the original time domain signal to obtain the reconstructed data.

[0147] In a specific embodiment, the process of executing step 105 may specifically include the following steps:

[0148] The output data of the production process parameter model are grouped according to different production lines or different batches to obtain multiple subtask data sets, and a subpopulation is constructed for each subtask data set. The initial population size is set to 100, and each individual represents a set of production process parameters to obtain an initial multi-task population.

[0149] Based on the initial multi-task population, a feature extractor of the ResNet structure enhanced by the attention mechanism is constructed, which includes 9 residual blocks. Each residual block contains two convolutional layers and a short-circuit connection, and a self-attention layer is added between the residual blocks to obtain a feature extraction network.

[0150] The output of the feature extraction network is input into the variational autoencoder, where both the encoder and decoder use a three-layer fully connected network with hidden layer sizes of 64 and 32 respectively, and the latent space dimension is set to 16 to obtain the task feature representation;

[0151] Based on the task feature representation, the similarity matrix between different tasks is calculated, and the cosine similarity is used as the metric to obtain the task relevance matrix. According to the task relevance matrix, the features of each subtask are weighted and fused, and the weight is the softmax value of the similarity to obtain the fused feature vector.

[0152] The fused feature vector is input into the production process parameter model to obtain the objective function value of each subtask, including product quality index, production efficiency and energy consumption. These objectives are normalized and weighted summed to obtain a comprehensive evaluation index.

[0153] Based on the comprehensive evaluation index, tournament selection, simulated binary crossover and polynomial mutation operations are performed on the multi-task population to obtain the target production process parameters.

[0154] Specifically, the output data of the production process parameter model is effectively grouped, and the output data is grouped according to different production lines or different production batches to form multiple subtask data sets. For example, if there are multiple production lines, the process parameters of each production line form an independent subtask data set. On this basis, a subpopulation is constructed for each subtask data set, and the initial population size is set to 100. Each individual represents a specific set of production process parameters, and an initial multi-task population is obtained. Construct a feature extraction network. A residual network structure enhanced by an attention mechanism is adopted, and the network contains 9 residual blocks. Each residual block contains two convolutional layers and a short-circuit connection. This structure can effectively alleviate the gradient vanishing problem in deep network training. The output of the residual block is expressed as:

[0155] ;

[0156] in, is the output of the residual block, It is the feature obtained by two convolutional layers. is the input feature. In order to improve the performance of the model, a self-attention layer is added between the residual blocks to make the model pay more attention to the important parts of the input features and improve the efficiency and accuracy of feature extraction. The output of the feature extraction network is input to the variational autoencoder. The encoder and decoder both use a three-layer fully connected network with hidden layer sizes of 64 and 32 respectively, and the latent space dimension is set to 16. Through this structure, the encoder compresses the input features into the latent space representation , the specific calculation is as follows:

[0157] ;

[0158] in, is the latent space distribution, and are the parameters learned by the network, is the unit matrix, and the task feature representation is obtained for subsequent similarity calculation. Based on the task feature representation, the similarity matrix between different tasks is calculated. In order to quantify the similarity between tasks, cosine similarity is used as the metric, and the calculation formula is:

[0159] ;

[0160] in, and are two different task feature vectors. In this way, a task correlation matrix is ​​obtained to reflect the similarity between each subtask. Based on this matrix, the features of each subtask are weighted and fused. The weight is the softmax value of the similarity. The calculation formula is as follows:

[0161] ;

[0162] in, It is The weight of the task, is the total number of tasks, and is the reference feature vector. Through weighted fusion, a fused feature vector is obtained, which represents the comprehensive characteristics of different subtasks. The fused feature vector is input into the production process parameter model to calculate the objective function value of each subtask. The objective function value includes product quality indicators, production efficiency and energy consumption, etc. The target is normalized and weighted summed to obtain a comprehensive evaluation index. Based on the comprehensive evaluation index, tournament selection, simulated binary crossover and polynomial mutation operations are performed on the multi-task population to obtain the target production process parameters. Tournament selection selects the best individual by comparing the fitness of the individual, simulated binary crossover generates new offspring individuals between two parent individuals, and polynomial mutation is to increase the diversity of the population. Through these operations, a set of optimized production process parameters is generated. Improve production efficiency and product quality.

[0163] In a specific embodiment, the execution step is based on the task feature representation, calculates the similarity matrix between different tasks, uses cosine similarity as a metric, obtains the task relevance matrix, and performs weighted fusion on the features of each subtask according to the task relevance matrix, with the weight being the softmax value of the similarity, and the process of obtaining the fused feature vector can specifically include the following steps:

[0164] Perform L2 normalization on the task feature representation to obtain a feature representation in the form of a unit vector, and based on the feature representation in the form of a unit vector, calculate the dot product between any two task feature vectors to obtain the original similarity matrix;

[0165] Symmetrically process the original similarity matrix and take (S+S T ) / 2 as the symmetric similarity matrix, where S is the original similarity matrix, S T Transpose its matrix;

[0166] Apply the softmax function to each row of the symmetric similarity matrix to convert the similarity value into a probability distribution to obtain the task relevance matrix. Based on the task relevance matrix, perform weighted summation on the feature vector of each subtask, with the weight being the element value of the corresponding row in the relevance matrix to obtain the initial fusion feature.

[0167] The initial fusion features are subjected to layer normalization, the mean and variance are calculated, and standardized to obtain standardized fusion features. The standardized fusion features are residually connected with the original task features, that is, the two are added together to obtain the fused feature representation. The ReLU activation function is applied to the fused feature representation, and a nonlinear transformation is performed to obtain a fused feature vector.

[0168] Specifically, the task feature representation is L2 normalized to obtain a feature representation in the form of a unit vector. The calculation method of L2 normalization is:

[0169] ;

[0170] in, represents the normalized eigenvector, is the original eigenvector, and is the L2 norm, which indicates the length of the feature vector. After processing, the length of each task feature vector is standardized to 1, making the subsequent similarity calculation more efficient. Based on the feature representation in the form of unit vectors, the dot product between any two task feature vectors is calculated to obtain the original similarity matrix The calculation formula is as follows:

[0171] ;

[0172] in, It is and The similarity between the task features, is the dimension of the feature. In this way, we construct an original similarity matrix , where each element reflects the similarity between the corresponding task features. In order to eliminate the asymmetry in the matrix, the original similarity matrix is ​​symmetricized and the symmetric similarity matrix is ​​calculated by the following formula

[0173] ;

[0174] in, is the transpose of the original similarity matrix. This step ensures the symmetry of the similarity matrix so that the similarity between each pair of tasks is consistent, thereby enhancing the stability and reliability of the model. A softmax function is applied to each row of the symmetric similarity matrix to convert the similarity value into a probability distribution. The softmax function is defined as:

[0175] ;

[0176] in, It's a task For the task In this way, we get a task relevance matrix , which reflects the relative importance of tasks. Based on the task relevance matrix, the feature vector of each subtask is weighted and summed, and the weight is the element value of the corresponding row in the relevance matrix. The calculation formula of this process is:

[0177] ;

[0178] in, It's a task The initial fusion features, It is The feature vector of each task is obtained by weighted summation, and the feature information of each task is integrated to form a new feature representation. The layer normalization operation is applied to the initial fusion feature, the mean and variance are calculated, and the standardization is performed. The formula is as follows:

[0179] ;

[0180] in, and Respectively represent the mean and standard deviation of the initial fusion features. Through the formula, the features are mapped to a unified scale, which is helpful for subsequent model training and optimization. The standardized fusion features are residually connected with the original task features. The calculation formula of the residual connection is:

[0181] ;

[0182] in, is the fused feature representation, is the original task feature. The residual connection can effectively combine the initial feature information with the processed feature information, thereby improving the feature expression ability and the robustness of the model. The ReLU activation function is applied to the fused feature representation for nonlinear transformation. The calculation formula is:

[0183] ;

[0184] in, It is the output value after being processed by the ReLU activation function. By introducing nonlinear activation functions, the expressive power of the model can be enhanced, enabling it to capture more complex feature relationships and obtain a fused feature vector.

[0185] The above describes the intelligent temperature-controlled lead-free solder bar production method in the embodiment of the present invention. The following describes the intelligent temperature-controlled lead-free solder bar production device in the embodiment of the present invention. Figure 2 , an embodiment of the intelligent temperature-controlled lead-free solder bar production device in the embodiment of the present invention includes:

[0186] The acquisition module 201 is used to collect and pre-process the data of multiple sensors on the production line to obtain a time series data set including temperature, pressure, flow rate and composition;

[0187] An interpolation module 202 is used to perform attention-guided position-sensitive multiple interpolation on missing values ​​in a time series data set to obtain a target data set;

[0188] A construction module 203 is used to construct and train a fuzzy neural network with a self-organizing loop structure based on a target data set to obtain an anomaly recognition model;

[0189] An analysis module 204 is used to analyze the normal production data screened by the abnormal recognition model and obtain a production process parameter model based on an implicit subspace recognition recurrent neural network decomposed and reconstructed by an orthogonal basis;

[0190] The optimization module 205 is used to optimize the parameters based on the production process parameter model using a multi-task feature fusion framework to obtain target production process parameters.

[0191] Through the collaborative cooperation of the above-mentioned components, the position-sensitive multiple interpolation method guided by attention effectively solves the problem of missing data in the production process and improves data quality and reliability. The fuzzy neural network with a self-organizing cyclic structure is used to construct an abnormality recognition model, which can accurately detect abnormal situations in the production process and reduce production risks. The implicit subspace recognition regression neural network based on orthogonal basis decomposition and reconstruction realizes the accurate modeling of production process parameters and provides a reliable basis for parameter optimization. The multi-task feature fusion framework is introduced to consider the differences between different production lines or batches, and more accurate and personalized process parameter optimization is achieved. The ResNet structure and variational autoencoder enhanced by the attention mechanism are used to improve the effect of feature extraction and representation, and enhance the generalization ability of the model. Through the multi-objective optimization method, multiple objectives such as product quality, production efficiency and energy consumption are considered at the same time, and the comprehensive optimization of the production process is achieved. From data acquisition, preprocessing, abnormality detection to parameter optimization, the present invention comprehensively improves the intelligent level of lead-free tin bar production.

[0192] above Figure 2 The intelligent temperature-controlled lead-free tin bar production device in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The intelligent temperature-controlled lead-free tin bar production equipment in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0193] Figure 3 3 is a schematic diagram of the structure of an intelligent temperature-controlled lead-free tin bar production device provided by an embodiment of the present invention. The intelligent temperature-controlled lead-free tin bar production device 300 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage device terminals) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 can be short-term storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the intelligent temperature-controlled lead-free tin bar production device 300. Furthermore, the processor 310 can be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the intelligent temperature-controlled lead-free tin bar production device 300 to implement the steps of the above-mentioned intelligent temperature-controlled lead-free tin bar production method.

[0194] The intelligent temperature-controlled lead-free solder bar production equipment 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. It can be understood by those skilled in the art that Figure 3 The structure of the intelligent temperature-controlled lead-free solder bar production equipment shown does not constitute a limitation on the intelligent temperature-controlled lead-free solder bar production equipment provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0195] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the intelligent temperature-controlled lead-free tin bar production method.

[0196] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0197] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0198] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for producing an intelligent temperature-controlled lead-free solder bar, characterized in that: The method comprises: Collect and preprocess data from multiple sensors on the production line to obtain a time series data set containing temperature, pressure, flow and composition; The missing values ​​in the time series data set are subjected to attention-guided position-sensitive multiple interpolation to obtain a target data set; specifically, the method comprises: scanning the time series data set to identify the position of the missing values ​​and the corresponding parameter types, including temperature, pressure, flow rate and composition parameters, to obtain missing value information; based on the missing value information, for each missing value, calculating the time interval and parameter correlation coefficient between the 10 non-missing values ​​before and after it and the position weight matrix; according to the position weight matrix, assigning a weight of 1.5 to the temperature parameter, a weight of 1.3 to the pressure parameter, a weight of 1.2 to the flow parameter, and a weight of 1.1 to the composition parameter to obtain a parameter attention distribution vector; based on the parameter attention distribution vector, performing a weighted average of the 10 non-missing values ​​before and after each missing value, with the weight being the attention value. The product of the force score and the inverse of the time interval is obtained to obtain an initial estimate, and a random perturbation obeying the normal distribution N(0,0.1) is added to the initial estimate to generate 50 candidate interpolation values ​​to obtain a candidate value set; the candidate value set is input into a pre-trained masked attention network, which includes 3 fully connected layers and a hidden layer size of 64. The ReLU activation function is used to calculate the attention score of each candidate value to obtain a candidate value weight vector; based on the candidate value weight vector, the candidate value set is weighted averaged, and the weight is the attention score after the softmax function is processed to obtain a final interpolation value, and the final interpolation value is filled into the missing position of the original time series data set, and the filled data is linearly interpolated and smoothed to obtain the target data set; Based on the target data set, a fuzzy neural network with a self-organizing loop structure is constructed and trained to obtain an anomaly recognition model; specifically, the target data set is segmented into time windows, the window size is set to 100 time steps, the step length is 10, and an input sequence set is obtained; a fuzzy neural network with a 5-layer structure is constructed, including an input layer, a fuzzification layer, a rule layer, a defuzzification layer and an output layer, wherein the number of nodes in the input layer is the same as the number of monitoring parameters, and the output layer is a single node representing the degree of anomaly; the input sequence set is fuzzified, and each input value is mapped to the interval [0,1] using a Gaussian membership function to obtain a fuzzified input; the initial number of fuzzy rules in the rule layer is set to 10, and each rule is represented by an IF-THEN statement, the IF part is a fuzzy set of input variables, and the THEN part is a fuzzy set of output variables; a recursive parameter α is introduced to balance the importance of the current moment and historical information, and the fuzzified input is spliced ​​with the α-fold result of the output of the previous moment to obtain an extended input, wherein 0<α<1; the extended input is passed into the rule layer, and the activation strength of each rule is calculated, using The algebraic product is used as the T-norm operator to obtain the rule activation. The center average method is used in the defuzzification layer to perform weighted average of the rule activation and the center value of the THEN part of each rule to obtain a clear output. The network structure adjustment indicators are defined, including coverage, discrimination and importance. The coverage threshold is initially set to 0.8, the discrimination threshold is 0.2, and the importance threshold is 0.

1. Based on the structure adjustment indicators, the number of network rules is dynamically adjusted: when the coverage is lower than the threshold, a new fuzzy rule is added; when the discrimination or importance is lower than the threshold, Redundant rules were deleted; after processing every 1000 samples, the coverage threshold was reduced by 0.01, and the discrimination and importance thresholds were increased by 0.01; the improved stochastic gradient descent method was used to optimize the network parameters, with the learning rate initially set to 0.01, and the learning rate decayed to 0.9 times the original value after every 1000 iterations. At the same time, the improved recursive least squares method was used to process the network output weights, and the iterative training was carried out until the loss function value on the validation set changed less than 0.001 for 10 consecutive iterations or reached the maximum number of iterations of 10,000, and the anomaly recognition model was obtained; According to the normal production data screened by the abnormality recognition model, the implicit subspace recognition regression neural network based on orthogonal basis decomposition and reconstruction is analyzed to obtain a production process parameter model; Based on the production process parameter model, a multi-task feature fusion framework is used to perform parameter optimization to obtain target production process parameters.

2. The method for producing an intelligent temperature-controlled lead-free solder bar according to claim 1, characterized in that: The data of multiple sensors on the production line are collected and preprocessed to obtain a time series data set containing temperature, pressure, flow and composition, including: Collecting data from temperature sensors, pressure sensors, flow sensors, and component analyzers on the production line to obtain an original data set, and sampling the original data set at a frequency of once per second to obtain a sampled data set; Eliminate data points in the sampled data set that are beyond a preset physical range to obtain a preliminary screening data set, apply a sliding window of 60 seconds to the preliminary screening data set, calculate the average value of the data in the window, and obtain a smoothed data set; Normalizing the smoothed data set to unify the numerical range to between 0 and 1 to obtain a normalized data set, and calculating the correlation coefficient matrix between the parameters based on the normalized data set to obtain correlation data; The normalized data set and the correlation data are combined according to timestamps to obtain a time series data set.

3. The method for producing an intelligent temperature-controlled lead-free solder bar according to claim 1, characterized in that: The candidate value set is input into a pre-trained masked attention network, which includes 3 fully connected layers, a hidden layer size of 64, and a ReLU activation function to calculate the attention score of each candidate value to obtain a candidate value weight vector, including: Normalizing the candidate value set, mapping each candidate value to a range of [-1, 1], and obtaining a normalized candidate value set; The normalized candidate value set is concatenated with the corresponding timestamp information to form an input vector, which is input into the first fully connected layer of the masked attention network to obtain the first layer output, and the ReLU activation function is applied to the first layer output to perform nonlinear transformation to obtain the first layer activation value; Inputting the first layer activation value into the second fully connected layer of the masked attention network to obtain the second layer output, and applying the ReLU activation function again to obtain the second layer activation value; Inputting the second layer activation value into the third fully connected layer of the masked attention network to obtain an original attention score, and applying a Softmax function to the original attention score to normalize the score to the range of [0, 1] to obtain an attention weight; Based on the attention weight, the masked attention score corresponding to each candidate value is calculated to obtain a candidate value attention score vector, and the candidate value attention score vector is normalized to obtain a candidate value weight vector.

4. The method for producing an intelligent temperature-controlled lead-free solder bar according to claim 1, characterized in that: The normal production data selected according to the abnormal recognition model is analyzed based on orthogonal basis decomposition and reconstruction of implicit subspace recognition regression neural network to obtain a production process parameter model, including: The target data set is input into the anomaly recognition model to obtain an anomaly degree score at each time point, and the anomaly degree score is thresholded, the threshold is set to 0.7, and the data points with an anomaly degree score lower than the threshold are marked as normal data to obtain normal production data; Perform sliding window segmentation on the normal production data, set the window size to 1000 time steps, and the step length to 100, to obtain a data window set, and perform fast Fourier transform on each window data in the data window set to obtain a frequency domain representation; Based on the frequency domain representation, the first K frequency components whose energy accounts for 90% are selected to obtain the frequency domain features after dimensionality reduction; An implicit subspace recognition recurrent neural network with an encoder-decoder structure is constructed. Both the encoder and decoder contain three fully connected layers, and the number of hidden layer nodes is 128, 64, and 32 respectively. Input the frequency domain features after dimensionality reduction into an encoder to obtain a latent space representation, and based on the latent space representation, construct a nonlinear autoregressive moving average model, the model order is determined by the Akaike information criterion, and a system dynamic model is obtained, and the output of the system dynamic model is reconstructed through a decoder to obtain time domain reconstructed data; The reconstruction error is calculated based on the time-domain reconstructed data and the original normal production data, and the reconstruction error is combined with the L1 regularization term as the loss function. The Adam optimizer is used for model training. The learning rate is initially set to 0.001 and decays by 10% every 50 epochs. The training is carried out until the loss on the validation set no longer decreases for 10 consecutive epochs or the maximum number of iterations is reached, thereby obtaining the production process parameter model.

5. The method for producing an intelligent temperature-controlled lead-free solder bar according to claim 4, characterized in that: The frequency domain features after dimensionality reduction are input into an encoder to obtain a latent space representation, and a nonlinear autoregressive moving average model is constructed based on the latent space representation, the model order is determined by the Akaike information criterion, and a system dynamic model is obtained. The output of the system dynamic model is reconstructed through a decoder to obtain time domain reconstructed data, including: Inputting the reduced-dimensional frequency domain features into a first fully connected layer of an encoder, performing a linear transformation using a first weight matrix and a first bias vector to obtain a first layer output, and applying a ReLU activation function to the first layer output to perform a nonlinear transformation to obtain a first layer activation value; Inputting the first layer activation value into the second fully connected layer of the encoder, performing linear transformation using a second weight matrix and a second bias vector, and applying a ReLU activation function to obtain a second layer activation value; Inputting the second layer activation value into the third fully connected layer of the encoder, performing linear transformation using a third weight matrix and a third bias vector to obtain a latent space representation; Based on the latent space representation, a nonlinear autoregressive moving average model is constructed, and the model order p is gradually increased from 1 to 10. The Akaike information criterion AIC value is calculated for each order, and the order with the smallest AIC value is selected as the optimal order p*, and a system dynamic model with the expression of NARMA(p*,q) is obtained, where q is the order of the moving average term, which is set to p* / 2; Using the system dynamics model to perform time series prediction on the latent space representation to obtain a prediction result, and inputting the prediction result into a first fully connected layer of a decoder, performing a linear transformation using a fourth weight matrix and a fourth bias vector, and applying a ReLU activation function to obtain a first layer activation value of the decoder; Inputting the activation value of the first layer of the decoder into the second fully connected layer of the decoder, performing a linear transformation using a fifth weight matrix and a fifth bias vector, and applying a ReLU activation function to obtain an activation value of the second layer of the decoder; The second layer activation value of the decoder is input into the third fully connected layer of the decoder, a sixth weight matrix and a sixth bias vector are used for linear transformation, and an inverse fast Fourier transform is applied to obtain time domain reconstructed data.

6. The method for producing an intelligent temperature-controlled lead-free solder bar according to claim 1, characterized in that: Based on the production process parameter model, the multi-task feature fusion framework is used to perform parameter optimization to obtain target production process parameters, including: The output data of the production process parameter model are grouped according to different production lines or different batches to obtain multiple subtask data sets, and a subpopulation is constructed for each subtask data set, the initial population size is set to 100, each individual represents a set of production process parameters, and an initial multi-task population is obtained; Based on the initial multi-task population, a feature extractor of a ResNet structure enhanced by an attention mechanism is constructed, comprising 9 residual blocks, each residual block comprising two convolutional layers and a short-circuit connection, and a self-attention layer is added between the residual blocks to obtain a feature extraction network; The output of the feature extraction network is input into a variational autoencoder, where both the encoder and the decoder use a three-layer fully connected network, with hidden layer sizes of 64 and 32 respectively, and the latent space dimension is set to 16, to obtain the task feature representation; Based on the task feature representation, the similarity matrix between different tasks is calculated, and the cosine similarity is used as a metric to obtain a task relevance matrix. According to the task relevance matrix, the features of each subtask are weighted and fused, and the weight is the softmax value of the similarity to obtain a fused feature vector. Input the fused feature vector into the production process parameter model to obtain the objective function value of each subtask, including product quality index, production efficiency and energy consumption, and normalize these objectives and perform weighted summation to obtain a comprehensive evaluation index; Based on the comprehensive evaluation index, tournament selection, simulated binary crossover and polynomial mutation operations are performed on the multi-task population to obtain the target production process parameters.

7. The method for producing an intelligent temperature-controlled lead-free solder bar according to claim 6, characterized in that: Based on the task feature representation, the similarity matrix between different tasks is calculated, and the cosine similarity is used as a metric to obtain a task relevance matrix. According to the task relevance matrix, the features of each subtask are weighted and fused, and the weight is the softmax value of the similarity to obtain a fused feature vector, including: Performing L2 normalization processing on the task feature representation to obtain a feature representation in the form of a unit vector, and based on the feature representation in the form of a unit vector, calculating the dot product between any two task feature vectors to obtain an original similarity matrix; Symmetric processing is performed on the original similarity matrix, and (S+S T ) / 2 as the symmetric similarity matrix, where S is the original similarity matrix, S T Transpose its matrix; Applying a softmax function to each row of the symmetric similarity matrix, converting the similarity value into a probability distribution, obtaining a task relevance matrix, and performing a weighted summation on the feature vector of each subtask based on the task relevance matrix, where the weight is the element value of the corresponding row in the relevance matrix, to obtain an initial fusion feature; A layer normalization operation is applied to the initial fusion feature, the mean and variance are calculated, and standardized processing is performed to obtain a standardized fusion feature, and the standardized fusion feature is residually connected with the original task feature, that is, the two are added to obtain a fused feature representation, and a ReLU activation function is applied to the fused feature representation to perform a nonlinear transformation to obtain a fused feature vector.

8. An intelligent temperature-controlled lead-free solder bar production device, characterized in that: The device is used to perform the intelligent temperature-controlled lead-free solder bar production method according to any one of claims 1 to 7, comprising: The acquisition module is used to collect and preprocess the data of multiple sensors on the production line to obtain a time series data set containing temperature, pressure, flow and composition; An interpolation module, configured to perform attention-guided position-sensitive multiple interpolation on missing values ​​in the time series dataset to obtain a target dataset; A construction module is used to construct and train a fuzzy neural network with a self-organizing loop structure based on the target data set to obtain an anomaly recognition model; An analysis module is used to analyze the normal production data screened by the abnormality recognition model and obtain a production process parameter model based on an implicit subspace recognition regression neural network decomposed and reconstructed by an orthogonal basis; The optimization module is used to optimize the parameters based on the production process parameter model using a multi-task feature fusion framework to obtain target production process parameters.

Images