A multi-station brazing jig and a brazing temperature intelligent control method

Abstract

The present application relates to the field of information technology, and discloses a multi-station brazing jig and a brazing temperature intelligent control method, which comprises the following steps: obtaining a surface ion distribution image of a contact interface between a target workpiece bottom surface and a support member through a collection device, extracting features of the surface ion distribution image by using a feature extraction model to obtain a first pollution feature vector; performing classification processing on the first pollution feature vector by using a classification model to obtain an initial pollution degree classification result of each target node contact interface and a corresponding first electroplating residue distribution density; determining a theoretical heat flow density of each target node within a preset heating time according to an interface thermal resistance value, calculating a difference between the theoretical heat flow density and a preset standard heat flow density to obtain a heat flow density deviation value; obtaining random difference features of each target node, and predicting by using a time series prediction model according to the heat flow density deviation value and the random difference features to obtain a dynamic compensation power of each target node at different heating stages.

Description

Technical Field

[0001] This invention relates to the field of information technology, and more particularly to brazing fixtures, specifically to a multi-station brazing fixture and an intelligent brazing temperature control method. Background Technology

[0002] Multi-station brazing technology is a core process for achieving mass production of complex components, playing a crucial role in structural connections in aerospace, electronic packaging, and automotive manufacturing. This process relies on a fixture to support multiple workpieces and uses heat transfer to melt the brazing filler metal and form a metallurgical bond. The consistency of the weld quality directly determines the reliability of the product and production efficiency. However, existing temperature control solutions generally employ passive strategies such as adding heat spreaders or extending the holding time to alleviate temperature unevenness. This approach cannot respond to thermal resistance drift caused by fixture oxidation, nor can it cope with thermal field disturbances caused by external environmental fluctuations. Ultimately, it still relies on sacrificing process time to achieve near-equilibrium temperature.

[0003] A long-overlooked technical contradiction exists in actual brazing production: before entering the brazing process, the workpieces are usually electroplated to improve the wettability of the brazing filler metal. However, metal ions and acid radicals in the electroplating solution residue adhere to the contact interface between the workpiece and the fixture in the form of liquid films or salt spots. During the heating process, these impurity ions not only form a high thermal resistance layer that hinders heat conduction, but also react chemically with the metal surface, locally altering the emissivity and contact thermal resistance characteristics of the interface.

[0004] More subtly, the types and quantities of residual ions carried by workpieces at different workstations vary randomly. Even if the temperature field of the fixture itself is uniformly distributed, the actual heat flux density received by the workpieces at each workstation will still produce unpredictable deviations due to inconsistencies in interface thermal resistance. This difference in the degree of interface contamination can lead to cold soldering at some workstations due to insufficient heating, while at others, heat accumulation can cause coarsening of the base material grains or overflow of the solder. Traditional temperature control logic only focuses on adjusting the temperature distribution of the fixture body, completely lacking the ability to perceive the ion contamination state of the workpiece surface, and also lacking means to collaboratively eliminate the interference of interface contamination on heat conduction during the brazing process. Therefore, how to accurately detect the degree of ion contamination on the bottom surface of the workpieces at each workstation before heating and implement targeted purification, while dynamically adjusting the differentiated heat compensation power according to the measured temperature of each workstation throughout the heating process, so that the temperature field is always maintained within the preset uniformity range, has become a key issue in improving the brazing yield and batch consistency of multi-station brazing. Summary of the Invention

[0005] This invention provides a multi-station brazing fixture and an intelligent brazing temperature control method, mainly including:

[0006] A surface ion distribution image of the contact interface between the bottom surface of the target workpiece and the support component is acquired using a data acquisition device. A feature extraction model is used to extract features from the surface ion distribution image to obtain a first contamination feature vector. Based on the first contamination feature vector, a classification model is used for classification processing to obtain the initial contamination level classification result of each target node contact interface and the corresponding first electroplating residue distribution density. If the initial contamination level classification result is greater than a preset contamination threshold, a surface cleaning operation is performed on the corresponding target node contact interface to acquire a second surface ion distribution image of the cleaned target node contact interface. The feature extraction model is used to extract features from the second surface ion distribution image to obtain a second contamination feature vector. The classification model is used to classify the second contamination feature vector to obtain a second electroplating residue distribution density. If the initial contamination level classification result is not greater than the preset contamination threshold, the first contamination feature vector is determined to be less than or equal to the first contamination threshold. The pollution feature vector is the second pollution feature vector, and the distribution density of the first electroplating residue is determined as the second electroplating residue distribution density. Based on the second pollution feature vector and the second electroplating residue distribution density, the interface thermal resistance value of the contact interface of each target node is calculated. Based on the interface thermal resistance value, the theoretical heat flux density of each target node within a preset heating time is determined, and the difference between the theoretical heat flux density and the preset standard heat flux density is calculated to obtain the heat flux density deviation value. The random difference characteristics of each target node are obtained. Based on the heat flux density deviation value and the random difference characteristics, a time-series prediction model is used to predict and obtain the dynamic compensation power of each target node at different heating stages. The real-time temperature data of each target node is obtained. If the difference between the real-time temperature data and the preset temperature threshold is greater than the preset temperature deviation threshold, the energy source output of the corresponding target node is adjusted according to the dynamic compensation power to obtain the final temperature field distribution of the target workpiece. Furthermore, the step of using a feature extraction model to extract features from the surface ion distribution image to obtain a first contamination feature vector includes: processing the surface ion distribution image using a grayscale conversion algorithm to obtain an initial grayscale matrix; calculating the grayscale feature value of each pixel based on the initial grayscale matrix; if the grayscale feature value is greater than a preset threshold, determining that the corresponding pixel belongs to a contamination area, and obtaining a binarized image containing multiple contamination areas; extracting connected components from the binarized image; combining the initial grayscale matrix to obtain the spatial coordinates and grayscale mapping values ​​of pixels within each connected component; and extracting the pixel with the largest grayscale mapping value in each connected component as a local extremum point, obtaining a local extremum set composed of multiple local extremum points; constructing a feature matrix using the spatial coordinates and grayscale mapping values ​​in the local extremum set; and performing dimensionality reduction processing on the feature matrix using a singular value decomposition algorithm to determine the first contamination feature vector.Further, the step of classifying the first contamination feature vector using a classification model to obtain the initial contamination level classification result and the corresponding first electroplating residue distribution density of each target node contact interface includes: dividing the first contamination feature vector using a support vector machine algorithm to obtain the initial contamination level classification result and the first electroplating residue distribution density of each target node contact interface; locating the corresponding region in the surface ion distribution image according to the first electroplating residue distribution density, extracting the flux crystal morphology of the corresponding region, obtaining the boundary contour of the flux crystal morphology, and calculating the contour curvature of the boundary contour; if the contour curvature is greater than a preset curvature threshold, it is determined that there is abnormal accumulation in the boundary contour, obtaining the pixel coordinates of the abnormal accumulation region, and calculating the corresponding density gradient according to the first electroplating residue distribution density of each pixel in the abnormal accumulation region; obtaining the spliced ​​data of the density gradient and the initial contamination level classification result, constructing a contamination fusion matrix, performing linear mapping processing on the contamination fusion matrix, and determining the final contamination state label of the target node contact interface; Furthermore, the step of using the feature extraction model to extract features from the second surface ion distribution image to obtain a second contamination feature vector, and using the classification model to classify the second contamination feature vector to obtain a second electroplating residue distribution density, includes: extracting the gray-level co-occurrence matrix of the spectral reflectance image; calculating texture feature values ​​based on the gray-level co-occurrence matrix; constructing a surface roughness matrix from the texture feature values; concatenating the second contamination feature vector with the surface roughness matrix to obtain a fused feature; using the classification model to classify the fused feature to obtain a second contamination level classification result and a second electroplating residue distribution density; calculating the ratio of the second electroplating residue distribution density to a preset standard residue density to obtain an energy attenuation coefficient; if the energy attenuation coefficient is greater than a preset attenuation threshold, then generating a secondary evaluation instruction for the contact interface of the target node; Furthermore, the step of calculating the interfacial thermal resistance value of each target node contact interface based on the second contamination feature vector and the second electroplating residue distribution density includes: obtaining the roughness and gap thickness of each target node contact interface based on the second contamination feature vector and the second electroplating residue distribution density; obtaining the heat flux density of each target node contact interface by combining the roughness and the gap thickness with the pre-acquired thermal conductivity and contact area of ​​each target node contact interface; and determining the interfacial thermal resistance value of each target node contact interface by using the heat flux density and the pre-established pressure distribution of each target node contact interface.The initial pollution level classification result, the first pollution feature vector, and the first electroplating residue distribution density are obtained. If the initial pollution level classification result is not greater than the preset pollution threshold, the first pollution feature vector is determined to be the second pollution feature vector, the first electroplating residue distribution density is determined to be the second electroplating residue distribution density, and the thermal conduction state of the target node contact interface is updated using the interface thermal resistance value. Further, determining the theoretical heat flux density of each target node within a preset heating time based on the interface thermal resistance value includes: acquiring the interface thermal resistance value and the pre-collected temperature gradient of the contact interface of each target node; constructing a heat capacity matrix of the contact interface of each target node using the interface thermal resistance value, the temperature gradient, and the specific heat capacity value of the contact interface of each target node as matrix elements; extracting the eigenvalues ​​of the heat capacity matrix; calculating the heat conduction of the contact interface of each target node based on the eigenvalues ​​and the preset heating time; performing a weighted summation of the heat conduction and the pre-collected radiation coefficient of the contact interface of each target node to determine the theoretical heat flux density of each target node within the heating time; acquiring a preset standard heat flux density and the deformation parameters of the contact interface of each target node; calculating the initial difference between the theoretical heat flux density and the standard heat flux density; if the initial difference is greater than a preset difference threshold, then processing the deformation parameters through a preset mapping function to generate a compensation coefficient. Furthermore, the step of calculating the difference between the theoretical heat flux density and the preset standard heat flux density to obtain the heat flux density deviation value includes: obtaining the deformation parameters of the contact interface between the preset standard heat flux density and each target node; calculating the initial difference between the theoretical heat flux density and the standard heat flux density; if the initial difference is greater than a preset difference threshold, then processing the deformation parameters through a preset mapping function to generate a compensation coefficient, and using the compensation coefficient to perform a product correction on the initial difference to obtain the heat flux density deviation value; obtaining the workstation environment data and historical power data of each target node, extracting random difference features from the workstation environment data, obtaining the heat flux density data of each target node, calculating the heat flux density deviation value between the heat flux density data and the preset standard heat flux density, and performing sequence splicing based on the random difference features and the heat flux density deviation value, combined with the historical power data, to construct a multi-dimensional time series sequence; Furthermore, the step of using a time-series prediction model to predict the dynamic compensation power of each target node at different heating stages based on the heat flux density deviation value and the random difference characteristics includes: constructing a multi-dimensional time-series sequence by combining the historical power data with the random difference characteristics and the heat flux density deviation value; and using a long short-term memory network to extract features and perform time-series analysis on the multi-dimensional time-series sequence to obtain the predicted power values ​​of each target node at different heating stages.A preset power reference is obtained. If the predicted power value is greater than the power reference, the dynamic compensation power is determined based on the difference between the predicted power value and the power reference. Compensation instructions for each target node are generated based on the dynamic compensation power. The real-time temperature value of the sensor of each target node is obtained. The difference between the real-time temperature value and the preset baseline is calculated. If the difference is greater than the preset deviation limit, the state value of the corresponding target node is extracted based on the difference. Furthermore, the step of adjusting the energy source output of the corresponding target node according to the dynamic compensation power includes: acquiring the real-time temperature value of each target node sensor, calculating the difference between the real-time temperature value and a preset baseline; if the difference is greater than a preset deviation limit, extracting the state value of the corresponding target node based on the difference; constructing a state action matrix based on the state value and a pre-acquired compensation amount; using a reinforcement learning algorithm to iterate the state action matrix to obtain the adjustment amount of the corresponding target node; generating a control command for the energy source based on the adjustment amount; controlling the output of the energy source through the control command; acquiring multi-node temperature data after the output amount is applied to the target workpiece, constructing a final temperature field based on the multi-node temperature data, and updating the temperature distribution state of the target workpiece using the final temperature field; Furthermore, the step of constructing the final temperature field based on the multi-node temperature data includes: acquiring multi-node temperature data after the output is applied to the target workpiece; extracting spatial distribution features and temporal evolution features from the multi-node temperature data; smoothing the multi-node temperature data using an interpolation algorithm based on the spatial distribution features and the temporal evolution features to obtain a continuous temperature distribution surface; extracting the extreme points and gradient rate of change of the temperature distribution surface; constructing a temperature field topology based on the extreme points and the gradient rate of change; comparing the temperature field topology with a preset standard temperature field model to calculate the topological similarity; if the topological similarity is greater than a preset similarity threshold, determining that the temperature field topology meets the processing requirements; generating the final temperature field based on the temperature field topology; and outputting the final temperature field to the control terminal.

[0007] A multi-station brazing fixture includes a base plate, an elastic clamping mechanism disposed on the base plate, and a heat compensation structure. The heat compensation structure is detected and controlled by an intelligent brazing temperature control method.

[0008] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0009] This invention discloses an intelligent temperature field control method for multi-station brazing, addressing the problem of uneven heat conduction caused by electroplating residues at the fixture contact interface during brazing. The method acquires surface ion distribution images of the workpiece bottom surface at the fixture contact interface using an image acquisition device. Feature extraction and classification models are used to identify the degree of contamination and the density of electroplating residue distribution. When contamination exceeds a threshold, laser cleaning is automatically performed and a reassessment is conducted. Based on the contamination feature vector and residue distribution density after cleaning, the thermal resistance values ​​of each station interface are calculated, thereby determining the deviation between the theoretical heat flux density and the standard heat flux density. Combining the random differences between each station, a time-series prediction model is used to dynamically predict the compensation power for different heating stages. Temperature data is monitored in real time, and the heating source output is automatically adjusted according to the deviation. This invention effectively solves the problem of uneven temperature in multi-station brazing through closed-loop control of contamination identification and cleaning, thermal resistance calculation, and dynamic power compensation, achieving precise temperature field distribution control and improving the consistency and stability of brazing quality. Attached Figure Description

[0010] Figure 1 This is a flowchart of a smart brazing temperature control method in this embodiment.

[0011] Figure 2 This is a flowchart illustrating step 3 in this embodiment.

[0012] Figure 3 This is a schematic diagram of the overall structure of the brazing fixture in the first direction of this embodiment.

[0013] Figure 4 This is a schematic diagram of the overall structure of the brazing fixture in the second direction in this embodiment. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] like Figures 1-2 The intelligent brazing temperature control method for a multi-station brazing fixture in this embodiment may specifically include:

[0016] Step S101: Obtain a surface ion distribution image of the interface between the bottom surface of the workpiece to be brazed and the fixture using an image acquisition device, and extract features from the surface ion distribution image using a feature extraction model to obtain a first contamination feature vector.

[0017] An image of the surface ion distribution at the interface between the bottom surface of the workpiece to be brazed and the fixture is acquired using an image acquisition device. In one embodiment, a grayscale conversion algorithm is used to process the surface ion distribution image to obtain an initial grayscale matrix. The grayscale feature value of each pixel is calculated based on the initial grayscale matrix. If the grayscale feature value is greater than a preset threshold, the corresponding pixel is determined to belong to a contaminated area, resulting in a binarized image containing multiple contaminated areas. Connected components are extracted from the binarized image. The spatial coordinates and grayscale mapping values ​​of pixels within each connected component are obtained by combining the initial grayscale matrix. The pixel with the largest grayscale mapping value in each connected component is extracted as a local extremum point, resulting in a local extremum set composed of multiple local extremum points. The spatial coordinates and grayscale mapping values ​​in the local extremum set are converted into a feature matrix. A singular value decomposition algorithm is used to reduce the dimensionality of the feature matrix to determine the first contamination feature vector. In another implementation, a convolutional neural network feature extraction model based on a deep residual network is used to extract features from the surface ion distribution image and directly output the first pollution feature vector. This model has learned the spatial distribution characteristics and gray-scale variation patterns of the pollution area through pre-training.

[0018] In one embodiment, an image of the surface ion distribution at the interface between the bottom surface of the workpiece to be brazed and the fixture is acquired using an image acquisition device.

[0019] Understandably, during the brazing process, the interface between the workpiece bottom surface and the fixture is prone to adhering to microscopic contaminants such as oxides and residual processing fluid. These contaminants manifest as abnormal aggregation of specific ions at the microscopic level. Therefore, obtaining an image of the surface ion distribution at this interface is a fundamental prerequisite for achieving high-precision contamination detection.

[0020] Specifically, the aforementioned image acquisition device can employ a laser-induced breakdown spectroscopy imaging apparatus. During the acquisition of the surface ion distribution image, this apparatus emits a high-energy pulsed laser beam towards the contact interface between the bottom surface of the workpiece to be brazed and the fixture. The laser beam is focused on the surface layer of the contact interface, causing trace amounts of material on the surface to vaporize instantaneously and form plasma. As the plasma cools, atoms and ions in the excited state transition back to the ground state, emitting characteristic spectra with specific wavelengths. The spectrometer inside the image acquisition device receives these characteristic spectral signals and, based on the wavelength and intensity information of the spectra, analyzes the types and relative concentrations of target ions such as carbon ions, oxygen ions, and sodium ions on the contact interface. Subsequently, by controlling the laser beam to perform a two-dimensional scan of the contact interface, the ion concentration data at each scan point are spatially mapped, thereby constructing a surface ion distribution image reflecting the spatial distribution of the microscopic chemical composition of the contact interface. This imaging method can non-destructively and intuitively present the microscopic ion contamination state that conventional optical cameras cannot capture. After obtaining the surface ion distribution image, in order to eliminate image noise caused by ambient light interference and dark current of the device, it is usually necessary to perform conventional image preprocessing operations such as filtering and smoothing on the image to improve the signal-to-noise ratio of the image, thereby providing a high-quality data source for subsequent feature extraction.

[0021] For example, a feature extraction model is used to extract features from the surface ion distribution image to obtain a first contamination feature vector. This feature extraction model can be a convolutional neural network model based on a deep residual network architecture. This model contains multiple cascaded convolutional layers, pooling layers, and residual connection modules. During feature extraction, the surface ion distribution image is input into the model. Shallow convolutional kernels first capture changes in ion concentration gradients in the image, extracting shallow texture features such as the edge contours of ion aggregation regions and local concentration abrupt changes. As the network depth increases, deep convolutional kernels further nonlinearly combine and abstract these shallow features to extract deep semantic features that characterize specific contamination patterns, such as a continuous high-concentration carbon ion distribution pattern corresponding to large-area oil residue, or a speckled aggregation pattern of oxygen ions corresponding to deep oxidation. The residual connection module effectively avoids feature loss during multi-layer transmission. Finally, the fully connected layers of the model reduce and fuse these multi-dimensional spatial and semantic features, outputting a fixed-dimensional one-dimensional numerical sequence, i.e., the first contamination feature vector.

[0022] It should be noted that the obtained first contamination feature vector highly condenses key information such as the type, distribution pattern, and degree of contaminants on the interface between the bottom surface of the workpiece to be brazed and the fixture. In this way, the complex two-dimensional ion distribution image is transformed into structured feature data that is easy for computers to process efficiently, providing reliable data support for subsequent accurate identification of contamination types and assessment of brazing quality risks.

[0023] Step S102: Based on the first pollution feature vector, a classification model is used for classification processing to obtain the initial pollution degree classification results of each workstation contact interface and the corresponding first electroplating residue distribution density.

[0024] The first contamination feature vector is divided using a support vector machine algorithm to obtain the initial contamination level classification results and the first electroplating residue distribution density of each workstation contact interface. The classification results are categorized into four levels: slight, moderate, moderate, and severe contamination. Based on the first electroplating residue distribution density, candidate aggregation regions with a density higher than 0.6 are located in the surface ion distribution image. The Canny operator is used to extract the boundary contours of electroplating residue patches in the candidate aggregation regions, and the contour curvature of the boundary contours is calculated. If the contour curvature is greater than a preset curvature threshold of 0.35, it is determined that there is abnormal accumulation in the boundary contour. The pixel coordinates of the abnormal accumulation region are obtained, and the corresponding density gradient g is calculated based on the first electroplating residue distribution density of each pixel in the abnormal accumulation region, where g = (d1 - d2) / L, d1 and d2 are the residue density values ​​of adjacent pixels, and L is the pixel spacing. The density gradient and the initial pollution level classification result are concatenated row-wise to construct a pollution fusion matrix M. The pollution fusion matrix is ​​then calibrated using a linear mapping y = W*M + b, where W is a 2*n dimensional weight matrix and b is a bias vector. W and b are pre-fitted using historical labeled samples within a least-squares framework. The output dimension is aligned to the 1-4 level range of the initial pollution level classification result. The calibrated classification level and its corresponding first electroplating residue distribution density are used as the final output of this step for subsequent steps.

[0025] In one embodiment, based on the structured feature data obtained in the aforementioned steps, a classification model is used to classify the data according to the first contamination feature vector, thereby obtaining the initial contamination level classification results of each workstation contact interface and the corresponding first electroplating residue distribution density.

[0026] Understandably, the first pollution feature vector contains spatial and semantic information about the microscopic chemical composition of the contact interface, providing a basic input for subsequent classification and quantitative assessment.

[0027] Specifically, the aforementioned classification model can employ a deep neural network model based on a multi-task learning architecture. This model primarily consists of three core modules: a shared representation layer, a discrete classification branch, and a continuous regression branch. During the classification process, the first pollution feature vector is input to the shared representation layer. This layer uses multiple fully connected networks to perform nonlinear mapping and feature reconstruction on the input feature vector, extracting common high-order feature representations suitable for multiple task objectives. Subsequently, these common high-order features are simultaneously passed to the discrete classification branch and the continuous regression branch. The discrete classification branch mainly consists of a multilayer perceptron and a normalized exponential function layer, and its function is to perform pattern recognition and probability calculation on the overall pollution state of the contact interface. The continuous regression branch contains a series of linear mapping layers, specifically used for numerical fitting of quantitative indicators for specific pollutants. Through this multi-branch parallel network architecture, the classification model can simultaneously handle qualitative classification and quantitative regression tasks. Further, the discrete classification branch processes the common high-order features and outputs the initial pollution level classification results for each workstation's contact interface. Here, the initial pollution level classification results refer to dividing the overall pollution state of the contact interface into multiple preset level categories.

[0028] For example, based on the range and concentration gradient of abnormal ion aggregation, the degree of contamination can be classified into four levels: no contamination, light contamination, moderate contamination, and heavy contamination. Light contamination typically corresponds to sporadic ion spots at the contact interface, while heavy contamination is characterized by a large-area continuous distribution of high ion concentration. The classification model calculates the posterior probability of the input features belonging to each level category and selects the category with the highest probability value as the initial contamination level classification result for the contact interface at that workstation.

[0029] It should be noted that in the actual production process of brazing, in addition to understanding the overall contamination level of the contact interface, accurate quantitative data on specific types of contaminants are also of crucial guiding significance for subsequent process parameter adjustments.

[0030] For example, the continuous regression branch is used to output the corresponding first electroplating residue distribution density. The first electroplating residue distribution density refers to the degree of aggregation of specific metal ions left over from the previous electroplating process within a unit area at the contact interface between the bottom surface of the workpiece to be brazed and the fixture. During the processing of the workpiece, trace amounts of electroplating solution components, such as nickel or copper ions, often remain on the surface. These electroplating residues are highly susceptible to adverse metallurgical reactions with the solder during high-temperature brazing, leading to incomplete soldering or porosity defects. The first electroplating residue distribution density objectively reflects the aggregation pattern of electroplating residues by quantifying the spatial density and average concentration of these specific ions at the contact interface. The higher the density value, the higher the enrichment degree of electroplating residues in the local area.

[0031] In one embodiment, after receiving the shared high-order features from the shared representation layer, the continuous regression branch performs weight allocation and numerical calculation on the dimensions related to electroplating residues in the features through an internal linear mapping layer. This branch transforms the implicit representation in the high-dimensional feature space into a one-dimensional continuous numerical output, thereby predicting the distribution density of the first electroplating residue on the current workstation interface. This process achieves a direct mapping from abstract ion distribution features to specific physical quantitative indicators.

[0032] For example, in a multi-station continuous brazing production line, the contamination status of the contact interfaces of the fixtures at each station varies due to differences in usage frequency and environment. By inputting the first contamination feature vectors collected and extracted from each station into the aforementioned classification model, the model can simultaneously output the initial contamination level classification result and the first electroplating residue distribution density for each independent station, thereby achieving refined monitoring of the fixture status at each station of the entire production line.

[0033] Preferably, to ensure the accuracy of the classification model's output, the model needs to be jointly trained before practical application. During training, a sample dataset containing a large number of historical first-pollution feature vectors, along with corresponding manually labeled pollution levels and measured residue density values, is used. The cross-entropy loss of the classification branch and the mean squared error loss of the regression branch are calculated, and these two are combined into a total loss function according to a preset weight ratio. The gradient descent algorithm is then used to iteratively update the network parameters within the model until convergence, thereby enabling the classification model to possess accurate classification and density prediction capabilities.

[0034] Step S103: If the initial contamination level classification result is greater than the preset contamination threshold, then perform laser cleaning operation on the corresponding workstation contact interface, obtain the second surface ion distribution image of the corresponding workstation contact interface after cleaning, use the feature extraction model to extract features from the second surface ion distribution image to obtain the second contamination feature vector, use the classification model to classify the second contamination feature vector to obtain the second electroplating residue distribution density.

[0035] If the initial contamination level classification result is greater than the preset contamination threshold of 0.6, a laser cleaning operation is performed on the corresponding workstation contact interface, and a second surface ion distribution image of the corresponding workstation contact interface after cleaning is acquired. A preset feature extraction model is used to extract features from the second surface ion distribution image to obtain a second contamination feature vector. The classification model is then used to classify the second contamination feature vector to obtain a second contamination level classification result and a second electroplating residue distribution density. If the second contamination level classification result is still greater than the contamination threshold of 0.6, the corresponding workstation contact interface is marked as needing repeated cleaning, and the laser power is increased by 10% from the original set value before cleaning is performed again until the second contamination level classification result is no higher than the contamination threshold. The ratio of the second electroplating residue distribution density to the preset standard residue density is calculated to obtain the residue proportion coefficient k, k = d2 / d0, where d2 is the second electroplating residue distribution density and d0 is the standard residue density. If k is greater than the preset residue threshold of 0.15, a secondary evaluation instruction is generated for the workstation contact interface. The secondary evaluation command is sent to the on-site testing terminal, triggering the contact resistance measurement and appearance re-inspection process of the contact interface of the workstation. The test results are sent back to the process management platform, and the release, rework or scrapping is decided based on the returned results, forming a closed-loop record of cleaning and evaluation.

[0036] In one implementation, after obtaining the initial contamination level classification results of the contact interfaces at each workstation, the system compares them with a preset contamination threshold. This contamination threshold is a baseline level limit set based on historical brazing quality data and process specifications.

[0037] For example, moderate contamination can be set as the critical threshold for triggering cleaning. When the initial contamination level classification result of a certain workstation reaches or exceeds the preset contamination threshold, the system determines that the current state of the contact interface can no longer meet the process requirements of high-quality brazing, and then automatically generates a cleaning command and sends it to the corresponding actuator.

[0038] Specifically, if the initial contamination level classification result is greater than a preset contamination threshold, a laser cleaning operation is performed on the corresponding workstation contact interface. Laser cleaning is a non-contact surface treatment technology. Its core lies in using a high-energy laser beam to irradiate the fixture contact interface, causing the surface deposits to absorb laser energy and undergo instantaneous thermal expansion and vaporization or peeling. In actual brazing production lines, after receiving a cleaning command, the laser cleaning equipment automatically adjusts the laser output power, pulse frequency, and scanning speed according to the initial contamination level classification result. For heavily contaminated areas, the equipment employs a high-power, high-overlap laser scanning strategy to ensure that deeply enriched metal ions and oxides are thoroughly removed. For areas with slightly excessive contamination, gentler cleaning parameters are used to remove contaminants while avoiding thermal damage to the fixture substrate. This precise energy delivery mechanism effectively removes stubborn electroplating solution components left over from previous processes. Furthermore, during the laser cleaning process, an auxiliary air blowing device can be simultaneously activated, using high-pressure inert gas to quickly blow away the peeled-off microparticles from the contact interface, preventing secondary adhesion. The entire cleaning process requires no chemical solvents, thus avoiding interference from chemical residues on the brazing metallurgical reaction and meeting the environmental requirements of green manufacturing. Furthermore, after completing the laser cleaning operation, the contact interface is scanned using optical detection equipment to obtain an image of the second surface ion distribution at the corresponding workstation contact interface after cleaning.

[0039] It should be noted that after acquiring the second surface ion distribution image, the aforementioned feature extraction model is used to extract features from the second surface ion distribution image to obtain the second contamination feature vector. The feature extraction model here is consistent with the model used in the previous steps, and it contains multiple convolutional and pooling layers. This model extracts high-dimensional structured data reflecting the ion distribution morphology, cluster size, and edge contours after cleaning by capturing spatial features layer by layer from the second surface ion distribution image. This high-dimensional data is flattened and stitched together to finally form the second contamination feature vector, providing a standardized input basis for secondary assessment.

[0040] In one embodiment, the classification model is used to classify the second contamination feature vector to obtain the second electroplating residue distribution density. This classification model also reuses the aforementioned multi-task learning architecture network. The second contamination feature vector is input to the shared representation layer of the classification model, and common high-order features after cleaning are extracted through nonlinear mapping. Subsequently, these features are passed to a continuous regression branch. The linear mapping layer inside the continuous regression branch recalculates the dimensions related to electroplating residues in the cleaned features, outputting the second electroplating residue distribution density. The second electroplating residue distribution density characterizes the degree of residual aggregation of specific metal ions per unit area on the contact interface after laser cleaning intervention. By obtaining this value, the system can accurately quantify the actual decontamination effect of laser cleaning. If the second electroplating residue distribution density is significantly reduced and within a safe range, it indicates that the cleaning operation is successful, and the fixture can continue to the next round of brazing. If the density value is still too high, it may be necessary to trigger secondary cleaning or manual intervention inspection.

[0041] For example, in a multi-station continuous brazing production scenario, the above-described process, from determining the contamination detection threshold to laser cleaning and secondary density prediction, forms a complete closed-loop control mechanism. For each independent station, the system can not only identify potential contamination risks and take timely physical removal measures, but also strictly control the interface state after cleaning through secondary feature extraction and model regression. This data-driven dynamic intervention strategy ensures that every contact interface entering the brazing heating stage has a high degree of cleanliness and consistency.

[0042] Step S104: If the initial pollution level classification result is not greater than the preset pollution threshold, then the first pollution feature vector is determined as the second pollution feature vector, the first electroplating residue distribution density is determined as the second electroplating residue distribution density, and the interface thermal resistance value of each workstation contact interface is calculated based on the second pollution feature vector and the second electroplating residue distribution density.

[0043] The first contamination feature vector is constructed by scanning the contact interfaces of each workstation with an infrared spectrometer to extract the absorption peak intensity, distribution uniformity, and coverage of contaminant components. Simultaneously, the first electroplating residue distribution density is obtained by measuring the number of electroplating residue particles per unit area using a laser confocal microscope. The first contamination feature vector is input into a pre-trained convolutional neural network model, which outputs an initial contamination level classification score from 0 to 10. If this score is not greater than the contamination threshold of 5.0, the first contamination feature vector is determined as the second contamination feature vector, and the first electroplating residue distribution density is determined as the second electroplating residue distribution density. Based on the coverage value in the second contamination feature vector, a preset surface roughness reference table is consulted. When the coverage is 10% to 30%, the corresponding roughness Ra is 0.8 μm to 1.6 μm; when the coverage is 30% to 60%, the corresponding roughness Ra is 1.6 μm to 3.2 μm. The gap thickness of each workstation contact interface is calculated by dividing the second electroplating residue distribution density by the bulk density of the residue material, which is 7.8 g / cm³. The temperature difference ΔT between the upper and lower surfaces of the contact interface at each workstation is collected at the operating temperature. Combined with the thermal conductivity λ (corrected for roughness) and the contact area A, the heat flux density q is obtained by dividing q = λ × A × ΔT by the gap thickness d. A pressure sensor array is used to measure the pressure distribution P at each workstation contact interface in real time. When the pressure value is between 1 MPa and 3 MPa, the interface thermal resistance R is equal to ΔT divided by q and multiplied by a correction factor of 0.85; when the pressure value is between 3 MPa and 6 MPa, the correction factor is 0.72; and when the pressure value is greater than 6 MPa, the correction factor is 0.60. This segmented correction method is used to determine the interface thermal resistance value of each workstation contact interface.

[0044] In one embodiment, after obtaining the initial pollution level classification result, the system compares it with a preset pollution threshold.

[0045] It is understood that the preset contamination threshold represents the upper limit of permissible slight contamination in the electroplating process. If the initial contamination level classification result is not greater than the preset contamination threshold, it indicates that the contamination situation at the current workstation is within a controllable range and has not caused serious damage to the physical morphology of the contact interface. Therefore, there is no need to perform complex nonlinear compensation or correction on the extracted features. Based on the above judgment result, the system directly determines the first contamination feature vector as the second contamination feature vector and determines the first electroplating residue distribution density as the second electroplating residue distribution density.

[0046] Specifically, the second contamination feature vector contains information on the physical properties of the contaminants, such as key parameters like the material's thermal conductivity, average thickness, and surface roughness. The second electroplating residue distribution density reflects the spatial aggregation of residues at the contact interface, i.e., the proportion of residue covering a unit area. These two parameters together form the basis for subsequent thermal analysis. Further, based on the second contamination feature vector and the second electroplating residue distribution density, the interfacial thermal resistance of each workstation's contact interface is calculated.

[0047] For example, interfacial thermal resistance refers to the resistance encountered by heat when flowing through two contacting solid surfaces. In the calculation process, the system first extracts the contaminant thickness *d* and thermal conductivity *k* from the second contamination feature vector, and obtains the basic thermal resistance value *R0* by dividing *d* by *k*. Subsequently, the system calculates the contact area attenuation coefficient *α* based on the distribution density of the second electroplating residue. This coefficient ranges from 0.3 to 1.0; the higher the distribution density, the smaller the value of *α*, reflecting the degree of reduction in the actual effective contact area. The system multiplies the basic thermal resistance value *R0* by the attenuation coefficient *α* to obtain the corrected thermal resistance value. Due to the discrete distribution of residues leading to a heat flow contraction effect, the system further superimposes the additional thermal resistance *Rg* from the micro-gap. This value is determined by the surface roughness parameter, ranging from 0.0001 to 0.001 square meters Kelvin per watt. Finally, the interfacial thermal resistance *R* is equal to the sum of the corrected thermal resistance value and the additional thermal resistance, i.e., *R* equals *R0* multiplied by *α* plus *Rg*, thus obtaining the accurate interfacial thermal resistance value for each contact interface.

[0048] It should be noted that the above calculation process establishes an accurate mathematical relationship between microscopic contaminant characteristics and macroscopic thermal conductivity. This approach, even with low levels of contamination, ensures the accuracy of thermal resistance calculations while avoiding excessive system resource consumption due to over-calculation, providing reliable data support for subsequent optimization of electroplating process parameters or equipment heat dissipation design.

[0049] Step S105: Based on the interface thermal resistance value, determine the theoretical heat flux density of each station within the preset heating time, calculate the difference between the theoretical heat flux density and the preset standard heat flux density, and obtain the heat flux density deviation value.

[0050] The interface thermal resistance value and the pre-collected temperature gradient of each workstation contact interface are obtained. The product of the interface thermal resistance and the temperature gradient is used as the thermal conductivity coefficient, which is then multiplied by the specific heat capacity value of each workstation contact interface to construct the thermal capacity matrix of each workstation contact interface. This matrix is ​​an n-dimensional diagonal matrix, where n is the number of workstations, and the i-th diagonal element represents the thermal capacity characteristic parameter of the i-th workstation. The eigenvalue λ of the thermal capacity matrix is ​​extracted, and the eigenvalue is multiplied by the preset heating time t to obtain the cumulative heat transfer capacity of each workstation. This cumulative heat transfer capacity is then divided by the contact area A and the heating time t to calculate the thermal flux density q1 = λ / A of each workstation contact interface. Deformation parameters of the contact interfaces at each workstation are acquired beforehand. These parameters characterize the actual contact rate of the contact surface. The thermal conductivity heat flux density q1 and the radiative heat flux density q2 of the contact interfaces at each workstation are weighted and summed with weights of 0.7 and 0.3, respectively, to determine the theoretical heat flux density q0 = 0.7 × q1 + 0.3 × q2 for each workstation during the heating time. A preset standard heat flux density of 5000 W / m² is obtained, and the initial difference Δq between the theoretical heat flux density q0 and the standard heat flux density is calculated. If the initial difference is greater than a preset difference threshold of 500 W / m², the deformation parameter ε is processed by a preset linear mapping function to generate a compensation coefficient k = 1 + 0.5 × ε. The initial difference Δq is corrected by multiplying the compensation coefficient k, resulting in a heat flux density deviation value of k × Δq.

[0051] Furthermore, after obtaining the interfacial thermal resistance values ​​of the contact interfaces at each workstation, the system will conduct a dynamic assessment of the heat conduction state based on this thermal resistance data. This process aims to transform static physical resistance parameters into dynamic thermodynamic indicators, thereby determining the theoretical heat flux density of each workstation within a preset heating time.

[0052] Specifically, theoretical heat flux density characterizes the amount of heat transferred per unit time through a unit contact interface area under specific heating conditions. In the heating stage of the electroplating process, heat is transferred from the heating source to the electroplated part through the contact interface. When calculating the theoretical heat flux density, the system first obtains the output power of the heating source at the current station and the set target heating temperature, and collects the initial ambient temperature of the electroplated part on the other side of the contact interface. Subsequently, the system calculates the temperature difference between the target heating temperature and the initial ambient temperature. Based on the evolution model of Fourier's law of thermal conductivity, the system divides this temperature difference by the previously calculated interface thermal resistance value to obtain the transient heat flux rate. On this basis, the system introduces a preset heating time as an integral variable in the time dimension, calculates the total amount of heat transferred within this time period, and then derives the average heat flux density per unit area, i.e., the theoretical heat flux density.

[0053] It should be noted that the preset heating time is not an arbitrarily set constant, but rather a standard heating cycle strictly determined according to the process specifications of the current electroplating procedure. The required preset heating time varies in different electroplating stages, such as pretreatment degreasing heating or electroplating solution isothermal heating. This time parameter determines whether the heat conduction process can reach a steady state, directly affecting the boundary conditions for calculating the theoretical heat flux density.

[0054] In one embodiment, the system internally stores a preset standard heat flux density. This preset standard heat flux density is a baseline heat flux density value obtained through experimental calibration or simulation modeling under ideal, pollution-free, absolutely smooth contact interfaces and no electroplating residues.

[0055] For example, after calculating the theoretical heat flux density, the system executes deviation comparison logic to calculate the difference between the theoretical heat flux density and the preset standard heat flux density, obtaining the heat flux density deviation value. In actual business processing, the system extracts the preset standard heat flux density as the minuend and uses the theoretical heat flux density calculated at the current workstation as the subtrahend for difference calculation. Due to the unavoidable presence of minor contamination features and electroplating residues at the actual workstation, which increases the interfacial thermal resistance, the theoretical heat flux density is usually lower than the preset standard heat flux density. The heat flux density deviation value obtained through this difference calculation objectively quantifies the extent of the decrease in heat conduction capacity at the current contact interface due to changes in physical morphology.

[0056] Understandably, the heat flux density deviation value is a scalar data point that includes both direction and magnitude. The magnitude of this value directly maps to the physical distance between the current heat conduction state and the ideal state, constituting a core data node reflecting the microscopic thermal changes in the electroplating process.

[0057] Step S106: Obtain the random difference characteristics of each workstation. Based on the heat flux density deviation value and the random difference characteristics, use a time-series prediction model to predict the dynamic compensation power of each workstation at different heating stages.

[0058] The process involves acquiring environmental data and historical power data for each workstation, extracting random variation features from the environmental data, obtaining heat flux density data for each workstation, and calculating the heat flux density deviation between the heat flux density data and a preset standard heat flux density. Based on the random variation features and the heat flux density deviation, and combined with the historical power data, a multi-dimensional time-series sequence is constructed. A Long Short-Term Memory (LSTM) network is used to extract features and perform time-series analysis on the multi-dimensional time-series sequence to obtain predicted power values ​​for each workstation at different heating stages. A preset power benchmark is obtained. If the predicted power value is greater than the power benchmark, a dynamic compensation power is determined based on the difference between the predicted power value and the power benchmark. Compensation instructions for each workstation are then generated based on the dynamic compensation power.

[0059] Based on the random difference characteristics, heat flux density deviation values, and historical power data obtained from S105, the three types of data are time-aligned according to a unified sampling period, for example, one sampling point every 5 seconds, and grouped by workstation number to form a three-channel vector sequence with timestamp as the primary key, serving as a multi-dimensional time series sequence. This sequence is divided by a sliding window, for example, with a window length of 60 sampling points and a step size of 10 sampling points, and input into a Long Short-Term Memory (LSTM) network. The input gate, forget gate, and output gate extract the time-series dependencies of different heating stages such as preheating, isothermal control, and heating up. The output layer provides the predicted power value P_pre for each workstation at the next moment. A preset power benchmark P_std is taken. If P_pre > P_std, the difference ΔP = P_pre - P_std is calculated, and a compensation coefficient k is introduced, with a value ranging from 0.8 to 1.2, which is dynamically adjusted according to the current heating stage to obtain the dynamic compensation power P_comp = k * ΔP. The compensation command is packaged with the workstation number and the effective time. For example, if a workstation has P_pre of 1200 watts, P_std of 1000 watts, and k of 1.0, the issued compensation command will add 200 watts of power output to that workstation at the specified time, thereby achieving real-time adjustment of the heating power.

[0060] In one implementation, for a multi-station rapid thermal processing equipment for semiconductor wafers, in order to achieve precise and coordinated temperature control at each station, the system first needs to acquire the random difference characteristics of each station. These characteristics objectively reflect the non-systematic physical state changes generated during long-term high-intensity operation of the equipment, which are important latent factors leading to inconsistent temperatures at multiple stations.

[0061] Specifically, the random difference characteristics refer to the nonlinear characteristics caused by hardware micro-differences and environmental disturbances during the heating process of wafers from the same or different batches at each workstation.

[0062] For example, random difference features may include the local aging decay rate of heating elements, the processing thickness tolerance of workstation trays, and the minute turbulence disturbance values ​​of the circulating liquid flow in the electroplating tank. The system collects operational status data in real time through a multimodal sensor array deployed at each workstation, and combines it with the equipment's historical maintenance records to perform dimensionality reduction processing on the high-dimensional status data, thereby extracting random difference feature vectors that can characterize the unique physical state of the current workstation. Based on the random difference features obtained in the above steps, the system further synchronizes the heat flux density deviation value calculated in the previous step with the random difference features in the time dimension. Through this data fusion operation, a joint feature sequence containing multidimensional status information is constructed, providing a complete data foundation for subsequent predictive analysis. In a preferred embodiment, the time-series prediction model is constructed using a long short-term memory network. The model receives the joint feature sequence as input, where the input vector X(t) contains the heat flux density deviation value Δq(t) at time t and the random difference feature vector F(t), and the model output is the predicted future power demand P(t+Δt). The model learns the temporal correlation between heat flux density deviation and random difference characteristics within a historical time window through training, and establishes a nonlinear mapping relationship between input features and future power demand.

[0063] It should be noted that the wafer heating process is typically divided into several distinct heating stages, such as a rapid heating stage, a steady-state isothermal stage, and a controlled cooling stage. The influence weights of the heat flux density variation gradient and stochastic variability characteristics differ significantly across these stages. By introducing an attention mechanism, the time-series prediction model can adaptively identify the current heating stage and dynamically adjust the focus on historical sequence data, thereby predicting the power variation trend of each workstation over multiple consecutive time steps in the future.

[0064] For example, based on the output of the time-series prediction model, the dynamic compensation power of each plating station during different heating stages such as heating up, constant temperature, and cooling is obtained. The dynamic compensation power refers to the transient electrical power value injected into or reduced by the heating unit of each station beyond the preset basic heating power formula to offset the temperature drift of the plating solution caused by heat flux density deviation and random differences. Its value typically ranges from -15% to +15% of the basic power, with a single adjustment step not exceeding 3% of the basic power. The adjustment cycle is consistent with the prediction step size, and is set to 500ms. After the above multi-feature fusion and time-series prediction processing, the temperature deviation of the plating solution at each station can be controlled within ±0.5℃, the uniformity deviation of the plating thickness on the surface of the plated parts can be controlled within 5%, and the appearance defect rate such as pinholes and pits in the plating layer is correspondingly reduced.

[0065] Step S107: Obtain real-time temperature data for each workstation. If the difference between the real-time temperature data and the preset temperature threshold is greater than the preset temperature deviation threshold, adjust the output of the heating source for the corresponding workstation according to the dynamic compensation power to obtain the final temperature field distribution of the multi-workstation brazing.

[0066] Real-time temperature values ​​from sensors at each workstation are acquired, and the temperature deviation ΔT between the real-time temperature value and a preset temperature threshold is calculated. If the absolute value of the temperature deviation ΔT is greater than the temperature deviation threshold of 5℃, the state value of the corresponding workstation is extracted based on the temperature deviation ΔT. The state value is divided into 5 levels according to the temperature deviation range: 5℃ to 10℃ is state 1, 10℃ to 15℃ is state 2, 15℃ to 20℃ is state 3, 20℃ to 25℃ is state 4, and greater than 25℃ is state 5. A state-action matrix is ​​constructed based on the state values ​​and the compensation amount obtained through training with historical brazing data. Each state in the state-action matrix corresponds to a different heating power adjustment action. A deep Q-network algorithm is used to iterate the strategy of the state-action matrix. The reward function R is set as the weighted difference between the decrease in temperature deviation and the increase in energy consumption. The state transition probability is obtained by statistically analyzing the temperature response data during actual brazing. After iterative training, the optimal adjustment amount for the corresponding workstation is obtained. The adjustment amount is converted into a power adjustment percentage of the heating source according to a linear mapping relationship, and a control command for the heating source is generated. The output power of the heating source is controlled by the control command. Multi-node temperature data is acquired after the output power is applied to the brazed component. The multi-nodes are temperature measurement points set every 50 mm along the length of the brazed component, totaling 20 nodes. The final temperature field distribution is constructed based on the multi-node temperature data using Kriging interpolation.

[0067] During the operation of the multi-station brazing system, the control center continuously acquires real-time temperature data for each station using high-frequency infrared temperature probes deployed at each station, with a sampling rate of 50Hz. The system compares the acquired real-time temperature data with a preset temperature threshold, setting the standard melting temperature threshold for the current brazing process to 650.5℃. Subsequently, the system calculates the absolute difference between the real-time temperature data and this preset temperature threshold, and compares it with a preset temperature deviation threshold, set at 2.0℃. When the real-time temperature at a station reaches 653.5℃, the absolute difference of 3.0℃ is significantly greater than the 2.0℃ temperature deviation threshold, and the system immediately triggers a closed-loop control mechanism. The control unit multiplies this absolute difference by the station's thermal inertia coefficient and then adds the coupling heat flux correction from adjacent stations to obtain the dynamic compensation power for the corresponding station; in this case, the result is 120.5W. The compensation power is input as a setpoint to the PID control module described above. The module uses a discrete incremental approach, with the output u = Kp*e + Ki*∑e + Kd*Δe, where e is the deviation between the compensation power and the current output power, and Kp, Ki, and Kd are set to 0.35, 0.08, and 0.02, respectively. The engineering tuning is based on the Ziegler-Nichols method. Subsequently, u is converted into a duty cycle increment according to a linear mapping relationship, that is, every 10W power change corresponds to a 0.375% duty cycle adjustment. Based on this, the duty cycle of the solid-state relay pulse width modulation signal is precisely increased from 60.0% to 64.5%, thereby changing the output current of the induction heating coil. After completing the synchronous adjustment of the output of the heating source at each station, the system collects the updated discrete temperature node data, uses the Kriging space interpolation algorithm described above to complete the gridding process, calculates the estimated temperature values ​​of the unmeasured points, and finally reconstructs the continuous three-dimensional temperature field distribution of the entire processing area.

[0068] Reference Figures 3-4 This embodiment provides a multi-station brazing fixture, including a base plate 100, an elastic clamping mechanism 200 disposed on the base plate, and a heat compensation structure 300. The heat compensation structure is detected and controlled using an intelligent brazing temperature control method.

[0069] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for intelligent control of brazing temperature, characterized in that, include: A surface ion distribution image of the contact interface between the bottom surface of the target workpiece and the support component is acquired using a data acquisition device. A feature extraction model is used to extract features from the surface ion distribution image to obtain a first contamination feature vector. Based on the first contamination feature vector, a classification model is used for classification processing to obtain the initial contamination level classification result and the corresponding first electroplating residue distribution density of each target node contact interface. If the initial contamination level classification result is greater than a preset contamination threshold, a surface cleaning operation is performed on the corresponding target node contact interface to acquire a second surface ion distribution image of the cleaned target node contact interface. The feature extraction model is used to extract features from the second surface ion distribution image to obtain a second contamination feature vector. The classification model is used to classify the second contamination feature vector to obtain a second electroplating residue distribution density. If the initial contamination level classification result is not greater than the preset contamination threshold, the first contamination feature vector is determined as the second contamination feature vector, and the first electroplating residue distribution density is determined as the second electroplating residue distribution density. Based on the second contamination feature vector and the second electroplating residue distribution density, the interface thermal resistance value of each target node contact interface is calculated. Based on the interface thermal resistance value, the theoretical heat flux density of each target node within a preset heating time is determined, and the difference between the theoretical heat flux density and the preset standard heat flux density is calculated to obtain the heat flux density deviation value. Random difference characteristics of each target node are acquired, and based on the heat flux density deviation value and the random difference characteristics, a time-series prediction model is used to predict the dynamic compensation power of each target node at different heating stages. Real-time temperature data of each target node is acquired. If the difference between the real-time temperature data and a preset temperature threshold is greater than a preset temperature deviation threshold, the energy source output of the corresponding target node is adjusted according to the dynamic compensation power to obtain the final temperature field distribution of the target workpiece.

2. The method as described in claim 1, characterized in that, The step of using a feature extraction model to extract features from the surface ion distribution image to obtain a first contamination feature vector includes: processing the surface ion distribution image using a grayscale conversion algorithm to obtain an initial grayscale matrix; calculating the grayscale feature value of each pixel based on the initial grayscale matrix, and determining that the corresponding pixel belongs to a contamination area if the grayscale feature value is greater than a preset threshold, thereby obtaining a binarized image containing multiple contamination areas; extracting connected components from the binarized image, obtaining the spatial coordinates and grayscale mapping values ​​of pixels within each connected component by combining the initial grayscale matrix, and extracting the pixel with the largest grayscale mapping value in each connected component as a local extremum point, thereby obtaining a local extremum set composed of multiple local extremum points; constructing a feature matrix using the spatial coordinates and grayscale mapping values ​​in the local extremum set, and performing dimensionality reduction processing on the feature matrix using a singular value decomposition algorithm to determine the first contamination feature vector.

3. The method as described in claim 1, characterized in that, The step of classifying the first contamination feature vector using a classification model to obtain the initial contamination level classification result and the corresponding first electroplating residue distribution density of the contact interface of each target node includes: dividing the first contamination feature vector using a support vector machine algorithm to obtain the initial contamination level classification result and the first electroplating residue distribution density of the contact interface of each target node; locating the corresponding region in the surface ion distribution image according to the first electroplating residue distribution density, extracting the flux crystal morphology of the corresponding region, obtaining the boundary contour of the flux crystal morphology, and calculating the contour curvature of the boundary contour; if the contour curvature is greater than a preset curvature threshold, determining that there is abnormal accumulation in the boundary contour, obtaining the pixel coordinates of the abnormal accumulation region, and calculating the corresponding density gradient according to the first electroplating residue distribution density of each pixel in the abnormal accumulation region; obtaining the concatenated data of the density gradient and the initial contamination level classification result, constructing a contamination fusion matrix, performing linear mapping processing on the contamination fusion matrix, and determining the final contamination state label of the contact interface of the target node.

4. The method as described in claim 1, characterized in that, The process of extracting features from the second surface ion distribution image using the feature extraction model to obtain a second contamination feature vector, and classifying the second contamination feature vector using the classification model to obtain a second electroplating residue distribution density, includes: extracting the gray-level co-occurrence matrix of the spectral reflectance image, calculating texture feature values ​​based on the gray-level co-occurrence matrix, and constructing a surface roughness matrix from the texture feature values; concatenating the second contamination feature vector with the surface roughness matrix to obtain a fused feature, classifying the fused feature using the classification model to obtain a second contamination level classification result and a second electroplating residue distribution density; calculating the ratio of the second electroplating residue distribution density to a preset standard residue density to obtain an energy attenuation coefficient, and if the energy attenuation coefficient is greater than a preset attenuation threshold, generating a secondary evaluation instruction for the contact interface of the target node.

5. The method as described in claim 1, characterized in that, The step of calculating the interfacial thermal resistance value of each target node contact interface based on the second contamination feature vector and the second electroplating residue distribution density includes: obtaining the roughness and gap thickness of each target node contact interface based on the second contamination feature vector and the second electroplating residue distribution density; obtaining the heat flux density of each target node contact interface by combining the roughness and the gap thickness with the pre-acquired thermal conductivity and contact area of ​​each target node contact interface; determining the interfacial thermal resistance value of each target node contact interface using the heat flux density and the pre-established pressure distribution of each target node contact interface; obtaining the initial contamination degree classification result, the first contamination feature vector, and the first electroplating residue distribution density; if the initial contamination degree classification result is not greater than a preset contamination threshold, then determining the first contamination feature vector as the second contamination feature vector, determining the first electroplating residue distribution density as the second electroplating residue distribution density, and updating the thermal conduction state of the target node contact interface using the interfacial thermal resistance value.

6. The method as described in claim 1, characterized in that, The step of determining the theoretical heat flux density of each target node within a preset heating time based on the interface thermal resistance value includes: acquiring the interface thermal resistance value and the pre-collected temperature gradient of the contact interface of each target node; constructing a heat capacity matrix of the contact interface of each target node using the interface thermal resistance value, the temperature gradient, and the specific heat capacity value of the contact interface of each target node as matrix elements; extracting the eigenvalues ​​of the heat capacity matrix; calculating the heat conduction of the contact interface of each target node based on the eigenvalues ​​and the preset heating time; weighting and summing the heat conduction and the pre-collected radiation coefficient of the contact interface of each target node to determine the theoretical heat flux density of each target node within the heating time; acquiring a preset standard heat flux density and the deformation parameters of the contact interface of each target node; calculating the initial difference between the theoretical heat flux density and the standard heat flux density; if the initial difference is greater than a preset difference threshold, processing the deformation parameters through a preset mapping function to generate a compensation coefficient.

7. The method as described in claim 1, characterized in that, The calculation of the difference between the theoretical heat flux density and the preset standard heat flux density to obtain the heat flux density deviation value includes: obtaining the deformation parameters of the interface between the preset standard heat flux density and each target node, and calculating the initial difference between the theoretical heat flux density and the standard heat flux density; if the initial difference is greater than a preset difference threshold, then processing the deformation parameters through a preset mapping function to generate a compensation coefficient, and using the compensation coefficient to multiply and correct the initial difference to obtain the heat flux density deviation value; obtaining the workstation environment data and historical power data of each target node, extracting the random difference features in the workstation environment data, obtaining the heat flux density data of each target node, calculating the heat flux density deviation value between the heat flux density data and the preset standard heat flux density, and constructing a multi-dimensional time series sequence by combining the random difference features and the heat flux density deviation value with the historical power data.

8. The method as described in claim 1, characterized in that, The step of predicting the dynamic compensation power of each target node at different heating stages using a time-series prediction model based on the heat flux density deviation value and the random difference characteristics includes: constructing a multi-dimensional time-series sequence by concatenating the historical power data with the random difference characteristics and the heat flux density deviation value; performing feature extraction and time-series analysis on the multi-dimensional time-series sequence using a long short-term memory network to obtain the predicted power value of each target node at different heating stages; obtaining a preset power benchmark, and if the predicted power value is greater than the power benchmark, determining the dynamic compensation power based on the difference between the predicted power value and the power benchmark, and generating compensation instructions for each target node based on the dynamic compensation power; obtaining the real-time temperature value of the sensor of each target node, calculating the difference between the real-time temperature value and the preset benchmark, and if the difference is greater than a preset deviation limit, extracting the state value of the corresponding target node based on the difference.

9. The method as described in claim 1, characterized in that, The step of adjusting the energy source output of the corresponding target node according to the dynamic compensation power includes: acquiring the real-time temperature value of each target node sensor and calculating the difference between the real-time temperature value and a preset baseline; if the difference is greater than a preset deviation limit, extracting the state value of the corresponding target node based on the difference; constructing a state-action matrix based on the state value and a pre-acquired compensation amount; using a reinforcement learning algorithm to iterate the strategy of the state-action matrix to obtain the adjustment amount of the corresponding target node; generating a control command for the energy source based on the adjustment amount; controlling the output of the energy source through the control command; acquiring multi-node temperature data after the output amount is applied to the target workpiece, constructing a final temperature field based on the multi-node temperature data, and extracting the spatial distribution features and temporal evolution features from the multi-node temperature data. Based on the spatial distribution characteristics and the temporal evolution characteristics, an interpolation algorithm is used to smooth the multi-node temperature data to obtain a continuous temperature distribution surface. The extreme points and gradient rates of change of the temperature distribution surface are extracted, and a temperature field topology is constructed based on these extreme points and gradient rates of change. The temperature field topology is compared with a preset standard temperature field model to calculate the topological similarity. If the topological similarity is greater than a preset similarity threshold, the temperature field topology is determined to meet the processing requirements. The final temperature field is then generated based on the temperature field topology and output to the control terminal. The temperature distribution state of the target workpiece is updated using the final temperature field.

10. A multi-station brazing fixture, characterized in that, It includes a base plate, an elastic clamping mechanism disposed on the base plate, and a heat compensation structure, wherein the heat compensation structure is detected and controlled by an intelligent brazing temperature control method according to any one of claims 1-9.

Images