A method and system for optimizing a pressing process of a composite floor

Abstract

The present application relates to the technical field of composite floor manufacturing process optimization, in particular to a kind of composite floor pressing process optimization method and system, comprising: by collecting multidimensional process data in production line, including temperature, coating image, moisture content and pressure data.Characteristic analysis is carried out respectively, and temperature uniformity atlas, coating quality vector and moisture content dynamic distribution field are generated.These features are spatially aligned and fused to form a panoramic feature set.A pre-trained neural network model is called to identify abnormalities and analyze parameter sensitivity, generating an influence factor matrix, and then outputting a process parameter adjustment scheme through reverse optimization.The method realizes accurate perception and dynamic control of the coupling effect of multiple physical fields in the pressing process, effectively improving the uniformity, stability and yield of the pressing quality.

Description

Technical Field

[0001] This invention relates to the field of composite flooring manufacturing process optimization technology, and in particular to a method and system for optimizing the composite flooring pressing process. Background Technology

[0002] In the production of laminated flooring, precise control of process parameters directly determines the product's adhesive strength, dimensional stability, and surface quality. Current technologies commonly employ methods such as monitoring key process parameters at single points or by average value. For example, this involves installing several thermocouples at different locations on the press to monitor temperature, randomly inspecting the adhesive patch area after the coating process, or sampling and testing the moisture content of the substrate upon its entry into the production line. These parameters are typically recorded and evaluated independently, and process windows are set based on experience.

[0003] Existing technical solutions have shortcomings. Isolated parameter monitoring cannot reveal the true uniformity of the heat field distribution on the press's hot plate, the actual coverage texture and continuity of the adhesive on the board surface, or the spatial gradient of the substrate's internal moisture content. The complex coupling relationship between temperature, pressure, moisture content, and adhesive application state during the pressing process is simplified into the control of a few independent variables, resulting in insufficient basis for process adjustment and difficulty in solving quality defects such as warping and delamination caused by local overheating, uneven adhesive distribution, or uneven moisture migration. Treating the substrate moisture content as a static threshold indicator and only focusing on whether its average value meets the requirements ignores the dynamic process of moisture migration and escape during hot pressing and its fundamental impact on the adhesive curing rate and the formation of internal stress in the board.

[0004] A method is needed that can integrate multi-dimensional information on the thermal field, adhesive field, and moisture field during the lamination process and characterize their spatiotemporal dynamic interaction. Simultaneously, moisture content must be transformed from a static indicator into a dynamic field variable that describes its changes over time and space throughout the entire board material, thus providing a more fundamental and precise input for process optimization. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an optimized method and system for composite flooring pressing process.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an optimized method for composite flooring pressing process, comprising: Collect multi-dimensional process data in the composite flooring pressing production line. The multi-dimensional process data includes press temperature timing data, adhesive coating image data, substrate moisture content measurement data, and pressure distribution matrix data. The press temperature time series data is processed by regional thermal field feature analysis to generate a temperature uniformity feature map. The adhesive coating image data is processed by texture and coverage analysis to generate a coating quality evaluation vector. The substrate moisture content measurement data is processed by spatiotemporal interpolation and gradient calculation to generate a dynamic moisture content distribution field. The temperature uniformity feature map, the coating quality evaluation vector and the moisture content dynamic distribution field are spatially aligned and feature fused to generate a panoramic feature set of composite bonding conditions. The pre-trained process neural network model is invoked to perform abnormal pattern recognition and key parameter sensitivity analysis on the panoramic feature set of the composite bonding process, and a process parameter influence factor matrix is ​​generated. Based on the process parameter influence factor matrix, the multidimensional process data is subjected to reverse parameter optimization processing to generate an optimized pressing process parameter adjustment scheme.

[0007] As a further aspect of the present invention, the step of performing regional thermal field feature analysis processing on the compressor temperature time series data to generate a temperature uniformity feature map includes: The press temperature time series data is sliced ​​according to the physical partition of the heating plate to obtain independent temperature sequences of multiple heating zones; The independent temperature sequence of each heating region is processed by extracting the steady-state holding segment and calculating the volatility to generate a temperature stability index for the heating region. The synchronization difference calculation is performed on the independent temperature sequences of adjacent heating regions to generate the temperature synchronization difference coefficient between regions. By combining the temperature stability index and the temperature synchronization difference coefficient between regions, a two-dimensional feature matrix is ​​constructed to characterize the overall thermal field uniformity of the compressor. The two-dimensional feature matrix is ​​spatially interpolated and fitted with isotherms using a thermal conduction simulation model, ultimately generating a visualized and machine-readable temperature uniformity feature map.

[0008] As a further aspect of the present invention, the step of performing texture and coverage analysis on the adhesive coating image data to generate a coating quality evaluation vector includes: The adhesive coating image data is preprocessed, including grayscale conversion, noise filtering, and edge enhancement. An improved local binary pattern algorithm is used to extract texture features from the preprocessed image to generate a texture feature descriptor of the adhesive distribution. An adaptive threshold segmentation algorithm is applied to separate the coated area from the background substrate, and the pixel coverage area and the regularity of the coating contour are calculated. The texture feature descriptor, the pixel coverage area, and the regularity of the coating contour are normalized and concatenated to form a multidimensional coating quality evaluation vector.

[0009] As a further aspect of the present invention, the step of performing spatiotemporal interpolation and gradient calculation processing on the moisture content measurement data of the substrate to generate a dynamic moisture content distribution field includes: Acquire the moisture content measurement data of the substrate at multiple sampling points on the production line at different times; Based on the spatial relationship of the sampling points, the Kriging spatial interpolation algorithm is used to perform spatial surface interpolation on the discrete moisture content data at the same time to generate a spatial distribution map of moisture content at each sampling time. The spatial distribution map of moisture content in the continuous time series is smoothed and differencing in the time dimension to calculate the gradient change of moisture content over time. By fusing spatial distribution information with temporal gradient information, a three-dimensional data field is constructed that can characterize the dynamic migration and distribution of moisture content within the substrate, namely the dynamic moisture content distribution field.

[0010] As a further aspect of the present invention, the step of spatially aligning and fusing the temperature uniformity feature map, the coating quality evaluation vector, and the moisture content dynamic distribution field to generate a panoramic feature set for composite bonding conditions includes: Establish a unified spatial coordinate system based on the surface of the floor substrate; The temperature uniformity feature map and the dynamic distribution field of moisture content are resampled and mapped to the unified spatial coordinate system to ensure that the spatial positions correspond one-to-one. The coating quality evaluation vector is expanded into a spatialized coating quality distribution map with the same resolution as the unified spatial coordinate system based on its corresponding image acquisition location. A multi-source feature fusion network is used to perform feature-level fusion on the mapped temperature uniformity feature map, the spatialized coating quality distribution map, and the mapped dynamic moisture content distribution field. The multi-source feature fusion network calculates the correlation weights between different feature maps through a cross-attention mechanism, performs weighted summation and feature enhancement, and finally outputs a panoramic feature set of the composite bonding condition that includes comprehensive information on temperature, coating, and moisture content.

[0011] As a further aspect of the present invention, the step of calling a pre-trained process neural network model to perform abnormal pattern recognition and key parameter sensitivity analysis on the panoramic feature set of the composite bonding condition, and generating a process parameter influence factor matrix, includes: The panoramic feature set of the composite bonding working condition is input into the feature encoder of the process neural network model. The feature encoder performs dimensionality reduction and abstraction on the input features to extract high-order working condition feature vectors. The high-order operating condition feature vector is input into the anomaly detection branch of the process neural network model. The anomaly detection branch identifies abnormal operating condition feature segments that deviate from the standard pattern by comparing them with a pre-stored standard operating condition feature pattern library. The high-order working condition feature vector is simultaneously input into the sensitivity analysis branch of the process neural network model. The sensitivity analysis branch calculates the gradient magnitude of the change in each input feature dimension on the final bonding quality prediction value through the gradient backpropagation method. By integrating the location information of the abnormal operating condition feature fragments with the gradient magnitude information of each feature dimension, a matrix is ​​constructed with process parameters as rows and quality indicators as columns. The value of each element in the matrix represents the influence weight of a specific process parameter on a specific quality indicator, i.e., the process parameter influence factor matrix. The process parameter influence factor matrix describes the non-linear correlation weights between the bonding quality index and each process parameter. The steps for constructing the process neural network model include: Collect composite lamination working condition data and corresponding lamination quality test results data accumulated in the historical production process to construct a training sample set; Design a multi-task neural network architecture that includes a feature encoder, an anomaly detection branch, and a sensitivity analysis branch; The feature encoder employs a multi-layer convolutional neural network structure to extract high-order features from the input panoramic feature set of composite bonding conditions. The anomaly detection branch adopts an autoencoder structure and identifies abnormal operating conditions by reconstructing errors. The sensitivity analysis branch adopts a fully connected network structure and analyzes parameter sensitivity through forward propagation and gradient calculation; The multi-task neural network is trained end-to-end using the training sample set, and the network weights are optimized using the backpropagation algorithm. During training, a dynamic weight balancing strategy is used to coordinate the loss functions of different task branches; Once the model's performance on the validation set reaches the preset standard, the final network parameters are saved to obtain the pre-trained process neural network model.

[0012] As a further aspect of the present invention, the step of performing reverse parameter optimization processing on the multidimensional process data based on the process parameter influence factor matrix to generate an optimized pressing process parameter adjustment scheme includes: Set the bonding quality target values, which include the expected range of flatness, bonding strength and formaldehyde release; Using the bonding quality target value as a constraint and the process parameter influence factor matrix as a guide, a target loss function is defined; An evolutionary optimization algorithm is used to iteratively optimize the adjustable process parameters in the currently collected multidimensional process data; In each iteration, the exploration step size and direction of different parameters are adjusted according to the process parameter influence factor matrix, with priority given to adjusting parameters with large influence weights; When the target loss function converges or reaches the preset number of iterations, a set of optimized process parameters that makes the predicted quality closest to the target value of the bonding quality is output, which serves as the core of the bonding process parameter adjustment scheme. The pressing process parameter adjustment scheme is used to guide the coordinated adjustment of the press temperature curve, coating rate and pressure distribution.

[0013] As a further aspect of the present invention, the pressing process parameter adjustment scheme is used to guide the coordinated adjustment of the press temperature profile, coating rate, and pressure distribution, specifically including: The optimized process parameter combination is analyzed, and the optimized setpoint sequence for the press temperature is extracted. Based on the optimized setpoint sequence and combined with the physical inertia model of the press heating plate, a smooth and executable press temperature rise and fall curve is generated. The optimized process parameter combination is analyzed to extract the optimized setting value for the adhesive coating rate; Based on the optimized set values ​​and combined with the mechanical response model of the coating roller, a speed control command sequence for the coating motor is generated. The optimized process parameter combination is analyzed to extract the optimized setpoint matrix for pressure distribution; Based on the optimized setpoint matrix and combined with the hydraulic pressure zone control system model, a pressure adjustment command sequence for each pressure zone is generated. The press temperature rise / fall curve, the speed control command sequence, and the pressure adjustment command sequence are synchronized and conflict checked on the time axis to ensure that the press temperature rise / fall curve, the speed control command sequence, and the pressure adjustment command sequence are coordinated without conflict, thus forming a complete pressing process parameter adjustment scheme.

[0014] As a further aspect of the present invention, the method further includes: A real-time control command stream is generated based on the aforementioned bonding process parameter adjustment scheme, and the real-time control command stream is sent to the actuator of the bonding production line to complete the dynamic closed-loop adjustment of the process parameters, specifically including: The time-sequential control commands in the pressing process parameter adjustment scheme are converted into standard data packets under a specific industrial communication protocol. The standard data messages are timestamped and prioritized to generate the real-time control command stream; Establish real-time data links with the press temperature controller, coating motor driver, and hydraulic pressure controller; According to the time sequence and priority of the real-time control command stream, the control commands are sequentially sent to the corresponding actuators through the real-time data link. During execution, new multidimensional process data are continuously collected, fed back, and compared with the expected adjustment target. If the deviation exceeds the threshold, a new round of optimization calculation and instruction generation is triggered, forming a dynamic closed-loop adjustment.

[0015] As a further aspect of the present invention, the present invention also includes a composite flooring pressing process optimization system, the system including a processor and a memory, the memory storing a computer program, the processor executing the computer program to implement the composite flooring pressing process optimization method as described above.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By analyzing the regional thermal field characteristics of press temperature time-series data to generate a temperature uniformity feature map, performing texture and coverage analysis on adhesive coating images to generate a coating quality evaluation vector, and performing spatiotemporal interpolation and gradient calculation on substrate moisture content measurement data to generate a dynamic moisture content distribution field, these three types of heterogeneous features are spatially aligned and fused to construct a panoramic feature set for composite pressing conditions. This process breaks down the barriers between different process parameter data, integrating previously isolated thermal, visual, and humidity information into a digital model that can simultaneously reflect the synergistic state of multiple physical fields—heat, adhesive, and moisture—on the surface and inside of the board. This transforms process condition diagnosis from alarms for single parameter exceedances to assessments of comprehensive operating conditions under multi-field coupling, enabling the identification of complex process anomalies caused by the interplay of multiple factors.

[0017] The substrate moisture content measurement data is processed through spatiotemporal interpolation and gradient calculation to generate a dynamic moisture content distribution field. This method no longer relies on the static average value of a few detection points, but instead constructs a dynamic field that can describe the differences in the spatial distribution of moisture within the board and its migration over time during hot pressing. This distribution field reveals the differences in heat transfer caused by moisture content gradients at different locations within the board and the asynchronous curing of the adhesive. Based on this dynamic distribution field, process control can predict the trend of moisture migration's impact on pressing quality, thereby enabling targeted adjustments to the pressure and temperature curves to balance the internal stress of the board and suppress defects such as blistering and delamination caused by excessive local moisture or poor migration. Attached Figure Description

[0018] Figure 1 This is a flowchart of the optimized method for composite flooring pressing process described in this invention; Figure 2 A flowchart for generating a dynamic moisture content distribution field; Figure 3 Analysis of dimensionality changes in convolutional modules; Figure 4 Simulated isotherm diagrams of the uniformity of temperature distribution in the thermal field of the press; Figure 5 Heat map of the factors affecting the composite flooring pressing process parameters. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] See Figure 1The method for optimizing the composite flooring pressing process involves the following steps: Collecting multi-dimensional process data from the composite flooring pressing production line, including press temperature time-series data, adhesive coating image data, substrate moisture content measurement data, and pressure distribution matrix data; performing regional thermal field feature analysis on the press temperature time-series data to generate a temperature uniformity feature map; performing texture and coverage analysis on the adhesive coating image data to generate a coating quality evaluation vector; performing spatiotemporal interpolation and gradient calculation on the substrate moisture content measurement data to generate a dynamic moisture content distribution field; spatially aligning and fusing the temperature uniformity feature map, the coating quality evaluation vector, and the dynamic moisture content distribution field to generate a panoramic feature set of the composite pressing condition; calling a pre-trained process neural network model to perform abnormal pattern recognition and key parameter sensitivity analysis on the panoramic feature set of the composite pressing condition to generate a process parameter influence factor matrix; and performing reverse parameter optimization on the multi-dimensional process data based on the process parameter influence factor matrix to generate an optimized pressing process parameter adjustment scheme.

[0022] In one embodiment of the present invention, the press temperature time series data originates from thermocouple sensor arrays installed on each physical partition of the press heating plate. These sensors record temperature values ​​at a fixed sampling frequency, forming a time series dataset. In an example scenario, a composite flooring pressing production line is equipped with 16 independently temperature-controlled heating zones, each zone having 3 thermocouple sensors. The collected raw temperature time series data is sliced ​​according to the physical partitions of the heating plate. Sensor data belonging to the same heating zone are grouped into independent temperature sequences for that zone. Data slicing is completed based on sensor numbers and a zone mapping table. Each independent temperature sequence contains all temperature sampling points from the start to the end of the pressing process. Steady-state holding segment extraction and volatility calculation are performed on the independent temperature sequences of each heating zone. Steady-state holding segment extraction is achieved through sliding window variance analysis. When the variance of the data within the window is continuously lower than a set threshold, it is determined to be a steady-state segment. Volatility calculation calculates the standard deviation of the data within the steady-state segment, generating a temperature stability index for that heating zone. The formula for calculating the temperature stability index is: Where: symbol Indicates the first Temperature stability index of each heating zone, symbol Indicates the number of steady-state segments extracted, symbol Indicates the number of sampling points within a single steady-state segment, symbol Indicates the first The heating zone is in the first Within the steady-state segment, the first... Temperature value at each sampling time, symbol Indicates the first The heating zone is in the first The average temperature within a steady-state segment is calculated. Synchronization difference calculation is performed on independent temperature sequences of adjacent heating regions. This calculation uses a dynamic time warping algorithm to calculate the minimum cumulative distance between two sequences on the time axis, and this distance is normalized to serve as the inter-regional temperature synchronization difference coefficient. This coefficient reflects the temporal alignment of temperature changes in different heating regions. In some embodiments, a two-dimensional feature matrix characterizing the overall thermal field uniformity of the compressor is constructed by combining the temperature stability index and the inter-regional temperature synchronization difference coefficient. The rows of the two-dimensional feature matrix correspond to the heating regions, the first column stores the temperature stability index of each region, and the second column stores the average temperature synchronization difference coefficient between that region and its adjacent regions. A heat conduction simulation model is used to spatially interpolate and fit isotherms to the two-dimensional feature matrix. The heat conduction simulation model discretizes the heating plate into a grid based on the finite difference method, using the data in the two-dimensional feature matrix as boundary or initial conditions for iterative calculation to simulate the spatial distribution of the thermal field. Finally, a visualized and machine-readable temperature uniformity feature map is generated. The temperature uniformity feature map is stored in the form of an image matrix, where each pixel value represents the simulated temperature value at that location.

[0023] In practice, adhesive coating image data is acquired immediately after the coating process using a high-resolution industrial camera. The image covers the entire surface of the flooring substrate. Preprocessing of the adhesive coating image data includes grayscale conversion, noise filtering, and edge enhancement. Grayscale conversion uses a weighted average method to convert the color image to grayscale. Noise filtering uses a median filter to remove salt-and-pepper noise from the image. Edge enhancement uses the Laplacian operator to highlight the boundary information of the coating area. An improved local binary pattern algorithm is used to extract texture features from the preprocessed image. This improved algorithm changes the circular neighborhood of the traditional local binary pattern to an ellipse to adapt to the adhesive flow direction. It also uses a rotation-invariant unified pattern to statistically analyze the histogram of the binary patterns within the neighborhood, generating a texture feature descriptor for the adhesive distribution. This texture feature descriptor is a feature vector containing 256 dimensions.

[0024] An adaptive threshold segmentation algorithm is applied to separate the coated area from the background substrate. This algorithm dynamically calculates the segmentation threshold for each pixel based on the local grayscale mean and standard deviation. For the segmented binary image, the pixel coverage area and the regularity of the coating contour are calculated. The pixel coverage area is the ratio of the total number of white pixels in the coated area to the total number of pixels in the image. The regularity of the coating contour is quantified by calculating the Fourier descriptor of the coating contour and taking the variance of its low-order components. Essentially, the texture feature descriptor, pixel coverage area, and coating contour regularity are normalized and concatenated. Normalization uses a min-max scaling method to map each feature value to the zero-to-one range. The concatenation operation connects these normalized values ​​sequentially to form a multi-dimensional coating quality evaluation vector. Optionally, in a data comparison scenario, for a batch of production data, processing one hundred coated images yields one hundred coating quality evaluation vectors. These vectors can be input into subsequent analysis modules for comparison with a standard vector library to identify abnormal coating patterns.

[0025] In some embodiments, the processes of generating temperature uniformity feature maps and coating quality evaluation vectors can be executed in parallel. In terms of computational resource allocation, the temperature data processing thread and the image processing thread run independently, exchanging intermediate results through shared memory. It is understood that the generation of temperature uniformity feature maps depends on the accurate extraction of steady-state segments. In specific implementations, the variance threshold needs to be determined based on the statistical distribution of historical data to avoid misjudging normal fluctuations as non-steady-state conditions. Optionally, in the texture feature extraction step of the coating quality evaluation vector, the neighborhood radius and number of sampling points of the improved local binary mode algorithm can be adjusted according to the image resolution to balance feature discriminative power and computational efficiency. In specific implementations, all generated feature data, including temperature uniformity feature maps and coating quality evaluation vectors, are stored in a real-time database in a structured array format for subsequent feature fusion processing steps.

[0026] In one embodiment of the present invention, see [reference] Figure 2In practice, the substrate moisture content measurement data at multiple sampling points on the production line at different times are acquired. This data is collected via an online near-infrared moisture sensor in the conveyor section before the substrate enters the lamination station. The sensor has five sampling points spaced at fixed intervals along the substrate width. Each sampling point continuously measures and records the moisture content value at a one-second interval. Within one lamination cycle, each sampling point generates a time series containing measurement values ​​from start to finish, forming the original substrate moisture content measurement dataset. Based on the spatial relationship of the sampling points, a Kriging spatial interpolation algorithm is used to perform spatial surface interpolation on the discrete moisture content data at the same time. The Kriging spatial interpolation algorithm uses a semi-variogram model to characterize the spatial correlation between sampling points. For a specific time, the algorithm uses the moisture content measurement data from the five sampling points as known points to predict the moisture content at other locations on the substrate plane, generating a spatial distribution map of the moisture content at each sampling time. This spatial distribution map is stored in a raster matrix format with a spatial resolution of one centimeter. The spatial distribution map of moisture content in a continuous time series is smoothed and differentially processed in the time dimension. The smoothing process employs median filtering within a sliding time window to reduce measurement noise. The differential processing calculates the numerical difference between corresponding grid cells of the spatial distribution map of moisture content at adjacent time stamps, thus obtaining the gradient change of moisture content over time. The gradient change matrix reflects the diffusion or evaporation rate of moisture within the substrate. The spatial distribution information and temporal gradient information are fused to construct a three-dimensional data field that characterizes the dynamic migration and distribution of moisture content within the substrate. The first two dimensions of the three-dimensional data field correspond to the spatial coordinates of the substrate plane, and the third dimension corresponds to the time coordinates. Each data cell stores the moisture content value and its temporal gradient value at that location at the corresponding time. This three-dimensional data field is the dynamic distribution field of moisture content, and the formula for constructing the dynamic distribution field of moisture content is expressed as: Where: symbol Represents spatial coordinates and time The eigenvector of the dynamic distribution field of water content at a given location, with the symbol... This represents the spatial coordinates obtained through Kriging interpolation. and time Moisture content value at the location, symbol This represents the spatial coordinates obtained through time difference calculation. and time The time gradient of moisture content at a given location.

[0027] In the specific implementation, a unified spatial coordinate system is established with the surface of the floor substrate as the reference. The origin of the unified spatial coordinate system is the midpoint of the front end of the substrate when it enters the press, with the length direction as the X-axis and the width direction as the Y-axis. A two-dimensional planar rectangular coordinate system is established, and the resolution of the coordinate system is consistent with the grid resolution of the moisture content spatial distribution map. The temperature uniformity feature map and the moisture content dynamic distribution field are resampled and mapped to the unified spatial coordinate system. The temperature uniformity feature map is derived from the coordinate system of the press heating plate. Through the known spatial position mapping relationship between the heating plate and the substrate surface, an affine transformation is performed to transform its pixel coordinates to the unified spatial coordinate system. For non-integer grid positions that may appear after the coordinate transformation, a bilinear interpolation algorithm is used to resample to fill the data. The moisture content dynamic distribution field itself is generated in the substrate planar coordinate system, but its original timestamp sequence may not be completely synchronized with the timestamp sequence of the temperature data. Therefore, a linear interpolation method on the time axis is used to interpolate the moisture content dynamic distribution field in the time dimension to a timestamp sequence that is completely consistent with the temperature uniformity feature map, ensuring that the spatial position corresponds one-to-one and the time is synchronized.

[0028] In some embodiments, the coating quality evaluation vector is expanded into a spatialized coating quality distribution map with the same resolution as the unified spatial coordinate system based on its corresponding image acquisition location. Each adhesive coating image records its corresponding substrate location interval during acquisition. The various feature values ​​(such as the statistics of texture feature descriptors, pixel coverage area, and regularity) in the calculated coating quality evaluation vector are used as constants and filled into all grid cells corresponding to the substrate location interval covered by the image. For grid cells not completely covered by any coating image, nearest neighbor interpolation is performed using the coating quality evaluation vector data of neighboring images. Finally, a spatialized coating quality distribution map with the same spatial resolution as the unified spatial coordinate system and where each grid cell contains complete coating quality evaluation vector information is generated. A multi-source feature fusion network is used to perform feature-level fusion of the mapped temperature uniformity feature map, the spatialized coating quality distribution map, and the mapped moisture content dynamic distribution field. The multi-source feature fusion network contains three parallel feature encoding branches, processing data from three modalities: temperature, coating, and moisture content, respectively. Each branch consists of several convolutional layers to extract local spatial features.

[0029] The multi-source feature fusion network calculates the association weights between different feature maps through a cross-attention mechanism. This mechanism operates in the form of query-key-value pairs; for example, the temperature feature map is used as the query, and the coating quality distribution feature map is used as the key and value. An attention weight matrix is ​​calculated, indicating the association strength between each location in the temperature field and all locations in the coating quality distribution. Based on this weight, the coating quality feature values ​​are weighted and summed to generate a new feature map fused with the temperature feature map. This process is performed pairwise between the feature maps of the three modalities, with weighted summation and feature enhancement, ultimately outputting a comprehensive feature set of composite lamination conditions containing integrated information on temperature, coating, and moisture content. This comprehensive feature set of composite lamination conditions can be understood as a four-dimensional tensor, where the dimensions represent batch size, number of feature channels, spatial height, and spatial width, respectively. Optionally, in a data comparison scenario, three-modal data corresponding to five consecutive production cycles of a batch are input. After the spatial alignment and feature fusion processing, five corresponding tensors of the comprehensive feature set of composite lamination conditions are output. These tensors can be directly used as input to subsequent process neural network models. In some embodiments, for substrate edge regions not directly measured, extrapolation boundary conditions are used during Kriging space interpolation and subsequent spatial mapping to maintain the spatial integrity of the data field. It is understood that timestamp synchronization is a prerequisite for successful fusion. In specific implementations, the clocks of all sensors and data acquisition devices are precisely synchronized through an industrial network to ensure the comparability of data from different sources on the timeline. Optionally, the multi-source feature fusion network needs to be jointly trained with the subsequent process neural network model during the training phase to learn the most effective feature fusion method for the final quality prediction task.

[0030] In one embodiment of the present invention, in a specific implementation, the composite lamination condition panoramic feature set is fed into the process neural network model as input. The composite lamination condition panoramic feature set is a four-dimensional tensor, whose dimensions represent batch size, number of feature channels, spatial height, and spatial width, respectively. In the example scenario, a batch of input contains composite lamination condition panoramic feature sets corresponding to 8 consecutive production cycles. The spatial resolution of each feature set is 128×128 pixels, and the number of feature channels is 64. The composite lamination condition panoramic feature set is input into the feature encoder of the process neural network model. The feature encoder performs dimensionality reduction and abstraction on the input features, extracting high-order condition feature vectors. The feature encoder consists of 4 concatenated convolutional modules. Each convolutional module contains a convolutional layer, a batch normalization layer, and a ReLU activation function. The stride of the convolutional layer is set to 2 to achieve gradual reduction of spatial size. After processing by the feature encoder, the input 128×128×64 feature map is converted into a 512-dimensional one-dimensional high-order condition feature vector.

[0031] The high-order operating condition feature vector is input into the anomaly detection branch of the process neural network model. The anomaly detection branch identifies abnormal operating condition feature segments that deviate from the standard pattern by comparing them with the pre-stored standard operating condition feature pattern library. The standard operating condition feature pattern library is a set of typical pattern vectors extracted from the high-order operating condition feature vectors of historical high-quality production batches through cluster analysis. The anomaly detection branch calculates the cosine similarity between the input high-order operating condition feature vector and all typical pattern vectors in the pattern library. If the maximum similarity is lower than a preset threshold, the current input feature is determined to be abnormal, and the time position index of the abnormal feature segment in the input batch is recorded.

[0032] The high-order working condition feature vector is simultaneously input into the sensitivity analysis branch of the process neural network model. The sensitivity analysis branch calculates the gradient magnitude of the change in each input feature dimension with respect to the final bonding quality prediction value through the gradient backpropagation method. The sensitivity analysis branch is a multi-layer fully connected network, and its output layer nodes correspond to multiple bonding quality indicators. After the network is forward propagated to obtain the quality prediction value, the partial derivative of the prediction value with respect to each component in the input high-order working condition feature vector is calculated through automatic differentiation. The absolute values ​​of these partial derivatives constitute a gradient vector, which characterizes the sensitivity of each input feature dimension to the quality prediction. By integrating the location information of abnormal operating condition feature fragments with the gradient magnitude information of each feature dimension, a matrix is ​​constructed with process parameters as rows and quality indicators as columns. The value of each element in the matrix represents the influence weight of a specific process parameter on a specific quality indicator, i.e., the process parameter influence factor matrix. The construction process maps the dimensions of the high-order operating condition feature vector back to the original process parameter space. For example, by analyzing the response region of the feature encoder convolution kernel, it is determined which feature dimensions mainly respond to temperature, coating, or moisture content information. The calculated gradient vector values ​​are allocated and accumulated at the intersection of the corresponding process parameter row and quality indicator column according to this mapping relationship. The process parameter influence factor matrix describes the nonlinear correlation weight between the bonding quality indicators and each process parameter, and its mathematical expression is: Where: symbol This represents the process parameter influence factor matrix, with symbols... This represents the mapping matrix from the higher-order feature vector dimensions to the original process parameter categories, with the symbol... The transpose operation of a matrix is ​​represented by the symbol [symbol missing]. This represents the vector of predicted quality values ​​calculated by the sensitivity analysis branch. For high-order operating condition feature vectors The gradient vector, symbol This represents the absolute value operation, with the symbol... This represents matrix multiplication.

[0033] The construction steps of the process neural network model include multiple stages. The process involves collecting composite pressing condition data and corresponding pressing quality test results data accumulated in the historical production process, constructing a training sample set, and the composite pressing condition data is the aforementioned composite pressing condition panoramic feature set. The pressing quality test results data includes flatness, bonding strength and formaldehyde release values ​​measured in the laboratory. Each training sample is a paired data, containing a composite pressing condition panoramic feature set tensor and a corresponding quality test result vector.

[0034] Design a multi-task neural network architecture comprising a feature encoder, an anomaly detection branch, and a sensitivity analysis branch. The feature encoder employs a multi-layer convolutional neural network structure to extract high-order features from the input composite overlay condition panoramic feature set, specifically consisting of four convolutional layers with a stride of 2 and channel numbers of 64, 128, 256, and 512, respectively. The anomaly detection branch uses an autoencoder structure to identify abnormal condition patterns through reconstruction error. The autoencoder consists of an encoder and a decoder, with the encoder connected to a shared feature encoder layer. The decoder attempts to reconstruct the high-order feature representation of the input from the encoded features, and the reconstruction error is used to measure the degree of anomaly of the input features. The sensitivity analysis branch employs a fully connected network structure to analyze parameter sensitivity through forward propagation and gradient calculation. This branch consists of three fully connected layers, with the number of neurons in the last layer equal to the number of quality indicators. The multi-task neural network is trained end-to-end using a training sample set, and the network weights are optimized using a backpropagation algorithm. The training loss function is a weighted sum of the losses of the three branches, including the mean squared error of feature extraction, the reconstruction error of anomaly detection, and the mean squared error of quality prediction.

[0035] During training, a dynamic weight balancing strategy is employed to coordinate the loss functions of different task branches. This strategy automatically adjusts the weight coefficients of each branch in the total loss based on the rate of change of the loss in recent training batches. Once the model's performance on the validation set reaches a preset standard, the final network parameters are saved, resulting in a pre-trained process neural network model. The preset standard refers to a determination coefficient between the prediction results of the quality prediction branch and the actual quality detection results on the independent validation set being no less than 0.9 for 10 consecutive training cycles. In some embodiments, the construction of the standard operating condition feature pattern library is independent of the training process of the process neural network model. After model training is complete, high-order operating condition feature vectors obtained from all training samples through the feature encoder are collected. K-means clustering is used to cluster these vectors into several categories, with the center vector of each category being a standard operating condition feature pattern, which is then stored in the pattern library. It is understood that gradient calculation for the sensitivity analysis branch needs to be performed after each forward propagation. During the model deployment and application phase, automatic differentiation needs to be enabled to calculate the process parameter influence factor matrix in real time. Optionally, the process parameter influence factor matrix can be periodically fine-tuned and updated using new production data to reflect slow changes in process conditions or material properties. In some embodiments, the anomaly detection branch, in addition to outputting anomaly identifiers, can also output similarity vectors between anomaly feature fragments and various patterns in a standard pattern library, allowing operators to further analyze anomaly types. It can be understood that although the feature encoder, anomaly detection branch, and sensitivity analysis branch are architecturally separate, their parameters are jointly optimized during training, making the extracted high-order operating condition feature vectors beneficial for both anomaly detection and sensitivity analysis tasks. Optionally, in data comparison scenarios, a batch of data containing normal and abnormal operating conditions is input into the trained process neural network model. The model can not only output corresponding quality prediction values ​​but also generate a process parameter influence factor matrix for each sample and mark the locations of abnormal samples and their abnormal feature fragments, achieving the integration of pattern recognition and attribution analysis.

[0036] See Figure 3In the feature encoder processing flow of the process neural network model, the dynamic changes in the number of feature channels and the total number of spatial pixels at each stage were visualized. The panoramic feature set of the composite bonding process received by the input layer is a four-dimensional tensor. In the example scene, its initial spatial resolution is 128×128 pixels, and the number of feature channels is 64. As the signal is passed sequentially through the input layer, convolutional modules 1 to 4, and the higher-order feature vector stage, the total number of spatial pixels decreases exponentially: the 16384 pixels (128×128) in the input layer, after downsampling with a stride of 2 in each convolutional module, is reduced to about 256 pixels (8×8) in the convolutional module 4 stage, and finally compressed into 1 point in the higher-order feature vector stage. At the same time, the number of feature channels increases stepwise: from 64 dimensions in the input layer, it remains unchanged through convolutional module 1, increases to 128 dimensions in convolutional module 2, increases to 256 dimensions in convolutional module 3, and reaches 512 dimensions in both convolutional module 4 and the higher-order feature vector stage. This dimensionality-changing strategy achieves spatial dimensionality reduction and channel dimensionality enhancement of features. By downsampling through convolutional layers, redundant spatial details are gradually discarded, while the number of channels is increased to aggregate more abstract high-order operating condition features, providing a compact and discriminative feature representation for subsequent anomaly detection and sensitivity analysis.

[0037] In one embodiment of the present invention, in specific implementation, a bonding quality target value is set, which includes the expected range of flatness, bonding strength, and formaldehyde emission. The flatness target range is set to be no greater than 0.5 mm / m, the bonding strength target range is set to be no less than 2.5 MPa, and the formaldehyde emission target range is set to be no greater than 0.05 mg / m³. These target values ​​serve as constraints that must be met during the optimization process. Using the bonding quality target value as a constraint and guided by the process parameter influence factor matrix, a target loss function is defined. The target loss function quantifies the comprehensive deviation between the quality value predicted by the model under the current process parameters and the bonding quality target value. The function also incorporates elements of the process parameter influence factor matrix as weights, giving higher attention to process parameters that have a greater impact on the final quality indicators during optimization. The formula for the target loss function is expressed as: Where: symbol Indicates a combination of process parameters The target loss function value for the variable, with sign Indexes representing quality indicators, symbols Indicates the first The importance weight of each quality indicator, symbol The neural network model representing the process uses parameters Time The predicted value of each quality indicator, symbol Indicates the first The bonding quality target value for each quality indicator, symbol Represents the regularization coefficient, with the symbol... The indices represent the process parameters and quality indicators, respectively, and the symbols are... The first element in the process parameter influence factor matrix represents the... Line number Column element values, symbols Indicates the first Candidate values ​​for each process parameter, symbol Indicates the first The current collected value of each process parameter.

[0038] An evolutionary optimization algorithm is employed to iteratively optimize the adjustable process parameters in the currently collected multidimensional process data. The algorithm utilizes a covariance matrix adaptive evolution strategy, which maintains a multivariate normal distribution to generate a population of candidate solutions in the parameter space. In each iteration, the exploration step size and direction of different parameters are adjusted based on the process parameter influence factor matrix, prioritizing parameters with high influence weights. Specifically, the norm of the column vector corresponding to the parameter row in the process parameter influence factor matrix is ​​normalized, and this norm is used as a scaling factor for the diagonal elements of the covariance matrix during initialization. This ensures that parameters with high combined influence weights on multiple quality indicators have larger variation steps in the early stages of evolution. When the target loss function converges or reaches the preset number of iterations, a set of optimized process parameter combinations that best approximates the predicted quality to the bonding quality target value is output. The convergence condition is set as the decrease in the target loss function value over 20 consecutive iterations being less than 0.0001. The output optimized process parameter combination serves as the core of the bonding process parameter adjustment scheme.

[0039] In practical implementation, after generating the optimized process parameter combination, it is parsed and converted into executable control commands. The optimized process parameter combination is parsed to extract the optimized setpoint sequence for the press temperature. This sequence includes temperature setpoints for each heating zone of the press at ten key time points during the pressing process. Based on this optimized setpoint sequence, and combined with the physical inertia model of the press heating plate, a smooth and executable press temperature rise / fall curve is generated. The physical inertia model is a first-order hysteresis system model, and its time constant is experimentally determined. When generating the temperature curve, this model is used to smooth and filter the step-change setpoint sequence, ensuring that the rate of change of the temperature command is within the trackable range of the heating plate physical system. The optimized process parameter combination is then parsed to extract the optimized setpoint for the adhesive coating rate. This optimized setpoint is a scalar representing the target linear velocity of the coating roller. Based on the optimized setpoints and combined with the mechanical response model of the coating roller, a speed control command sequence for the coating motor is generated. The mechanical response model describes the dynamic process from the issuance of the speed command to the stabilization of the actual rotational speed of the coating roller. The speed control command sequence is generated with the optimized setpoints as the target, through feedforward and pre-compensation methods to offset the response lag of the mechanical system.

[0040] The process parameter combination is analyzed and optimized, and an optimized setpoint matrix for pressure distribution is extracted. The rows of the optimized setpoint matrix correspond to the pressure zone numbers of the press, and the columns correspond to the stage numbers of the pressing process. The matrix element values ​​are the target pressure values ​​for each zone at each stage. Based on the optimized setpoint matrix and combined with the hydraulic pressure zone control system model, a pressure regulation command sequence for each pressure zone is generated. The hydraulic pressure zone control system model describes the nonlinear relationship between the hydraulic valve core opening and the output pressure. The pressure regulation command sequence converts the target pressure value into the duty cycle control signal of the corresponding hydraulic valve. The press temperature rise / fall curve, speed control command sequence, and pressure regulation command sequence are synchronized and conflict checked on the time axis. Synchronization mapping maps all command sequences to the timeline of a unified high-precision industrial timing controller. Conflict checking verifies whether the commands of different actuators are physically feasible at any given time, such as checking whether the heating power requirement matches the total power load capacity, forming a complete pressing process parameter adjustment scheme. Table 1 shows the core data of an exemplary optimized process parameter combination.

[0041] Table: Optimized Process Parameter Combinations In some embodiments, the population size of the evolutionary optimization algorithm is set to 50, the upper limit of the number of iterations is set to 200, and the internal parameters of the covariance matrix adaptive evolutionary strategy are initialized according to standard methods. It is understood that the physical inertia model of the press heating plate, the mechanical response model of the coating roller, and the hydraulic pressure zoning control system model all need to be pre-acquired through system identification methods during the production equipment commissioning phase and input as known conditions into the optimization system. Optionally, if a conflict is detected in the conflict checking step, the execution of the pressure adjustment command sequence is prioritized, while the temperature rise / fall curve or speed control command sequence is slightly shifted or scaled until the conflict is resolved. In some embodiments, the pressing process parameter adjustment scheme is ultimately encapsulated into a structured data file containing all control commands sorted by timestamp and their corresponding actuator address information. It is understood that in data comparison scenarios, for the same initial working condition, multiple runs of reverse parameter optimization processing may produce slightly different combinations of optimized process parameters due to the randomness of the evolutionary algorithm, but the corresponding predicted quality results all meet the constraints of the pressing quality target value. Optionally, after the bonding process parameter adjustment scheme is generated, it will first be virtually tested in a simulation environment based on the actual equipment model to verify its logical rationality and safety, and then it will be sent to the actual production line.

[0042] See Figure 4 In the thermal field characteristic analysis of the composite flooring pressing process, the temperature uniformity distribution of the substrate surface under the action of the press heating plate is visually presented. The figure uses the physical coordinates of the substrate in the X and Y directions as axes, and constructs a continuous temperature field visualization map through isotherms and pseudo-color mapping. The central region shows a high value of yellow (approximately 187.5℃), gradually transitioning to orange, red, and purple towards the periphery, until a low value (approximately 135.0℃) in the edge region, clearly reflecting the gradient decay law of the thermal field from the center to the edge. This concentric isotherm distribution indicates that the heat conduction characteristics of the press heating plate exhibit a central symmetry pattern, and also reveals the temperature loss effect in the edge region. The generation of the map is based on the regional thermal field characteristic analysis of the press temperature time series data: firstly, the press temperature data is sliced ​​according to the heating plate region, and the steady-state temperature sequence and stability index of each region are extracted. Then, spatial interpolation and isotherm fitting are performed through a heat conduction simulation model, finally obtaining this machine-readable and visualized temperature uniformity characteristic map. It not only provides a spatialized temperature benchmark for subsequent multi-source feature fusion, but also provides direct visual evidence for identifying the typical abnormal condition of "insufficient edge temperature" in the sensitivity analysis of process parameters.

[0043] In one embodiment of the present invention, in a specific implementation, the time-sequential control instructions in the bonding process parameter adjustment scheme are converted into standard data messages under a specific industrial communication protocol. The bonding process parameter adjustment scheme includes a sequence of instructions ordered by timestamps. Each instruction includes the actuator address, instruction type, and target value. In the example scenario, the Modbus TCP / IP protocol is used as the communication standard. The conversion process maps each temperature setting instruction to a message with function code 16 written to the holding register, the speed control instruction to a message with function code 6 written to a single register, and the pressure adjustment instruction to a message with function code 15 written to multiple coils. The payload of the standard data message is filled strictly according to the protocol format, including the slave address, function code, starting address, data quantity, and specific instruction data bytes. Standard data messages are timestamped and prioritized to generate a real-time control command stream. Each standard data message is timestamped with a high-precision timestamp, which indicates the absolute time when the message should be executed or the time offset relative to the start of the batch. Priority is based on the urgency and logical dependencies of the commands. For example, pressure safety interlock commands have the highest priority, followed by temperature adjustment commands, and then speed fine-tuning commands. The sorted commands are encapsulated into an ordered command list, i.e., the real-time control command stream.

[0044] A real-time data link is established with the press temperature controller, coating motor driver, and hydraulic pressure controller. Each actuator is assigned an independent IP address and port within the industrial control network. A persistent TCP connection is established via socket programming. For the press temperature controller, the data link is based on the Modbus TCP protocol; for the coating motor driver, the data link supports both Modbus TCP and CANopen protocols, with protocol conversion handled by a gateway; for the hydraulic pressure controller, the data link uses the real-time Ethernet protocol. Control commands are sequentially sent to the corresponding actuators according to the time sequence and priority of the real-time control command stream via the real-time data link. A central scheduling program reads the real-time control command stream and determines the sending timing based on the command's timestamp and the current system clock. When the system time reaches the command's timestamp, the scheduling program sends a standard data packet through the corresponding real-time data link and waits for an acknowledgment response from the actuator. If no acknowledgment is received or an error response is received within the timeout period, the scheduling program will resend the command or trigger an alarm according to a preset retry strategy.

[0045] During execution, new multi-dimensional process data is continuously collected and fed back, then compared with the expected adjustment target. After the command is issued, the system continuously acquires new press temperature timing data, adhesive coating image data, substrate moisture content measurement data, and pressure distribution matrix data from the sensor network. These new multi-dimensional process data undergo the same processing flow as in the aforementioned embodiments to generate a new panoramic feature set of composite pressing conditions. This set is then input into a pre-trained process neural network model to obtain predicted values ​​of key quality indicators under the current conditions. The predicted values ​​are compared with the expected quality target values ​​in the pressing process parameter adjustment scheme to calculate the comprehensive deviation index. If the deviation exceeds a threshold, a new round of optimization calculation and command generation is triggered, forming a dynamic closed-loop adjustment. The deviation threshold is a preset scalar. When the comprehensive deviation index is greater than this threshold, the system determines that the current process state deviates significantly from the expected target. At this time, a new round of optimization calculation is automatically started. The optimization calculation uses the latest multi-dimensional process data as input and re-executes the entire process from generating the process parameter influence factor matrix to generating a new pressing process parameter adjustment scheme. A new real-time control command stream is generated and sent to the actuator. The formula for calculating the deviation is: Where: symbol Indicates the comprehensive deviation index, symbol Indexes representing quality indicators, symbols The neural network model representing the process is applied to the first... The latest predicted value of each quality indicator, symbol This indicates the setting of the first step in the bonding process parameter adjustment scheme. The expected target values ​​for each quality indicator. In some embodiments, the generation of standard data packets takes into account the bandwidth and frame length limitations of industrial networks; longer instruction sequences are split into multiple packets conforming to the maximum transmission unit size for transmission. It is understood that the priority ordering logic of real-time control instruction flow needs to avoid priority inversion. For example, if a low-priority coating speed instruction is a prerequisite for the execution of a high-priority pressure instruction, its actual scheduling priority needs to be temporarily increased. Optionally, when establishing a real-time data link, a link health check is performed, heartbeat messages are sent periodically to ensure a smooth connection, and automatic reconnection is attempted in the event of a disconnection.

[0046] In some embodiments, the comprehensive deviation index calculation in the feedback comparison stage can assign different weights to different quality indicators, with the weights configured according to product grade requirements. It is understood that the threshold setting for triggering a new round of optimization needs to balance system sensitivity and stability, avoiding frequent re-optimization due to measurement noise or minor fluctuations. Typically, the threshold is determined through historical data analysis. Optionally, in data comparison scenarios, the number of times dynamic closed-loop adjustments are triggered, the reasons (i.e., which quality indicator deviation dominates), and the convergence of the deviation after adjustment are recorded for ten consecutive production batches to evaluate the effectiveness and stability of the closed-loop system. In some embodiments, all issued instructions, received feedback data, and triggered optimization events are recorded in a time-series database for production traceability and process analysis.

[0047] See Figure 5 In the characteristic integration stage of the composite flooring pressing process, the process parameter influence factor matrix clearly quantifies the nonlinear influence weights of three key process parameters—press temperature, coating rate, and hydraulic pressure—on three core quality indicators: flatness, bonding strength, and formaldehyde emission. Specifically, the coating rate has the highest influence weight on bonding strength, reaching 0.89, making it the most critical parameter determining bonding performance; the press temperature has a weight of 0.85 on flatness, making it a core factor in ensuring board surface flatness; and the hydraulic pressure has a weight of 0.82 on bonding strength, serving as an important parameter to assist in improving bonding performance. From the perspective of quality indicators, bonding strength is most significantly affected by the combined influence of all process parameters, with corresponding influence weights all at a high level (0.78~0.89); flatness is mainly dominated by press temperature and coating rate; while formaldehyde emission has the lowest influence weight among all process parameters (0.49~0.62), indicating that this indicator is less sensitive to parameter fluctuations in the current process system. The matrix uses a visualization method that combines color gradients and numerical labels to intuitively present the correlation strength between process parameters and quality indicators, providing accurate quantitative basis for subsequent reverse parameter optimization and process adjustment.

[0048] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An optimized method for composite flooring pressing process, characterized in that, The method includes: Collect multi-dimensional process data in the composite flooring pressing production line. The multi-dimensional process data includes press temperature timing data, adhesive coating image data, substrate moisture content measurement data, and pressure distribution matrix data. The press temperature time series data is processed by regional thermal field feature analysis to generate a temperature uniformity feature map. The adhesive coating image data is processed by texture and coverage analysis to generate a coating quality evaluation vector. The substrate moisture content measurement data is processed by spatiotemporal interpolation and gradient calculation to generate a dynamic moisture content distribution field. The temperature uniformity feature map, the coating quality evaluation vector and the moisture content dynamic distribution field are spatially aligned and feature fused to generate a panoramic feature set of composite bonding conditions. The pre-trained process neural network model is invoked to perform abnormal pattern recognition and key parameter sensitivity analysis on the panoramic feature set of the composite bonding process, and a process parameter influence factor matrix is ​​generated. Based on the process parameter influence factor matrix, the multidimensional process data is subjected to reverse parameter optimization processing to generate an optimized pressing process parameter adjustment scheme.

2. The method for optimizing the composite flooring pressing process according to claim 1, characterized in that, The step of performing regional thermal field feature analysis on the time-series temperature data of the compressor to generate a temperature uniformity feature map includes: The press temperature time series data is sliced ​​according to the physical partition of the heating plate to obtain independent temperature sequences of multiple heating zones; The independent temperature sequence of each heating region is processed by extracting the steady-state holding segment and calculating the volatility to generate a temperature stability index for the heating region. The synchronization difference calculation is performed on the independent temperature sequences of adjacent heating regions to generate the temperature synchronization difference coefficient between regions. By combining the temperature stability index and the temperature synchronization difference coefficient between regions, a two-dimensional feature matrix is ​​constructed to characterize the overall thermal field uniformity of the compressor. The two-dimensional feature matrix is ​​spatially interpolated and fitted with isotherms using a thermal conduction simulation model, ultimately generating a visualized and machine-readable temperature uniformity feature map.

3. The method for optimizing the composite flooring pressing process according to claim 1, characterized in that, The process of performing texture and coverage analysis on the adhesive coating image data to generate a coating quality evaluation vector includes: The adhesive coating image data is preprocessed, including grayscale conversion, noise filtering, and edge enhancement. An improved local binary pattern algorithm is used to extract texture features from the preprocessed image to generate a texture feature descriptor of the adhesive distribution. An adaptive threshold segmentation algorithm is applied to separate the coated area from the background substrate, and the pixel coverage area and the regularity of the coating contour are calculated. The texture feature descriptor, the pixel coverage area, and the regularity of the coating contour are normalized and concatenated to form a multidimensional coating quality evaluation vector.

4. The method for optimizing the composite flooring pressing process according to claim 1, characterized in that, The step of performing spatiotemporal interpolation and gradient calculation on the moisture content measurement data of the substrate to generate a dynamic moisture content distribution field includes: Acquire the moisture content measurement data of the substrate at multiple sampling points on the production line at different times; Based on the spatial relationship of the sampling points, the Kriging spatial interpolation algorithm is used to perform spatial surface interpolation on the discrete moisture content data at the same time to generate a spatial distribution map of moisture content at each sampling time. The spatial distribution map of moisture content in the continuous time series is smoothed and differencing in the time dimension to calculate the gradient change of moisture content over time. By fusing spatial distribution information with temporal gradient information, a three-dimensional data field is constructed that can characterize the dynamic migration and distribution of moisture content within the substrate, namely the dynamic moisture content distribution field.

5. The method for optimizing the composite flooring pressing process according to claim 1, characterized in that, The step of spatially aligning and fusing the temperature uniformity feature map, the coating quality evaluation vector, and the moisture content dynamic distribution field to generate a panoramic feature set for composite bonding conditions includes: Establish a unified spatial coordinate system based on the surface of the floor substrate; The temperature uniformity feature map and the dynamic distribution field of moisture content are resampled and mapped to the unified spatial coordinate system to ensure that the spatial positions correspond one-to-one. The coating quality evaluation vector is expanded into a spatialized coating quality distribution map with the same resolution as the unified spatial coordinate system based on its corresponding image acquisition location. A multi-source feature fusion network is used to perform feature-level fusion on the mapped temperature uniformity feature map, the spatialized coating quality distribution map, and the mapped dynamic moisture content distribution field. The multi-source feature fusion network calculates the correlation weights between different feature maps through a cross-attention mechanism, performs weighted summation and feature enhancement, and finally outputs a panoramic feature set of the composite bonding condition that includes comprehensive information on temperature, coating, and moisture content.

6. The method for optimizing the composite flooring pressing process according to claim 1, characterized in that, The pre-trained process neural network model is invoked to perform abnormal pattern recognition and key parameter sensitivity analysis on the panoramic feature set of the composite bonding condition, generating a process parameter influence factor matrix, including: The panoramic feature set of the composite bonding working condition is input into the feature encoder of the process neural network model. The feature encoder performs dimensionality reduction and abstraction on the input features to extract high-order working condition feature vectors. The high-order operating condition feature vector is input into the anomaly detection branch of the process neural network model. The anomaly detection branch identifies abnormal operating condition feature segments that deviate from the standard pattern by comparing them with a pre-stored standard operating condition feature pattern library. The high-order working condition feature vector is simultaneously input into the sensitivity analysis branch of the process neural network model. The sensitivity analysis branch calculates the gradient magnitude of the change in each input feature dimension on the final bonding quality prediction value through the gradient backpropagation method. By integrating the location information of the abnormal operating condition feature fragments with the gradient magnitude information of each feature dimension, a matrix is ​​constructed with process parameters as rows and quality indicators as columns. The value of each element in the matrix represents the influence weight of a specific process parameter on a specific quality indicator, i.e., the process parameter influence factor matrix. The process parameter influence factor matrix describes the non-linear correlation weights between the bonding quality index and each process parameter. The steps for constructing the process neural network model include: Collect composite lamination working condition data and corresponding lamination quality test results data accumulated in the historical production process to construct a training sample set; Design a multi-task neural network architecture that includes a feature encoder, an anomaly detection branch, and a sensitivity analysis branch; The feature encoder employs a multi-layer convolutional neural network structure to extract high-order features from the input panoramic feature set of composite bonding conditions. The anomaly detection branch adopts an autoencoder structure and identifies abnormal operating conditions by reconstructing errors. The sensitivity analysis branch adopts a fully connected network structure and analyzes parameter sensitivity through forward propagation and gradient calculation; The multi-task neural network is trained end-to-end using the training sample set, and the network weights are optimized using the backpropagation algorithm. During training, a dynamic weight balancing strategy is used to coordinate the loss functions of different task branches; Once the model's performance on the validation set reaches the preset standard, the final network parameters are saved to obtain the pre-trained process neural network model.

7. The method for optimizing the composite flooring pressing process according to claim 6, characterized in that, The step of performing reverse parameter optimization on the multidimensional process data based on the process parameter influence factor matrix to generate an optimized pressing process parameter adjustment scheme includes: Set the bonding quality target values, which include the expected range of flatness, bonding strength and formaldehyde release; Using the bonding quality target value as a constraint and the process parameter influence factor matrix as a guide, a target loss function is defined; An evolutionary optimization algorithm is used to iteratively optimize the adjustable process parameters in the currently collected multidimensional process data; In each iteration, the exploration step size and direction of different parameters are adjusted according to the process parameter influence factor matrix, with priority given to adjusting parameters with large influence weights; When the target loss function converges or reaches the preset number of iterations, a set of optimized process parameters that makes the predicted quality closest to the target value of the bonding quality is output, which serves as the core of the bonding process parameter adjustment scheme. The pressing process parameter adjustment scheme is used to guide the coordinated adjustment of the press temperature curve, coating rate and pressure distribution.

8. The method for optimizing the composite flooring pressing process according to claim 7, characterized in that, The pressing process parameter adjustment scheme is used to guide the coordinated adjustment of the press temperature profile, coating rate, and pressure distribution, specifically including: The optimized process parameter combination is analyzed, and the optimized setpoint sequence for the press temperature is extracted. Based on the optimized setpoint sequence and combined with the physical inertia model of the press heating plate, a smooth and executable press temperature rise and fall curve is generated. The optimized process parameter combination is analyzed to extract the optimized setting value for the adhesive coating rate; Based on the optimized set values ​​and combined with the mechanical response model of the coating roller, a speed control command sequence for the coating motor is generated. The optimized process parameter combination is analyzed to extract the optimized setpoint matrix for pressure distribution; Based on the optimized setpoint matrix and combined with the hydraulic pressure zone control system model, a pressure adjustment command sequence for each pressure zone is generated. The press temperature rise / fall curve, the speed control command sequence, and the pressure adjustment command sequence are synchronized and conflict checked on the time axis to ensure that the press temperature rise / fall curve, the speed control command sequence, and the pressure adjustment command sequence are coordinated without conflict, thus forming a complete pressing process parameter adjustment scheme.

9. The method for optimizing the composite flooring pressing process according to claim 8, characterized in that, The method further includes: A real-time control command stream is generated based on the aforementioned bonding process parameter adjustment scheme, and the real-time control command stream is sent to the actuator of the bonding production line to complete the dynamic closed-loop adjustment of the process parameters, specifically including: The time-sequential control commands in the pressing process parameter adjustment scheme are converted into standard data packets under a specific industrial communication protocol. The standard data messages are timestamped and prioritized to generate the real-time control command stream; Establish real-time data links with the press temperature controller, coating motor driver, and hydraulic pressure controller; According to the time sequence and priority of the real-time control command stream, the control commands are sequentially sent to the corresponding actuators through the real-time data link. During execution, new multidimensional process data are continuously collected, fed back, and compared with the expected adjustment target. If the deviation exceeds the threshold, a new round of optimization calculation and instruction generation is triggered, forming a dynamic closed-loop adjustment.

10. A composite flooring pressing process optimization system, characterized in that, The system includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement a composite flooring bonding process optimization method as described in any one of claims 1 to 9.

Images