Dynamic optimization control method and system for process parameters of hot compress patch

Abstract

The application provides a hot compress patch process parameter dynamic optimization control method and system, relates to the fields of industrial control and digital processing, and comprises the following steps: collecting point cloud data sets and infrared thermal image time series, and constructing thermodynamic state data sets; generating a thermal state sequence through a joint transformation function, extracting high-dimensional data features by using a physical prior processing unit containing a static space bias matrix, and outputting a comprehensive performance score through a linear data decoding function and a performance score calculation model; when the score is lower than a threshold value, fusing dynamic thermal performance features and initial static features on a process graph data structure, and outputting a fault probability distribution through a multi-round state iteration calculation method; identifying a fault node based on the fault probability distribution, generating an early warning instruction through rule base matching, constructing a hybrid optimization objective function in combination with a prediction and a target thermal performance feature vector, and outputting an optimal control instruction. The application fuses multi-modal perception to evaluate product thermal performance, supports root cause tracing, and realizes process parameter dynamic optimization control.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial control and digital processing, in particular to a hot compress patch process parameter dynamic optimization control method and system. BACKGROUND

[0002] The precise quality control and dynamic process optimization of the production process of precision thermal control products such as hot compress patches have become key technologies to ensure the consistency of their core performance and the safety of their use. How to achieve scientific optimization of product heating uniformity and stability in a complex and variable production environment, and avoid product failure caused by local overheating or insufficient heating and other defects, has become an important issue that needs to be addressed in the intelligent production and quality control upgrade of this field.

[0003] Chinese patent application with publication number CN120370717A provides a patch machine high-precision mounting head adaptive thermal compensation system and method. The method includes: collecting temperature and deformation data of the mounting head components in real time through a temperature-deformation sensing unit, reconstructing a three-dimensional temperature field based on a finite element algorithm by a thermal-mechanical coupling analysis unit, dynamically calculating the thermal expansion distribution, solving the roughness problem of the traditional single-point temperature measurement model, predicting the thermal drift within the next 5ms through online parameter identification and a long short-term memory network model, combining with the motion condition to adaptively adjust the compensation strategy, and decomposing the compensation amount into displacement and torsion correction instructions by a closed-loop execution unit to realize precise offset of thermal deformation and build a whole-process thermal compensation closed loop.

[0004] However, the current technology still faces many challenges. In the large-scale continuous production process of hot compress patches, when the key process parameters are affected by equipment aging or working condition fluctuations and produce slight drift, the traditional quality control method relies on single-point temperature measurement and offline sampling, which is limited by low spatial coverage and non-continuous time, making it difficult to timely and comprehensively monitor the early degradation of the drift on the product surface temperature field uniformity. Once this early degradation is not discovered in time, the process parameter deviation will continue to accumulate and amplify, eventually leading to the production of a large number of critical defect products or even the entire batch being scrapped, causing material waste and economic losses. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a hot compress patch process parameter dynamic optimization control method and system, and the specific technical solutions are as follows:

[0006] The hot compress patch process parameter dynamic optimization control method comprises:

[0007] Collecting point cloud data sets and infrared thermal image time series, mapping the infrared thermal image time series to a three-dimensional grid model constructed by the point cloud data sets based on the spatial transformation relationship between sensors, and constructing a thermodynamic state data set;

[0008] adopting a joint transformation function based on a thermodynamic state dataset to generate a thermodynamic state sequence, performing matrix calculation on the thermodynamic state sequence by a physical prior processing unit introducing a static spatial bias matrix, outputting high-dimensional data features, mapping the high-dimensional data features to corresponding performance indicators through a parallel linear data decoding function, inputting the performance indicators into a performance score calculation model, and outputting a comprehensive performance score;

[0009] judging whether the comprehensive performance score is lower than a preset quality threshold to activate a root cause diagnosis process, if the root cause diagnosis process is activated, fusing a dynamic thermal performance feature vector with initial static feature vectors of each node in a pre-constructed production process graph data structure to generate node input features, and iteratively updating the node input features by a multi-round state iteration calculation method to output fault probability distributions of all nodes;

[0010] identifying fault nodes based on the fault probability distributions by a maximum a posteriori probability decision method, and generating production warning instructions by data matching in a preset diagnosis-disposal rule library, and constructing and solving a feed-forward-feedback coupled hybrid optimization objective function by comprehensively utilizing the fault probability distributions, a predicted thermal performance feature vector generated based on online sensor data, and a preset target thermal performance feature vector, and outputting optimal control instructions.

[0011] Further, the construction method of the thermodynamic state dataset comprises:

[0012] collecting an initial three-dimensional geometric appearance before the hot compress patch is activated to generate a point cloud dataset representing surface geometric features of the hot compress patch;

[0013] periodically collecting time-series infrared radiation intensity signals within a whole thermal life cycle of the hot compress patch, and generating infrared thermal image time series representing surface temperature fields of the hot compress patch by real-time mapping of the infrared radiation intensity signals combined with a calibration function;

[0014] constructing a three-dimensional grid model based on the point cloud dataset by Poisson surface reconstruction, and solving spatial transformation relationships between heterogeneous coordinate systems by joint calibration of sensors, mapping the infrared thermal image time series to the three-dimensional grid model according to the spatial transformation relationships and an internal parameter matrix of the thermal imager, and constructing a thermodynamic state dataset fusing dynamic thermodynamic behavior.

[0015] Further, the output step of the comprehensive performance score comprises:

[0016] concatenating static spatial positions and dynamic instantaneous temperatures of each three-dimensional grid vertex in the thermodynamic state dataset into a state vector, performing nonlinear data transformation on the state vector by a joint transformation function, arranging the transformed state vectors in time sequence, and constituting a thermodynamic state sequence.

[0017] A static spatial bias matrix is calculated based on geometric features between each three-dimensional grid vertex in a three-dimensional grid model, the static spatial bias matrix is taken as an additive bias term, matrix calculation is performed on the thermal state sequence, dynamic thermodynamic dependence modeling guided by physical prior is realized, and high-dimensional data features are output;

[0018] Global average pooling is performed on the high-dimensional data features along the space-time dimension to aggregate into a context vector, the context vector is input into a group of parallel preset linear data decoding functions, mapped to corresponding performance indicators, a plurality of performance indicators are spliced along the feature dimension to construct a dynamic thermal performance feature vector;

[0019] The dynamic thermal performance feature vector is input into a performance score calculation model based on supervised training of a training data set labeled with dynamic thermal performance feature vectors and authoritative quality scores, and a comprehensive performance score is output.

[0020] Further, the static spatial bias matrix is calculated by a learnable spatial bias encoding function, which takes the Euclidean distance and Z-axis height difference between three-dimensional grid vertices as input features to generate corresponding bias matrix elements.

[0021] Further, the performance indicators include a thermal dispersion gradient vector, an entropy increase uniformity index, and an abnormal hot spot space-time trajectory matrix. The thermal dispersion gradient vector is used to quantify the macroscopic gradient distribution of the hot compress patch surface temperature field in space and its evolution trend over time. The entropy increase uniformity index is used to globally quantify the uniformity of the heating process. The abnormal hot spot space-time trajectory matrix is used to parameterize the motion trajectory of the key hot spot in the two-dimensional space and the change of the temperature value.

[0022] Further, the output step of the failure probability distribution comprises:

[0023] A structured data modeling method is used to abstract key factors affecting the thermal performance of the hot compress patch in the production process as entity nodes, and the dependency relationship between the entity nodes is defined as a directed edge to construct a production process graph data structure;

[0024] The dynamic thermal performance feature vector and the initial static feature vector of each node in the production process graph data structure are spliced to generate a node input feature, and a multi-round state iteration calculation method is used to iteratively update the production process graph data structure to obtain the final feature state vector of each node. The final feature state vector of each node is input into a failure probability function to construct the failure probability distribution of all nodes.

[0025] Further, the entity nodes are divided into material attribute nodes, process parameter nodes, equipment state nodes, and intermediate quality nodes; wherein the material attribute nodes are used to represent the key physical and chemical properties of the input materials, the process parameter nodes are used to represent the set and controlled operation parameters in the production and manufacturing process, the equipment state nodes are used to represent the working state parameters of the production equipment in the running process, and the intermediate quality nodes are used to represent the intermediate product properties that have an impact on the subsequent process.

[0026] Further, the directed edges are divided into causal relationship edges, time sequence relationship edges, and subordinate relationship edges; wherein the causal relationship edges represent that the change of one node state will cause the change of another node state, the time sequence relationship edges represent the sequence of the production and manufacturing process, and the subordinate relationship edges represent that the process parameter nodes are inherent attributes of the material attribute nodes or the equipment state nodes.

[0027] Further, the generation of the production early warning instruction includes: identifying a fault node based on the fault probability distribution using a maximum a posteriori probability decision method, performing data matching in a preset diagnosis-disposal rule library taking the fault node as a query index, generating a production early warning instruction, and pushing the production early warning instruction to a monitoring terminal to trigger a corresponding disposal process.

[0028] Further, the output step of the optimal control instruction includes:

[0029] Real-time acquisition of online sensor data, input of the online sensor data into a pre-trained performance predictor model, and generation of a predicted thermal performance feature vector;

[0030] Based on the predicted thermal performance feature vector, the fault probability distribution, and a preset target thermal performance feature vector, a feedforward-feedback coupled mixed optimization objective function is constructed and solved to determine the optimal control instruction.

[0031] Further, the performance predictor model is preferably a multilayer perceptron MLP or a simplified graph neural network, which learns the complex nonlinear mapping relationship between the online sensor data that is easy to collect in real time and the final product performance features that are difficult to measure directly in real time by training on a large amount of historical data.

[0032] Further, the mixed optimization objective function is coupled by a performance error term and a diagnosis regularization term; wherein the performance error term is based on a pre-trained forward process model, performs forward inference on the process parameters after applying a hypothetical parameter adjustment vector, predicts the predicted thermal performance feature vector of the future product, and minimizes the parameter adjustment vector by calculating the square of the Euclidean distance between the predicted thermal performance feature vector and the target thermal performance feature vector; and the diagnosis regularization term takes the fault probability distribution as an adaptive penalty weight to constrain the parameter adjustment vector.

[0033] The hot compress patch process parameter dynamic optimization control system comprises a multi-modal information acquisition module, a data scoring module, a root cause reasoning module and a control decision module.

[0034] The multi-modal information acquisition module is used for acquiring a point cloud data set and an infrared thermal image time sequence, mapping the infrared thermal image time sequence to a three-dimensional grid model constructed by the point cloud data set based on a spatial transformation relationship between sensors, and constructing a thermodynamic state data set.

[0035] The data scoring module generates a thermal state sequence by using a joint transformation function based on the thermodynamic state data set, performs matrix calculation on the thermal state sequence by using a physical prior processing unit introducing a static spatial bias matrix, outputs high-dimensional data features, maps the high-dimensional data features to corresponding performance indicators by using a parallel linear data decoding function, inputs the performance indicators into a performance score calculation model, and outputs a comprehensive performance score.

[0036] The root cause reasoning module judges whether the comprehensive performance score is lower than a preset quality threshold to activate a root cause diagnosis process. If the root cause diagnosis process is activated, a dynamic thermal performance feature vector is fused with an initial static feature vector of each node in a production process graph data structure to generate node input features, the node input features are iteratively updated by using a multi-round state iteration calculation method, and fault probability distributions of all nodes are output.

[0037] The control decision module identifies fault nodes by using a maximum posterior probability decision method based on the fault probability distributions, matches data in a preset diagnosis-disposal rule library to generate a production warning instruction, constructs and solves a feedforward-feedback coupled hybrid optimization objective function by comprehensively using the fault probability distributions, a predicted thermal performance feature vector generated based on online sensor data and a preset target thermal performance feature vector, and outputs an optimal control instruction.

[0038] Compared with the prior art, the present application has the following advantages:

[0039] The present application fuses three-dimensional point cloud data representing static geometric topology and infrared thermal image time sequences representing dynamic thermodynamic behavior as two independent and complementary information sources for modeling, constructs a unified thermodynamic state data set, and avoids the problem of inaccurate evaluation of heat emission uniformity and inability to trace defect root causes caused by ignoring the physical impact of product surface micro-topography on heat transfer in traditional detection methods.

[0040] The application inputs a space bias matrix representing static geometric topology as a physical prior into a dynamic correlation calculation of a physical prior processing unit, performs matrix calculation on a thermal state sequence, outputs high-dimensional data features, learns and represents a heat conduction law more consistent with a real physical process, and improves the problem that a traditional black-box deep learning model lacks physical constraints and thus cannot learn a complex spatiotemporal evolution law sufficiently.

[0041] The application combines a graph data structure containing production process causal relationships with a multi-round state iteration calculation method, iteratively updates a feature vector representing product performance abnormalities on the topology structure of the graph data structure for attribution reasoning, improves the problem that when a traditional fault diagnosis is based on statistical correlation, a diagnosis result is poor in interpretability and is prone to false correlation due to lack of structured cognition of the overall causal chain of the production process.

[0042] The application constructs a hybrid optimization objective function coupling feedforward performance prediction and feedback fault diagnosis, can pre-compensate future performance deviations caused by process disturbances when calculating process parameter adjustment amounts, and can constrain adjustment behavior on the most likely fault root cause diagnosed from historical data, thereby improving the problem that a traditional single feedback control response is lagging or a single feedforward control lacks robustness, and realizing the unity of speed and accuracy of control. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0044] Figure 1 is a principle flowchart of the process parameter dynamic optimization control method of the hot compress patch of the present application;

[0045] Figure 2 is a functional module diagram of the process parameter dynamic optimization control system of the hot compress patch of the present application. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0047] Embodiment 1:

[0048] Please refer to Figure 1 The embodiment shown provides a dynamic optimization control method for heat patch process parameters, comprising:

[0049] S1000, collecting a point cloud dataset and an infrared thermal image time series , based on the spatial transformation relationship between sensors, mapping the infrared thermal image time series to a three-dimensional grid model constructed by the point cloud dataset , constructing a thermodynamic state dataset .

[0050] Specifically, this step aims to systematically capture and deeply integrate the static three-dimensional geometric features and full-cycle dynamic surface temperature field evolution behavior of the heat patch physical entity through the collaborative application of multi-modal sensing technology, thereby constructing a thermodynamic state dataset .

[0051] Further, step S1000 comprises:

[0052] Step S1100, collecting the initial three-dimensional geometric morphology of the heat patch before activation, generating a point cloud dataset representing the surface geometric features of the heat patch .

[0053] Specifically, this step aims to obtain the macro three-dimensional geometric morphology of the heat patch in the resting state before activation, i.e., the point cloud dataset This morphology data is the unified spatial reference for subsequent multi-modal data fusion, and also contains surface micro-undulation information caused by differences in internal heating material filling density, compaction process, etc., which has prior value for analyzing heating uniformity and process stability.

[0054] In the specific implementation process, a non-contact high-precision structured light three-dimensional scanner is used to perform three-dimensional scanning of the upper surface of the heat patch in the full field range, obtaining the original point cloud data of the heat patch. During the scanning process, a pre-encoded grating pattern is projected onto the surface of the heat patch sample, and the deformed encoded pattern caused by the undulation of the heat patch sample surface is captured by an image sensor. Subsequently, based on the triangulation principle, the captured pattern is analyzed, and the three-dimensional coordinate point set of the heat patch covering the full field of view is calculated.

[0055] To improve the signal-to-noise ratio and reduce the computational load of subsequent processing, comprehensive filtering preprocessing is performed on the original point cloud data , which is realized by constructing a comprehensive filtering operator , which filters out noise and redundant information from the original point cloud data ​mapping to an optimized point cloud dataset .

[0056] wherein the comprehensive filtering operator is a set of one or more point cloud preprocessing algorithms, for example, a Statistical Outlier Removal (SOR) algorithm for removing isolated outliers caused by environmental noise or reflections, and a Voxel Grid Downsampling (VGD) algorithm for reducing data density and achieving data uniformization while maintaining key geometric profiles. represents a point cloud dataset, for reconstructing a digitized object in a computer in the form of a discrete point set to a high-fidelity three-dimensional surface model; represents a point cloud dataset the i-th three-dimensional spatial coordinate vector, containing all spatial information of a specific sampling point position on the surface of the hot compress patch sample, i.e., the three coordinate components of the sampling point in a three-dimensional Cartesian coordinate system ; represents the Cartesian coordinate components of the three-dimensional spatial coordinate vector , the values of which are high-precision floating-point numbers, wherein respectively represent the projection values of the sampling point on the X-axis, Y-axis and Z-axis of the measurement device coordinate system, which together determine the accurate position of the sampling point in three-dimensional space; i represents the index of the point cloud dataset , used to uniquely identify and traverse all three-dimensional spatial coordinate vectors in the point cloud dataset , which takes a continuous positive integer value from 1 to N; N represents the total number of point clouds of three-dimensional spatial coordinate vectors in the point cloud dataset .

[0057] Step S1200, periodically collecting the time series infrared radiation intensity signals of the hot compress patch throughout its entire thermal life cycle, and combining the calibration function to map the infrared radiation intensity signals in real time to generate an infrared thermal image time series representing the surface temperature field of the hot compress patch.

[0058] Specifically, this step aims to capture the dynamic thermodynamic behavior of the hot compress patch throughout its entire life cycle in a non-contact, full-field imaging manner, to obtain the spatio-temporal evolution data of its surface temperature field, i.e., the infrared thermal image time series . The infrared thermal image time series is not only the basis for subsequent objective quantitative evaluation of heating performance and intelligent diagnosis of thermal anomaly patterns, but also the fundamental input data for establishing the dynamic correlation between process parameters and heating performance.

[0059] In the implementation process, the activated hot compress patch is placed in a test chamber that can accurately control the environmental parameters to simulate the real use scenario of the hot compress patch. For example, when simulating the human body surface environment, the environmental temperature is constantly controlled at , and the relative humidity is maintained at . A high-precision infrared thermal imager calibrated by a standard blackbody radiation source is used to continuously and uninterruptedly record the infrared thermal imaging of the upper surface of the hot compress patch sample at a preset time sampling frequency and sampling time interval , until the average surface temperature of the hot compress patch is lower than the preset effective working threshold , and the recording is terminated.

[0060] At each preset sampling time t, the raw data collected by the detector array of the infrared thermal imager is the raw infrared radiation intensity signal . In order to convert this raw infrared radiation intensity signal into a pixel point temperature value with a specific and explicit physical meaning, a calibration function is needed for mapping and conversion. The calibration function is a multivariate physical model based on the physical law of thermal radiation, which comprehensively introduces the core factors affecting thermal radiation measurement. The specific process formula is as follows:

[0061] ;

[0062] wherein represents the pixel point temperature value, which is an accurate temperature scalar obtained after radiation calibration and environmental correction, and its physical dimension is Celsius or Kelvin, used to represent the instantaneous thermodynamic state of the hot compress patch surface at position and time t; represents the discrete sampling time, which is a timestamp or index used to identify a specific data point in the time series; represents the raw infrared radiation intensity signal, which is the raw physical quantity collected by the sensing unit of the detector array of the infrared thermal imager at the mth row and nth column at time t; represents the emissivity of the hot compress patch sample surface, which ranges from 0 to 1, and is used to correct the infrared radiation intensity deviation caused by the difference in material properties; represents the environmental temperature compensation parameter, which is an accurate temperature value, used to eliminate the interference of environmental stray radiation on the measurement results; represents the atmospheric path compensation parameter, used to correct the energy attenuation caused by atmospheric absorption and scattering during the propagation of the infrared signal.

[0063] At the sampling time t, the raw infrared radiation intensity signal Each sensing unit performs the above temperature calculation process in parallel and outputs the calculated pixel temperature values. Constructed as a two-dimensional temperature field matrix This refers to the single-frame heatmap matrix. This matrix is ​​a discretized spatial representation of the instantaneous thermodynamic state of the heat patch surface at sampling time t. Here, m and n represent the values ​​in the single-frame heatmap matrix, respectively. The row and column indices, with values ​​ranging from 1 to H and 1 to W respectively, together form a coordinate pair. Used in In a two-dimensional array, W and H uniquely identify the position of a pixel; W and H respectively represent the position of a single frame in the heatmap matrix. The number of pixels in the horizontal and vertical directions together define the spatial sampling density and data dimension of the temperature field in a single frame.

[0064] Over time, at fixed sampling intervals The above single-frame heatmap matrix is ​​continuously generated and recorded. This constitutes a complete time series recording the thermodynamic behavior of the heat therapy patch sample throughout its entire life cycle, namely, an infrared thermogram time series. The specific formula is as follows:

[0065] ;

[0066] in, Representing an infrared thermal image time series, it is a matrix of single-frame thermal images. The resulting ordered set provides a data foundation for subsequent dynamic performance evaluation and fault diagnosis; The sampling time interval is the time step between two consecutive heatmap matrices, and its reciprocal is the dynamic detection frequency. The end time of the test is the point at which the heating process of the heat patch stops and data sampling ceases.

[0067] Step S1300, based on point cloud dataset A 3D mesh model is constructed by reconstructing the Poisson surface. Furthermore, by jointly calibrating the sensors to solve the spatial transformation relationship between heterogeneous coordinate systems, and based on the spatial transformation relationship and the thermal imager's intrinsic parameter matrix, the infrared thermal image time series is generated. Mapped to the three-dimensional mesh model Construct a thermodynamic state dataset that integrates dynamic thermodynamic behavior .

[0068] Specifically, this step aims to address the inconsistencies in coordinate system, resolution, and sampling dimension between three-dimensional geometric data from heterogeneous sensors and two-dimensional time-series temperature data. A thermodynamic state dataset is generated by registering and mapping dynamic thermodynamic information onto a static three-dimensional geometric model. .

[0069] In practice, this step is implemented through the following three consecutive technical phases:

[0070] Phase 1: Constructing the digital geometric basis. This involves using the point cloud dataset output from step S1100. A standardized three-dimensional mesh model representing the macroscopic and microscopic geometry of the thermal patch was constructed using the Poisson Surface Reconstruction (PSR) algorithm. The three-dimensional mesh model It is the static spatial basis that carries subsequent temperature field data, and its specific expression formula is as follows:

[0071] ;

[0072] in, Representing a 3D mesh model The set of vertices constitutes the three-dimensional mesh model. An ordered list of all vertices, used for discretizing the 3D mesh model. The geometric surface; Representing a 3D mesh model A set of triangular facets used to describe a 3D mesh model. Surface continuity.

[0073] The second stage involves establishing the spatial mapping relationship for multimodal data. To ensure the accuracy of the fusion between temperature data and the geometric model, systematic errors introduced by differences in sensor layout must be eliminated. This is achieved by jointly calibrating the 3D scanner and the infrared thermal imager, solving for the coordinates from the infrared thermal imager. To the coordinate system of the 3D scanner The spatial transformation relationship is defined by a 4×4 homogeneous transformation matrix. This homogeneous transformation matrix is ​​uniquely described. The rotation matrix R and the translation vector t are coupled, and the specific formula is as follows:

[0074] ;

[0075] in, This represents a homogeneous transformation matrix, used to unify the coordinate space of data from different sensors. The coordinate system of the 3D scanner is a three-dimensional Cartesian coordinate system established based on the 3D scanner. The coordinate system of the infrared thermal imager is a three-dimensional Cartesian coordinate system with the optical center of the infrared thermal imager as the origin. This represents the rotation matrix, used to describe the coordinate system of the infrared thermal imager. Relative to the coordinate system of the 3D scanner The rotational posture; This represents the translation vector, used to describe the coordinate system of the infrared thermal imager. Origin relative to the coordinate system of the 3D scanner Spatial displacement of the origin.

[0076] The third stage: Projecting the dynamic temperature field and constructing a spatiotemporal dataset. This involves the infrared thermal image time series constructed in step S1200. As a dynamic data source, each frame contains a single-frame heatmap matrix. Projected onto the aforementioned constructed 3D mesh model Above. Specifically, for three-dimensional mesh models. Any three-dimensional mesh vertex on Using the known thermal imager intrinsic parameter matrix and the spatial transformation relationships already obtained Calculate vertices using perspective projection model Projected pixel coordinates on a two-dimensional heatmap The specific process formula is as follows:

[0077] ;

[0078] in, Representing a 3D mesh model vertex set The j-th vertex of the 3D mesh; These represent the j-th 3D mesh vertex. The coordinates of the three-dimensional mesh on the X, Y, and Z axes; j represents the vertex of the three-dimensional mesh. The index is used to traverse the vertex set. ; This represents the thermal imager intrinsic parameter matrix, the values ​​of which are obtained in advance through camera calibration and are used to project three-dimensional points in the camera coordinate system onto the two-dimensional image plane. This represents the scale factor, used to convert homogeneous coordinates to non-homogeneous coordinates. This represents the coordinates of the projected pixels, used to locate the temperature interpolation region.

[0079] Considering the projected pixel coordinates Due to the non-integer nature of the graph, a bilinear interpolation algorithm is used, based on the projected pixel coordinates. Calculate the coordinates of the projected pixel by taking the temperature values ​​and weights of its four neighboring pixels. The precise interpolated temperature is used as the vertex of the 3D mesh. The instantaneous temperature at time t By traversing all vertices of the 3D mesh And all time frames, generate a four-dimensional state-space representation that fully describes the four-dimensional thermodynamic behavior of the heat patch, i.e., a thermodynamic state dataset. The specific formula is as follows:

[0080] ;

[0081] in, Represents the vertices of a 3D mesh The instantaneous temperature at time t is a temperature scalar calculated through interpolation, used to characterize the vertices of the 3D mesh. The thermodynamic state at time t.

[0082] S2000, based on thermodynamic state dataset Joint transformation functions are used to generate thermodynamic state sequences. By introducing a static space bias matrix The physical prior processing unit processes the thermodynamic state sequence. Perform matrix calculations and output high-dimensional data features. The high-dimensional data features are decoded using a parallel linear data decoding function. Mapping these metrics to corresponding performance indicators, inputting these performance indicators into the performance scoring calculation model, and outputting a comprehensive performance score. .

[0083] Specifically, this step aims to utilize an end-to-end deep learning model to process the high-dimensional and complex thermodynamic state dataset generated in step S1300. Deep spatiotemporal feature extraction and information compression are performed to transform them into a low-dimensional structured feature vector that can comprehensively characterize the dynamic thermal performance of the heat therapy patch sample, namely, the dynamic thermal performance feature vector. Furthermore, the dynamic thermal performance feature vector is... Transformed into a comprehensive performance score that can be directly used for quality control and process optimization. This is a key process that connects intelligent analysis and intelligent control.

[0084] Further, step S2000 includes:

[0085] Step S2100, the thermodynamic state dataset is... Each 3D mesh vertex The static spatial position and the dynamic instantaneous temperature are spliced into a state vector, a joint transformation function is used for nonlinear data transformation of the state vector, the state vector after transformation is arranged in time sequence, and a thermal state sequence is formed .

[0086] Specifically, the present step aims to convert the thermodynamic state data set into a structured and serialized expression suitable for processing by a time series deep learning model, i.e., a thermal state sequence in a high-dimensional feature space .

[0087] In the specific implementation process, for any three-dimensional grid vertex at any time t in the thermodynamic state data set , the complete instantaneous state is composed of two parts of information: one is a static spatial position vector that does not change with time, used to represent the fixed position of the three-dimensional grid vertex in the three-dimensional grid model; the other is a dynamic instantaneous temperature that changes with time, used to represent the thermodynamic state of the three-dimensional grid vertex at time t.

[0088] To realize the unified feature expression of the above heterogeneous information, the present step uses a joint transformation function composed of a series of multi-level nonlinear data transformations, which receives the state vector spliced from the static spatial position vector and the dynamic instantaneous temperature in the thermodynamic state data set , and learns the complex coupling relationship between the spatial position and the temperature state through multi-layer nonlinear data transformation. Finally, a high-dimensional and unified input token is output to represent the comprehensive state feature of the corresponding space-time point.

[0089] At any time t, the input tokens of all three-dimensional grid vertices constitute the token set of the full-field state at time t. Further, the token sets at all times are arranged in time sequence to form a thermal state sequence embedded in a high-dimensional feature space. The thermal state sequence not only retains the spatio-temporal evolution law of the surface temperature field of the hot compress patch, but also can be used as the input of the time series attention network for subsequent dynamic thermal performance feature extraction and process parameter optimization control.

[0090] Step S2200, calculating a static spatial bias matrix based on the geometric features between the three-dimensional grid vertices in the three-dimensional grid model , and using the static spatial bias matrix As an additive bias term, for the thermodynamic state sequence Perform matrix calculations to achieve dynamic thermodynamic dependency modeling guided by physical priors and output high-dimensional data features. .

[0091] Specifically, this step aims to propose an improved self-attention encoder layer (Physics-Informed Self-Attention, PISA) that incorporates prior physical knowledge to achieve high-precision modeling of the dynamic evolution of the temperature field on the surface of a heat-treated patch. This encoder layer serves as the core computational unit, and multiple layers can be stacked to construct a complete multi-layer temporal attention network. Its core lies in the deep coupling of two heterogeneous information elements: one is the topological information representing the static geometric structure, derived from the point cloud dataset output in step S1100. The three-dimensional mesh model constructed in step S1300 Secondly, there is time-series data characterizing dynamic thermodynamic behavior, namely the thermodynamic state sequence output in step S2100. By fusing cross-modal features, a dynamic system identification model that conforms to real physical processes is constructed, outputting high-dimensional data features. .

[0092] In its implementation, while the standard self-attention mechanism possesses strong long-distance dependency modeling capabilities in sequence data processing, its inherent permutation invariance prevents it from directly perceiving stable physical spatial topological relationships between input heat tokens. To overcome this deficiency, this step proposes a matrix computation method that incorporates prior physical information into the core computation layer. The implementation of this step includes the following three stages:

[0093] Phase 1: Quantization and encoding of static geometric associations. Based on the point cloud dataset output from step S1100. Constructed 3D mesh model A static spatial bias matrix is ​​pre-calculated. The static spatial bias matrix bias matrix elements Used to quantify the Three-dimensional mesh vertices and the Three-dimensional mesh vertices The static geometric correlation between them. This static geometric correlation can be encoded by a learnable spatial bias coding function. The calculated encoding function is... Based on the Euclidean distance between the vertices of the 3D mesh Z-axis height difference Using geometric features as input features, the corresponding bias matrix elements are generated. The specific process formula is as follows:

[0094] ;

[0095] wherein, represents a static spatial bias matrix The bias matrix element in the static spatial bias matrix is the scalar value at the th row and the th column; respectively represent the vertex indices of the three-dimensional mesh vertices, which are used to uniquely identify a specific vertex in the vertex set , and also correspond to the row index and the column index of the static spatial bias matrix respectively; represents a spatial bias encoding function, which is used to convert the geometric relationship into a bias value; represents the non-negative scalar of the straight-line distance between the three-dimensional mesh vertices and in the three-dimensional space, i.e., the Euclidean distance; represents the L2 norm operator, which is used to quantify the physical proximity of two three-dimensional mesh vertices; represents the non-negative scalar of the separation degree between the three-dimensional mesh vertices and in the vertical direction, i.e., the Z-axis height difference, which is used to reflect the local concave-convex shape or thickness difference on the surface of the hot compress patch.

[0096] Second stage: dynamic attention calculation guided by spatial bias. The heat state sequence is mapped to the query matrix Q, the key matrix K, and the value matrix V through linear transformation. When calculating the attention weight, the pre-computed static spatial bias matrix is directly injected as an additive bias into the dot product similarity calculation between the query matrix Q and the key matrix K. This operation enables the consideration of the fixed physical spatial relationship between any two three-dimensional mesh vertices when evaluating the dynamic thermodynamic correlation between them, thereby realizing the modeling of dynamic thermodynamic dependence guided by physical prior. Finally, the attention feature matrix after spatial information modulation is output.

[0097] Third stage: layer stacking and generation of final hidden state. After the input attention feature matrix is processed by the feedforward neural network, residual connection, and layer normalization, the final output of the PISA encoder layer is generated. All PISA encoder layers are stacked in sequence to build a multi-layer temporal attention network, where the output of the previous layer is used as the input of the next layer. When the data flows through all the encoder layers for processing, the high-dimensional feature representation output by the last encoder layer is the high-dimensional data feature of the multi-layer temporal attention network .

[0098] Step S2300, the high-dimensional data features Global average pooling along the space-time dimension to aggregate into a context vector, input the context vector into a set of parallel preset linear data decoding functions, map to the corresponding performance indicators, concatenate multiple performance indicators along the feature dimension to construct a dynamic thermal performance feature vector .

[0099] Specifically, the step aims to decode and map the high-dimensional abstract features output by the multi-layer time series attention network in step S2200, i.e., the high-dimensional data features , into a set of low-dimensional performance indicators with clear physical or statistical significance and interpretability, i.e., a dynamic thermal performance feature vector .

[0100] In specific implementation, the high-dimensional data features is a high-dimensional tensor, containing comprehensive and highly abstract representation of the thermal patch in the entire space-time evolution process. To extract specific performance indicators for engineering application, information aggregation and decoding of the high-dimensional data features is required. By applying global average pooling (GAP) operation along the space and time dimensions, the high-dimensional data features are aggregated into a fixed-dimensional context vector.

[0101] Then, the context vector is parallelly input to multiple independent and preset linear data decoding functions. Each linear data decoding function is an independent linear transformation operation, with internal parameters including weight matrix and bias vector, obtained through supervised learning training on large-scale labeled data sets, rather than preset fixed values. During training, these parameters are iteratively optimized through backpropagation algorithm, so that any context vector is mapped to the corresponding specific performance indicator. Since the physical connotations of various performance indicators are different, each linear data decoding function has independent and exclusive weight and bias parameters.

[0102] The performance indicators include thermal dispersion gradient vector, entropy increase uniformity index, and abnormal hotspot space-time trajectory matrix. Among them, the thermal dispersion gradient vector is used to quantify the macroscopic gradient distribution of the thermal patch surface temperature field in space and its evolution trend over time, i.e., the trend of heat diffusion. The entropy increase uniformity index is used to globally quantify the uniformity of the heating process as a scalar, i.e., whether the overall heating is uniform. The abnormal hotspot space-time trajectory matrix is used to parameterize the motion trajectory of the key hotspot in the two-dimensional space and the change of the temperature value, so as to monitor whether an abnormal hotspot and its motion trajectory appear.

[0103] Finally, the performance indicators output by all linear data decoding functions are spliced along the feature dimension to form a dynamic thermal performance feature vector .

[0104] In step S2400, the dynamic thermal performance feature vector is input into a performance score calculation model trained based on the dynamic thermal performance feature vector and a training data set labeled with authoritative quality scores, to output a comprehensive performance score .

[0105] Specifically, this step aims to input the dynamic thermal performance feature vector extracted in step S2300 into the performance score calculation model and map it into a single and standardized comprehensive performance score . The comprehensive performance score achieves nonlinear mapping from a multi-dimensional feature space to a one-dimensional quality evaluation space, providing intuitive and quantifiable feedback evaluation indicators for automatic quality grading, real-time process monitoring, and dynamic optimization of process parameters of hot compress patches.

[0106] In the specific implementation process, a pre-trained lightweight performance score calculation model is used to process the high-dimensional dynamic thermal performance feature vector and output a comprehensive performance score . The performance score calculation model is a calculation structure of a multilayer perceptron, and the internal parameters are pre-determined through supervised learning training on a large number of labeled samples.

[0107] The training data set is composed of a series of feature vector-expert score pairs . Among them, is the dynamic thermal performance feature vector of the kth training sample; is the authoritative quality score for the kth training sample obtained by a domain expert or standardized experimental determination. During the training process, the performance score calculation model receives the dynamic thermal performance feature vector of the training sample and outputs a predicted score . The error between the predicted score and the authoritative quality score is quantified by a loss function, and the internal parameters of the network are iteratively optimized using a backpropagation algorithm until the mapping relationship that can reproduce the expert evaluation criteria is learned, achieving a stable mapping relationship from the dynamic thermal performance feature vector to the performance score.

[0108] In actual application, the dynamic thermal performance feature vector is input into the pre-trained performance score calculation model, and after one forward propagation calculation, a standardized comprehensive performance score is output. The comprehensive performance score is a normalized value between 0 and 1, and the higher the value, the better the comprehensive heating performance of the hot compress patch. The comprehensive performance score can be used as a basis for quality grading to achieve automatic grading management of products and rapid identification and data rejection of unqualified products.

[0109] S3000, based on the comprehensive performance score determine whether it is lower than the preset quality threshold to activate the root cause diagnosis process, and if the root cause diagnosis process is activated, the pre-constructed production process graph data structure , the dynamic thermal performance feature vector is fused with the initial static feature vector of each node in the production process graph data structure to generate node input features, and a multi-round state iteration calculation method is used to iteratively update the node input features to output the fault probability distribution of all nodes .

[0110] Specifically, this step aims to define a condition-triggered intelligent diagnosis mechanism. The mechanism is based on the comprehensive performance score output in step S2400, and only when the comprehensive performance score is lower than the preset quality threshold , it indicates that the hot compress patch sample has potential thermal performance defects, thereby starting the intelligent diagnosis mechanism of this step. The intelligent diagnosis mechanism is based on a multi-round state iteration calculation method, which fuses and reasons the dynamic thermal performance feature vector output in step S2300 with the pre-constructed production process graph data structure to output a fault probability distribution pointing to specific process parameter deviations, and realizes fault probability attribution of each process parameter deviation for the current thermal performance anomaly.

[0111] Further, step S3000 includes:

[0112] Step S3100, using a structured data modeling method, abstracting key factors affecting the thermal performance of the hot compress patch in the production process as entity nodes, defining the dependency relationship between the entity nodes as directed edges, and constructing a production process graph data structure .

[0113] Specifically, this step aims to formalize and structure the unstructured prior process knowledge contained in production procedures, expert experience and bill of materials for data extraction and structured modeling to construct a production process graph data structure that can be understood and automatically processed by machines.

[0114] ​In the implementation process, this step is an offline construction process, and a structured data modeling method is used to deconstruct and formally represent the complete production and manufacturing process of the hot compress patch. The core of the deconstruction and formalization process is to define the basic elements of the production process graph data structure , namely nodes and edges.

[0115] In the node definition stage, all key factors that may affect the thermal performance of the final hot compress patch product in the production and manufacturing process are abstracted as entity nodes in the production process graph data structure . To ensure the completeness of the production process graph data structure , entity nodes are divided into four categories, namely material attribute nodes, process parameter nodes, equipment state nodes, and intermediate quality nodes.

[0116] Among them, the material attribute node is used to represent the key physical and chemical properties of the input material, such as "iron powder mesh", "activated carbon moisture content", and "adhesive viscosity", etc.; the process parameter node is used to represent the operation parameters that can be set and controlled in the production and manufacturing process, such as "mixing time", "rolling pressure", and "drying temperature curve", etc.; the equipment state node is used to represent the working state parameters of the production equipment during operation, such as "mixer cylinder temperature" and "roller gap", etc.; the intermediate quality node is used to represent the properties of the intermediate product that have important influence on the subsequent process, such as "mixture uniformity" and "medicine package compaction density", etc.

[0117] In the edge definition stage, the physical, chemical, logical or time sequence relationship between the above nodes is abstracted as a directed edge in the production process graph data structure , to depict the dependency relationship between knowledge units. To accurately express the cause and effect and constraint relationship, each edge is assigned different semantic types, including cause and effect relationship edge, time sequence relationship edge, and subordinate relationship edge.

[0118] Among them, the cause and effect relationship edge represents that the change of the state of a node will cause the change of the state of another node, for example: the influence of the "mixing time" process parameter node on the "mixture uniformity" intermediate quality node; the time sequence relationship edge represents the sequence of the production and manufacturing process, for example, the "mixing process" node precedes the "rolling process" node; the subordinate relationship edge represents that the process parameter node is the inherent attribute of a material attribute node or equipment state node, for example, the "rolling pressure" process parameter node is subordinate to the "rolling equipment" equipment state node.

[0119] By systematically defining the above nodes and edges, a production process graph data structure that fully and deeply reflects the real production and manufacturing process is constructed.

[0120] Step S3200, concatenating the dynamic thermal performance feature vector and the initial static feature vector of each node in the production process graph data structure to generate node input features, iteratively updating the production process graph data structure using a multi-round state iterative calculation method to obtain the final feature state vector of each node, inputting the final feature state vector of each node into the failure probability function, and constructing the failure probability distribution of all nodes .

[0121] Specifically, this step aims to analyze the dynamic thermal performance feature vector of the hot compress patch using a multi-round state iterative calculation method based on the production process graph data structure constructed in step S3100, locate the process parameters most likely to cause thermal performance abnormalities, and output the corresponding failure probability distribution to achieve automated root cause diagnosis and quantitative evaluation.

[0122] In the specific implementation process, this step specifically includes the following stages:

[0123] First stage: node feature initialization and diagnostic information input. Initialize the initial static feature vector for each node in the production process graph data structure . This initial static feature vector is used to encode the prior process information of each node. For example, the process parameter node can encode the design nominal value and tolerance range, the material attribute node can encode the key technical indicators, and the equipment state node can encode the operating parameter range, etc.

[0124] Subsequently, the dynamic thermal performance feature vector characterizing the abnormal thermal performance of the current hot compress patch sample is input into the multi-round state iterative calculation method as the initial symptom information for this diagnosis. By concatenating the dynamic thermal performance feature vector with the initial static feature vector of each node in the production process graph data structure , a node input feature vector that integrates static attributes and dynamic states is generated, providing more abundant input data for the initial calculation of the multi-round state iterative calculation method.

[0125] Second stage: multi-round iterative information propagation and data aggregation. The multi-round state iterative calculation method iterates through an iterative process consisting of L calculation cycles in the production process graph data structure The information is propagated and aggregated on the graph to update the feature state vector of each node. In the (l+1)th iteration, the feature state vector of each node is calculated by an aggregation function, which collects the feature state information of all its neighbor nodes in the lth iteration and combines it with its own feature state information in the lth iteration, and then is transformed via a learnable weight matrix and a nonlinear activation function to form the updated node feature state vector, where, , is the total number of iterations.

[0126] After L rounds of iteration calculation, the final feature state vector of each node not only contains its own initial static information, but also deeply integrates its contextual associated information in the entire process chain and the current product performance abnormal symptom information, forming a high-level feature representation containing rich diagnostic clues, i.e., the final feature state vector.

[0127] The third stage: fault probability decoding and output. The final feature state vector of all nodes in the production process graph data structure is input into the fault probability function to calculate the probability of each node becoming the root cause of the current fault. The probabilities of all nodes together constitute the final fault probability distribution . The fault probability distribution quantifies and ranks the potential process problems, which is the direct input for the upper dynamic optimization control system to make decisions, so that it can prioritize adjusting the most suspicious process parameters according to the fault probability distribution, achieving precise control and real-time quality improvement of the production process.

[0128] S4000, based on the fault probability distribution , the fault node is identified by the maximum a posteriori probability decision method, and data matching is performed in the preset diagnosis-disposal rule library to generate immediate feedback production warning instructions , the fault probability distribution , the predicted thermal performance feature vector generated based on online sensor data , and the preset target thermal performance feature vector are comprehensively utilized to construct and solve a mixed optimization objective function with feedforward-feedback coupling, and the optimal control instructions are output.

[0129] Specifically, this step aims to build a mixed intelligent control strategy with feedforward and feedback coupling, which converts the diagnostic and prediction information generated in the previous steps into adaptive regulation of the production process, forming a complete technical closed loop. This strategy comprehensively utilizes three types of information, one is the feedback diagnostic information, i.e., the fault probability distribution output in step S3200; two is the feedforward disturbance information, i.e., the online sensor data collected by real-time sensors on the production line ; the third is the control target, i.e. the target thermal performance feature vector representing the ideal performance pre-set by the expert . Through the above multi-source information, a constrained optimization problem is solved to output the optimal control instruction for guiding production adjustment .

[0130] Further, the step S4000 comprises:

[0131] Step S4100, based on the fault probability distribution The maximum a posteriori probability decision method is used to identify the fault node, and the fault node is taken as a query index to perform data matching in the pre-set diagnosis-action rule base to generate a production warning instruction , and the production warning instruction is pushed to a monitoring terminal to trigger a corresponding treatment process.

[0132] Specifically, this step aims to convert the quantitative fault probability distribution output by the previous step into a human-machine readable warning or intervention instruction, i.e. a production warning instruction , which has a clear guiding significance, constitutes a feedback control and decision support link in the control system, and provides production management personnel or an automatic system with immediate and executable fault positioning and processing suggestions, so as to realize rapid response and timely adjustment to quality deviation.

[0133] In the specific implementation process, the input fault probability distribution is analyzed, and the process node with the highest probability value is found and determined through the maximum a posteriori probability (Maximum a Posteriori Probability, MAP) decision method. The determined process node is determined as the most likely fault root cause, i.e. the fault node. The system identifies the fault node as a query index to perform data matching in the diagnosis-action rule base (Diagnosis-Action Rule Base) in the pre-set process knowledge base. The rule base pre-defines the mapping relationship between each potential fault node and a set of recommended operations, maintenance suggestions or parameter calibration range. According to the matched rule, the system generates a structured production warning instruction . The production warning instruction not only contains fault positioning information, but also provides specific treatment guidance.

[0134] Finally, the production warning instruction is pushed to a plurality of pre-set monitoring terminals, such as the human-machine interaction interface of the operator control panel, the manufacturing execution system (MES) of the factory or the mobile device of the related engineer, to trigger a corresponding manual or automatic treatment process.

[0135] Step S4200, real-time collection of online sensor data , inputting the online sensor data into a pre-trained performance predictor model to generate a predicted thermal performance feature vector .

[0136] Specifically, this step aims to use the online sensor data easily and quickly acquired on the production line to build an advanced prediction model for the key performance of the final product, outputting a predicted thermal performance feature vector of the final product, so as to achieve early sensing and active compensation for process disturbances such as raw material fluctuations and environmental changes during product manufacturing, thereby improving the response speed and robustness of the control system.

[0137] The online sensor data is collected by various non-destructive real-time sensors arranged at key process nodes on the production line, including but not limited to feeding ports, mixing sections, coating stations, etc. These sensors can cover different technical equipment, including but not limited to: near-infrared spectrometers for analyzing the chemical composition of materials, machine vision systems for monitoring the mixing form of materials, and process sensors for reading the running state of equipment. The system regularly collects the output signals of these heterogeneous sensors and pre-processes and fuses the data to finally generate a unified process state vector, i.e., the online sensor data .

[0138] In the specific implementation process, traditional feedback control usually relies on the detection results of the final product, inevitably resulting in inherent time lag. To overcome this defect, this step introduces a feedforward prediction mechanism.

[0139] The feedforward prediction mechanism uses a pre-trained and lightweight performance predictor model. The performance predictor model is preferably a multi-layer perceptron (MLP) or a simplified graph neural network, which learns the complex nonlinear mapping relationship from the online sensor data easily and quickly collected in real time to the performance features of the final product that are difficult to measure directly in real time by training on a large amount of historical data.

[0140] In actual application, the system real-time collects the online sensor data of the current production batch and inputs it into the performance predictor model. Through one forward propagation calculation, the predicted thermal performance feature vector of the final performance of the product is obtained before the product is finally formed.

[0141] It is crucial that the data structure and dimension of the predicted thermal performance feature vector are the same as those of the dynamic thermal performance feature vector and the target thermal performance feature vector used in the subsequent steps are completely consistent. Ensuring the mathematical comparability between the vectors, the system can directly quantify the deviation between the predicted performance and the target performance in step S4300, thereby calculating the control instruction for advance compensation.

[0142] Step S4300, based on the predicted thermal performance feature vector , the fault probability distribution and the preset target thermal performance feature vector , a feedforward-feedback coupled mixed optimization objective function is constructed and solved to determine the optimal control instruction .

[0143] Specifically, this step aims to construct and solve a feedforward-feedback coupled mixed optimization objective function based on the fault probability distribution output in step S3200, the predicted thermal performance feature vector output in step S4200, and the preset target thermal performance feature vector , to calculate the optimal process parameter adjustment, i.e., the optimal control instruction , to intelligently balance the minimization of future performance deviation in the hot patch production process, i.e., the feedforward control objective, and the adjustment robustness based on historical diagnosis, i.e., the feedback constraint.

[0144] In the specific implementation process, this step is the core calculation link for dynamic optimization. Let the current production process parameter vector be , and the parameter adjustment vector to be solved be , where is the optimization variable that is updated iteratively during optimization. The received predicted thermal performance feature vector reveals the expected performance endpoint of the product under the current process parameters and the disturbance influence. When there is a deviation between the predicted thermal performance feature vector and the target thermal performance feature vector , this step uses a feedforward-feedback coupled mixed optimization objective function to calculate the final optimal control instruction to compensate for the deviation. The optimal control instruction contains specific adjustment amounts for each process parameter, which are used to directly issue to the underlying process control system (PCS) or programmable logic controller (PLC) for execution of the corresponding adjustment. The specific process formula is as follows:

[0145] ;

[0146] wherein, denotes a minimum parameter solver, which is used to find a parameter adjustment vector that can minimize the subsequent objective function expression and returns it as a result.

[0147] The hybrid optimization objective function is coupled by the following two key parts, as follows:

[0148] The first term is a performance error term . This term is based on the core idea of Model Predictive Control (MPC). Through the pre-trained forward process model , the performance vector of the future product, i.e., the second predicted thermal performance feature vector, is predicted by performing forward inference on the process parameters after applying the hypothetical parameter adjustment vector . This term minimizes the parameter adjustment vector by calculating the squared Euclidean distance between the second predicted thermal performance feature vector and the expected target thermal performance feature vector . Wherein, denotes the square of the L2 norm, which is a mathematical measure for measuring the distance between two vectors, and is used to calculate the squared Euclidean distance between the second predicted thermal performance feature vector and the target thermal performance feature vector as a scalarized performance error.

[0149] The second term is a diagnostic regularization term . This term takes the fault probability distribution , i.e., the feedback diagnostic result, as an adaptive penalty weight to constrain the parameter adjustment vector and improve the robustness of the control process. Wherein, denotes the regularization coefficient, which is a non-negative hyperparameter set by the user to balance the weight between performance error and adjustment penalty. denotes the diagnostic regularization function.

[0150] Embodiment 2:

[0151] This embodiment is based on Embodiment 1 and provides a hot compress patch process parameter dynamic optimization control system, as shown in Figure 2 , which includes a multi-modal information acquisition module, a data scoring module, a root cause reasoning module, and a control decision module.

[0152] The multi-modal information acquisition module: is used to acquire a point cloud data set representing the static geometric features of the hot compress patch Infrared thermogram time series of dynamic thermodynamic behavior Based on the spatial transformation relationship between sensors, the infrared thermal image time series is... Mapped to the point cloud dataset Constructed 3D mesh model Above, construct a thermodynamic state dataset. ;

[0153] The data scoring module is based on a thermodynamic state dataset. Joint transformation functions are used to generate thermodynamic state sequences. By introducing a static space bias matrix The physical prior processing unit processes the thermodynamic state sequence. Perform matrix calculations and output high-dimensional data features. The high-dimensional data features are decoded using a parallel linear data decoding function. Mapping these metrics to corresponding performance indicators, inputting these performance indicators into the performance scoring calculation model, and outputting a comprehensive performance score. ;

[0154] The root cause reasoning module is based on a comprehensive performance score. Determine if the quality is below the preset threshold. To activate the root cause diagnosis process, if the root cause diagnosis process is activated, then in the pre-built production process diagram data structure Above, the dynamic thermal performance feature vector With the aforementioned production process diagram data structure The initial static feature vectors of each node are fused to generate node input features. A multi-round state iteration calculation method is used to iteratively update these node input features, outputting the fault probability distribution of all nodes. ;

[0155] The control decision module is based on the fault probability distribution. Fault nodes are identified using the maximum a posteriori probability decision method, and data is matched against a pre-defined diagnosis-handling rule base to generate real-time production early warning instructions. Comprehensively utilize the aforementioned fault probability distribution Based on online sensor data The generated predictive thermal performance feature vector and the preset target thermal performance feature vector Construct and solve the feedforward-feedback coupled hybrid optimization objective function, and output the optimal control command. .

[0156] The parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0157] The specific embodiments described above further illustrate the objects, technical solutions, and beneficial effects of the present application. It should be understood that the above description is merely specific embodiments of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for dynamically optimizing process parameters of a hot compress patch, characterized in that, The method comprises the following steps: Collecting point cloud data sets and infrared thermal image time series, mapping the infrared thermal image time series to a three-dimensional grid model constructed by the point cloud data sets based on the spatial transformation relationship between sensors, and constructing a thermodynamic state data set; Based on the thermodynamic state data set, a joint transformation function is used to generate a thermal state sequence, a physical prior processing unit introducing a static spatial bias matrix is used to perform matrix calculation on the thermal state sequence, high-dimensional data features are output, the high-dimensional data features are mapped to corresponding performance indicators through a parallel linear data decoding function, the performance indicators are input into a performance score calculation model, and a comprehensive performance score is output; Based on the comprehensive performance score, it is judged whether it is lower than a preset quality threshold to activate a root cause diagnosis process. If the root cause diagnosis process is activated, a dynamic thermal performance feature vector is fused with an initial static feature vector of each node in the production process graph data structure to generate a node input feature. A multi-round state iteration calculation method is used to iteratively update the node input feature, and the fault probability distribution of all nodes is output. The output step of the fault probability distribution comprises: Using a structured data modeling method, abstracting key factors affecting the thermal performance of the thermal patch in the production process as entity nodes, and defining the dependency relationship between the entity nodes as directed edges to construct a production process graph data structure; The dynamic thermal performance feature vector and the initial static feature vector of each node in the production process graph data structure are spliced to generate a node input feature. A multi-round state iteration calculation method is used to iteratively update the production process graph data structure to obtain the final feature state vector of each node. The final feature state vector of each node is input into a fault probability function to construct the fault probability distribution of all nodes. Based on the fault probability distribution, a fault node is identified by a maximum a posteriori probability decision method, and a production warning instruction is generated by matching data in a preset diagnosis-disposal rule library. The fault probability distribution, the predicted thermal performance feature vector generated based on the online sensor data, and the preset target thermal performance feature vector are comprehensively utilized to construct and solve a feedforward-feedback coupled mixed optimization objective function, and an optimal control instruction is output.

2. The method according to claim 1, wherein, The construction method of the thermodynamic state data set comprises: Collecting the initial three-dimensional geometric appearance of the thermal patch before activation to generate a point cloud data set representing the geometric features of the thermal patch surface; Periodically collecting time-series infrared radiation intensity signals during the entire thermal life cycle of the thermal patch, and combining a calibration function to map the infrared radiation intensity signals in real time to generate an infrared thermal image time series representing the temperature field of the thermal patch surface; Based on the point cloud data set, a three-dimensional grid model is constructed by Poisson surface reconstruction, and the spatial transformation relationship between heterogeneous coordinate systems is solved by joint calibration of sensors. According to the spatial transformation relationship and the thermal imager intrinsic matrix, the infrared thermal image time series is mapped to the three-dimensional grid model to construct a thermodynamic state data set that integrates dynamic thermodynamic behavior.

3. The method of claim 1, wherein the method further comprises: The output step of the comprehensive performance score comprises: The static spatial position and dynamic instantaneous temperature of each three-dimensional grid vertex in the thermodynamic state data set are spliced into a state vector, a joint transformation function is used for nonlinear data transformation of the state vector, the transformed state vector is arranged in time sequence, and a thermodynamic state sequence is formed; A static spatial bias matrix is calculated based on the geometric characteristics between each three-dimensional grid vertex in the three-dimensional grid model, the static spatial bias matrix is used as an additive bias term, a matrix calculation is performed on the thermodynamic state sequence, dynamic thermodynamic dependence modeling guided by physical prior is realized, and high-dimensional data features are output; The high-dimensional data features are globally averaged and pooled along the space-time dimension to aggregate into a context vector, the context vector is input into a group of parallel preset linear data decoding functions, mapped to corresponding performance indicators, a plurality of performance indicators are spliced along the feature dimension to construct a dynamic thermal performance feature vector; The dynamic thermal performance feature vector is input into a performance score calculation model based on supervised training of a training data set labeled with dynamic thermal performance feature vectors and authoritative quality scores, and a comprehensive performance score is output.

4. The method according to claim 3, wherein, The entity nodes are divided into material attribute nodes, process parameter nodes, equipment state nodes, and intermediate quality nodes; wherein the material attribute nodes are used to represent the key physical and chemical properties of the input materials, the process parameter nodes are used to represent the set and controlled operation parameters in the production and manufacturing process, the equipment state nodes are used to represent the working state parameters of the production equipment during operation, and the intermediate quality nodes are used to represent the intermediate product attributes that affect subsequent processes.

5. The method according to claim 4, wherein, The directed edges are divided into causal relationship edges, time sequence relationship edges, and subordinate relationship edges; wherein the causal relationship edges represent that the change of one node state will cause the change of another node state, the time sequence relationship edges represent the sequence of the production and manufacturing process, and the subordinate relationship edges represent that the process parameter nodes are inherent attributes of the material attribute nodes or the equipment state nodes.

6. The method of claim 1, wherein, The generation of the production warning instruction includes: identifying a fault node based on the fault probability distribution using a maximum a posteriori probability decision method, matching data in a preset diagnosis-disposal rule library using the fault node as a query index, generating a production warning instruction, and pushing the production warning instruction to a monitoring terminal to trigger a corresponding disposal process.

7. The method of claim 1, wherein the method further comprises: The output step of the optimal control instruction includes: Real-time acquisition of online sensor data, input of the online sensor data into a pre-trained performance predictor model, generation of a predicted thermal performance feature vector; Based on the predicted thermal performance feature vector, the fault probability distribution, and the preset target thermal performance feature vector, a feedforward-feedback coupled mixed optimization objective function is constructed and solved to determine the optimal control instruction.

8. The method according to claim 7, wherein, The mixed optimization objective function is coupled by a performance error term and a diagnostic regularization term; wherein the performance error term is based on a pre-trained forward process model, performs forward inference on the process parameters after applying the assumed parameter adjustment vector, predicts the predicted thermal performance feature vector of the future product, and minimizes the parameter adjustment vector by calculating the square of the Euclidean distance between the predicted thermal performance feature vector and the target thermal performance feature vector; the diagnostic regularization term takes the fault probability distribution as an adaptive penalty weight to constrain the parameter adjustment vector.

9. A hot compress patch process parameter dynamic optimization control system for implementing the hot compress patch process parameter dynamic optimization control method of any one of claims 1-8, characterized in that, The system comprises a multi-modal information acquisition module, a data scoring module, a root cause reasoning module and a control decision module. The multi-modal information acquisition module is used to acquire a point cloud data set and an infrared thermal image time sequence, map the infrared thermal image time sequence to a three-dimensional grid model constructed by the point cloud data set based on the spatial transformation relationship between sensors, and construct a thermodynamic state data set. The data scoring module generates a thermodynamic state sequence based on the thermodynamic state data set using a joint transformation function, performs matrix calculation on the thermodynamic state sequence using a physical prior processing unit introducing a static spatial bias matrix, outputs high-dimensional data features, maps the high-dimensional data features to corresponding performance indicators through a parallel linear data decoding function, inputs the performance indicators into a performance score calculation model, and outputs a comprehensive performance score. The root cause reasoning module determines whether the comprehensive performance score is lower than a preset quality threshold to activate a root cause diagnosis process, and if the root cause diagnosis process is activated, fuses a dynamic thermal performance feature vector and an initial static feature vector of each node in a production process graph data structure to generate node input features, iteratively updates the node input features using a multi-round state iteration calculation method, and outputs fault probability distributions of all nodes. The control decision module identifies fault nodes based on the fault probability distribution through a maximum a posteriori probability decision method, matches data in a preset diagnosis-disposal rule library to generate a production warning instruction, constructs and solves a mixed optimization objective function of feedforward-feedback coupling by comprehensively utilizing the fault probability distribution, a predicted thermal performance feature vector generated based on online sensor data and a preset target thermal performance feature vector, and outputs an optimal control instruction.

Images