A multimodal model mechanism constraint and cross-domain self-adaption method and related device

Abstract

The application discloses a multimodal model mechanism constraint and cross-domain self-adaption method and related device, which is used for improving the stability of production line mechanism constraint in the intelligent manufacturing scene. The method comprises the following steps: receiving production line data, and constructing multimodal operation data according to the production line data; inputting the multimodal operation data into an initial model to generate an initial output mechanism; calculating a mechanism standard vector and a residual error of the initial output mechanism to obtain a mechanism residual error vector; calculating a domain offset vector through a working condition standard vector and the multimodal operation data; projecting the mechanism residual error vector and the domain offset vector to a model space corresponding to the initial output mechanism to construct a correction factor set and a conflict marker quantity; and correcting the initial output mechanism through an asymmetric shunt gate combined with the correction factor set and the conflict marker quantity to obtain a target constraint mechanism.

Description

Technical Field

[0001] This application relates to the field of mechanistic constraints, and in particular to a method and apparatus for mechanistic constraints and cross-domain adaptation of multimodal models. Background Technology

[0002] In smart manufacturing scenarios, the operational status of production lines typically requires comprehensive judgment based on multi-source data, including images, sound, vibration, current, voltage, equipment control signals, process documentation, and quality inspection results. With the application of multimodal models in industrial settings, they can jointly analyze different types of data and be used for equipment status identification, process parameter judgment, quality anomaly detection, and production process optimization.

[0003] In order to improve the reliability of model output results, existing technologies usually introduce physical laws or process rules in the manufacturing process, such as thermal balance relationships, mass conservation relationships, equipment response laws, process window ranges, etc., to reduce the situation in model output results that obviously do not conform to manufacturing laws and improve the stability of production line analysis results.

[0004] In actual production line operation, model output deviations may originate from two types of factors simultaneously: one is that the production process has deviated from the established technological rules, and the other is normal production variations caused by changes in equipment, materials, environment, or operating conditions. Existing methods typically treat these two types of factors uniformly. When the constraint rules are set too strongly, they tend to suppress normal production variations, leading to a decrease in the model's adaptability to new production lines, new equipment, or new material batches. Conversely, when the constraint rules are set too weakly, they tend to amplify output results that do not conform to manufacturing rules, affecting the reliability of the model's judgments. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a multimodal model mechanism constraint and cross-domain adaptive method and related apparatus to improve the stability of production line mechanism constraints in intelligent manufacturing scenarios.

[0006] The technical solution provided in this application is described below:

[0007] The first aspect of this application provides a multimodal model mechanism constraint and cross-domain adaptation method, including: Receive production line data and construct multimodal operation data based on the production line data, wherein the production line data includes deployment data and operation data; The multimodal operating data is input into the initial model to generate the initial output mechanism; The residual vector of the mechanism is obtained by calculating the standard vector of the mechanism and the residual of the initial output mechanism; The domain offset vector is calculated using the standard operating condition vector and the multimodal operating data. The mechanism residual vector and the domain offset vector are projected onto the model space corresponding to the initial output mechanism to construct a set of correction factors and a conflict label quantity. The set of correction factors includes mechanism factors, domain offset factors, overlap and conflict factors, and neutral factors. The initial output mechanism is modified by combining the set of correction factors and the conflict marker quantity with asymmetric diversion gating to obtain the target constraint mechanism.

[0008] Optionally, the step of projecting the mechanism residual vector and the domain offset vector onto the model space corresponding to the initial output mechanism to construct a set of correction factors and a conflict marker includes: The mechanism residual vector and the domain offset vector are projected onto the model space corresponding to the initial output mechanism, and the mechanism residual direction basis and the domain offset direction basis are obtained by spatial coordinate correlation calculation. Calculate the directional phase between the mechanistic residual directional basis and the domain offset directional basis, and generate a conflict marker quantity based on the directional phase; A subspace is constructed based on the described mechanism residual direction basis and the described domain offset direction basis; The initial output mechanism is decomposed into subspace conflicts based on the phase relationship of the subspace to obtain a set of correction factors, which includes mechanism factors, domain offset factors, overlap conflict factors, and neutral factors.

[0009] Optionally, the step of modifying the initial output mechanism by combining the set of correction factors and the conflict marker quantity with asymmetric splitting gating to obtain the target constraint mechanism includes: The offset direction of each factor in the set of correction factors is determined based on the conflict marker amount; The correction strength is determined based on the absolute value of each factor in the set of correction factors; The initial output mechanism is corrected based on the correction intensity and the offset direction to obtain the target constraint mechanism.

[0010] Optionally, the initial output mechanism is modified according to the correction intensity and the offset direction to obtain the target constraint mechanism, including: By associating the correction intensity with the offset direction, an offset vector is obtained; The initial output mechanism is decomposed based on the offset vector, and the initial output mechanism after the decomposed mechanism is corrected item by item by the offset vector to obtain the target constraint mechanism.

[0011] Optionally, the step of calculating the domain offset vector using the standard operating condition vector and the multimodal operating data includes: Extract operating condition parameters from the multimodal operating data; Construct the current operating condition vector based on the operating condition parameters; The difference between the current operating condition vector and the standard operating condition vector is calculated to obtain the domain offset vector.

[0012] Optionally, the step of inputting the multimodal operating data into the initial model to generate the initial output mechanism includes: Build a time window using the initial model; Establish the time series association between the multimodal operation data and the time window to obtain the time series operation vector; The initial model is used to analyze the correlation state of the time-series running vectors to obtain the initial output mechanism.

[0013] Optionally, calculating the residual between the mechanism standard vector and the initial output mechanism to obtain the mechanism residual vector includes: The initial output mechanism is vectorized to obtain an initial mechanism vector, which is used to align the vector dimensions of the initial output mechanism with those of the standard mechanism vector. The difference between the initial mechanism vector and the standard mechanism vector is calculated to obtain the mechanism residual vector.

[0014] Optionally, receiving production line data and constructing multimodal operation data based on the production line data includes: Receive production line data, which includes deployment data and operational data; Generate a mechanism standard vector and an operating condition standard vector based on the deployment data; The runtime data is lexicalized to obtain multimodal runtime data.

[0015] A second aspect of this application provides a multimodal model mechanism constraint and cross-domain adaptive system, the system comprising: A receiving unit is used to receive production line data and construct multimodal operation data based on the production line data, wherein the production line data includes deployment data and operation data; The input unit is used to input the multimodal operation data into the initial model and generate the initial output mechanism; The first calculation unit is used to calculate the residual between the mechanism standard vector and the initial output mechanism to obtain the mechanism residual vector; The second calculation unit is used to calculate the domain offset vector using the standard working condition vector and the multimodal operating data. A construction unit is used to project the mechanism residual vector and the domain offset vector onto the model space corresponding to the initial output mechanism, and construct a set of correction factors and a conflict marker quantity. The set of correction factors includes mechanism factors, domain offset factors, overlap conflict factors, and neutral factors. The correction unit is used to correct the initial output mechanism by combining the correction factor set and the conflict marker quantity with asymmetric splitting gating to obtain the target constraint mechanism.

[0016] Optionally, the building unit is specifically used for: The mechanism residual vector and the domain offset vector are projected onto the model space corresponding to the initial output mechanism, and the mechanism residual direction basis and the domain offset direction basis are obtained by spatial coordinate correlation calculation. Calculate the directional phase between the mechanistic residual directional basis and the domain offset directional basis, and generate a conflict marker quantity based on the directional phase; A subspace is constructed based on the described mechanism residual direction basis and the described domain offset direction basis; The initial output mechanism is decomposed into subspace conflicts based on the phase relationship of the subspace to obtain a set of correction factors, which includes mechanism factors, domain offset factors, overlap conflict factors, and neutral factors.

[0017] Optionally, the correction unit is specifically used for: The offset direction of each factor in the set of correction factors is determined based on the conflict marker amount; The correction strength is determined based on the absolute value of each factor in the set of correction factors; The initial output mechanism is corrected based on the correction intensity and the offset direction to obtain the target constraint mechanism.

[0018] Optionally, the correction unit is further configured to: By associating the correction intensity with the offset direction, an offset vector is obtained; The initial output mechanism is decomposed based on the offset vector, and the initial output mechanism after the decomposed mechanism is corrected item by item by the offset vector to obtain the target constraint mechanism.

[0019] Optionally, the second computing unit is specifically used for: Extract operating condition parameters from the multimodal operating data; Construct the current operating condition vector based on the operating condition parameters; The difference between the current operating condition vector and the standard operating condition vector is calculated to obtain the domain offset vector.

[0020] Optionally, the input unit is specifically used for: Build a time window using the initial model; Establish the time series association between the multimodal operation data and the time window to obtain the time series operation vector; The initial model is used to analyze the correlation state of the time-series running vectors to obtain the initial output mechanism.

[0021] Optionally, the first computing unit is specifically used for: The initial output mechanism is vectorized to obtain an initial mechanism vector, which is used to align the vector dimensions of the initial output mechanism with those of the standard mechanism vector. The difference between the initial mechanism vector and the standard mechanism vector is calculated to obtain the mechanism residual vector.

[0022] Optionally, the receiving unit is specifically used for: Receive production line data, which includes deployment data and operational data; Generate a mechanism standard vector and an operating condition standard vector based on the deployment data; The runtime data is lexicalized to obtain multimodal runtime data.

[0023] A third aspect of this application provides a multimodal model mechanism constraint and cross-domain adaptation device, the device comprising: Processor, memory, input / output units, and bus; The processor is connected to the memory, the input / output unit, and the bus; The memory stores a program, which the processor invokes to execute the first aspect and any one of the optional methods in the first aspect.

[0024] A fourth aspect of this application provides a computer-readable storage medium storing a program that, when executed on a computer, performs the methods of the first aspect and any one of the first aspects.

[0025] As can be seen from the above technical solutions, this application has the following advantages: By constructing multimodal operating data and generating initial output mechanisms, combining the mechanism standard vector to obtain the mechanism residual vector, and combining the operating condition standard vector to obtain the domain offset vector, and then projecting the two onto the model space to form a set of correction factors and conflict markers, the initial output mechanism can be differentiated through asymmetric diversion gating, reducing the erroneous corrections caused by the mutual interference between the mechanism residual vector and the domain offset vector, and improving the mechanism consistency and cross-domain adaptive capability of the target constraint mechanism. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a schematic flowchart of an embodiment of the multimodal model mechanism constraint and cross-domain adaptive method in this application; Figure 2a This is a schematic flowchart of an embodiment of the first stage of the multimodal model mechanism constraint and cross-domain adaptive method in this application; Figure 2b This is a schematic flowchart of an embodiment of the second stage of the multimodal model mechanism constraint and cross-domain adaptive method in this application; Figure 3 This is a schematic diagram of an embodiment of the multimodal model mechanism constraint and cross-domain adaptive system in this application; Figure 4 This is a schematic diagram of an embodiment of the multimodal model mechanism constraint and cross-domain adaptive device in this application. Detailed Implementation

[0028] The execution subject of the method provided in this embodiment can be an electronic device with data processing capabilities, a server, an edge computing device, an industrial control device, a production line management platform, an industrial internet platform, or a computing system that deploys a multimodal model. This embodiment does not specifically limit this.

[0029] In one optional implementation, the execution entity can be located on the production line side to receive production line data and process the production line's operating status in real time. In another optional implementation, the execution entity can also be located on the cloud or server side to centrally process production line data corresponding to one or more production lines. In yet another optional implementation, the execution entity can also be implemented collaboratively by production line-side equipment and cloud equipment, where the production line-side equipment is used to collect or preprocess production line data, and the cloud equipment is used to perform model calculations and mechanistic constraint corrections.

[0030] For ease of explanation, the following embodiments use a computing device as the execution subject to illustrate the multimodal model mechanism constraint and cross-domain adaptive method provided in the embodiments of this application. This example does not constitute a limitation on the scope of protection of this application.

[0031] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0032] Please see Figure 1 This application first provides an embodiment of a multimodal model mechanism constraint and cross-domain adaptive method, which includes: S101. Receive production line data and construct multimodal operation data based on the production line data, wherein the production line data includes deployment data and operation data; The computing device receives production line data, which includes deployment data and operational data. Deployment data provides baseline information established during the production line deployment phase, while operational data provides real-time or phased information generated during the production line operation phase. Operational data covers data content from different sources, in different collection formats, and with different business implications within the production line, enabling the computing device to acquire the production line's operational status from multiple dimensions.

[0033] The computing equipment organizes production line data in a unified manner, transforming it into multimodal operational data suitable for initial model processing. This multimodal operational data preserves the correspondence between different data sources and enables visual data, sensor data, equipment status data, process text data, and quality inspection result data to participate in subsequent model processing. Through this processing, data generated scattered throughout the production line is organized into a unified input object, providing a data foundation for generating the initial output mechanism.

[0034] S102. Input the multimodal operation data into the initial model to generate the initial output mechanism; The computing device inputs multimodal operating data into the initial model. The initial model fuses the production line status, process information, equipment information, and quality information from the multimodal operating data, and generates an initial output mechanism based on the correlation between different information in the multimodal operating data.

[0035] The initial output mechanism represents the mechanistic analysis results of the initial model on the production line operation process. This initial output mechanism reflects the relationship between the operating state, process state, and equipment state corresponding to the multimodal operating data, providing the constraint objects for subsequent mechanism residual vector calculation and correction processing.

[0036] In this embodiment, the model space corresponding to the initial output mechanism is the intermediate computational space formed after the initial model undergoes multimodal fusion processing. This model space is located after the multimodal operational data has been fused and parsed, and before the lightweight domain adaptation processing, enabling the subsequent mechanism residual vector and domain offset vector to participate in the correction during the intermediate computation stage of the initial model. The computing device retains the operational relationships, equipment relationships, process relationships, and quality relationships corresponding to the multimodal operational data in this model space, allowing the initial output mechanism to serve as a common carrier for mechanism constraint correction and cross-domain adaptive correction.

[0037] S103. The residual of the initial output mechanism is calculated by calculating the standard vector of the mechanism and the residual of the mechanism; The computing device acquires the mechanism standard vector and uses it as a benchmark to measure whether the initial output mechanism conforms to the production line mechanism constraints. The mechanism standard vector corresponds to the mechanism constraint information formed by the production line during the deployment phase or under baseline operating conditions.

[0038] The computing device performs residual calculations on the mechanism standard vector and the initial output mechanism to obtain the mechanism residual vector. The mechanism residual vector represents the degree of deviation of the initial output mechanism from the mechanism standard vector, and this degree of deviation is used to indicate the part of the initial output mechanism that needs to be corrected for mechanism constraints.

[0039] S104. Calculate the domain offset vector using the standard working condition vector and the multimodal operation data; The computing equipment acquires the standard operating condition vector and performs offset calculations on the current operating condition corresponding to the multimodal operating data based on the standard operating condition vector. The standard operating condition vector corresponds to the operating condition baseline formed by the production line under standard operating conditions or source domain operating conditions.

[0040] The computing device obtains the domain offset vector by analyzing the difference between the standard operating condition vector and the multimodal operating data. The domain offset vector represents the offset of the current production line, current equipment, current material batch, or current operating environment relative to the standard operating condition. This offset is used to distinguish the source of influence between normal cross-domain changes and mechanistic deviations.

[0041] S105. Project the mechanism residual vector and the domain offset vector onto the model space corresponding to the initial output mechanism to construct a set of correction factors and a conflict marker quantity. The set of correction factors includes mechanism factors, domain offset factors, overlap conflict factors, and neutral factors. The computing device maps the mechanism residual vector and the domain offset vector to the model space corresponding to the initial output mechanism, enabling the mechanism residual vector and the domain offset vector to participate in the construction of correction information within the same computational space. The model space corresponds to the computational representation space of the initial output mechanism and is used to carry the influence relationship between mechanism deviation and operating condition deviation on the initial output mechanism.

[0042] The computing device constructs a set of correction factors and conflict markers based on the relationship between the mechanism residual vector and the domain offset vector in the model space. The set of correction factors includes mechanism factors, domain offset factors, overlapping conflict factors, and neutral factors. Mechanism factors represent correction components caused by mechanism residuals, domain offset factors represent adaptation components caused by operating condition offsets, overlapping conflict factors represent components jointly affected by mechanism residuals and operating condition offsets, and neutral factors represent components that maintain the stability of the initial output mechanism. Conflict markers represent the influence relationship between the mechanism residual vector and the domain offset vector, providing a control basis for subsequent asymmetric split gating.

[0043] S106. The initial output mechanism is modified by combining the set of correction factors and the conflict marker quantity with asymmetric diversion gating to obtain the target constraint mechanism.

[0044] The computing device processes the set of correction factors and the amount of conflict markers through asymmetric split gating. The asymmetric split gating differentiates the different factors in the set of correction factors based on the amount of conflict markers, so that the mechanism factor, domain offset factor, overlapping conflict factor, and neutral factor participate in the adjustment of the initial output mechanism with different correction effects.

[0045] The computing device uses the correction results of the asymmetric shunt gating output to modify the initial output mechanism, thereby obtaining the target constraint mechanism. The target constraint mechanism retains the effective operational relationships in the initial output mechanism and combines the mechanism residual vector and the domain offset vector to complete the constraint adjustment, so that the output result has both mechanism consistency and cross-domain adaptability.

[0046] It should be noted that, in this embodiment, the constraint mechanism is used to constrain the initial output mechanism generated by the initial model based on multimodal operating data. Specifically, the constraint object is the computational component in the initial output mechanism that is inconsistent with the production line mechanism rule. The computing device determines the constraint range corresponding to the production line mechanism rule through the mechanism standard vector, and locates the component in the initial output mechanism that deviates from the constraint range through the mechanism residual vector. For components that deviate from thermal balance constraints, mass conservation constraints, equipment response constraints, or process window boundary constraints, the computing device applies a correction towards the mechanism standard vector through mechanism factors. For effective operating condition offset components formed by the current production line, equipment status, material batch, or operating environment, the computing device retains its adaptation direction through domain offset factors. For components that are simultaneously affected by the mechanism residual vector and the domain offset vector, the computing device controls the correction direction and correction intensity of the overlapping conflict factor according to the conflict marker amount, so that the component is split and corrected between mechanism constraints and operating condition adaptation. Through the above processing, the constraint mechanism is incorporated into the correction process of the initial output mechanism, constraining and converging the components that violate the production line mechanism rules, preserving the differences in effective operating conditions, and controlling the adjustment of conflicting components.

[0047] This embodiment constructs multimodal operating data and generates an initial output mechanism. It combines the mechanism standard vector to obtain the mechanism residual vector and the operating condition standard vector to obtain the domain offset vector. Then, it projects the two onto the model space to form a set of correction factors and conflict markers. It can differentiate the initial output mechanism through asymmetric diversion gating, reduce the erroneous correction caused by the mutual interference between the mechanism residual vector and the domain offset vector, and improve the mechanism consistency and cross-domain adaptive capability of the target constraint mechanism.

[0048] Please see Figures 2a to 2bThis application provides another embodiment of the multimodal model mechanism constraint and cross-domain adaptive method, which includes: S201. Receive production line data, the production line data including deployment data and operation data; The computing device receives production line data. This production line data includes deployment data and operational data. Deployment data originates from baseline information generated during the production line deployment phase, while operational data originates from real-time or periodic information generated during the production line operation phase.

[0049] Deployment data includes production line baseline configuration, equipment baseline status, process baseline conditions, material batch baseline information, and mechanistic constraint baseline information. Operational data includes visual data, sensor sequence data, equipment status data, process text data, and quality inspection result data collected during production line operation. The computing device uses the deployment and operational data as the input foundation for constructing multimodal operational data.

[0050] S202. Generate a mechanism standard vector and an operating condition standard vector based on the deployment data; The computing device generates mechanistic standard vectors and operating condition standard vectors based on the deployment data. It extracts mechanistic constraint baseline information from the deployment data and converts this information into mechanistic standard vectors. Mechanism constraint baseline information includes, but is not limited to, thermal balance constraints, mass conservation constraints, kinematic constraints, heat conduction constraints, equipment transfer function constraints, and process window boundary constraints. The mechanistic standard vectors carry the mechanistic constraint relationships of the production line under baseline conditions and are used in subsequent calculations of mechanistic residual vectors.

[0051] The computing device extracts baseline information from the deployment data and converts it into a standard vector. This standard vector carries the production line's operating condition under baseline conditions and is used in subsequent domain offset vector calculations. Through this process, the mechanistic baseline and operating condition baseline in the deployment data are respectively transformed into the standard vectors required for subsequent calculations.

[0052] S203. Lexicalize the running data to obtain multimodal running data.

[0053] The computing device performs lexicalization on the operational data to obtain multimodal operational data. The computing device converts the visual data, sensor sequence data, equipment status data, process text data, and quality inspection result data in the operational data into lexical forms that the initial model can process, and organizes the lexicals from different sources according to the correspondence in the production line operation process.

[0054] During the lexicalization process, the computing device retains information on modal origin, acquisition time, equipment status, and process stage from the operational data, enabling multimodal operational data to simultaneously carry the operational content of different modalities and their interrelationships. The resulting multimodal operational data serves as input to the initial model, used to generate the initial output mechanism.

[0055] The computing device also generates modal reliability information based on the data integrity, acquisition stability, and temporal continuity of each modality in the operational data. This modal reliability information corresponds to the reliability of visual data, vibration data, acoustic data, PLC status data, and process text data at the current moment, and is incorporated into subsequent processing along with the multimodal operational data. When constructing the correction factor set and asymmetric shunt gating correction, the computing device uses the modal reliability information as an auxiliary basis for adjusting the intensity of different modal influences.

[0056] S204. Construct a time window using the initial model; After the computing device inputs the multimodal operation data into the initial model, the initial model constructs a time window based on the acquisition time, operation cycle, and production line event boundaries in the multimodal operation data. The time window covers the multimodal data generated within the same production line operation segment, forming a unified processing range for different sampling frequencies, acquisition delays, and data sources.

[0057] When constructing the time window, the initial model grouped visual data, sensor sequence data, equipment status data, process text data, and quality inspection result data into the same operational segment, and organized the data into windows according to the production line's operating rhythm. After time window processing, the discretely acquired information in the multimodal operational data was organized into window data that could reflect the stage-specific status of the production line.

[0058] When constructing the time window, the computing device also determines the range of data to be calculated within the window based on production line-side inference latency, computing resource consumption, and the scale of the lightweight adaptation path. For mechanism direction information, equipment model offset information, and material batch offset templates already formed during the deployment phase, the computing device calls the corresponding cached results within the current time window and calculates the residual correction amount related to the current operating state during the online processing phase. Through this process, the amount of computation within the time window is kept in line with the requirements of online inference.

[0059] S205. Establish the time series association between the multimodal operation data and the time window to obtain the time series operation vector; The computing device establishes a time-series association between multimodal operational data and the time window based on the data arrangement within the time window. This time-series association maps data from different modalities to corresponding positions within the same time window according to the chronological order of the production line operation, thus establishing a correspondence between data from different sources in the time dimension.

[0060] The computing device vectorizes the multimodal operational data according to the associated time sequence, resulting in a time-series operational vector. The time-series operational vector preserves the sequential, synchronous, and event relationships of the multimodal operational data within a time window, enabling the initial model to perform subsequent analysis based on continuous changes during production line operation.

[0061] S206. The correlation state of the time-series running vector is analyzed through the initial model to obtain the initial output mechanism.

[0062] The computing device uses an initial model to analyze the relationships in the time-series operation vector. Based on the time correspondence and state relationships between different modes in the time-series operation vector, the initial model analyzes the connections between production line operation status, equipment response status, process execution status, and quality change status.

[0063] The computing device generates an initial output mechanism based on the correlation state analysis results. The initial output mechanism carries the mechanistic output results formed by multimodal operating data within the time window, and serves as the processing object for subsequent mechanism residual vector calculation, domain offset vector calculation, and asymmetric shunt gating correction.

[0064] S207. Vectorize the initial output mechanism to obtain an initial mechanism vector, which is used to align the vector dimensions of the initial output mechanism with the standard mechanism vector. The computing device vectorizes the initial output mechanism, converting the production line operation relationships, equipment response relationships, process state relationships, and quality state relationships in the initial output mechanism into a unified vector form, resulting in the initial mechanism vector. The initial mechanism vector and the standard mechanism vector use the same vector dimension, enabling residual calculations to be performed under the same computational dimension.

[0065] During the vectorization process, the computing device arranges and maps the initial output mechanisms according to the dimensional structure of the standard mechanism vector, so that each mechanism calculation result in the initial output mechanism enters the corresponding vector position. After this processing, the initial output mechanism is converted into an initial mechanism vector that can be directly compared with the standard mechanism vector.

[0066] S208. Calculate the difference between the initial mechanism vector and the mechanism standard vector to obtain the mechanism residual vector.

[0067] The computing device calculates the difference between the initial mechanism vector and the standard mechanism vector to obtain the mechanism residual vector. The computing device calculates the deviation of the initial mechanism vector from the standard mechanism vector item by item according to the vector dimension, so that the mechanism deviation in the initial output mechanism is converted into a residual vector that can participate in subsequent model space processing.

[0068] When the mechanistic residual vector is used in subsequent processing, it serves as the basis for calculating the mechanistic constraint direction. The computing device allocates time to different residual components within the mechanistic residual vector based on its changes over a continuous time window. For residual components that change rapidly with the current operational impact, the computing device includes them in the current time window; for residual components that change slowly with temperature changes, heat accumulation, or equipment wear, the computing device includes them in the residual echo information across time windows; for residual components formed by material batch drift or switching to new operating conditions, the computing device includes them in the observation confirmation state. Through this processing, the deviation of the initial output mechanism from the standard mechanistic vector is clearly defined in the vector dimension, forming mechanistic residual inputs that can participate in subsequent model space processing.

[0069] S209. Extract operating condition parameters from the multimodal operation data; The computing device extracts data items related to the current operating condition from the multimodal operating data. The multimodal operating data includes the current production line operating status, equipment status, material batch status, environmental status, and process execution status. The computing device extracts operating condition parameters that reflect the current operating conditions from the above data according to the operating condition identification requirements.

[0070] When extracting operating condition parameters, the computing device incorporates the current production line identifier, equipment operating status, equipment aging status, material batch information, process window status, environmental temperature and humidity changes, sensor drift information, and status information from a small sample operating condition window into the range of operating condition parameters. These operating condition parameters collectively reflect the sources of change in the current operating data relative to the baseline operating condition, enabling the subsequent domain offset vector to reflect the differences in operating conditions caused by different production lines, different equipment, different material batches, and different operating environments.

[0071] S210. Construct the current working condition vector based on the working condition parameters; The computing device vectorizes the operating condition parameters to obtain the current operating condition vector. Based on the type and numerical meaning of the operating condition parameters, the computing device converts discrete operating condition parameters into coded values, continuous operating condition parameters into normalized values, and state-type operating condition parameters into state variables, and arranges the operating condition parameters within the same time range into a unified vector structure.

[0072] Once the current operating condition vector is formed, the computing device unifies the operating condition information corresponding to the current production line, current equipment, current material batch, and current operating environment into a single vector object. Through this process, the current operating condition vector can perform difference calculations with the standard operating condition vector, allowing changes in the current operating condition relative to the baseline operating condition to participate in subsequent processing within the same dimension.

[0073] S211. Calculate the difference between the current working condition vector and the working condition standard vector to obtain the domain offset vector.

[0074] The computing device aligns the current operating condition vector with the standard operating condition vector in terms of dimensions and calculates the difference between the two to obtain the domain offset vector. The standard operating condition vector is formed from the baseline operating condition corresponding to the deployed data, while the current operating condition vector is formed from the real-time operating condition corresponding to the multimodal operating data. The difference between the two reflects the offset of the current operating condition relative to the baseline operating condition.

[0075] The computing device generates a domain offset vector based on the difference results. When this domain offset vector is processed in the subsequent model space, it serves as input for cross-domain adaptive operating condition changes, participating in the construction of the correction factor set and conflict marker quantity. In this way, operational differences caused by the current production line, equipment aging status, material batch changes, environmental changes, and small sample operating condition windows are converted into domain offset information that can participate in model correction. This approach follows the technical chain described in the disclosure document, where the domain offset vector is calculated from the current production line, material batch, equipment aging status, and small sample operating condition windows.

[0076] S212. Project the mechanism residual vector and the domain offset vector onto the model space corresponding to the initial output mechanism, and calculate the mechanism residual direction basis and the domain offset direction basis through spatial coordinate correlation. After acquiring the mechanism residual vector and domain offset vector, the computing device determines the model space corresponding to the initial output mechanism. This model space is an intermediate computing space formed when the initial model generates the initial output mechanism, and it is located before the multimodal fusion processing results enter the lightweight domain adaptation path. Within this model space, the computing device carries the correlation between production line operating status, mechanism deviation information, and operating condition offset information, enabling the mechanism residual vector and domain offset vector to participate in directional calculations within the same intermediate computing space.

[0077] The computing device projects the mechanism residual vector onto the model space and determines the mechanism residual direction basis based on the coordinate distribution of the mechanism residual vector in the model space. The mechanism residual direction basis is used to represent the direction in which the mechanism residual vector exerts a constraint effect on the initial output mechanism, thus transforming the mechanism residual vector from a single residual quantity into directional information that can participate in model space correction.

[0078] The computing device projects the domain offset vector into the model space and determines the domain offset direction basis based on the coordinate distribution of the domain offset vector in the model space. The domain offset direction basis is used to represent the direction in which the domain offset vector has a cross-domain adaptation effect on the initial output mechanism, so that the domain offset vector is transformed from a single operating condition offset into directional information that can participate in model space correction.

[0079] When determining the mechanistic residual direction basis, the computing device extracts the dominant direction of the mechanistic residual vector in the model space based on the local variation relationship between the mechanistic residual vector and each calculated component in the initial output mechanism. This dominant direction is formed by the local variation relationship of the mechanistic residual with respect to the predicted variable, the sensitivity relationship between mechanistic constraint nodes, the response relationship after unit normalization of physical quantities, or the local variation relationship of the equipment transfer function. When determining the domain offset direction basis, the computing device extracts the dominant change direction of the current production line, equipment status, material batch, and small sample operating window in the model space based on the offset distribution of the current operating condition vector relative to the standard operating condition vector.

[0080] The mechanistic residual directional basis and the domain offset directional basis reside in the same model space, enabling the computing device to identify the relationship between the mechanistic constraint influence and the cross-domain offset influence in the same spatial coordinates. Through this process, the constraint direction corresponding to the mechanistic residual vector and the adaptation direction corresponding to the domain offset vector are unified into the model space corresponding to the initial output mechanism, providing a directional basis for subsequent conflict marker generation and correction factor set construction.

[0081] S213. Calculate the directional phase between the mechanism residual directional basis and the domain offset directional basis, and generate a conflict marker quantity based on the directional phase; The computing device calculates the directional phase between the mechanistic residual directional basis and the domain offset directional basis. The directional phase is used to represent the relative directional relationship between the mechanistic residual directional basis and the domain offset directional basis in the model space. This relative directional relationship includes directional overlap, directional cancellation, and directional separation.

[0082] When calculating the directional phase, the computing device uses the directional angle, directional correlation, or directional overlap between the mechanistic residual directional basis and the domain offset directional basis as the basis for calculation. The directional overlap is obtained from the inner product relationship or equivalent correlation relationship between the mechanistic residual directional basis and the domain offset directional basis. The directional phase tends to overlap as the directional overlap increases and tends to separate as the directional overlap decreases, and forms a canceling state when the two acting directions have opposite trends.

[0083] When the directional phase representation mechanism residual directional basis and the domain offset directional basis have directional overlap in model space, it indicates that the constraint influence corresponding to the mechanism residual vector and the adaptation influence corresponding to the domain offset vector act on the similar computational regions in the initial output mechanism. The computing device then generates a conflict marker quantity to record the overlap relationship.

[0084] When the directional phase representation mechanism residual directional basis and the domain offset directional basis have a directional cancellation relationship in the model space, it indicates that the constraint influence corresponding to the mechanism residual vector and the adaptation influence corresponding to the domain offset vector have a mutual weakening trend in the initial output mechanism. The computing device then generates a conflict marker quantity to record the cancellation relationship.

[0085] When the directional phase representation mechanism residual directional basis and the domain offset directional basis have a directional separation relationship in the model space, it indicates that the constraint influence corresponding to the mechanism residual vector and the adaptation influence corresponding to the domain offset vector act on different computational regions in the initial output mechanism, respectively. The computing device then generates a conflict marker quantity to record the separation relationship.

[0086] After generating conflict marker quantities through directional phase, the computing device converts the spatial relationship between the mechanistic residual directional basis and the domain offset directional basis into control information that can be used by subsequent asymmetric shunt gating, so that the subsequent correction process can distinguish the influence of mechanistic constraints, cross-domain adaptation influence, and the conflict influence between the two.

[0087] S214. Construct a subspace based on the mechanism residual direction basis and the domain offset direction basis; The computing device constructs a subspace based on the distribution relationship of the mechanism residual direction basis and the domain offset direction basis in the model space. The subspace is used to hold the computational components in the initial output mechanism corresponding to different sources of influence, so that the initial output mechanism can complete the spatial-level influence division before correction.

[0088] The computing device determines the spatial region corresponding to the influence of mechanistic constraints based on the mechanistic residual direction basis. This spatial region is used to accommodate the computational components in the initial output mechanism that are mainly affected by the mechanistic residual vector. The computing device also determines the spatial region corresponding to the cross-domain adaptation influence based on the domain offset direction basis. This spatial region is used to accommodate the computational components in the initial output mechanism that are mainly affected by the domain offset vector.

[0089] The computing device determines overlapping conflict regions based on the combined influence of the mechanistic residual directional basis and the domain offset directional basis. These overlapping conflict regions are used to accommodate computational components in the initial output mechanism that are simultaneously affected by both the mechanistic residual vector and the domain offset vector. The computing device also determines neutral regions in the model space based on regions with low correlation to both the mechanistic residual directional basis and the domain offset directional basis. These neutral regions are used to accommodate computational components in the initial output mechanism that remain stable.

[0090] The computing device forms corresponding projection relationships based on the aforementioned spatial regions. Mechanism constraints influence the formation of mechanistic projection relationships for corresponding spatial regions, cross-domain adaptation influences the formation of domain offset projection relationships for corresponding spatial regions, overlapping and conflicting regions form conflict projection relationships, and neutral regions form neutrality-preserving projection relationships. These projection relationships distribute the computational components in the initial output mechanism to different subspaces, enabling asymmetric flow gating to apply different correction rules to different subspaces.

[0091] The subspace constructed in this way can correspond to the effects of mechanistic constraints, cross-domain adaptation, overlapping conflict, and neutrality preservation. Through this subspace, the initial output mechanism is organized in the model space into a structure suitable for diversion correction, providing a spatial basis for the generation of subsequent correction factor sets.

[0092] S215. The initial output mechanism is decomposed into subspace conflict through the phase relationship of the subspace to obtain a set of correction factors, which includes mechanism factors, domain offset factors, overlap conflict factors, and neutral factors.

[0093] The computing device performs subspace conflict decomposition on the initial output mechanism based on the phase relationship of the subspace. The phase relationship of the subspace is formed by the directional phase transfer between the mechanism residual direction basis and the domain offset direction basis, and is used to determine the correspondence between each computational component in the initial output mechanism and the mechanism residual vector and the domain offset vector.

[0094] The computing device decomposes the computational components in the initial output mechanism that fall within the spatial region affected by the mechanism constraints into mechanism factors. The mechanism factors are used to represent the components in the initial output mechanism that are affected by the mechanism residual vector and need to participate in the mechanism constraint correction.

[0095] The computing device decomposes the computational components in the initial output mechanism that fall into the spatial region affected by cross-domain adaptation into domain offset factors. The domain offset factor is used to represent the components in the initial output mechanism that are affected by the domain offset vector and need to participate in cross-domain adaptation correction.

[0096] The computing device decomposes the computational components in the initial output mechanism that fall into overlapping conflict regions into overlapping conflict factors. The overlapping conflict factor is used to represent the components in the initial output mechanism that are simultaneously affected by the mechanism residual vector and the domain offset vector. This component corresponds to the part where there is a conflict between the mechanism constraints and the cross-domain adaptation.

[0097] The computing device decomposes the computational components in the initial output mechanism that fall into the neutral region into neutral factors. The neutral factors represent components in the initial output mechanism that are less correlated with the mechanism residual vector and the domain offset vector, and these components are used to preserve stable information in the initial output mechanism.

[0098] Mechanism factors, domain offset factors, overlap and conflict factors, and neutral factors together form a set of correction factors. The set of correction factors, together with the conflict marker, is input into the subsequent asymmetric split gating, enabling the asymmetric split gating to differentiate the initial output mechanism based on the source and conflict relationships of different factors.

[0099] S216. Determine the offset direction of each factor in the set of correction factors based on the conflict marker amount; The computing device inputs conflict markers into the asymmetric splitting gating. The conflict markers record the directional relationship between the mechanistic residual direction basis and the domain offset direction basis. The asymmetric splitting gating determines the offset direction of each factor in the correction factor set based on this directional relationship and determines the adaptation state of the lightweight domain adaptation path. The adaptation states include mechanistic locked state, domain released state, and glide state. The mechanistic locked state corresponds to the constraint direction of the mechanistic factor, the domain released state corresponds to the adaptation direction of the domain offset factor, and the glide state corresponds to the low-speed update direction of the overlapping conflict factor. The computing device determines the adaptation state based on the mechanistic residual vector, the domain offset vector, the direction phase, and the modal confidence information, and uses the adaptation state as the control condition for determining the offset direction of each factor.

[0100] When the conflict marker indicates a separation effect between the mechanism residual vector and the domain offset vector, asymmetric split gating determines the offset direction of the mechanism factor as the constraint direction tending towards the mechanism standard vector, and determines the offset direction of the domain offset factor as the adaptation direction that retains the changes in the current operating condition. This process allows the mechanism factor and the domain offset factor to participate in the correction according to their respective corresponding influence directions.

[0101] When the conflict marker indicates an overlapping effect between the mechanism residual vector and the domain offset vector, the asymmetric split gating switches the lightweight domain adaptation path to a sliding state and determines the offset direction of the overlapping conflict factor as the sliding correction direction. The sliding correction direction performs a low-speed coupling adjustment along the direction of the combined action of the mechanism residual vector and the domain offset vector, allowing the overlapping conflict factor to participate in subsequent corrections in a weak update manner, while retaining the effective changes in the target condition that need to be further observed and confirmed.

[0102] When the conflict marker indicates a cancellation effect between the mechanism residual vector and the domain offset vector, the asymmetric splitting gating determines the offset direction of the overlapping conflict factor based on the conflict marker and limits the offset directions of the mechanism factor and the domain offset factor, ensuring that the mechanism constraint effect and the cross-domain adaptation effect remain distinguishable during the correction process. The offset direction of the neutral factor is determined as the stability-maintaining direction, used to preserve components in the initial output mechanism that are less correlated with the mechanism residual vector and the domain offset vector.

[0103] S217. Determine the correction strength based on the absolute value of each factor in the set of correction factors; The computing device calculates the absolute value of each factor in the set of correction factors and determines the corresponding correction intensity based on the absolute value. The correction intensity is used to represent the magnitude of the effect of the mechanistic factor, domain offset factor, overlap conflict factor, and neutral factor when participating in the initial output mechanistic correction.

[0104] The computing device determines the intensity of mechanism constraint correction based on the absolute value of the mechanism factor. The absolute value of the mechanism factor represents the degree to which the initial output mechanism is affected by the mechanism residual vector. The intensity of mechanism constraint correction increases with the degree of this influence, so that components with larger deviations from the mechanism receive stronger constraint correction.

[0105] The computing device determines the cross-domain adaptation correction strength based on the absolute value of the domain offset factor. The absolute value of the domain offset factor represents the degree to which the initial output mechanism is affected by the domain offset vector. The cross-domain adaptation correction strength increases with the degree of this influence, so that the components corresponding to the changes in the current operating conditions receive the corresponding adaptation correction.

[0106] The computing device determines the conflict adjustment intensity based on the absolute value of the overlap conflict factor. The absolute value of the overlap conflict factor represents the degree to which the mechanism residual vector and the domain offset vector jointly act on the initial output mechanism, and the conflict adjustment intensity is used to control the correction magnitude of the overlap conflict components between mechanism constraints and cross-domain adaptation.

[0107] The computing device also adjusts the correction intensity based on modal reliability information, current operating condition safety status, residual echo information, and adaptation status. When the current operating condition is in a critical safety state, the computing device increases the correction intensity corresponding to the mechanistic factor; when the current operating condition exhibits stable material batch drift or stable equipment state drift, the computing device reduces the direct constraint intensity in the overlapping conflict factor; when the data reliability of a certain mode decreases, the computing device reduces the impact of the residual component corresponding to that mode on the correction intensity. For mechanistic factors corresponding to instantaneous residuals, the computing device allocates the corresponding correction intensity to the current time window; for mechanistic factors corresponding to delayed residuals, the computing device allocates the corresponding correction intensity to subsequent time windows; for mechanistic factors corresponding to cumulative residuals, the computing device accumulates the corresponding correction intensity over multiple time windows; for overlapping conflict factors corresponding to observed and confirmed residuals, the computing device reduces the direct correction intensity within the current time window and retains the overlapping conflict factor to continue participating in the conflict marker update in subsequent time windows. When the adaptation state is mechanistic locked, the computing device increases the correction intensity corresponding to the mechanistic factor; when the adaptation state is domain released, the computing device increases the correction intensity corresponding to the domain offset factor; when the adaptation state is slipping, the computing device decreases the update rate of the overlap conflict factor. Through this process, the correction intensity is jointly determined by the correction factor magnitude, conflict relationship, modal confidence, current operating state, residual time attribute, and adaptation state.

[0108] S218. Associate the correction intensity and the offset direction to obtain the offset vector; The computing device correlates the correction intensity and offset direction to obtain an offset vector. The offset vector is used to simultaneously represent the correction direction and correction magnitude corresponding to each factor in the correction factor set, enabling asymmetric shunt gating to unify directional control and intensity control into a correction amount for the initial output mechanism.

[0109] The computing device correlates the offset direction of the mechanistic factor with the correction strength of the mechanistic constraint to obtain the mechanistic offset component. The mechanistic offset component is used to represent the correction amount that needs to be adjusted along the mechanistic constraint direction for the component in the initial output mechanism affected by the mechanistic residual vector.

[0110] The computing device correlates the offset direction of the domain offset factor with the cross-domain adaptation correction strength to obtain the domain offset component. The domain offset component is used to represent the correction amount that the components affected by the domain offset vector in the initial output mechanism need to be adjusted along the current operating condition adaptation direction.

[0111] The computing device correlates the offset direction of the overlap conflict factor with the conflict adjustment intensity to obtain the conflict offset component. The conflict offset component represents the correction amount required to adjust the components in the initial output mechanism affected by both the mechanism residual vector and the domain offset vector according to the controlled adjustment direction. The computing device correlates the offset direction of the neutral factor with the stability holding intensity to obtain the neutral holding component. The mechanism offset component, domain offset component, conflict offset component, and neutral holding component together constitute the offset vector.

[0112] S219. The initial output mechanism is decomposed according to the offset vector, and the initial output mechanism after the decomposed mechanism is corrected item by item by the offset vector to obtain the target constraint mechanism.

[0113] The computing device loads the offset vector into the model space corresponding to the initial output mechanism and performs correlation matching on each computational component in the initial output mechanism according to the coordinate distribution of the offset vector in the model space. For computational components that are correlated with the coordinates of the mechanism offset component, the computing device classifies them as mechanism-correlated components; for computational components that are correlated with the coordinates of the domain offset component, the computing device classifies them as domain-offset-correlated components; for computational components that are correlated with both the mechanism offset component and the domain offset component, the computing device classifies them as conflict-correlated components; for computational components whose overall correlation with the offset vector is lower than the preset correlation condition, the computing device classifies them as neutral-correlated components.

[0114] The computing device performs constraint correction on the mechanism-related components based on the mechanism offset components, causing the mechanism-related components to converge to the constraint range corresponding to the mechanism standard vector along the reverse correction direction corresponding to the mechanism residual vector. For mechanism-related components corresponding to instantaneous residuals, the computing device completes the constraint correction within the current time window; for mechanism-related components corresponding to delayed residuals, the computing device delays the corresponding correction amount according to the residual echo information; for mechanism-related components corresponding to cumulative residuals, the computing device superimposes the residual effects within multiple time windows before participating in the correction. The computing device performs adaptation correction on the domain offset components based on the domain offset components, ensuring that the domain offset components retain the effective differences formed by the current production line, equipment, material batch, and operating environment along the direction of operating condition changes corresponding to the domain offset vector. The computing device performs sliding correction on the conflict-related components based on the conflict offset components, causing the conflict-related components to undergo low-speed weak updates in the direction jointly influenced by the mechanism residual vector and the domain offset vector, and retaining this part of the calculated components for further confirmation in subsequent time windows. The computing device performs hold correction on the neutral-related components based on the neutral-holding components, ensuring that the neutral-related components maintain their original calculation state during the correction process.

[0115] The computing device recombines the corrected mechanism-related components, domain offset-related components, conflict-related components, and neutral-related components to form the target constraint mechanism. This target constraint mechanism achieves constraint convergence in the mechanism deviation direction, domain adaptation preservation in the operating condition offset direction, and slip correction in the overlapping conflict direction, while maintaining stable computational components of the initial output mechanism in the neutral direction. After the target constraint mechanism enters the lightweight domain adaptation path or the subsequent processing layer of the initial model, the lightweight domain adaptation path performs subsequent processing according to the adaptation state: suppressing physical violation directions in the target constraint mechanism in the mechanism-locked state, preserving the target operating condition change direction in the target constraint mechanism in the domain-released state, and performing low-speed updates on the overlapping conflict directions in the target constraint mechanism in the slip state.

[0116] The computing device inputs the target constraint mechanism into the subsequent processing layer of the lightweight domain adaptation path or the initial model. The lightweight domain adaptation path includes, but is not limited to, Adapter, LoRA, Prefix tuning, Prompt tuning, low-rank gating matrix, hybrid expert router, and sparse activation sub-network. The computing device determines the activation level of the lightweight domain adaptation path based on the mechanism-related components, domain offset-related components, and conflict-related components in the target constraint mechanism, ensuring that the target constraint mechanism continues to participate in the domain adaptation processing under the target operating condition after the mechanism constraint correction is completed.

[0117] This embodiment generates an initial output mechanism by constructing multimodal operating data and separately calculates the mechanism residual vector and domain offset vector, enabling the differentiation between mechanism deviation and operating condition deviation in the initial output mechanism. By projecting the mechanism residual vector and domain offset vector onto the model space, a set of correction factors and conflict markers are generated, and asymmetric split gating is used for item-by-item correction. This reduces erroneous corrections caused by mutual interference between the mechanism residual vector and the domain offset vector, improving the mechanism consistency and cross-domain adaptive capability of the target constraint mechanism.

[0118] The above provides a detailed description of the multimodal model mechanism constraint and cross-domain adaptive method in the embodiments of this application. The following will provide a detailed description of the multimodal model mechanism constraint and cross-domain adaptive system and device.

[0119] Please see Figure 3 This application provides an embodiment of a multimodal model mechanism constraint and cross-domain adaptive system, which includes: The receiving unit 301 is used to receive production line data and construct multimodal operation data based on the production line data, wherein the production line data includes deployment data and operation data; Input unit 302 is used to input the multimodal operation data into the initial model to generate the initial output mechanism; The first calculation unit 303 is used to calculate the residual between the mechanism standard vector and the initial output mechanism to obtain the mechanism residual vector; The second calculation unit 304 is used to calculate the domain offset vector using the standard working condition vector and the multimodal operation data. The construction unit 305 is used to project the mechanism residual vector and the domain offset vector onto the model space corresponding to the initial output mechanism, and construct a set of correction factors and a conflict marker quantity. The set of correction factors includes mechanism factors, domain offset factors, overlap conflict factors, and neutral factors. The correction unit 306 is used to correct the initial output mechanism by combining the correction factor set and the conflict marker quantity with asymmetric split gating to obtain the target constraint mechanism.

[0120] Optionally, the building unit 305 is specifically used for: The mechanism residual vector and the domain offset vector are projected onto the model space corresponding to the initial output mechanism, and the mechanism residual direction basis and the domain offset direction basis are obtained by spatial coordinate correlation calculation. Calculate the directional phase between the mechanistic residual directional basis and the domain offset directional basis, and generate a conflict marker quantity based on the directional phase; A subspace is constructed based on the described mechanism residual direction basis and the described domain offset direction basis; The initial output mechanism is decomposed into subspace conflicts based on the phase relationship of the subspace to obtain a set of correction factors, which includes mechanism factors, domain offset factors, overlap conflict factors, and neutral factors.

[0121] Optionally, the correction unit 306 is specifically used for: The offset direction of each factor in the set of correction factors is determined based on the conflict marker amount; The correction strength is determined based on the absolute value of each factor in the set of correction factors; The initial output mechanism is corrected based on the correction intensity and the offset direction to obtain the target constraint mechanism.

[0122] Optionally, the correction unit 306 is further configured to: By associating the correction intensity with the offset direction, an offset vector is obtained; The initial output mechanism is decomposed based on the offset vector, and the initial output mechanism after the decomposed mechanism is corrected item by item by the offset vector to obtain the target constraint mechanism.

[0123] Optionally, the second calculation unit 304 is specifically used for: Extract operating condition parameters from the multimodal operating data; Construct the current operating condition vector based on the operating condition parameters; The difference between the current operating condition vector and the standard operating condition vector is calculated to obtain the domain offset vector.

[0124] Optionally, the input unit 302 is specifically used for: Build a time window using the initial model; Establish the time series association between the multimodal operation data and the time window to obtain the time series operation vector; The initial model is used to analyze the correlation state of the time-series running vectors to obtain the initial output mechanism.

[0125] Optionally, the first computing unit 303 is specifically used for: The initial output mechanism is vectorized to obtain an initial mechanism vector, which is used to align the vector dimensions of the initial output mechanism with those of the standard mechanism vector. The difference between the initial mechanism vector and the standard mechanism vector is calculated to obtain the mechanism residual vector.

[0126] Optionally, the receiving unit is specifically used for: Receive production line data, which includes deployment data and operational data; Generate a mechanism standard vector and an operating condition standard vector based on the deployment data; The runtime data is lexicalized to obtain multimodal runtime data.

[0127] In this embodiment, the functions of each unit are the same as those described above. Figures 1 to 2b The steps in the illustrated embodiments are the same and will not be repeated here.

[0128] Please see Figure 4 This application provides an embodiment of a multimodal model mechanism constraint and cross-domain adaptive device, including: Processor 401, memory 402, input / output unit 403, bus 404; The processor 401 is connected to the memory 402, the input / output unit 403 and the bus 404; The processor 401 specifically executes... Figures 1 to 2b The specific operations corresponding to the steps in the method will not be elaborated here.

[0129] This application also relates to a computer-readable storage medium on which a program is stored, which, when run on a computer, causes the computer to perform any of the methods described above.

[0130] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0131] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0132] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0133] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0134] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A multimodal model mechanism constraint and cross-domain adaptive method, characterized in that, The method includes: Receive production line data and construct multimodal operation data based on the production line data, wherein the production line data includes deployment data and operation data; The multimodal operating data is input into the initial model to generate the initial output mechanism; The residual vector of the mechanism is obtained by calculating the standard vector of the mechanism and the residual of the initial output mechanism; The domain offset vector is calculated using the standard operating condition vector and the multimodal operating data. The mechanism residual vector and the domain offset vector are projected onto the model space corresponding to the initial output mechanism to construct a set of correction factors and a conflict label quantity. The set of correction factors includes mechanism factors, domain offset factors, overlap and conflict factors, and neutral factors. The initial output mechanism is modified by combining the set of correction factors and the conflict marker quantity with asymmetric diversion gating to obtain the target constraint mechanism.

2. The method according to claim 1, characterized in that, The step of projecting the mechanism residual vector and the domain offset vector onto the model space corresponding to the initial output mechanism to construct a set of correction factors and a conflict marker includes: The mechanism residual vector and the domain offset vector are projected onto the model space corresponding to the initial output mechanism, and the mechanism residual direction basis and the domain offset direction basis are obtained by spatial coordinate correlation calculation. Calculate the directional phase between the mechanistic residual directional basis and the domain offset directional basis, and generate a conflict marker quantity based on the directional phase; A subspace is constructed based on the described mechanism residual direction basis and the described domain offset direction basis; The initial output mechanism is decomposed into subspace conflicts based on the phase relationship of the subspace to obtain a set of correction factors, which includes mechanism factors, domain offset factors, overlap conflict factors, and neutral factors.

3. The method according to claim 1, characterized in that, The initial output mechanism is modified by combining the asymmetric splitting gating with the set of correction factors and the conflict marker quantity to obtain the target constraint mechanism, including: The offset direction of each factor in the set of correction factors is determined based on the conflict marker amount; The correction strength is determined based on the absolute value of each factor in the set of correction factors; The initial output mechanism is corrected based on the correction intensity and the offset direction to obtain the target constraint mechanism.

4. The method according to claim 3, characterized in that, The initial output mechanism is corrected based on the correction intensity and the offset direction to obtain the target constraint mechanism, including: By associating the correction intensity with the offset direction, an offset vector is obtained; The initial output mechanism is decomposed based on the offset vector, and the initial output mechanism after the decomposed mechanism is corrected item by item by the offset vector to obtain the target constraint mechanism.

5. The method according to claim 1, characterized in that, The calculation of the domain offset vector using the standard operating condition vector and the multimodal operating data includes: Extract operating condition parameters from the multimodal operating data; Construct the current operating condition vector based on the operating condition parameters; The difference between the current operating condition vector and the standard operating condition vector is calculated to obtain the domain offset vector.

6. The method according to any one of claims 1 to 5, characterized in that, The mechanism of inputting the multimodal operating data into the initial model to generate the initial output includes: Build a time window using the initial model; Establish the time series association between the multimodal operation data and the time window to obtain the time series operation vector; The initial model is used to analyze the correlation state of the time-series running vectors to obtain the initial output mechanism.

7. The method according to any one of claims 1 to 5, characterized in that, The calculation of the residual between the mechanism standard vector and the initial output mechanism to obtain the mechanism residual vector includes: The initial output mechanism is vectorized to obtain an initial mechanism vector, which is used to align the vector dimensions of the initial output mechanism with those of the standard mechanism vector. The difference between the initial mechanism vector and the standard mechanism vector is calculated to obtain the mechanism residual vector.

8. The method according to any one of claims 1 to 5, characterized in that, The process of receiving production line data and constructing multimodal operation data based on the production line data includes: Receive production line data, which includes deployment data and operational data; Generate a mechanism standard vector and an operating condition standard vector based on the deployment data; The runtime data is lexicalized to obtain multimodal runtime data.

9. A multimodal model mechanism constraint and cross-domain adaptive system, characterized in that, The system includes: A receiving unit is used to receive production line data and construct multimodal operation data based on the production line data, wherein the production line data includes deployment data and operation data; The input unit is used to input the multimodal operation data into the initial model and generate the initial output mechanism; The first calculation unit is used to calculate the residual between the mechanism standard vector and the initial output mechanism to obtain the mechanism residual vector; The second calculation unit is used to calculate the domain offset vector using the standard working condition vector and the multimodal operating data. A construction unit is used to project the mechanism residual vector and the domain offset vector onto the model space corresponding to the initial output mechanism, and construct a set of correction factors and a conflict marker quantity. The set of correction factors includes mechanism factors, domain offset factors, overlap conflict factors, and neutral factors. The correction unit is used to correct the initial output mechanism by combining the correction factor set and the conflict marker quantity with asymmetric split gating to obtain the target constraint mechanism.

10. A multimodal model mechanism constraint and cross-domain adaptive device, characterized in that, The device includes: Processor, memory, input / output units, and bus; The processor is connected to the memory, the input / output unit, and the bus; The memory stores a program, which the processor invokes to perform the method as described in any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains a program that, when executed on a computer, performs the method as described in any one of claims 1 to 8.

Images