Solder paste stirring intelligent temperature control system and method based on temperature sensor

Abstract

The invention relates to the technical field of intelligent temperature control, and discloses a solder paste stirring intelligent temperature control system and method based on temperature sensors, and the method comprises the steps: carrying out the preprocessing of temperature data collected by a plurality of temperature sensors in a solder paste stirring device, and obtaining a structured temperature data set; performing multi-scale feature extraction to obtain a temperature feature set; performing regression prediction on the temperature feature set to obtain temperature change features; performing temperature anomaly detection and positioning based on the temperature change characteristics to obtain a temperature anomaly score graph and an abnormal area mask; performing feature integration on the temperature change features and the abnormal region masks to obtain comprehensive feature representation; the temperature control parameters are subjected to staged differential optimization according to the comprehensive feature representation, an intelligent temperature control strategy in the solder paste stirring process is obtained, then multi-source sensing data are comprehensively utilized, the method has the self-adaptive learning capacity, accurate intelligent temperature control can be achieved, and the automation level of the solder paste stirring technology is improved.

Description

Technical Field

[0001] The present application relates to the field of intelligent temperature control technology, and in particular to a solder paste stirring intelligent temperature control system and method based on a temperature sensor. Background Art

[0002] Traditional solder paste stirring temperature control methods mainly rely on fixed parameter settings and manual experience, which is difficult to cope with the complex and changing production environment and the differences in characteristics of different batches of solder paste. As electronic products develop towards miniaturization and high density, the requirements for welding accuracy continue to increase, and the limitations of traditional temperature control methods are becoming increasingly prominent.

[0003] Existing intelligent temperature control systems still have problems such as low efficiency and insufficient precision when processing high-dimensional, nonlinear temperature data. In particular, there is still much room for improvement in capturing the dynamic characteristics of temperature changes, identifying minor anomalies, and adaptively adjusting control strategies according to different production stages. Summary of the invention

[0004] The present application provides a solder paste stirring intelligent temperature control system and method based on a temperature sensor, which comprehensively utilizes multi-source sensor data, has adaptive learning capabilities, and can achieve precise intelligent temperature control, thereby improving the automation level of the solder paste stirring process.

[0005] In a first aspect, the present application provides a solder paste stirring intelligent temperature control method based on a temperature sensor, and the solder paste stirring intelligent temperature control method based on a temperature sensor comprises: Preprocessing the temperature data collected by multiple temperature sensors in the solder paste stirring device to obtain a structured temperature data set; Inputting the structured temperature data set into an improved pyramid visual transformer to extract multi-scale features to obtain a temperature feature set; Using a polarized self-attention module to perform regression prediction on the temperature feature set to obtain temperature change features; Based on the temperature change characteristics, an efficient phase perception module is used to detect and locate temperature anomalies, and a temperature anomaly score map and an abnormal area mask are obtained; Based on the cross-layer fusion and propagation module, the temperature change feature and the abnormal area mask are integrated to obtain a comprehensive feature representation; According to the comprehensive characteristic representation, the temperature control parameters are optimized in stages and differentiated to obtain an intelligent temperature control strategy for the solder paste stirring process.

[0006] The second aspect of the present application provides a solder paste stirring intelligent temperature control system based on a temperature sensor, and the solder paste stirring intelligent temperature control system based on a temperature sensor comprises: An acquisition module is used to pre-process the temperature data collected by multiple temperature sensors in the solder paste stirring device to obtain a structured temperature data set; An extraction module, used for inputting the structured temperature data set into an improved pyramid visual transformer for multi-scale feature extraction to obtain a temperature feature set; A prediction module, used for performing regression prediction on the temperature feature set using a polarized self-attention module to obtain a temperature change feature; A detection module, used to detect and locate temperature anomalies based on the temperature change characteristics using an efficient phase perception module to obtain a temperature anomaly score map and an abnormal area mask; An integration module, used for integrating the temperature change feature and the abnormal area mask based on the cross-layer fusion and propagation module to obtain a comprehensive feature representation; The optimization module is used to perform differentiated optimization of the temperature control parameters in stages according to the comprehensive feature representation to obtain an intelligent temperature control strategy for the solder paste stirring process.

[0007] The third aspect of the present application provides an electronic 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 electronic device executes the above-mentioned temperature sensor-based solder paste stirring intelligent temperature control method.

[0008] A fourth aspect of the present application provides a computer-readable storage medium, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer executes the above-mentioned intelligent temperature control method for solder paste stirring based on a temperature sensor.

[0009] Compared with the prior art, the present application has the following beneficial effects: by using an improved pyramid visual transformer to extract multi-scale features, the variation characteristics of temperature data at different spatiotemporal scales can be fully captured, and the recognition accuracy of temperature distribution patterns and the sensitivity of the system to small temperature changes can be improved. The polarization self-attention module is used to predict temperature changes, which effectively integrates amplitude and phase information, enhances the model's ability to understand dynamic temperature changes, and improves the accuracy and robustness of predictions. Based on the efficient phase perception module, temperature anomaly detection and positioning can quickly identify and locate temperature anomaly areas, and improve the system's response speed and processing accuracy to abnormal situations. The temperature change features and abnormal area masks are integrated through the cross-layer fusion and propagation module, which realizes the effective fusion of multi-source information and improves the comprehensiveness and accuracy of feature representation. The temperature control parameters are optimized in stages according to the comprehensive feature representation, and the control strategy can be adaptively adjusted for different stages of the solder paste stirring process, which improves the accuracy and flexibility of temperature control. A multi-objective optimization model including temperature stability, energy consumption and product quality is constructed to achieve a comprehensive balance of multiple key indicators and improve the overall performance and practicality of the temperature control strategy. Adaptive weight adjustment is performed on the stability evaluation results of temperature control parameters to enhance the system's adaptability to different production conditions and improve the robustness and generalization of the temperature control strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0011] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0012] Figure 1 It is a flow chart of a solder paste stirring intelligent temperature control method based on a temperature sensor provided in an embodiment of the present invention; Figure 2 It is a schematic block diagram of the structure of a solder paste stirring intelligent temperature control system based on a temperature sensor provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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.

[0014] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0015] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0016] It should be further understood that the term "and / or" used in the specification and appended claims of this application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. Figure 1 In the embodiment of the present application, an embodiment of the intelligent temperature control method for solder paste stirring based on a temperature sensor includes: Step 100: preprocessing the temperature data collected by multiple temperature sensors in the solder paste stirring device to obtain a structured temperature data set; It is understandable that the execution subject of the present application can be a solder paste stirring intelligent temperature control system based on a temperature sensor, or a terminal or a server, which is not limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0017] Specifically, the median filtering method is applied to the raw temperature data collected by multiple temperature sensors to effectively remove outliers and transient noise in the data, and obtain smoother and more reliable filtered data. For the missing values ​​that may exist in the filtered data, the linear interpolation method is used to fill them. Through linear interpolation, a linear function is constructed between the known data points before and after the missing data to predict and fill the missing temperature values, and obtain a continuous and complete temperature data sequence. The temperature data sequence is associated with the corresponding timestamp information to form a time series temperature data set containing time information. The time series temperature data is mapped to three-dimensional spatial coordinates, and each temperature data point is matched with its spatial position when it is collected, so as to obtain a temperature data point set with spatial position information, which is helpful to identify the spatial pattern of temperature distribution. The temperature data point set is normalized so that the value range of all data points is within a standard scale, eliminating the measurement deviation between different sensors and the difference in data dimensions, and obtaining standardized temperature data. The standardized temperature data is time-aligned to obtain synchronized temperature data. The synchronized temperature data is processed by dimensionality reduction. Through principal component analysis, the original high-dimensional temperature data is mapped to a low-dimensional space to retain the main information while reducing redundant data, and obtain the feature data after dimensionality reduction. Cluster analysis is performed on the feature data after dimensionality reduction. Using clustering algorithms such as K-means or DBSCAN, temperature data points with similar characteristics are divided into different categories to identify different temperature distribution patterns. Each clustering result represents a temperature distribution pattern and is assigned an independent label, and finally a structured temperature data set containing temperature distribution pattern labels is obtained.

[0018] Step 200: Input the structured temperature data set into the improved pyramid visual transformer to perform multi-scale feature extraction to obtain a temperature feature set; Specifically, the structured temperature data set is reorganized into a two-dimensional numerical matrix, in which the rows and columns represent the spatial coordinates of the temperature data, and each element value of the matrix corresponds to the temperature value at the spatial position, forming a temperature distribution matrix. The temperature distribution matrix is ​​subjected to multi-scale decomposition processing, and the original temperature distribution matrix is ​​decomposed according to different resolution levels to obtain a multi-scale temperature representation containing information of different scales. Each decomposition level corresponds to a specific resolution level, which jointly reflects the characteristics of the temperature data at different spatial scales. The multi-scale temperature representation of each level is adaptively adjusted for the receptive field. According to the characteristics of different levels, the perception range is dynamically adjusted to obtain global information at a lower resolution level and capture local detail features at a higher resolution level, thereby obtaining a feature representation with a dynamic perception range. The feature representation with a dynamic perception range is input into the improved visual transformer encoder, and the input feature representation is encoded and processed at multiple levels to obtain a multi-level global-local feature representation. The multi-level global-local feature representation is subjected to temporal information fusion, and the changing trend of the temperature data at different time points is combined with its spatial distribution characteristics to form a four-dimensional feature tensor containing time dimension information. A multi-scale feature aggregation network is constructed based on the four-dimensional feature tensor to comprehensively process information at different scale levels. A comprehensive feature representation containing information at different scales is obtained. The feature selection mechanism that guides the attention of the comprehensive feature representation selectively highlights the key temperature patterns by assigning different attention weights, while weakening irrelevant or redundant features, to obtain a more refined feature representation. The refined features that highlight the key temperature patterns are input into the lightweight decoder network for feature processing. While ensuring the lightweight of the model, the input features are effectively decoded and processed, and finally a temperature feature set is obtained.

[0019] Step 300: Use the polarized self-attention module to perform regression prediction on the temperature feature set to obtain temperature change features; It should be noted that the temperature feature set is feature decomposed to decompose the input temperature feature into two parts: amplitude feature and phase feature. The amplitude feature represents the intensity or size information of the temperature data, while the phase feature reflects the change trend and relative relationship of the temperature data in the spatial or temporal dimension. Based on the amplitude feature obtained by decomposition, its amplitude attention weight is calculated, and the amplitude feature is used to determine the importance of each data point in the overall feature set. By calculating the attention weight of each amplitude feature, an amplitude attention matrix is ​​obtained, which reflects the importance of different amplitude features in the prediction of temperature changes. At the same time, a phase difference-sensitive attention calculation is performed on the phase feature. The phase difference in the phase feature reflects the relative change of the temperature data at different positions or time points. Through the phase difference-sensitive attention calculation, a phase attention matrix is ​​generated to reflect the correlation and dependency between different phase features. The amplitude attention matrix and the phase attention matrix are adaptively weighted fused, so that the model can autonomously adjust the contribution of the amplitude and phase features in the overall comprehensive representation according to the specific application scenario and data characteristics, and obtain a comprehensive attention representation. The comprehensive attention representation is residually connected, and the comprehensive attention representation is fused with the original temperature feature information to obtain an enhanced feature representation. Perform time series modeling on the enhanced features, input the enhanced feature representation into the time series model, capture the law and trend of temperature data changing over time, and obtain time series features that can reflect the law of temperature time change. Input the time series features into the multi-step prediction head network to predict the temperature separately. The multi-step prediction head network is a model structure for time series data prediction, which can predict temperature data at multiple time steps and obtain temperature distribution prediction results at multiple time points. Perform Bayesian reasoning on the temperature distribution prediction results. Through the Bayesian reasoning method, based on the prediction results, combined with prior knowledge and data uncertainty, the temperature change characteristics are inferred and optimized to obtain the temperature change characteristic prediction results, and the uncertainty in the prediction results can be quantified.

[0020] Step 400: Based on the temperature change characteristics, the efficient phase sensing module is used to detect and locate the temperature anomaly, and a temperature anomaly score map and an abnormal area mask are obtained; Specifically, a multi-scale feature pyramid is constructed for the temperature change feature, and the temperature change feature is decomposed into representations of different scales to form a feature pyramid, so that the model can capture the change pattern of temperature data from multiple scale levels. A phase-sensitive convolution operation is applied to the constructed feature pyramid to capture the phase information in the temperature feature, enhance the sensitivity of the phase change in the feature map, and obtain a phase-enhanced feature map. Global average pooling and one-dimensional convolution operations are performed based on the phase-enhanced feature map. The purpose of global average pooling is to average all position features in the phase-enhanced feature map to obtain a global feature representation and summarize the spatial distribution information of the entire temperature data. A one-dimensional convolution operation is applied to encode the global features to obtain context features containing global temperature distribution information. The context features are fused with the phase-enhanced feature map by channel attention. The channel attention mechanism assigns different weights according to the importance of different channel features to obtain an enhanced feature representation that fuses global and local information. A depth-separable convolution operation is performed on the enhanced feature representation to extract more detailed feature representations while maintaining efficient calculations, and obtain a multi-channel anomaly score map, where the score map of each channel corresponds to different types of temperature anomalies. The multi-channel anomaly score map is segmented by adaptive threshold. The threshold is dynamically adjusted according to the statistical characteristics of the score map of each channel to separate the abnormal area from the normal area, and a preliminary abnormal area binary mask is obtained. The preliminary abnormal area binary mask is subjected to morphological opening and closing operations. The morphological opening operation can remove small noise points in the binary mask, while the morphological closing operation can fill the small holes that may exist in the mask and make the boundary of the mask smoother. After morphological operations, the abnormal area mask after denoising and smoothing the boundary is obtained. The abnormal area mask after denoising and smoothing the boundary is subjected to connected domain analysis and statistical feature extraction. The connected domain analysis can identify the abnormal areas connected to each other in the mask, and each connected area is regarded as an independent temperature abnormal area. Through statistical feature extraction, the statistical information such as shape, area, and position of each connected domain is extracted. Combining the results of the connected domain analysis with the multi-channel anomaly score map, the temperature anomaly score map and the abnormal area mask are obtained. The temperature anomaly score map can quantitatively reflect the degree of temperature anomaly at each location, while the abnormal area mask indicates the specific area where temperature anomaly occurs during solder paste stirring.

[0021] Among them, Gaussian pyramid downsampling is applied to the temperature change feature, and the original temperature change feature map is downsampled according to different scales to generate a series of multi-scale feature atlases with different spatial resolutions. On the basis of the multi-scale feature atlas, a Laplacian pyramid is constructed. The Laplacian pyramid is a differential image obtained by subtracting the lower resolution level features from each scale level of the Gaussian pyramid. Its main function is to extract edge information at each scale. Through the construction of the Laplacian pyramid, a set of differential feature atlases that highlight edge information at different scales is obtained. The gradient directional convolution kernel is applied to the differential feature atlas to capture the gradient information in the image, that is, the first-order derivative information of the temperature data, and obtain a directional feature map reflecting the temperature change gradient. The directional feature map is input into the Laplacian convolution kernel for convolution operation. The Laplacian convolution kernel is used to extract the second-order derivative information of the image, that is, the curvature feature. By convolving the directional feature map, a curvature feature map reflecting the curvature change of the temperature data is obtained, revealing more complex change patterns in the temperature field, such as the acceleration change of the temperature gradient. The directional feature map and the curvature feature map are channel-cascaded, and the two sets of feature maps are merged in the channel dimension to obtain a multi-channel phase-sensitive feature map. The multi-channel phase-sensitive feature map is processed, including adaptive perception range adjustment and feature enhancement operations, to obtain a phase-enhanced feature map with a dynamic perception range. A global average pooling operation is performed on each channel of the phase-enhanced feature map. All eigenvalues ​​in each channel are averaged to obtain a one-dimensional feature vector reflecting the global temperature distribution. The global average pooling result of each channel reflects the temperature distribution pattern in the corresponding scale and direction. The one-dimensional feature vector is input into a one-dimensional convolutional network for feature compression and nonlinear transformation. The one-dimensional convolutional network can extract and compress the input one-dimensional feature vector through operations such as convolution and activation functions, remove redundant information, and extract more discriminative features, and finally obtain context features that encode global temperature distribution information.

[0022] Step 500: integrating the temperature change features and the abnormal area mask based on the cross-layer fusion and propagation module to obtain a comprehensive feature representation; Specifically, the channel attention mechanism is applied to the temperature change feature and the abnormal region mask respectively, and a weight is assigned to each feature channel. The size of the weight reflects the importance of the channel in the overall feature representation. By applying the channel attention mechanism to the temperature change feature, the channels that play a key role in the temperature change process are identified. By applying the channel attention mechanism to the abnormal region mask, the channel information that is closely related to the temperature anomaly can be highlighted. Based on the channel importance weight, the temperature change feature and the abnormal region mask are selected for the channel, and the channels with higher weights and significant contributions to the subsequent tasks are selected, and the redundant channels with lower weights are removed to obtain the streamlined feature representation. The streamlined feature representation is input into the deformable convolutional network. On the basis of the traditional convolutional network, the position and shape of the convolution kernel are dynamically adjusted to achieve spatial adaptive alignment of the input feature map, and the spatial adaptive alignment feature is obtained. The spatial attention mechanism is applied to the spatial adaptive alignment feature. The spatial attention mechanism can assign different weights in the spatial dimension to highlight the regional information that plays a key role in the temperature control task, and obtain enhanced features containing key regional information. The enhanced features are input into the recursive refinement unit. The recursive refinement unit gradually refines and enhances the important information in the input features through multiple iterative interactions and refinement operations, while eliminating noise and unnecessary redundant information, and obtains the feature representation after multiple interactions and refinement. The feature representation after multiple interactions and refinement is residually connected, and the features that have been refined multiple times are added to the original features to obtain fused features that retain the original information, avoiding the phenomenon of excessive smoothing or information loss in the feature during multiple processing. The fused features are input into the autoencoder for feature compression and reconstruction. The autoencoder is an unsupervised neural network structure that can map the input high-dimensional features to a low-dimensional latent space to achieve feature compression. In the process of feature compression, the autoencoder can extract the core information in the input features, remove redundancy and noise, and obtain a low-dimensional feature embedding. The low-dimensional feature embedding is multimodally fused to integrate features from different sources or different properties, so as to contain richer multi-faceted information in a unified representation, and obtain a comprehensive feature representation that not only contains the law of temperature change but also reflects the distribution of abnormal areas.

[0023] Step 600: Perform differentiated optimization of the temperature control parameters in stages according to the comprehensive feature representation to obtain an intelligent temperature control strategy for the solder paste stirring process.

[0024] Specifically, a time series analysis is performed on the comprehensive feature representation. By analyzing the changing rules of temperature data in the time dimension during solder paste stirring, multiple different stages in the stirring process are identified. Based on the stage division results and the comprehensive feature representation, a temperature distribution probability model is constructed. The model obtains the distribution of temperature in different time periods and spatial positions by statistically analyzing the temperature data in each stage, forming a multi-source probability distribution function. The multi-source probability distribution function is fitted using the Kriging surrogate model. As a spatial statistical method, the Kriging model obtains the temperature distribution in the entire spatial range by fitting a small amount of sample data. The Kriging surrogate model is used to fit the relationship between the multi-source probability distribution function and the temperature control parameters, and obtains the mapping relationship between the temperature control reliability index and the temperature control parameters. Based on the mapping relationship between the temperature control reliability index and the temperature control parameters, a multi-objective optimization model is constructed. The model takes temperature stability, energy consumption and product quality as the optimization objective functions, and aims to find the optimal solution with the minimum temperature fluctuation, the lowest energy consumption and the best product quality under different combinations of temperature control parameters. The temperature stability objective function is used to evaluate the stability of temperature changes during solder paste stirring; the energy consumption objective function quantifies the amount of energy consumed under a specific temperature control strategy; and the product quality objective function evaluates the final quality performance of solder paste during the stirring process based on indicators such as solder paste stirring effect and uniformity. Through the comprehensive optimization of these three objectives, the stability, economy and product quality of temperature control are balanced under different temperature control parameters. In order to effectively solve this multi-objective optimization problem, an improved non-dominated sorting genetic algorithm II is used. This algorithm can handle complex multi-objective optimization problems and obtain the Pareto optimal solution set for each stage of the stirring process by simulating the selection, crossover and mutation operations in the biological evolution process. The Pareto optimal solution set means that there is no solution that can be superior to other solutions in all objectives at the same time between different objectives, so they constitute a set of optimal solution candidates. Based on the Pareto optimal solution set, a robustness analysis is performed, that is, the stability and adaptability of various temperature control parameter combinations under different stirring conditions and environmental changes are evaluated, and those temperature control parameter combinations that are stable under different working conditions and uncertainties are identified, and the stability evaluation results of the temperature control parameters are obtained. The stability evaluation results are adaptively weighted, and the weights of each objective function are adjusted according to the actual needs and importance of each stage to determine the optimal temperature control parameter combination for each stage. Based on the optimal temperature control parameter combination for each stage, an intelligent temperature control strategy for the solder paste mixing process is generated. This strategy can dynamically adjust the temperature control parameters at different stages of the mixing process to ensure that the temperature of the solder paste is always kept within the optimal range, thereby improving the uniformity and efficiency of mixing, and can minimize energy consumption while ensuring product quality.

[0025] In a specific embodiment, the process of executing step 100 may specifically include the following steps: Perform median filtering on the raw temperature data collected by multiple temperature sensors to obtain filtered data with outliers and transient noise removed; Linearly interpolate the missing values ​​in the filtered data to obtain a continuous and complete temperature data sequence, and associate the temperature data sequence with the corresponding timestamp information to obtain time series temperature data; Perform three-dimensional spatial coordinate mapping on the time series temperature data to obtain a temperature data point set with spatial position information, and perform normalization processing on the temperature data point set to obtain standardized temperature data; Performing time alignment on the standardized temperature data to obtain synchronized temperature data, and performing dimension reduction processing on the synchronized temperature data to obtain feature data after dimension reduction; Cluster analysis is performed on the feature data after dimensionality reduction to obtain a structured temperature dataset containing temperature distribution pattern labels.

[0026] Specifically, median filtering is performed on the raw temperature data collected by multiple temperature sensors to effectively remove sudden noise and outliers in the data. For each data point in a set of data, the median of the point and all data points in its neighborhood is used to replace the original data point to achieve the effect of filtering out isolated outliers. Assume that the original temperature data sequence is ,in Indicates The temperature sensor is in the The temperature values ​​collected at each time point. The filtered temperature data sequence is obtained by median filtering. ,in median , is the radius of the filter window, which indicates the range of the front and back neighborhoods considered during the filtering process. Fill missing values ​​in the filtered temperature data. Linear interpolation predicts and fills missing data points through the linear relationship between known data points. Assume that in the filtered temperature data sequence In the case of two known data points and There are missing values ​​between them, and the linear interpolation formula is used to estimate these missing values: ; in Represents the index of the missing data point, satisfying , is the interpolated temperature data point. In this way, a continuous and complete temperature data sequence is obtained. , ensuring that each time point corresponds to a temperature value. The corresponding timestamp information Associate to form time series temperature data . The time series temperature data is mapped to three-dimensional spatial coordinates, and each temperature data point is combined with its collection location and time to form a temperature data point set with spatial location information. Assume that the temperature data is collected on a two-dimensional plane, and the collection location is represented by two-dimensional coordinates. To represent, each temperature data point is mapped to ,in represents the position of the sensor in the plane, is the timestamp, is the corresponding temperature value. Through three-dimensional space coordinate mapping, a temperature data point set with position and time information is obtained. The temperature data point set is normalized to eliminate the differences in different data dimensions and numerical ranges, improve the comparability between different features, and reduce the deviation caused by dimension differences during data analysis. The standardized temperature data is obtained. It has zero mean and unit variance. Perform time alignment on the standardized temperature data to align the data from different sensors to the same time base. Assume Indicates the time base, Represents the actual acquisition time of a sensor, and is converted into The corresponding temperature value is mapped to the reference time The synchronized temperature data is obtained after time alignment. The synchronized temperature data is subjected to dimensionality reduction processing, and the high-dimensional temperature data is mapped to a low-dimensional space, retaining the most important features of the data while reducing redundant information. Assume that the standardized temperature data is , whose covariance matrix is , use principal component analysis to solve the following eigenvalue problem: ; in, is the eigenvalue, is the corresponding eigenvector. The eigenvector corresponding to the largest eigenvalue is used to obtain the feature data after dimensionality reduction. Cluster analysis is performed on the feature data after dimensionality reduction. Cluster analysis is an unsupervised learning method that aims to group similar temperature data in order to identify different temperature distribution patterns. The K-means clustering algorithm is used to cluster the temperature data after dimensionality reduction. Clustering is performed. Assume that the number of clusters is , then the goal of the K-means algorithm is to minimize the sum of the squares of the distances from each data point to the center of its cluster: ; in, represents a temperature data point, Indicates Clusters, is the center of the cluster. By iteratively updating the cluster center and cluster assignment, we can get temperature distribution pattern labels to form a structured temperature dataset containing temperature distribution pattern labels.

[0027] In a specific embodiment, the process of executing step 200 may specifically include the following steps: Reorganize the structured temperature data set into a two-dimensional numerical matrix, where the rows and columns represent spatial coordinates and the matrix element values ​​represent temperatures, and obtain a temperature distribution matrix; Perform multi-scale decomposition on the temperature distribution matrix to obtain a multi-scale temperature representation containing different resolution levels, and perform adaptive receptive field adjustment on each level of the multi-scale temperature representation to obtain a feature representation with a dynamic perception range; The feature representation with dynamic perceptual range is input into the improved visual transformer encoder to obtain a multi-level global-local feature representation, and the multi-level global-local feature representation is fused with temporal information to obtain a four-dimensional feature tensor containing time dimension information; A multi-scale feature aggregation network is constructed based on the four-dimensional feature tensor to obtain a comprehensive feature representation that integrates information at different scales. A feature selection mechanism is then implemented to guide the comprehensive feature representation, thereby obtaining refined features that highlight key temperature patterns. The refined features highlighting the key temperature patterns are input into the lightweight decoder network for feature processing to obtain the temperature feature set.

[0028] Specifically, the temperature data is associated with its corresponding spatial coordinates. Suppose there is a set of temperature data sets ,in Indicates the position in space The temperature value at Represents the total number of sensors. Map the temperature data to a two-dimensional matrix The rows and columns of the matrix represent the spatial coordinates and , the element values ​​in the matrix It represents the temperature value at the corresponding coordinate position. Through mapping, a complete temperature distribution matrix is ​​obtained to show the temperature distribution in the entire spatial area. The temperature distribution matrix is ​​decomposed at multiple scales to represent the original temperature data at different resolution levels and capture the change pattern of temperature distribution at different spatial scales. The Gaussian pyramid decomposition method is used to decompose the original temperature matrix. Downsampling is performed to generate temperature matrices with gradually decreasing resolutions layer by layer. Matrices of different resolutions form multi-scale temperature representations. For example, the first layer is the original matrix , the second layer is downsample The third layer is downsample , and so on. Each layer of temperature matrix can reflect the characteristics of temperature distribution at its corresponding scale. A Laplacian pyramid is constructed based on the Gaussian pyramid. By performing Gaussian filtering and smoothing operations on the matrices of each layer of the Gaussian pyramid, and subtracting the matrix of each layer from its downsampled version, edge information matrices at different scales are obtained. Adaptive receptive field adjustment is performed on each level of the multi-scale temperature representation. The receptive field refers to the local area of ​​the input data perceived by the network during the feature extraction process. By adaptively adjusting the size of the receptive field, the feature representations at different levels can better capture local and global information. For each layer of the feature matrix , by introducing convolution kernels of different sizes to dynamically adjust the receptive field. For example, higher resolution layers (such as ) Use a smaller convolution kernel (such as ) to capture detailed information; while lower resolution layers (such as ) then use a larger convolution kernel (such as ) in order to capture global information. Through adaptive adjustment of the receptive field, a feature representation with a dynamic perceptual range is obtained. The feature representation with a dynamic perceptual range is input into the improved visual transformer encoder. The visual transformer is a neural network model based on an attention mechanism that can effectively process image data. The visual transformer divides the input image into a series of non-overlapping small blocks and processes each small block as an input token. Each matrix block in the multi-scale temperature representation is regarded as a token and input into the transformer. The improved visual transformer extracts global and local feature representations layer by layer through a multi-layer self-attention mechanism. Assume that the input temperature feature is ,in Represents the features after each block, then the output of the visual transformer is expressed as: Attention softmax ; in, denote query, key, and value matrices respectively, , is the linear transformation matrix, is the scaling factor. In this way, a multi-level global-local feature representation is obtained. The multi-level global-local feature representation is fused with the time series information to form a four-dimensional feature tensor containing time dimension information. Assume that the multi-level feature representations at different time points are , then construct a four-dimensional feature tensor Based on the four-dimensional feature tensor , construct a multi-scale feature aggregation network to effectively fuse information at different scale levels and obtain a comprehensive feature representation containing features of different scales. A multi-branch convolutional neural network is used, each branch processes features of a specific scale, and then the outputs of each branch are fused by element-by-element addition or concatenation. Assume that the features of different scales are represented as , the result of multi-scale feature aggregation is expressed as: concat . Through the fusion operation, a comprehensive feature representation that integrates information of different scales is obtained. Based on the comprehensive feature representation, an attention-guided feature selection mechanism is applied. The attention mechanism can assign different weights according to the importance of features, thereby highlighting key temperature pattern information. This feature selection is achieved through the self-attention mechanism, which calculates the attention weight of each position in the comprehensive feature representation, and uses these weights to weight the feature values ​​of different positions to obtain a refined feature representation. Assume that the comprehensive feature representation is , the attention weight is , then the refined feature representation is expressed as: ; in, is the attention weight calculated by the softmax function. The refined key temperature pattern features are input into the lightweight decoder network for feature processing. The lightweight decoder network is a neural network structure with high computational efficiency and few parameters, which can reduce computational overhead while maintaining model performance. Through multiple layers of deconvolution and linear transformation operations, the lightweight decoder can convert the input refined features into a set of temperature features.

[0029] In a specific embodiment, the process of executing step 300 may specifically include the following steps: Perform feature decomposition on the temperature feature set to obtain amplitude features and phase features, and calculate the amplitude attention weight based on the amplitude features to obtain the amplitude attention matrix; Perform phase difference-sensitive attention calculation on the phase feature to obtain the phase attention matrix, and perform adaptive weighted fusion of the amplitude attention matrix and the phase attention matrix to obtain the comprehensive attention representation; Perform residual connection on the comprehensive attention representation to obtain enhanced features that integrate the original temperature feature information, and perform time series modeling on the enhanced features to obtain time series features that capture the time variation law of temperature; The time series features are input into the multi-step prediction head network to predict the temperatures separately, and the temperature distribution prediction results are obtained. The temperature distribution prediction results are then subjected to Bayesian inference to obtain the temperature change characteristics.

[0030] Specifically, the temperature feature set is transformed from the time domain to the frequency domain through Fourier transform. The Fourier transform can decompose the temperature feature set into different frequency components, where each frequency component is represented by amplitude and phase. Assume that the temperature feature set is , is a story about time The function is expressed in the frequency domain by Fourier transform. ,Right now: ; in, Indicates frequency, is an imaginary unit, is a complex exponential function. Expressed as complex magnitude and phase: ; in, is the amplitude, indicating the signal at frequency The strength of is the phase, which indicates the phase offset of the frequency component. Through Fourier transform, the original temperature feature set is decomposed into amplitude features and phase characteristics Based on the amplitude characteristics Calculate the amplitude attention weight. The amplitude feature represents the intensity of the temperature data at different frequency components. The importance of different frequency components is measured through the attention mechanism. For each frequency component , calculate the attention weight of the amplitude feature: ; in, Indicates frequency The attention weight is the magnitude The ratio of the amplitude of all frequency components to the sum of the amplitudes. After normalization, the amplitude attention matrix is ​​obtained. , which reflects the relative importance of different frequency components in the entire temperature feature set. Perform phase difference sensitive attention calculation. The phase feature reflects the temporal variation pattern of the temperature data. When processing the phase feature, the phase difference between different frequency components is considered. The phase difference sensitive attention calculation is implemented by the following steps: Calculate the difference matrix of the phase feature, that is, the phase difference between different frequency components: ; in, Indicates frequency and The phase difference between them. Constructing phase attention weights based on the phase difference matrix , calculated using the following formula: ; in, is the reference frequency, usually the frequency component with the smallest phase difference is selected. is a hyperparameter that controls the sensitivity of phase difference. Through the phase difference sensitive attention calculation, the phase attention matrix is ​​obtained , which reflects the importance of phase consistency between different frequency components. and the phase attention matrix Perform adaptive weighted fusion to obtain comprehensive attention representation. The process of adaptive weighted fusion is expressed as: ; in, Indicates frequency The comprehensive attention weight of and are adaptive weights, which respectively represent the relative importance of amplitude attention and phase attention in the comprehensive attention representation. Weighted fusion can comprehensively consider the information of amplitude features and phase features on different frequency components to obtain a more comprehensive attention representation. Perform residual connection and combine it with the original temperature feature to obtain enhanced features that integrate the original temperature feature information. The residual connection is expressed by the following formula: ; in, represents the enhanced temperature characteristics, is the original temperature feature. Through the residual connection method, while retaining the original feature information, the attention mechanism is introduced to weight different frequency components to obtain a richer feature representation. Perform time series modeling to capture the changing rules of temperature features in the time dimension. Use time series models such as long short-term memory networks or gated recurrent units to process enhanced features. Assume that the input of the time series model is , the output is the time series feature , the recursive calculation process of the time series model is expressed as: LSTM 1)); in, Indicates at time The temporal characteristics of the moment, is the time series feature of the previous moment. Through recursive calculation, it captures the dynamic change pattern of temperature features over time. Input the multi-step prediction head network to predict the temperature separately. The multi-step prediction head network is a model structure for time series prediction, which can predict the future temperature at multiple time steps. Assume that the output of the multi-step prediction head network is the temperature distribution prediction result , expressed as: ForecastNet ; Among them, ForecastNet represents the multi-step prediction head network, It's in time Temperature prediction results at the moment. Temperature distribution prediction results Perform Bayesian inference to obtain the temperature change characteristics. Bayesian inference is a statistical method that combines prior information and observation data to update the distribution. Combining the temperature prediction results with the prior distribution can obtain a more reliable representation of the temperature change characteristics. Assume that the prior distribution is , the likelihood function of the observed data is , the posterior distribution of Bayesian inference is expressed as: ; in, represents the temperature change characteristic parameter, is the posterior distribution. Through Bayesian reasoning, the distribution of temperature prediction results is combined with prior information to obtain a more accurate and robust distribution of temperature change characteristics.

[0031] In a specific embodiment, the process of executing step 400 may specifically include the following steps: A multi-scale feature pyramid is constructed and a phase-sensitive convolution operation is performed on the temperature change features to obtain a phase-enhanced feature map. Global average pooling and one-dimensional convolution operations are performed based on the phase-enhanced feature map to obtain context features that encode global temperature distribution information. The context features are fused with the phase-enhanced feature map by channel attention to obtain an enhanced feature representation that integrates global and local information, and a depth-wise separable convolution operation is performed on the enhanced feature representation to obtain a multi-channel anomaly score map. Perform adaptive threshold segmentation on the multi-channel anomaly score map to obtain a preliminary abnormal region binary mask, and perform morphological opening and closing operations on the preliminary abnormal region binary mask to obtain a denoised and smoothed boundary abnormal region mask; The connected domain analysis and statistical feature extraction are performed on the abnormal area mask after denoising and smoothing the boundaries to obtain the temperature anomaly score map and abnormal area mask.

[0032] Specifically, multi-scale feature information is extracted from the temperature data to capture the temperature change patterns at different spatial and temporal scales. The temperature change features are constructed as a multi-scale feature pyramid. The multi-scale feature pyramid is a structure that decomposes the original feature data into different resolution levels, so that each level can reflect the feature information of the temperature data at different spatial scales. The construction process is implemented through a Gaussian pyramid, that is, the temperature feature map is downsampled at each layer to obtain a feature map set with gradually reduced resolution. Assume that the original temperature feature map is , then Layer feature map It is expressed as: downsample ; Where downsample(·) is a downsampling operation, which is to obtain a low-resolution feature representation by performing Gaussian filtering on the image and then downsampling it, forming a feature map with different resolutions. A multi-scale feature pyramid is composed of a feature map. Each feature map contains information about temperature change characteristics at different scales, which helps to capture local and global temperature change patterns. A phase-sensitive convolution operation is performed on the feature map of each level. By combining the phase information in the feature maps of different scales, the direction and trend of temperature change are sensitively captured. The design of the phase-sensitive convolution kernel contains amplitude information as well as phase difference information to better characterize the temperature change characteristics. For example, for the feature map Each position in , its phase-sensitive convolution operation is expressed as: ; in, is the phase enhancement feature map, is the convolution kernel weight, is the phase information, which indicates the phase offset of the convolution kernel at each position. Through the phase-sensitive convolution operation, the enhanced temperature feature map is obtained. Based on the phase-enhanced feature map, global average pooling and one-dimensional convolution operations are performed. The global information in the feature map is summarized to obtain the context information reflecting the overall temperature distribution characteristics. Global average pooling averages the feature values ​​of all positions in the phase-enhanced feature map to form a global context feature vector. Assume that the phase-enhanced feature map is , whose size is , then the global average pooling operation is expressed as: ; in, It is The context features of the layer feature map reflect the global distribution characteristics of the temperature data. Perform a one-dimensional convolution operation to extract the key features of the global temperature distribution information. The one-dimensional convolution operation obtains a more expressive global feature representation by performing linear transformation and nonlinear activation on the context features. Assume that the one-dimensional convolution kernel is , whose size is ,in is the dimension of the convolution kernel, then the one-dimensional convolution operation is expressed as: ; Among them, * represents the convolution operation, is the bias term, is a nonlinear activation function, such as ReLU. Through operation, the obtained feature representation Contains information about the global temperature distribution. Phase enhancement feature map Perform channel attention fusion. The channel attention mechanism can dynamically adjust the weights of different channels to highlight the channel information that is important to the task. The channel weight is obtained by calculating the channel attention of the context features.

[0033] softmax(FC ; Among them, FC(·) is the fully connected layer operation, and softmax(⋅) is used to normalize the channel weights. The calculated channel weights are compared with the phase enhancement feature map Perform weighted fusion to obtain enhanced feature representation that integrates global and local information The channel attention fusion method can dynamically adjust the importance of each local feature while combining global information to generate a more representative feature representation. Perform depthwise separable convolution. Depthwise separable convolution is an effective convolution operation that decomposes the traditional convolution into depthwise convolution and pointwise convolution, reducing the amount of calculation while retaining the expressive power of features. Perform a depthwise convolution operation, that is, perform convolution on each channel separately: DepthwiseConv ; in, It is the feature representation after deep convolution. Then perform point-by-point convolution to integrate the information of different channels: PointwiseConv ; in, It is the feature representation after the depthwise separable convolution. Through the depthwise separable convolution operation, an abnormal score map containing multi-channel information is obtained. , reflecting the possible abnormal patterns on different channels. Perform adaptive threshold segmentation. According to the distribution characteristics of different channels, the threshold is automatically adjusted to distinguish abnormal areas from normal areas. Assume that the score map of each channel is , whose threshold is , then the binary mask is expressed as: ; in, It is The binary mask of each channel. When the pixel value in the score map is greater than the threshold, the mask value is 1, indicating that the area is an abnormal area; otherwise, the mask value is 0, indicating that the area is a normal area. Perform morphological opening and closing operations on the preliminary abnormal area binary mask. The morphological opening operation is used to remove small noise points in the binary mask, while the closing operation is used to fill small holes that may exist in the binary mask. The opening operation is to erode first and then expand, which can remove small noise points; the closing operation is to expand first and then erode, which can fill small holes. After the opening and closing operations, the abnormal area mask after denoising and smoothing the boundaries is obtained. . Mask the abnormal area after denoising and smoothing the boundaries Connected domain analysis and statistical feature extraction are performed. Connected domain analysis is used to identify abnormal areas that are interconnected in the binary mask and mark them as different connected domains. Statistical feature extraction is to calculate the geometric features of each connected domain, including area, perimeter, centroid position, etc. Assume that the connected domain is marked as ,in represents the number of connected domains, then the area of ​​each connected domain is It is expressed as: ; By performing statistical analysis on the features of each connected domain, we can obtain the temperature anomaly score map and the final abnormal area mask. The temperature anomaly score map can quantify the degree of abnormality in each area, while the abnormal area mask marks the temperature abnormal areas during the solder paste stirring process.

[0034] In a specific embodiment, the execution step performs a multi-scale feature pyramid construction and a phase-sensitive convolution operation on the temperature change feature to obtain a phase-enhanced feature map, and performs a global average pooling and a one-dimensional convolution operation based on the phase-enhanced feature map to obtain a context feature encoding global temperature distribution information. The process may specifically include the following steps: The temperature change features are downsampled using a Gaussian pyramid to obtain a multi-scale feature atlas with different spatial resolutions. A Laplacian pyramid is constructed based on the multi-scale feature atlas to obtain a differential feature atlas that highlights edge information at different scales. Applying the gradient directional convolution kernel to the differential feature atlas to obtain the directional feature map that captures the first-order derivative information of the temperature gradient, and inputting the directional feature map into the Laplace convolution kernel for convolution operation to obtain the curvature feature map that captures the second-order derivative information of the temperature field; Channel-concatenating the directional feature map and the curvature feature map to obtain a multi-channel phase-sensitive feature map, and processing the multi-channel phase-sensitive feature map to obtain a phase-enhanced feature map with a dynamic perception range; Perform a global average pooling operation on each channel of the phase enhancement feature map to obtain a one-dimensional feature vector encoding the global temperature distribution; The one-dimensional feature vector is input into a one-dimensional convolutional network for feature compression and nonlinear transformation to obtain contextual features that encode global temperature distribution information.

[0035] Specifically, the original temperature change feature map is processed at multiple scales to capture the characteristic information of the temperature field at different spatial resolutions. Gaussian pyramid is a multi-scale analysis method commonly used in image processing. It generates a series of images with gradually reduced resolutions by smoothing and downsampling the image layer by layer. Assume that the temperature feature map is , and perform Gaussian filtering and smoothing to obtain , and then downsample it to get an image with half the resolution This process is expressed as: downsample(GaussianBlur ; in, Indicates Feature map of layer, GaussianBlur Indicates Gaussian filtering operation, downsample Represents the downsampling operation. Through this formula, multi-level feature graphs are generated in sequence. , where the resolution of each layer of feature maps is halved layer by layer to form a multi-scale feature map set. The Laplacian pyramid is constructed based on the Gaussian pyramid to extract edge features and detail information at different scales. The Laplacian pyramid is obtained by performing differential calculations on each layer of the feature map of the Gaussian pyramid and its downsampled version. Assume that The Gaussian pyramid feature map of the layer is , No. The downsampled Gaussian feature map of the layer is , then Laplacian eigenmap of the layer It is expressed as: upsample ; Among them, upsample Indicates that the lower layer feature map is upsampled to restore it to the same resolution as the previous layer. Through differential operation, prominent edge information at different scales is obtained to form a Laplacian pyramid . Apply the gradient direction convolution kernel to the differential feature map set of each layer in the Laplacian pyramid to extract the first-order derivative information of the temperature gradient. The gradient direction convolution kernel is a filter commonly used to capture the gradient changes of an image and extract the gradient features along a specific direction in the image. For the differential feature map in the temperature field, the gradient convolution kernels in the horizontal and vertical directions are designed, such as and , they are represented as: , ; For each differential feature map , which is and The gradients in the directions are expressed as: , ; Among them, * represents the convolution operation, and Respectively expressed in and Directional gradient map. By combining the gradient information in these two directions, a directional feature map is obtained: ; The directional characteristic map contains the first-order derivative information in the temperature change field and can reflect the gradient direction of the temperature change. Input the Laplace convolution kernel for convolution operation to capture the second-order derivative information of the temperature field and obtain the curvature feature map. The Laplace convolution kernel is a filter used to detect the second-order change information of the image, which can effectively extract the curvature information in the image. The Laplace convolution kernel is expressed as: ; Directional feature map Convolution with Laplacian kernel Perform convolution operation to obtain the curvature feature map It is expressed as: ; in, Represents the curvature feature map. This curvature feature map reflects the second-order derivative information in the temperature change field and can capture more complex curvature changes in the temperature field. And curvature feature map Perform channel cascade operation to obtain multi-channel phase-sensitive feature maps. Channel cascade is to splice two or more feature maps in the channel dimension to form a multi-channel feature representation. Assume that the sizes of the directional feature map and the curvature feature map are , phase sensitivity feature map after channel cascading It is expressed as: concat ; Among them, concat Indicates concatenation in the channel dimension. The resulting phase-sensitive feature map Contains information about first-order and second-order changes in the temperature field. Phase sensitivity characteristic diagram for multiple channels The phase enhancement feature map with dynamic perceptual range is processed. The perceptual range in the feature map is dynamically adjusted by the adaptive receptive field adjustment method. The adaptive receptive field adjustment is achieved by using convolution kernels of different sizes to dynamically adjust the size of the receptive field at different locations of the feature map to capture information at different scales. A global average pooling operation is performed on each channel of to obtain a one-dimensional feature vector encoding the global temperature distribution. Global average pooling averages the feature values ​​of all positions on each channel to obtain a global feature representation. Assuming the phase enhancement feature map The size is ,in and Represents the height and width respectively, represents the number of channels, then The global average pooling feature of each channel is expressed as: ; in, Indicates The global feature representation of the channels. The global feature vector is input into the one-dimensional convolutional network for feature compression and nonlinear transformation to obtain the contextual features that encode the global temperature distribution information. The one-dimensional convolutional network can effectively compress the dimension of the feature and extract contextual information with stronger expression ability by performing one-dimensional convolution operation and nonlinear activation on the global feature vector. Assume that the global feature vector is , the convolution kernel of the one-dimensional convolutional network is , whose size is ,in is the dimension of the convolution kernel, then the one-dimensional convolution operation is expressed as: ; in, is a context feature, is the bias term, is a nonlinear activation function, such as ReLU. The context features obtained in this way The information of global temperature distribution is encoded, which can provide effective feature representation for subsequent temperature analysis and decision-making.

[0036] In a specific embodiment, the process of executing step 500 may specifically include the following steps: Apply channel attention mechanism to temperature change features and abnormal area masks respectively to obtain their respective channel importance weights; Channel selection is performed on the temperature change features and abnormal area masks based on the channel importance weights to obtain a streamlined feature representation; The simplified feature representation is input into the deformable convolutional network to obtain spatially adaptive alignment features, and the spatial attention mechanism is applied to the spatially adaptive alignment features to obtain enhanced features that highlight the key area information; The enhanced features are input into the recursive refinement unit to obtain feature representations after multiple interactions and refinements, and residual connections are performed on the feature representations after multiple interactions and refinements to obtain fused features that retain the original information; The fused features are input into the autoencoder for feature compression and reconstruction to obtain low-dimensional feature embedding, and the low-dimensional feature embedding is subjected to multimodal fusion to obtain a comprehensive feature representation.

[0037] Specifically, the channel attention mechanism is applied to the temperature change feature and the abnormal area mask respectively. The weight of each channel is automatically adjusted according to the contribution of different channels in the feature expression. For a given temperature change feature map and abnormal area mask , respectively, with sizes of of shape, where and Represents the height and width respectively, represents the number of channels. The channel attention mechanism obtains the global feature representation of each channel by performing global average pooling on each channel. Assume that the first The channels are , then its global average pooling feature It is expressed as: ; in, Indicates The global information of each channel. For the abnormal area mask , calculate the global feature representation of each channel . The global eigenvector and Input them into the fully connected network respectively to calculate the importance weight of each channel. The fully connected network consists of two layers of fully connected layers. The output dimension of the first layer is usually smaller than the input dimension, which is used for dimensionality reduction. The second layer restores the dimension to the original number of channels to achieve nonlinear transformation. The calculation process of channel importance weight is expressed as: ; in, and Represent the first and second layer fully connected operations respectively, and They are activation functions (such as ReLU and Sigmoid), represents the channel weight of the temperature change feature map. Similarly, for the abnormal area mask, the channel weight obtained is . Channel selection is performed based on the channel importance weights for temperature change features and abnormal area masks. Set a threshold Select channels with higher weights, that is, retain channel information that is more important to the task. The selected channels are expressed as: , ; in, and They represent the indicator function of channel selection, that is, when the channel weight is greater than the threshold, the channel is retained, otherwise it is set to zero. In this way, the simplified feature representation is obtained and exception mask The simplified feature representation is and Input into the deformable convolutional network to obtain spatially adaptive alignment features. The deformable convolutional network is a model that enhances the receptive field and deformation ability of the traditional convolutional network. By introducing dynamically learned displacement in the standard convolution operation, the convolution kernel can dynamically adjust its sampling position according to the local structure of the feature map. For the input feature map , the output of the deformable convolution It is expressed as: ; in, is the convolution kernel at position The weight of the convolution kernel, R is the sampling area of ​​the convolution kernel, is the convolution kernel at position The offset is dynamically learned by the deformable convolution network. Deformable convolution enables different regions of the feature map to adaptively adjust the receptive field, thereby capturing more detailed local structural information in a complex temperature change field. After being processed by the deformable convolution network, the spatially adaptive alignment feature is obtained. . Apply the spatial attention mechanism to the spatially adapted alignment features O to highlight the key area information. The spatial attention mechanism is a mechanism for assigning weights in the spatial dimension, which can assign different weights to each position in the feature map and enhance the attention to important areas. Calculate the average pooling and maximum pooling feature representation of the feature map in the spatial dimension and , and then concatenate them and pass them through a convolutional layer to get the spatial attention map

[0038] Conv ; in, represents the concatenation operation in the channel dimension, Represents the Sigmoid function. The spatial attention map Alignment features that are adaptive to space Perform element-by-element multiplication to obtain enhanced features

[0039] ; Among them, © represents the element-by-element multiplication operation. In this way, the spatial attention mechanism can effectively highlight the regional information that contributes to the task and obtain more refined enhanced features. Input the recursive refinement unit to interact and refine the feature information multiple times. The recursive refinement unit gradually refines the input features through a multi-layer recurrent network structure, and each iteration refines the previous output. Assume that the input of the recursive refinement unit is , No. The output of the next refinement is , then the recursive refinement process is expressed as: RRU ; Among them, RRU represents the operation of recursively refining the unit, is the result of the previous refinement. After multiple recursive refinements, the feature representation after multiple interactions and refinements is obtained In order to avoid the loss of information during multiple recursive refinements, the feature representation after multiple interactions and refinements is Performing residual connection with the original features effectively alleviates the gradient vanishing problem in deep networks and retains the original information. The residual connection operation is expressed as: ; in, It is the fusion feature after residual connection, which combines the high-level information after multiple refinements and the original enhanced feature information, and has stronger expression ability. Input the autoencoder for feature compression and reconstruction. The autoencoder consists of two parts: the encoder and the decoder. The encoder is used to compress the input features into a low-dimensional latent space, while the decoder reconstructs the original features from the latent space. Assume that the encoder part of the autoencoder outputs a low-dimensional feature embedding , expressed as: ; in, Represents the encoder part of the autoencoder. The low-dimensional feature embedding z extracts the most core and useful information of the input features. The decoder part reconstructs the original features from the low-dimensional feature embedding z: Decoder ; in, is the reconstructed feature, Decoder Represents the decoder part of the autoencoder. For low-dimensional feature embedding Perform multimodal fusion to obtain comprehensive feature representation. Multimodal fusion jointly models feature representations from different sources or different characteristics to obtain richer feature representations.

[0040] In a specific embodiment, the process of executing step 600 may specifically include the following steps: The comprehensive feature representation is analyzed in time series to obtain the multiple stage division results of the solder paste stirring process. Based on the stage division results and the comprehensive feature representation, a temperature distribution probability model is constructed to obtain a multi-source probability distribution function. Based on the Kriging surrogate model, the multi-source probability distribution function is fitted to obtain the mapping relationship between the temperature control reliability index and the temperature control parameters. A multi-objective optimization model is constructed based on the mapping relationship to obtain the optimization objective function including temperature stability, energy consumption and product quality; The improved non-dominated sorting genetic algorithm II is applied to the optimization objective function to obtain the Pareto optimal solution set of each stage of the mixing process, and the robustness analysis is performed based on the Pareto optimal solution set to obtain the stability evaluation results of the temperature control parameters; The stability evaluation results are adaptively weighted to obtain the optimal temperature control parameter combination for each stage, and an intelligent temperature control strategy for the solder paste stirring process is generated based on the optimal temperature control parameter combination.

[0041] Specifically, the temperature data and time information of the solder paste stirring process are combined to establish a time series feature representation. Assume that the comprehensive feature representation is ,in Represents time. The comprehensive feature representation includes a variety of information such as temperature, stirring speed, energy consumption, etc., which can reflect the dynamic change pattern of the solder paste stirring process. In order to analyze the timing characteristics, a clustering algorithm (such as the dynamic time warping clustering algorithm) is used to cluster the characteristics of different time periods to obtain multiple stage division results in the stirring process. Each stage corresponds to a different stirring state, such as the preheating stage, the uniform stirring stage, and the cooling stage. A temperature distribution probability model is constructed based on the stage division results and the comprehensive feature representation. The temperature distribution probability model is a statistical model that describes the temperature changes in each stage, which can convert complex temperature feature information into a probability distribution representation. Assume that in the first Stage, comprehensive feature representation The temperature distribution is expressed as a probability density function To indicate that, is the temperature, It is By modeling the data of each stage, the temperature distribution probability model of each stage is obtained. In order to integrate the temperature distribution information of these different stages, a multi-source probability distribution function is constructed. This function combines the temperature distribution models of multiple stages to describe the temperature changes during the entire mixing process. , the Kriging proxy model is used for fitting, and the mapping relationship between the temperature control reliability index and the temperature control parameters is obtained. The Kriging model is a spatial statistical method that models and interpolates complex multidimensional functions with a small number of sampling points. Assume that the multi-source probability distribution function is ,in represents the parameter set of all stages, and the temperature control parameter is , find the mapping relationship between temperature control parameters and temperature control reliability index Through the Kriging proxy model, the following mapping relationship is established: ; in is the global trend function of the temperature control parameters. is a Gaussian process that satisfies zero mean and a specific covariance function. The mapping relationship can transform the temperature control parameters Reliability of temperature control Based on the mapping relationship obtained by the Kriging model, a multi-objective optimization model is constructed to seek the best combination of temperature control parameters. The goal of the multi-objective optimization model is to simultaneously optimize multiple objective functions such as temperature stability, energy consumption, and product quality. Assume that the objective function of temperature stability is , which measures the magnitude of temperature fluctuation during the mixing process; the objective function of energy consumption is , which measures the energy consumption under specific temperature control parameters; the objective function of product quality is , reflecting the uniformity and physical properties of the solder paste after stirring. The multi-objective optimization problem is expressed as: ,in, Represents the set of temperature control parameters. The multi-objective optimization model can find a balance between multiple objectives so that the combination of temperature control parameters can simultaneously meet the requirements of temperature stability, energy consumption and product quality. In order to effectively solve the multi-objective optimization problem, the improved non-dominated sorting genetic algorithm II (NSGA-II) is applied. NSGA-II is a multi-objective optimization algorithm that can find a set of Pareto optimal solutions among multiple objectives. The Pareto optimal solution set refers to the solution set that cannot further optimize a certain objective without sacrificing other objectives. NSGA-II searches for the Pareto optimal solution in the solution space through operations such as population initialization, selection, crossover and mutation. Assume that in the The individuals in the generation population are , its fitness function is: ; Through continuous evolution, we find a solution set that satisfies the following conditions: st and ,in, is a solution in the Pareto optimal solution set, and is the solution in the current population. Through evolutionary strategies, the optimal temperature control parameter combination is gradually approached. Robustness analysis is performed based on the Pareto optimal solution set at each stage of the mixing process to evaluate the stability of the temperature control parameters. The purpose of robustness analysis is to evaluate the performance of the Pareto optimal solution under different working conditions, that is, in practical applications, whether these temperature control parameters can still perform stably under uncertain conditions. By simulating the temperature control process under different working conditions, the performance fluctuations of each Pareto solution are calculated. Assume that a solution in the Pareto solution set is , the performance fluctuation under different working conditions is measured by standard deviation: ; in, represents the standard deviation of performance fluctuation, It is Performance indicators under the secondary working conditions, is the average performance under all working conditions. By calculating the standard deviation of each solution, we can determine its stability under different working conditions and obtain the stability evaluation results of the temperature control parameters. We can adaptively adjust the weights of the stability evaluation results to obtain the optimal temperature control parameter combination for each stage. For example, in the stage with large temperature fluctuations, we can increase the weight of temperature stability; in the stage with high energy consumption, we can increase the weight of energy consumption optimization. Assume that in the first The optimization objective function of each stage is: ; in Respectively represent By dynamically adjusting these weights, the optimal temperature control parameter combination is found in each stage. So as to formulate the best temperature control strategy. According to the best temperature control parameter combination in each stage , generating an intelligent temperature control strategy for the solder paste mixing process. This strategy can dynamically adjust the temperature control parameters at different mixing stages so that the temperature of the solder paste is always kept within the optimal range, thereby improving the uniformity and quality of mixing while reducing energy consumption and production costs.

[0042] In a specific embodiment, the execution step performs a time series analysis on the comprehensive feature representation to obtain multiple stage division results of the solder paste stirring process, and constructs a temperature distribution probability model based on the stage division results and the comprehensive feature representation. The process of obtaining the multi-source probability distribution function can specifically include the following steps: The comprehensive feature representation is transformed by wavelet packet to obtain a multi-scale time-frequency feature representation, and the dynamic time warping distance matrix is ​​calculated based on the multi-scale time-frequency feature representation to obtain the similarity measurement result of the stirring process; Perform spectral clustering on the similarity measurement results to obtain preliminary stage division results, and perform segmented statistical analysis on the comprehensive feature representation based on the preliminary stage division results to obtain the feature statistics of each stage; A Gaussian mixture model is constructed based on characteristic statistics to obtain the temperature distribution probability density function of each stage, and an adaptive kernel density estimation is performed on the temperature distribution probability density function to obtain a non-parametric probability distribution representation; A multidimensional joint probability distribution model is constructed based on the non-parametric probability distribution representation, and importance sampling is performed on the multidimensional joint probability distribution model to obtain a multi-source probability distribution function.

[0043] Specifically, for comprehensive features Perform wavelet packet transform to obtain multi-scale time-frequency features. Assume that a time series signal in the comprehensive feature representation is , through wavelet packet decomposition, we can get the coefficient matrix at different scales and frequencies ,in Indicates scale, Represents the frequency band. The dynamic time warping distance matrix is ​​calculated based on the multi-scale time-frequency feature representation. Dynamic time warping allows nonlinear deformation on the time axis to minimize the distance between two time series, which is suitable for comparative analysis of different stages in the mixing process. Assume that the two time-frequency feature sequences are and , corresponding to two different stirring stages, the distance of dynamic time warping is defined as: ; in, represents the time-aligned path, It is the local distance measure between two time-frequency feature points. By minimizing the cumulative distance , construct a similarity measurement matrix between different stirring stages. Using this matrix, the stirring process is divided into different stages, such as heating, stable stirring and cooling stages. Spectral clustering is performed on the similarity measurement results, and the stirring process is divided into stages through the spectral clustering algorithm. Spectral clustering is a clustering method based on graph theory. It regards the similarity matrix as the adjacency matrix of the graph, and discovers the clustering structure in the data by calculating the Laplacian matrix of the graph and performing eigenvalue decomposition on it. Assume that the similarity matrix is , the normalized Laplacian matrix of the computation graph: ; in, is the degree matrix, is the identity matrix. By changing the Laplace matrix Perform eigenvalue decomposition to obtain a set of eigenvectors, revealing the clustering structure in the similarity matrix. Apply the k-means algorithm to cluster the eigenvectors and obtain preliminary mixing stage division results. Each stage corresponds to a specific state in the mixing process, such as preheating, uniform mixing, and cooling. Based on the stage division results, perform segmented statistical analysis on the comprehensive feature representation in each stage and extract the feature statistics of each stage. Assume that in the first The comprehensive features are expressed as , calculate the characteristic mean, variance, kurtosis and other statistics in this stage as the characteristic description of this stage. Based on the statistics, a Gaussian mixture model is constructed to describe the temperature distribution probability density of each stage. The Gaussian mixture model describes complex probability distributions through the linear combination of multiple Gaussian distributions. Assume that in the The temperature distribution probability density function is: ; in, is the mixing coefficient, and They are The mean and covariance matrix of the Gaussian distribution are obtained. These parameters are estimated through the EM algorithm (expectation maximization algorithm) and the Gaussian mixture model of each stage is obtained. Adaptive kernel density estimation is performed on the basis of the Gaussian mixture model to construct a non-parametric probability distribution representation. Kernel density estimation obtains the probability density distribution of the entire data set by placing a kernel function on each data point and smoothing it. Assume that The temperature data for each stage is , the expression of kernel density estimation is: ; in, is a kernel function (such as a Gaussian kernel function), is the bandwidth parameter, which determines the degree of smoothing. , capturing the detailed characteristics of the temperature distribution and obtaining a more accurate probability density estimate. The non-parametric probability distribution representations of each stage are combined to construct a multi-dimensional joint probability distribution model. This model can describe the joint distribution relationship of the comprehensive characteristics of different stages. Assume that the temperature distribution probability of multiple stages is represented as , by combining these distributions, we construct a multi-source probability distribution function In order to efficiently extract representative samples from the joint distribution, importance sampling techniques are applied. Importance sampling works by sampling from an easily sampled distribution Sampling and based on the target distribution Adjust the weights of the samples to estimate the characteristics of the target distribution. The sampling weights are expressed as: ; in, is the sampling weight, is the multi-source probability distribution function, It is an easy-to-sample distribution. In this way, samples are effectively extracted from the joint distribution to obtain a multi-source probability distribution function that reflects the feature distribution of different stages.

[0044] The above describes the solder paste stirring intelligent temperature control method based on the temperature sensor in the embodiment of the present application. The following describes the solder paste stirring intelligent temperature control system 10 based on the temperature sensor in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the solder paste stirring intelligent temperature control system 10 based on a temperature sensor includes: The acquisition module 11 is used to pre-process the temperature data collected by multiple temperature sensors in the solder paste stirring device to obtain a structured temperature data set; An extraction module 12 is used to input the structured temperature data set into an improved pyramid visual transformer for multi-scale feature extraction to obtain a temperature feature set; A prediction module 13 is used to perform regression prediction on the temperature feature set using a polarized self-attention module to obtain temperature change features; The detection module 14 is used to detect and locate temperature anomalies based on temperature change characteristics using an efficient phase sensing module to obtain a temperature anomaly score map and an abnormal area mask; An integration module 15 is used to integrate the temperature change features and the abnormal area mask based on the cross-layer fusion and propagation module to obtain a comprehensive feature representation; The optimization module 16 is used to perform differentiated optimization of the temperature control parameters in stages according to the comprehensive feature representation to obtain an intelligent temperature control strategy for the solder paste stirring process.

[0045] Through the collaboration of the above components, the improved pyramid visual transformer is used to extract multi-scale features, which can fully capture the changing characteristics of temperature data at different spatiotemporal scales, improve the recognition accuracy of temperature distribution patterns and the sensitivity of the system to small temperature changes. The polarized self-attention module is used to predict temperature changes, which effectively integrates amplitude and phase information, enhances the model's understanding of dynamic temperature changes, and improves the accuracy and robustness of prediction. Based on the efficient phase perception module, temperature anomaly detection and positioning can quickly identify and locate temperature anomaly areas, improving the system's response speed and processing accuracy to abnormal situations. The cross-layer fusion and propagation module integrates the temperature change features and abnormal area masks, realizes the effective fusion of multi-source information, and improves the comprehensiveness and accuracy of feature representation. The temperature control parameters are optimized in stages according to the comprehensive feature representation, and the control strategy can be adaptively adjusted for different stages of the solder paste stirring process, which improves the accuracy and flexibility of temperature control. A multi-objective optimization model including temperature stability, energy consumption and product quality is constructed to achieve a comprehensive balance of multiple key indicators and improve the overall performance and practicality of the temperature control strategy. Adaptive weight adjustment is performed on the stability evaluation results of temperature control parameters to enhance the system's adaptability to different production conditions and improve the robustness and generalization of the temperature control strategy.

[0046] The present application also provides an electronic device, which includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the temperature sensor-based solder paste stirring intelligent temperature control method in the above-mentioned embodiments.

[0047] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein instructions are stored in the computer, and when the instructions are executed on a computer, the computer executes the steps of the temperature sensor-based solder paste stirring intelligent temperature control method.

[0048] 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.

[0049] 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 application is essentially or the part that contributes to the prior art or all 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 an electronic 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 application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0050] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. 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 application.

Claims

1. An intelligent temperature control method for solder paste stirring based on a temperature sensor, characterized in that: The method comprises: Preprocessing the temperature data collected by multiple temperature sensors in the solder paste stirring device to obtain a structured temperature data set; Inputting the structured temperature data set into an improved pyramid visual transformer to extract multi-scale features to obtain a temperature feature set; Using a polarized self-attention module to perform regression prediction on the temperature feature set to obtain temperature change features; Based on the temperature change characteristics, an efficient phase perception module is used to detect and locate temperature anomalies, and a temperature anomaly score map and an abnormal area mask are obtained; Based on the cross-layer fusion and propagation module, the temperature change feature and the abnormal area mask are integrated to obtain a comprehensive feature representation; According to the comprehensive characteristic representation, the temperature control parameters are optimized in stages and differentiated to obtain an intelligent temperature control strategy for the solder paste stirring process.

2. The intelligent temperature control method for solder paste stirring based on a temperature sensor according to claim 1, characterized in that: The temperature data collected by multiple temperature sensors in the solder paste stirring device are preprocessed to obtain a structured temperature data set, including: Performing median filtering on the original temperature data collected by the multiple temperature sensors to obtain filtered data with outliers and transient noise removed; Performing linear interpolation on missing values ​​in the filtered data to obtain a continuous and complete temperature data sequence, and associating the temperature data sequence with corresponding timestamp information to obtain time series temperature data; Performing three-dimensional spatial coordinate mapping on the time series temperature data to obtain a temperature data point set with spatial position information, and performing normalization processing on the temperature data point set to obtain standardized temperature data; Performing time alignment on the standardized temperature data to obtain synchronized temperature data, and performing dimensionality reduction processing on the synchronized temperature data to obtain feature data after dimensionality reduction; Cluster analysis is performed on the feature data after dimension reduction to obtain a structured temperature data set containing temperature distribution pattern labels.

3. The intelligent temperature control method for solder paste stirring based on a temperature sensor according to claim 1, characterized in that: The structured temperature data set is input into the improved pyramid visual transformer for multi-scale feature extraction to obtain a temperature feature set, including: Reorganizing the structured temperature data set into a two-dimensional numerical matrix, wherein rows and columns represent spatial coordinates and matrix element values ​​represent temperatures, to obtain a temperature distribution matrix; Performing multi-scale decomposition on the temperature distribution matrix to obtain a multi-scale temperature representation including different resolution levels, and adaptively adjusting the receptive field of each level of the multi-scale temperature representation to obtain a feature representation with a dynamic perception range; Inputting the feature representation with dynamic perception range into an improved visual transformer encoder to obtain a multi-level global-local feature representation, and fusing the multi-level global-local feature representation with temporal information to obtain a four-dimensional feature tensor containing time dimension information; A multi-scale feature aggregation network is constructed based on the four-dimensional feature tensor to obtain a comprehensive feature representation that integrates information at different scales, and an attention-guided feature selection mechanism is performed on the comprehensive feature representation to obtain refined features that highlight key temperature patterns; The refined features of the salient key temperature patterns are input into a lightweight decoder network for feature processing to obtain a temperature feature set.

4. The intelligent temperature control method for solder paste stirring based on a temperature sensor according to claim 1, characterized in that: The polarized self-attention module is used to perform regression prediction on the temperature feature set to obtain temperature change features, including: Performing feature decomposition on the temperature feature set to obtain amplitude features and phase features, and calculating amplitude attention weights based on the amplitude features to obtain an amplitude attention matrix; Performing phase difference-sensitive attention calculation on the phase feature to obtain a phase attention matrix, and adaptively weighted fusion of the amplitude attention matrix and the phase attention matrix to obtain a comprehensive attention representation; Performing residual connection on the comprehensive attention representation to obtain enhanced features that integrate original temperature feature information, and performing time series modeling on the enhanced features to obtain time series features that capture the time variation law of temperature; The time series features are input into a multi-step prediction head network to predict the temperatures respectively, to obtain a temperature distribution prediction result, and Bayesian reasoning is performed on the temperature distribution prediction result to obtain a temperature change feature.

5. The intelligent temperature control method for solder paste stirring based on a temperature sensor according to claim 1, characterized in that: Based on the temperature change characteristics, the efficient phase perception module is used to detect and locate temperature anomalies to obtain a temperature anomaly score map and an abnormal area mask, including: Performing a multi-scale feature pyramid construction and a phase-sensitive convolution operation on the temperature change feature to obtain a phase-enhanced feature map, and performing a global average pooling and a one-dimensional convolution operation based on the phase-enhanced feature map to obtain a context feature that encodes global temperature distribution information; Performing channel attention fusion on the context feature and the phase enhancement feature map to obtain an enhanced feature representation that fuses global and local information, and performing a depthwise separable convolution operation on the enhanced feature representation to obtain a multi-channel anomaly score map; Performing adaptive threshold segmentation on the multi-channel abnormal score map to obtain a preliminary abnormal region binary mask, and performing morphological opening and closing operations on the preliminary abnormal region binary mask to obtain a denoised and boundary-smoothed abnormal region mask; Connected domain analysis and statistical feature extraction are performed on the abnormal region mask after denoising and boundary smoothing to obtain a temperature anomaly score map and an abnormal region mask.

6. The intelligent temperature control method for solder paste stirring based on a temperature sensor according to claim 5, characterized in that: The temperature change feature is subjected to multi-scale feature pyramid construction and phase-sensitive convolution operation to obtain a phase-enhanced feature map, and global average pooling and one-dimensional convolution operation are performed based on the phase-enhanced feature map to obtain context features encoding global temperature distribution information, including: Performing Gaussian pyramid downsampling on the temperature change feature to obtain a multi-scale feature atlas containing different spatial resolutions, and constructing a Laplacian pyramid based on the multi-scale feature atlas to obtain a differential feature atlas highlighting edge information of different scales; Applying a gradient directional convolution kernel to the differential feature atlas to obtain a directional feature map that captures first-order derivative information of the temperature gradient, and inputting the directional feature map into a Laplace convolution kernel for convolution operation to obtain a curvature feature map that captures second-order derivative information of the temperature field; Performing channel cascading on the directional feature map and the curvature feature map to obtain a multi-channel phase-sensitive feature map, and processing the multi-channel phase-sensitive feature map to obtain a phase-enhanced feature map with a dynamic perception range; Performing a global average pooling operation on each channel of the phase enhancement feature map to obtain a one-dimensional feature vector encoding a global temperature distribution; The one-dimensional feature vector is input into a one-dimensional convolutional network for feature compression and nonlinear transformation to obtain context features that encode global temperature distribution information.

7. The intelligent temperature control method for solder paste stirring based on a temperature sensor according to claim 1, characterized in that: The step of integrating the temperature change feature and the abnormal region mask based on the cross-layer fusion and propagation module to obtain a comprehensive feature representation includes: Applying a channel attention mechanism to the temperature change feature and the abnormal region mask respectively to obtain respective channel importance weights; Performing channel selection on the temperature change feature and the abnormal region mask based on the channel importance weight to obtain a simplified feature representation; Inputting the simplified feature representation into a deformable convolutional network to obtain spatially adaptive alignment features, and applying a spatial attention mechanism to the spatially adaptive alignment features to obtain enhanced features that highlight key area information; Inputting the enhanced features into a recursive refinement unit to obtain feature representations after multiple interactions and refinement, and performing residual connections on the feature representations after multiple interactions and refinement to obtain fused features that retain original information; The fused features are input into the autoencoder for feature compression and reconstruction to obtain low-dimensional feature embedding, and the low-dimensional feature embedding is subjected to multimodal fusion to obtain a comprehensive feature representation.

8. The intelligent temperature control method for solder paste stirring based on a temperature sensor according to claim 1, characterized in that: The temperature control parameters are optimized in stages according to the comprehensive characteristic representation to obtain an intelligent temperature control strategy for the solder paste stirring process, including: Performing a time series analysis on the comprehensive feature representation to obtain multiple stage division results of the solder paste stirring process, and constructing a temperature distribution probability model based on the stage division results and the comprehensive feature representation to obtain a multi-source probability distribution function; Fitting the multi-source probability distribution function based on the Kriging proxy model to obtain a mapping relationship between the temperature control reliability index and the temperature control parameters; Building a multi-objective optimization model based on the mapping relationship to obtain an optimization objective function including temperature stability, energy consumption and product quality; Applying an improved non-dominated sorting genetic algorithm II to the optimization objective function to obtain a Pareto optimal solution set for each stage of the stirring process, and performing a robustness analysis based on the Pareto optimal solution set to obtain a stability evaluation result of the temperature control parameters; The stability evaluation result is adaptively weighted to obtain an optimal temperature control parameter combination for each stage, and an intelligent temperature control strategy for the solder paste stirring process is generated based on the optimal temperature control parameter combination.

9. The intelligent temperature control method for solder paste stirring based on a temperature sensor according to claim 8, characterized in that: The time series analysis of the comprehensive feature representation is performed to obtain multiple stage division results of the solder paste stirring process, and a temperature distribution probability model is constructed based on the stage division results and the comprehensive feature representation to obtain a multi-source probability distribution function, including: Performing wavelet packet transform on the comprehensive feature representation to obtain a multi-scale time-frequency feature representation, and calculating a dynamic time warping distance matrix based on the multi-scale time-frequency feature representation to obtain a similarity measurement result of the stirring process; Performing spectral clustering on the similarity measurement results to obtain preliminary stage division results, and performing segmented statistical analysis on the comprehensive feature representation based on the preliminary stage division results to obtain feature statistics of each stage; A Gaussian mixture model is constructed based on the characteristic statistics to obtain a temperature distribution probability density function at each stage, and an adaptive kernel density estimation is performed on the temperature distribution probability density function to obtain a non-parametric probability distribution representation; A multidimensional joint probability distribution model is constructed based on the non-parametric probability distribution representation, and importance sampling is performed on the multidimensional joint probability distribution model to obtain a multi-source probability distribution function.

10. An intelligent temperature control system for solder paste stirring based on a temperature sensor, characterized in that: Used to execute the intelligent temperature control method for solder paste stirring based on a temperature sensor according to any one of claims 1 to 9, the system comprises: An acquisition module is used to pre-process the temperature data collected by multiple temperature sensors in the solder paste stirring device to obtain a structured temperature data set; An extraction module, used for inputting the structured temperature data set into an improved pyramid visual transformer for multi-scale feature extraction to obtain a temperature feature set; A prediction module, used for performing regression prediction on the temperature feature set using a polarized self-attention module to obtain a temperature change feature; A detection module, used to detect and locate temperature anomalies based on the temperature change characteristics using an efficient phase perception module to obtain a temperature anomaly score map and an abnormal area mask; An integration module, used for integrating the temperature change feature and the abnormal area mask based on the cross-layer fusion and propagation module to obtain a comprehensive feature representation; The optimization module is used to perform differentiated optimization of the temperature control parameters in stages according to the comprehensive feature representation to obtain an intelligent temperature control strategy for the solder paste stirring process.

Images