A method and system for process knowledge enhancement of a model

Abstract

The application relates to the technical field of artificial intelligence and machine learning, and discloses a model process knowledge enhancement method and system. The method acquires training samples and process knowledge including process knowledge application conditions and process constraint relationships, determines a process state according to the training samples and matches the process knowledge, determines a dependency relationship between the process state used for judging the process knowledge application conditions and a model processing result of a to-be-trained model, verifies the process state dependent on the model processing result of the to-be-trained model, obtains an independent data support result, and determines whether the process knowledge is used for constraining model training according to the dependency relationship and the independent data support result. The to-be-trained model is trained by using the training samples, and when the process knowledge is used for constraining training, knowledge training constraints are applied according to the process constraint relationships. The application can reduce the risk of triggering false knowledge constraints by process states that have not been independently verified and strengthen the prediction deviation of a model, and improve the reliability of knowledge enhanced model training.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and machine learning technology, specifically to a method and system for enhancing the process knowledge of a model. Background Technology

[0002] With the application of artificial intelligence, machine learning, and neural network technologies in industrial production, training predictive models using production process data has become an important technical means for product quality prediction, process parameter prediction, equipment status analysis, and production process optimization. Existing model training typically determines the loss function based on training samples and corresponding supervised data, and updates model parameters through backpropagation and other methods. To reduce the potential for unreasonable predictions arising from models relying solely on training data, prior knowledge such as process procedures, expert experience, historical production patterns, and knowledge graphs can be introduced into model training. This knowledge constraint assists machine learning or neural network models in learning, ensuring that the model's prediction results conform as closely as possible to existing process patterns.

[0003] In industrial applications, process knowledge typically has specific conditions of applicability. For example, the same process parameter may have different impacts on product quality or remaining processing time at different processing stages, under different material conditions, or under different equipment operating conditions. Therefore, when using process knowledge to assist model training, it is usually necessary to first determine the current process state based on the training samples, and then determine whether the corresponding process knowledge is applicable to the current training samples. After the process knowledge is determined to be applicable, the target value, allowable range, or parameter variation law specified by the process knowledge is used to constrain the model training process.

[0004] However, in the training process of existing knowledge-enhanced models, the process states used to determine the applicability of process knowledge do not necessarily all originate from data directly recorded by sensors, detection equipment, or production control systems. Some process states are difficult to detect directly and require judgment based on the prediction results, intermediate features, or other data further calculated from the model's processing results. In such cases, if the applicability of process knowledge is still determined directly based on this process state, and the corresponding process knowledge is further used to constrain the same model under training, the model's own processing results may in turn become the basis for constraining model training.

[0005] For example, in the drying process, neural network models can be trained to predict the remaining processing time of materials, and there are process rules such as "after entering the later drying stage, increasing the processing temperature usually does not increase the remaining processing time." If the decision of "whether to enter the later drying stage" is also based on the model's predicted remaining processing time, when the model's current prediction is biased, it may first obtain an incorrect process state, then apply the corresponding process knowledge based on the incorrect process state, and continue to constrain the model parameter updates through knowledge training loss. This not only fails to effectively utilize prior knowledge to correct the model, but may also cause the model's original prediction bias to be continuously reinforced in subsequent training processes.

[0006] Existing technologies for knowledge reinforcement learning using process knowledge typically focus on whether training samples meet the applicability conditions of the process knowledge and whether the model's predictions conform to process constraints. However, they lack further distinction regarding whether the process states used to determine the applicability conditions of process knowledge are affected by the processing results of the model under training. Furthermore, when process states are influenced by the processing results of the model under training, there is a lack of a mechanism to determine whether the corresponding process state has independent basis. Therefore, how to prevent process states formed based on the processing results of the model under training from being used to trigger process knowledge constraints without valid confirmation, thereby reducing the impact of erroneous process knowledge constraints on the model training process, is a technical problem that needs to be solved in the training of process knowledge reinforcement models. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this application aims to provide a method and system for enhancing process knowledge in models. This method addresses the problem that, during knowledge enhancement learning in artificial intelligence, machine learning, or neural networks, when the process state used to determine the applicability of process knowledge is influenced by the model's own processing results, unverified process states may trigger erroneous process knowledge constraints, further reinforcing model prediction bias during training. The method identifies the dependency relationship between the process state and the model's processing results, verifies the dependent process states using independent data, and controls whether process knowledge is used to constrain the training of the model based on the dependency relationship and independent data support results. This improves the reliability of process knowledge-enhanced model training.

[0008] To achieve the above objectives, this application adopts the following technical solution: Firstly, this application provides a method for enhancing the process knowledge of a model, comprising: Acquire training samples and process knowledge used to assist in training the model to be trained. The process knowledge includes the applicable conditions of the process knowledge and the process constraint relationships that hold when the applicable conditions of the process knowledge are met. The corresponding process state is determined based on the training samples, and the process knowledge that matches the training samples is determined based on the process state and the applicable conditions of the process knowledge. Determine the dependency between the process state used to determine the applicability of process knowledge and the model processing results of the model to be trained; The process states that the process states depend on the model processing results of the model to be trained are validated, the validation results are obtained, and the independent data support results for the corresponding process states are determined based on the validation results. Based on the dependencies and the independent data support results corresponding to the process states of the model processing results that depend on the model to be trained, determine whether the process knowledge that matches the training samples is used to constrain the training of the model to be trained. The training samples are used to train the model to be trained, and when process knowledge is used to constrain the training of the model to be trained, knowledge training constraints are applied to the model to be trained according to the process constraint relationship.

[0009] Furthermore, the dependency between the process state used to determine the applicability of process knowledge and the model processing results of the model to be trained is determined, including: The process of generating data corresponding to the process status; When the data generation process uses the output results, intermediate processing results, or data obtained by further processing the model processing results of the model to be trained, it is determined that there is a dependency between the process state and the model processing results of the model to be trained. Otherwise, there is no dependency between determining the process state and the model processing results of the model to be trained.

[0010] Furthermore, the process state, which represents the process state depending on the model processing results of the model to be trained, is validated, including: Obtain independent data and the corresponding sample identifier, production batch, equipment identifier, material identifier, collection location or collection time for training samples and independent data; Determine the association between independent data and training samples based on sample identifier, production batch, equipment identifier, material identifier, collection location or collection time; The corresponding process state is validated based on independent data that are correlated with the training samples.

[0011] Furthermore, the corresponding process states are validated based on independent data that are correlated with the training samples, including: Obtain the process status verification conditions used to verify the corresponding process status; When independent data meets the process status verification conditions, a verification result indicating support for the corresponding process status is obtained; When independent data does not meet the process status verification conditions, a verification result is obtained indicating that the corresponding process status is not supported.

[0012] Furthermore, the process status verification conditions are determined based on the status judgment criteria in the established process specifications, actual production test results, or historical samples and corresponding independent data. When determining process status verification conditions based on historical samples and corresponding independent data, the process status corresponding to the historical samples is confirmed by manual testing, laboratory testing, or independent testing equipment.

[0013] Furthermore, the applicable conditions of process knowledge include multiple condition items that correspond to the process state, and the process knowledge also includes the logical relationship between multiple condition items; Based on the process status, determine whether multiple conditions are true or false, and based on the logical relationship between multiple conditions, determine one or more combinations of conditions from the conditions that are true, which can make the process knowledge applicable.

[0014] Furthermore, it also includes: Independent data supports the results, including both supported and unsupported data. Determine the condition confirmation result for each condition item in each condition combination; When the process state corresponding to a condition item is directly determined by data from the model processing results that do not depend on the model to be trained, and the condition item is determined to be valid based on the corresponding process state, the condition confirmation result is determined as direct confirmation. When the process state corresponding to the condition item depends on the model processing result of the model to be trained and the independent data support result is supported, the condition confirmation result will be determined as support confirmation. When the process state corresponding to a condition item depends on the model processing result of the model to be trained and the independent data support result is not supported, the condition confirmation result will be determined as unconfirmed.

[0015] Further, determining whether the process knowledge matched with the training samples is used to constrain the training of the model to be trained includes: When there exists a combination of conditions, and the confirmation results of all conditions in the combination are direct confirmations, process knowledge is determined to constrain the training of the model to be trained, and knowledge training constraints are applied to the model to be trained according to the process constraint relationship.

[0016] Further, determining whether the process knowledge matched with the training samples is used to constrain the training of the model to be trained includes: When there is no combination of conditions where all condition confirmation results are directly confirmed, but there is a combination of conditions that does not include unconfirmed condition items and includes condition items that support confirmation, process knowledge is determined to constrain the training of the model to be trained, and the constrained knowledge training constraint is adopted.

[0017] Further, determining whether the process knowledge matched with the training samples is used to constrain the training of the model to be trained includes: When all combinations of conditions include condition items whose confirmation result is unconfirmed, it is determined that process knowledge is not used to constrain the training of the model to be trained.

[0018] Furthermore, the process constraint relationship includes the target value or allowable range of the model prediction result, as well as the relationship between the set target process parameter and the change direction of the model prediction result. The change direction relationship is used to characterize the change direction of the model prediction result when the target process parameter increases or decreases. The constrained knowledge training constraint does not impose constraints based on the target value or the allowable range, but imposes knowledge training constraints on the model to be trained based on the relationship of the direction of change.

[0019] Furthermore, process constraints include the allowable range of model prediction results; The constrained knowledge training constraint applies knowledge training constraints to the model to be trained using an expanded allowable range, where the expanded allowable range includes the allowable range in the process constraint relationship.

[0020] Furthermore, the process constraint relationship includes multiple constraint items that correspond to different process states; The restricted knowledge training constraints include retaining the condition confirmation results of the corresponding process state from multiple constraint items as the constraint items that support confirmation, and applying knowledge training constraints to the model to be trained based on the retained constraint items.

[0021] Furthermore, knowledge training constraints are imposed on the model to be trained based on the direction of change, including: Obtain the applicable scope of process knowledge, the allowable adjustment range of equipment, and the allowable range of actual production corresponding to the target process parameters, and determine the range of variation of the target process parameters based on the applicable scope of process knowledge, the allowable adjustment range of equipment, and the allowable range of actual production. By changing the target process parameters in the training samples within the range of variation, a variable training sample is obtained. The training samples and the changed training samples are processed by the model to be trained respectively to obtain the original prediction results and the changed prediction results. The knowledge training constraints are determined based on whether the original prediction results and the changed prediction results satisfy the relationship of the change direction.

[0022] Furthermore, knowledge training constraints are imposed on the model to be trained based on the direction of change, including: Acquire multiple training samples and preset similarity conditions for comparing the corresponding process states of the multiple training samples; Identify two training samples from multiple training samples that meet preset similar conditions for process states but have different target process parameters; Two training samples are processed using the model to be trained to obtain two model prediction results. Knowledge training constraints are determined based on whether the two model prediction results satisfy the relationship of change direction.

[0023] Furthermore, it also includes: Record the process knowledge that was not used to constrain the training of the model to be trained, as well as the corresponding training samples and process states; Acquire new independent data associated with the training samples, and re-validate the process status based on the new independent data to obtain updated validation results; The independent data support results for the process state are updated based on the updated validation results, and the process knowledge is re-determined whether it should be used to constrain the subsequent training of the model to be trained based on the updated independent data support results.

[0024] Furthermore, when training samples are matched with multiple pieces of process knowledge, it also includes: Determine whether the process constraint relationships corresponding to multiple process knowledge can be valid simultaneously. When multiple process constraints corresponding to process knowledge can be simultaneously established, the knowledge training constraints corresponding to multiple process knowledge are used together to constrain the training of the model to be trained. When multiple process constraints corresponding to process knowledge cannot be simultaneously established, the knowledge training constraints corresponding to multiple process knowledge will not be used together to constrain the training of the model to be trained.

[0025] Secondly, this application provides a process knowledge enhancement system for a model, comprising: The knowledge acquisition module is used to acquire training samples and process knowledge to assist in training the model to be trained. The process knowledge includes the applicable conditions of the process knowledge and the process constraint relationships that hold when the applicable conditions of the process knowledge are met. The process state determination module is used to determine the corresponding process state based on the training samples, and to determine the process knowledge that matches the training samples based on the process state and the applicable conditions of process knowledge. The dependency determination module is used to determine the dependency between the process state used to determine the applicability of process knowledge and the model processing results of the model to be trained. The process state verification module is used to verify the process state that represents the process state's dependence on the model's processing results, obtain the verification results, and determine the independent data support results for the corresponding process state based on the verification results. The knowledge constraint determination module is used to determine whether the process knowledge matching the training samples is used to constrain the training of the model, based on the dependency relationship and the independent data support results corresponding to the process state that depends on the model processing results of the model to be trained. The model training module is used to train the model to be trained using training samples, and to apply knowledge training constraints to the model to be trained based on the process constraint relationship when process knowledge is used to constrain the training of the model to be trained.

[0026] Compared with the prior art, the beneficial effects achieved by this application are as follows: In the training process of process knowledge enhancement models based on artificial intelligence, machine learning, or neural networks, after determining the process state and matching process knowledge based on training samples, this application further determines the dependency relationship between the process state used to judge the applicability of process knowledge and the model processing results of the model to be trained. The process state that depends on the model processing results of the model to be trained is independently verified. Then, the process knowledge is determined to be included in the model training constraints by combining the dependency relationship and the results of independent data support. This can avoid directly triggering the corresponding process knowledge based solely on the process state formed by the output of the model to be trained or the intermediate processing results. This allows the knowledge training constraints to be established on the basis of process states with relatively independent verification basis, reducing the risk of erroneous process knowledge constraints reinforcing the model prediction bias in the knowledge enhancement learning process, and improving the reliability of the model training process and the process knowledge enhancement results. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating a method for enhancing the process knowledge of a model, as provided in an embodiment of this application.

[0028] Figure 2 This is a schematic diagram illustrating the implementation environment of a process knowledge enhancement method for a model provided in an embodiment of this application.

[0029] Figure 3 This is a flowchart illustrating the process status verification and process knowledge usage judgment provided in the embodiments of this application.

[0030] Figure 4 This is a schematic diagram illustrating the process of applying knowledge training constraints to the model to be trained based on the changing direction relationship, as provided in an embodiment of this application.

[0031] Figure 5 This is a schematic diagram of the module relationships of a process knowledge enhancement system for a model provided in an embodiment of this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and technical effects of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings; the described embodiments are used to illustrate the technical solutions of this application and do not limit the scope of protection of this application.

[0033] In the embodiments of this application, the first, second, and other sequential terms are used to distinguish different objects, different data, or different results; multiple means two or more.

[0034] Examples 2 to 7 further illustrate some of the processing procedures in Example 1; for processing procedures not specifically described in Examples 2 to 7, the corresponding processing procedures in Example 1 shall be executed; Example 8 is used to illustrate the combined application of the processing procedures in Examples 1 to 7 in an industrial production scenario; Example 9 is used to illustrate the process knowledge enhancement system of the model.

[0035] The training samples in this application embodiment are used to train the model to be trained; the training samples consist of production data generated during the industrial production process; the production data includes one or more of the following: temperature, pressure, flow rate, equipment speed, weight, image, spectrum, equipment operating status, material detection results, manual detection results, or product quality detection results; the production data is read from sensors, detection equipment, production control system, industrial database, or manual detection records.

[0036] The data fields in the training samples correspond to the input fields of the model to be trained; numerical production data are input into the model to be trained according to the data format already used before training; image or spectral data are input into the model to be trained according to the data reading method corresponding to the model to be trained; the ordinary data cleaning, missing value handling and data format conversion of the training samples are performed using the processing methods in the existing training process of the model to be trained.

[0037] The process knowledge in this application embodiment is used to assist in the training of the model to be trained; the process knowledge includes the applicable conditions of the process knowledge and the process constraint relationship that is established when the applicable conditions of the process knowledge are met; the process knowledge is read from process procedures, equipment operation specifications, expert experience, historical production data statistics, process knowledge base or knowledge graph, and stored according to the applicable conditions of the process knowledge and the process constraint relationship respectively.

[0038] The applicable conditions of process knowledge are used to specify the process states that must be met for the process knowledge to be valid. The process constraint relationships are used to specify the target values, allowable ranges, change direction relationships, or constraint items that the model prediction results must meet when the applicable conditions of process knowledge are valid. Each process constraint relationship also records the corresponding model prediction result field, and the change direction relationship also records the training sample input field corresponding to the target process parameters. Numerical comparison uses the original units of the model input field and the model prediction result field. If the units are inconsistent, they are converted to the same units according to the existing unit conversion relationship in the process specification or data source. When the applicable conditions of process knowledge include multiple conditions, the logical relationship between the multiple conditions is recorded at the same time. The logical relationship between the multiple conditions consists of AND logic, OR logic, and combinations of AND logic and OR logic.

[0039] In this embodiment, the process state is used to characterize the process stage, water content, reaction degree, maturation degree, equipment operating status, internal state of materials, or degree of processing completion of the industrial production process corresponding to the training sample; the process state is determined based on the production data in the training sample; when the temperature, pressure, and equipment operating status can be directly obtained from the sensor data or production control system records corresponding to the training sample, the corresponding data are directly read to determine the process state.

[0040] When the process state cannot be directly obtained from the production data corresponding to the training samples, the process state is determined according to the actual data generation process used when generating the process state. The data generation process may include inputting the training samples into the model to be trained to obtain the process state, obtaining the process state based on the intermediate processing results of the model to be trained, or inputting the model processing results of the model to be trained into other processing processes to obtain the process state. The corresponding data generation process is saved for different process states.

[0041] In this embodiment of the application, determining the process state directly from data that does not depend on the model processing results of the model to be trained means that the process of determining the process state does not use the output results, intermediate processing results, or data obtained by further processing the model processing results of the model to be trained. Determining the process state directly from data that does not depend on the model processing results of the model to be trained includes directly reading the corresponding data fields to determine the process state, processing the corresponding data according to pre-fixed calculation rules that do not use the model processing results of the model to be trained, or processing the corresponding data using other models with fixed parameters and whose input data does not contain the model processing results of the model to be trained, and then determining the process state.

[0042] In this embodiment, independent data is used to verify the process status that depends on the model processing results of the model to be trained. Independent data includes sensor detection data, detection equipment detection results, manual detection results, laboratory detection results, or production control system recorded data. The data generation process of independent data does not use the output results, intermediate processing results, or data obtained by further processing the model processing results of the model to be trained. The independent data used for process status verification can be a single independent data record or a set of independent data records that are associated with the same training sample. When the independent data source has a number for uniquely identifying the data record, the existing unique record number in the independent data source is associated with and saved with the corresponding independent data record.

[0043] Independent data support results include supported and unsupported. If an independent data record or set of independent data can form all the verification fields required for the process state verification condition, and all verification fields satisfy the corresponding process state verification condition, the independent data support result is determined to be supported. If an independent data record or set of independent data can form all the verification fields, but the verification result does not satisfy the process state verification condition, the independent data support result is determined to be unsupported. If no independent data with a correlation to the training samples is obtained, the correlated independent data cannot form all the verification fields, or the corresponding process state cannot form the process state verification condition, the independent data support result is determined to be unsupported, and the corresponding reason for unsupport is recorded. An independent data support result of unsupported indicates that the current independent data does not support the corresponding process state, and is not used to indicate that the corresponding process state is invalid.

[0044] The training model in this application embodiment is used to output model prediction results based on production data in the training samples; the training model can perform classification, regression, or prediction tasks; the specific network structure of the training model adopts the network structure that has been determined for the corresponding model task; when the model prediction result includes multiple output fields, each process constraint relationship selects the corresponding output field according to the model prediction result field recorded in the process knowledge, and the target value constraint, allowable range constraint, and change direction relationship use the model prediction result field that can be numerically compared; when this application enhances the training model with process knowledge, it does not change the original input fields and model output definitions of the training model.

[0045] In regression prediction tasks, training samples are input into the model to be trained to obtain model prediction results. The model prediction results are compared with the actual production results corresponding to the training samples, and the basic training loss is calculated. The basic training loss can be the square of the difference between the model prediction results and the actual production results. The model parameters of the model to be trained are updated according to the basic training loss based on the original backpropagation process of the model to be trained. For classification tasks, the basic training loss already used by the corresponding model task is adopted.

[0046] Process constraints are used to record the target values, allowable ranges, direction of change, or constraint items originally specified by process knowledge; knowledge training constraints are used to represent the process constraints actually selected and used for current training based on the results of using process knowledge; knowledge training loss is used to represent the numerical training loss obtained by converting knowledge training constraints; when the original training process of the model to be trained has already performed normalization, standardization, or loss scaling processing on the model output, the knowledge training loss adopts the same processing method; when the original training process of the model to be trained has not performed output scaling processing and there are multiple numerical model prediction result fields, the mean and standard deviation of each model prediction result field are calculated separately based on the training data, and the model prediction results, corresponding actual production results, target values, and allowable range boundaries are standardized using the same mean and standard deviation before calculating the corresponding basic training loss and knowledge training loss. When the standard deviation is 0, it is set to 1; after scaling processing, the basic training loss and knowledge training loss are combined.

[0047] See Figure 2 The implementation environment of this application embodiment includes a production data source, a process knowledge storage location, an independent data source, and a model training execution device; the production data source provides training samples to the model training execution device; the process knowledge storage location provides process knowledge to the model training execution device; the independent data source provides independent data for verifying the process state to the model training execution device; the model training execution device executes a process knowledge enhancement method for a model and saves the model parameters of the model to be trained.

[0048] Production data sources can be sensors, testing equipment, production control systems, or industrial databases; process knowledge storage locations can be process documents, process knowledge bases, or knowledge graph storage devices; independent data sources can be independent testing equipment, laboratory testing records, manual testing records, sensors, or production control systems; when the same equipment can provide both training samples and independent data simultaneously, the data generation processes of the training samples and independent data are checked separately. If only the data generation process of independent data does not use the model processing results of the model to be trained, the corresponding data is considered independent data.

[0049] Example 1: See Figure 1 This embodiment provides a method for enhancing the process knowledge of a model; the method for enhancing the process knowledge of a model is executed by a model training execution device; the method for enhancing the process knowledge of a model includes steps S101 to S106.

[0050] Step S101: Obtain training samples and process knowledge used to assist in training the model to be trained.

[0051] Read the production data that has been generated during the industrial production process and used to train the model to be trained. Combine the production data corresponding to a production process, a material object, a period of equipment operation, or a collection time into training samples according to the input format of the model to be trained. The data fields in the training samples correspond one-to-one with the input fields of the model to be trained. When the production data contains sample identifiers, production batches, equipment identifiers, material identifiers, collection locations, or collection times, the corresponding information is associated with and saved with the training samples.

[0052] Read the process knowledge used to assist in training the model to be trained; the process knowledge is obtained from process procedures, equipment operation specifications, expert experience records, historical production data statistics, process knowledge base or knowledge graph; record the content that limits the validity of process knowledge as the applicable conditions of process knowledge, and record the content that the model prediction results are required to meet when the applicable conditions of process knowledge are valid as process constraint relationships.

[0053] When the applicable conditions of process knowledge include one condition item, the process state corresponding to the condition item and the comparison condition for judging whether the process state meets the applicable requirements of process knowledge are read from the original content of process knowledge, and the comparison condition is recorded as the condition item's validity rule. When the applicable conditions of process knowledge include multiple condition items, the process state corresponding to each condition item, the condition item's validity rule, and the logical relationship between multiple condition items are recorded respectively. The process constraint relationship records the target value, allowable range, change direction relationship, or multiple constraint items corresponding to the process state of the model prediction result according to the original content of process knowledge, and records the model prediction result field corresponding to each process constraint relationship. The change direction relationship also records the training sample input field corresponding to the target process parameter.

[0054] When the original content of process knowledge limits the range, upper limit, or lower limit of the target process parameter's value when the relationship of change direction is valid, the corresponding limit is recorded as the scope of application of the process knowledge for the target process parameter; when the original content of process knowledge provides a complete range of values, the lower and upper limits of the range are recorded; when the original content of process knowledge only provides the upper limit, the target process parameter is recorded as not exceeding the corresponding upper limit; when the original content of process knowledge only provides the lower limit, the target process parameter is recorded as not falling below the corresponding lower limit; when the original content of process knowledge does not limit the scope of application of the process knowledge for the target process parameter, the processing method of determining knowledge training constraints based on change training samples is not adopted, and it is checked according to Example 5 whether the processing method of determining knowledge training constraints based on two training samples can be adopted.

[0055] Step S102: Determine the corresponding process state based on the training samples, and determine the process knowledge that matches the training samples based on the process state and the applicable conditions of the process knowledge.

[0056] Read the production data used to determine the process status from the training samples; when the process status can be directly represented by the production data in the training samples, directly read the corresponding production data to determine the process status; when the process status needs to be determined through a processing procedure, input the corresponding production data in the training samples into the processing procedure that has been used to generate the process status, and determine the process status according to the output of the processing procedure; save each process status corresponding to the training samples and the data generation procedure corresponding to each process status.

[0057] Read the applicable conditions of the process knowledge one by one, and match the condition items in the applicable conditions of the process knowledge with the corresponding process states of the training samples; compare the process states with the requirements of the condition items according to the rules for the condition items being met; when the applicable conditions of the process knowledge contain multiple condition items, determine whether the applicable conditions of the process knowledge are met according to the logical relationship between the multiple condition items; when the applicable conditions of the process knowledge are met, determine the corresponding process knowledge as the process knowledge that matches the training samples.

[0058] The process knowledge matched with the training sample indicates that the current process state of the training sample meets the applicable conditions of the process knowledge; the process knowledge matched with the training sample enters step S103 to further check the dependency relationship between the process state used to determine the applicable conditions of the process knowledge and the model processing result of the model to be trained; whether the process knowledge is used to constrain the training of the model to be trained is further determined by step S105.

[0059] Step S103: Determine the dependency between the process state used to determine the applicability of process knowledge and the model processing results of the model to be trained.

[0060] For each piece of process knowledge that matches the training samples, read all process states used to determine the applicability of the process knowledge and the data generation process corresponding to each process state; check the actual data source used step by step from the data generation process of the process state to the input data; the check includes the output results of the model to be trained, the intermediate processing results of the model to be trained, and the data obtained by further processing the model processing results of the model to be trained.

[0061] When the data generation process uses the output results, intermediate processing results, or data obtained by further processing the model processing results of the model to be trained, it is determined that there is a dependency between the corresponding process state and the model processing results of the model to be trained; when the data used in the data generation process all come from sensors, detection equipment, production control systems, manual inspection records, or processing processes that have no data transfer relationship with the model processing results of the model to be trained, it is determined that there is no dependency between the corresponding process state and the model processing results of the model to be trained.

[0062] Step S104: Verify the process state of the model processing result that represents the process state of the dependency relationship.

[0063] For process states where the dependency relationship indicates that the process state depends on the model processing results of the model to be trained, independent data, as well as the sample identifier, production batch, equipment identifier, material identifier, collection location, or collection time corresponding to the training samples and independent data, are read. The association relationship between the independent data and the training samples is determined based on one or more available association information. If no independent data with an association relationship with the training samples is obtained, the independent data support result of the corresponding process state is determined as unsupported, and the reason for unsupport is recorded as no associated independent data is obtained. After obtaining the associated independent data, an independent data record or set of independent data is determined for process state verification according to the screening, combination, and selection rules of Example 2.

[0064] When an independent data record or set of independent data used for process status verification can form all the verification fields required for process status verification conditions, all verification fields are compared with the process status verification conditions item by item. When all the verification requirements are met, a verification result indicating support for the corresponding process status is obtained, and the independent data support result for the corresponding process status is determined to be supported. When any of the required verification requirements are not met, a verification result indicating non-support for the corresponding process status is obtained, and the independent data support result for the corresponding process status is determined to be unsupported. When the associated independent data cannot form all the verification fields, the independent data support result is determined to be unsupported, and the reason for non-support is recorded as the lack of independent data fields required to complete the process status verification.

[0065] When the dependency relationship indicates that the process state does not depend on the model processing result of the model to be trained, the dependency relationship obtained in step S103 is retained, and the corresponding process state is directly used for the process knowledge use judgment in step S105; when the dependency relationship indicates that the process state depends on the model processing result of the model to be trained, the independent data support result obtained in step S104 is associated with the corresponding process state and saved, and provided to step S105.

[0066] Step S105: Based on the dependency relationship and the results of independent data support, determine whether the process knowledge matching the training samples is used to constrain the training of the model to be trained.

[0067] Read the process knowledge that matches the training samples, the process states used to determine the applicable conditions of the process knowledge, the dependencies of each process state, and the independent data support results corresponding to the process states that depend on the model processing results of the model to be trained; when the process state does not depend on the model processing results of the model to be trained, confirm that the corresponding process state can be used as the basis for process knowledge based on the results that the process state has met the applicable conditions of the process knowledge.

[0068] When the process state depends on the model processing result of the model to be trained, it is confirmed whether the process state has independent data support based on the independent data support result of the corresponding process state; when all process states in the process knowledge application conditions can form a confirmation result that meets the requirements for using process knowledge according to the dependency relationship or independent data support result, it is determined that the process knowledge is used to constrain the training of the model to be trained; when it cannot form a confirmation result that meets the requirements for using process knowledge, it is determined that the process knowledge is not used to constrain the training of the model to be trained.

[0069] When the applicable conditions for process knowledge include multiple condition items and logical relationships between multiple condition items, one or more condition combinations are determined according to Example 3, and the process knowledge usage result is determined according to the condition confirmation result obtained in Example 3; when using restricted knowledge training constraints, restricted knowledge training constraints are formed according to Example 4 and Example 5.

[0070] Step S106: Train the model to be trained using training samples.

[0071] Before each training batch enters the model parameter update stage, steps S102 to S105 are executed on the training samples in the training batch based on the current model parameters of the model to be trained. The process status, process knowledge matching results, independent data support results, and process knowledge usage results obtained in steps S102 to S105 are only used for this model parameter update. After completing steps S102 to S105, the training samples are input into the model to be trained to obtain the model prediction results, and the basic training loss is calculated according to the model task corresponding to the model to be trained. When multiple training samples constitute a training batch, the basic training loss of each training sample in the training batch is summarized according to the original training process of the model to be trained.

[0072] When determining in step S105 that process knowledge is used to constrain the training of the model to be trained, the actual knowledge training constraint used in the current training is first determined based on the results of the use of process knowledge, and then the knowledge training constraint is converted into knowledge training loss. When the aforementioned scaling method requires the conversion of the corresponding model prediction result field, the same scaling method is first used to convert the model prediction result, the target value, and the allowable range boundary, and then the knowledge training loss is calculated. When the knowledge training constraint is a target value constraint, the square of the difference between the model prediction result after scaling and the target value is calculated, and the square of the difference is used as the knowledge training loss. When the knowledge training constraint is an allowable range constraint, the knowledge training loss is recorded as 0 when the model prediction result is within the allowable range after scaling, and the square of the difference between the model prediction result and the allowable range boundary closest to the model prediction result is used as the knowledge training loss when the model prediction result is outside the allowable range.

[0073] When the knowledge training constraints include multiple constraint items, the model prediction result field corresponding to each constraint item is read, and the knowledge training loss is calculated according to the target value, allowable range, or change direction relationship recorded for each constraint item. When multiple constraint items are used for the current training, the knowledge training losses corresponding to multiple constraint items are added together to obtain the knowledge training loss corresponding to the current process knowledge. The knowledge training loss corresponding to the change direction relationship is determined according to Example 5. The allowable range in the normal knowledge training constraints adopts the original allowable range recorded by the process constraint relationship, and the second type of restricted knowledge training constraint in Example 4 adopts the expanded allowable range.

[0074] When multiple training samples constitute a training batch, first calculate the total knowledge training loss for each training sample separately, and then add the total knowledge training losses for the same training sample to obtain the sample knowledge training loss. If a training sample does not have a knowledge training loss for the current training, the corresponding sample knowledge training loss is recorded as 0. When the original training process of the model to be trained uses the average method to summarize the basic training loss of each training sample in the training batch, the knowledge training loss of each sample is also summarized by averaging. When the original training process uses the summation method to summarize, the knowledge training loss of each sample is also summarized by summation.

[0075] The aggregated basic training loss is added to the aggregated knowledge training loss to obtain the training loss used in the current training. Backpropagation is performed based on the training loss used in the current training, and the model parameters of the model to be trained are updated. After the model parameters are updated, when the same training sample participates in training again in subsequent training rounds, the process state that depends on the model processing results of the model to be trained is re-determined. If the re-determined process state changes, the corresponding process knowledge matching and independent data verification are re-executed. If the re-determined process state does not change and the associated independent data and process state verification conditions do not change, the corresponding independent data support results can be reused. The process state that does not depend on the model processing results of the model to be trained can reuse the already determined results. Subsequent training batches are processed in the above manner until the original training termination conditions of the model to be trained are met.

[0076] When step S105 determines that process knowledge is not used to constrain the training of the model to be trained, the basic training loss corresponding to the training sample is retained, and the model parameters of the model to be trained are updated only based on the basic training loss; the corresponding process knowledge does not form the knowledge training loss of the current training sample; the training sample still participates in the training of the model to be trained.

[0077] Through steps S101 to S106, after matching the process knowledge with the training samples, it is further checked whether the process state used to determine the applicability of the process knowledge depends on the model processing results of the model to be trained, and independent data validation is performed on the process state that depends on the model processing results of the model to be trained; when the process state is determined by the model processing results of the model to be trained, the knowledge training constraints formed by the process knowledge will participate in the model parameter update of the same model to be trained, and independent data validation adds the verification basis that there is no data dependency relationship with the model processing results of the model to be trained before the knowledge training constraints are formed; based on the dependency relationship and the independent data support results, it is controlled whether the process knowledge forms knowledge training constraints.

[0078] Example 2: See Figure 3 This embodiment, based on embodiment 1, further explains the dependency between the process state and the model processing results of the model to be trained, as well as the independent data verification process.

[0079] For each process state used to determine the applicability of process knowledge, read the actual input data and processing steps used when generating the process state; starting from the direct generation step of the process state, trace back according to the source of the input data; if the input data comes from another processing step, continue reading the input data used by that other processing step; if the input data comes directly from sensors, detection equipment, production control systems, manual inspection records, or the model to be trained, stop tracing the corresponding data branch.

[0080] The data generation process is composed of the traced input data sources and processing steps in the actual data transmission order. The data generation process is only used to determine whether there is a data dependency between the process state and the model processing results of the model to be trained. The data generation process can be obtained by reading the actual call relationship, data processing configuration or data field source record in the model training program, and does not require the use of a fixed data table format.

[0081] The data generation process for each process state is checked item by item. When the data generation process directly reads the final output of the model to be trained, it is determined that the data generation process uses the output result of the model to be trained. When the data generation process reads the output of any internal processing stage of the model to be trained and uses the output of the internal processing stage to generate the process state, it is determined that the data generation process uses the intermediate processing result of the model to be trained. When the data generation process inputs the model processing result of the model to be trained into the calculation rules or other models to obtain the process state, it is determined that the data generation process uses the model processing result of the model to be trained to continue processing the data obtained.

[0082] When any of the aforementioned situations occur during the data generation process, it is determined that there is a dependency between the corresponding process state and the model processing result of the model to be trained; when all inputs to the data generation process can be traced back to sensors, detection equipment, production control systems, manual inspection records, or processing processes that have no data transfer relationship with the model to be trained, it is determined that there is no dependency between the corresponding process state and the model processing result of the model to be trained.

[0083] When the process state is obtained from other models with fixed parameters, continue to check the input source of other models with fixed parameters; when the input of other models with fixed parameters includes the output result or intermediate processing result of the model to be trained, determine that there is a dependency between the corresponding process state and the model processing result of the model to be trained; when all the inputs of other models with fixed parameters come from data sources independent of the model to be trained, determine that there is no dependency between the corresponding process state and the model processing result of the model to be trained.

[0084] For process states that rely on the model processing results of the model to be trained, independent data that can be used to verify the corresponding process state is read; at the same time, sample identifiers, production batches, equipment identifiers, material identifiers, collection locations or collection times that can be obtained from training samples and independent data are read; before comparing related information, it is first determined whether the semantics of the fields are consistent, and only fields at the same level that represent the same material object or the same production process are compared; the equipment identifier of the independent detection equipment itself is not compared with the equipment identifier of the corresponding production equipment in the training sample; when the independent data record has the identifier of the production equipment to be detected, the identifier of the production equipment to be detected is compared with the equipment identifier of the corresponding production equipment in the training sample.

[0085] When both training samples and independent data have sample identifiers representing the same material object, if the sample identifiers are the same, the corresponding independent data will be used as candidate associated independent data. If there is no common sample identifier, information that can be obtained from production batch, material identifier, semantically consistent equipment identifier, and collection location will be compared item by item. If all the information obtained is the same, the information obtained will be combined as search conditions. If the search results can identify the same material object or the same production process, the independent data in the search results will be used as candidate associated independent data. If both training samples and candidate associated independent data have collection times, the collection time will continue to be used to determine the correspondence. If either training samples or candidate associated independent data lacks a collection time, the rule of not using collection time to establish the association relationship will be followed.

[0086] When both training samples and candidate associated independent data have acquisition times, the allowable time range for the corresponding process state is obtained first; when the process specification provides the effective time for the independent detection results for the corresponding process state, the corresponding effective time is used as the allowable time range; when the process specification does not provide the aforementioned effective time, and the independent detection specification stipulates that the independent detection results are effective for the corresponding process state within an independent detection cycle, the independent detection cycle is used as the allowable time range; when neither of the aforementioned two types of time can be obtained, no allowable time range is generated and the acquisition time comparison is stopped; the total duration of the process state is not directly used as the allowable time range.

[0087] When both the training samples and candidate associated independent data have collection times and the allowable time range has been obtained, first check whether the production process records can determine the start time, end time, or process stage interval of the corresponding process state; if the corresponding time interval can be determined, and the collection time of the training samples and the collection time of the candidate associated independent data are both within the same time interval, and the absolute time difference between the two collection times does not exceed the allowable time range, then the candidate associated independent data is determined as associated independent data; if any collection time is not within the same time interval or the absolute time difference exceeds the allowable time range, the candidate associated independent data is not used to verify the process state corresponding to the training samples.

[0088] When the production process record cannot determine the start time, end time, or process stage interval of the corresponding process state, the association relationship is determined only when the allowable time range comes from the valid time of the independent test results in the process specification record or the independent test cycle specified in the independent test procedure. The association relationship is determined based on the absolute time difference between the training sample collection time and the candidate associated independent data collection time. If the absolute time difference does not exceed the allowable time range and the discrete association information already used meets the corresponding matching requirements, the candidate associated independent data is determined as associated independent data. If the allowable time range cannot be obtained, the rule of not using the collection time to establish the association relationship is adopted.

[0089] When the association relationship is not established using the collection time, the corresponding independent data is identified as associated independent data only when the discrete association information can uniquely identify the same material object or the same production process, can uniquely identify the corresponding independent data, and the recording semantics of the independent data can determine the process stage of the training sample corresponding to the independent data, or can determine the corresponding process state remains unchanged during the period when the training sample and the independent data correspond, according to the process specification or production process record. If the above requirements cannot be met at the same time, the process state corresponding to the training sample is not verified based on the corresponding independent data, and the support result of the independent data is determined as unsupported. The reason for unsupported is recorded as the absence of associated independent data with a time-series correspondence with the process state corresponding to the training sample.

[0090] When discrete correlation information corresponds to multiple candidate independent data and the corresponding relationship cannot be further determined using the acquisition time, independent data is not selected from the multiple candidate independent data. When it can be determined from the process specifications or production process records that the corresponding process status remains unchanged during the coverage of all candidate independent data, and all candidate independent data belong to the same material object or the same production process, candidate independent data that can form a verification field can be combined into an independent data set. When it cannot be determined that the corresponding process status remains unchanged during the coverage of candidate independent data, the independent data support result is determined as unsupported.

[0091] After obtaining one or more related independent data, read all verification fields in the process status verification conditions; when a verification field can be directly read from a single related independent data, use the related independent data containing the corresponding verification field as the data source for the verification field; when a verification field needs to be calculated from multiple related independent data, read the calculation rules associated with the process status verification conditions, select the corresponding data from the related independent data according to the input fields, data selection methods, and calculation steps recorded in the calculation rules, and calculate the verification field; when all verification fields are formed by multiple related independent data, the related independent data that participate in the calculation and directly provide the verification fields are combined into an independent data set.

[0092] If multiple independent data points directly provide the verification field for the same verification field, and both the training sample and all candidate independent data points have a collection time, the independent data point with the smallest absolute time difference between its collection time and the training sample's collection time is selected. If there are identical minimum time differences, the independent data point with a collection time no later than the training sample's collection time is prioritized. If multiple independent data points still have the same collection time, the first independent data point is selected in ascending order of the unique record number already present in the independent data source. If the training sample or candidate independent data points lack a collection time, the corresponding independent data point is used only if the discrete association information can uniquely identify a corresponding independent data point, and the aforementioned requirement of maintaining the process stage or process state without using collection time to establish the association relationship is met. If the aforementioned requirements cannot be met simultaneously, the independent data support result is determined as unsupported.

[0093] When an independent data record or set of independent data can form all the verification fields required for the process status verification conditions, the process status verification begins; if all the verification fields cannot be formed, the independent data support result is determined to be unsupported, and the reason for unsupport is recorded as the lack of independent data fields required to complete the process status verification; if no independent data with a correlation to the training samples is obtained, the independent data support result for the corresponding process status is determined to be unsupported, and the reason for unsupport is recorded as the lack of associated independent data; after the independent data support result is recorded, the condition confirmation result determination process in Example 3 begins.

[0094] When determining the verification conditions for process status according to the status judgment criteria in the established process procedure, the judgment items and judgment requirements for the corresponding process status in the process procedure are read; when the judgment requirement is a numerical range, the numerical range is directly used as the verification range of the corresponding independent data field; when the judgment requirement is an upper limit or lower limit, the upper limit or lower limit is directly used as the comparison boundary of the corresponding independent data field; when the judgment requirement is a category status, the category specified in the process procedure is used as the verification condition for process status.

[0095] When determining process status verification conditions based on actual production test results, select production test samples whose process status has been confirmed through manual inspection, laboratory inspection, or independent inspection equipment; read the verification fields in the independent data corresponding to the production test samples; if the production test report has given the range of verification fields used to determine the corresponding process status, directly use the range of verification fields in the production test report; if the production test report gives the status determination boundary, directly use the status determination boundary to form the process status verification conditions.

[0096] When determining process status verification conditions based on historical samples and their corresponding independent data, first screen historical samples whose process status has been confirmed by manual inspection, laboratory inspection, or independent testing equipment; after excluding invalid data records according to existing quality control rules of the data source, read the independent data value corresponding to the screened historical samples for each verification field; when all required verification fields have valid data values, if the verification field is numerical data, the minimum value among all independent data values ​​is used as the lower limit of the verification range and the maximum value among all independent data values ​​is used as the upper limit of the verification range; when any required verification field does not have a valid data value, do not establish process status verification conditions based on historical samples, and instead use the status judgment criteria in the process specification or the results of actual production tests to establish process status verification conditions.

[0097] When the validation field is categorical data, the categories that have appeared in the historical samples after statistical filtering and correspond to the target process state are used as the validation conditions for the categorical process state. When the process state needs to be judged using multiple validation fields, validation requirements are formed for each validation field. When the process specification or production test specifies the judgment logic between multiple validation fields, multiple validation requirements are combined according to the corresponding judgment logic. When neither the process specification nor the production test specifies the judgment logic between multiple validation fields, multiple validation requirements are combined using AND logic. When all validation fields meet the corresponding validation requirements, the independent data is determined to meet the process state validation conditions. When any validation field does not meet the corresponding validation requirements, the independent data is determined not to meet the process state validation conditions.

[0098] After the process status verification conditions are completed, the process status name and verification field are associated and saved. When the verification field needs to be calculated from multiple independent data, the input fields to be read to form the verification field, the independent data selection method, and the calculation steps are also saved. The calculation steps are read from the process specification, production test records, or the original calculation definition of the independent data. When using historical samples to determine the process status verification conditions, the confirmation results of the corresponding process status of the historical samples come from manual testing, laboratory testing, or independent testing equipment. The number of historical samples shall be in accordance with the sample number requirements specified in the process specification or production test. If the process specification and production test do not specify the sample number, all historical samples that meet the screening conditions shall be used.

[0099] After using historical samples to form the range of verification fields, read historical samples that have been confirmed by manual testing, laboratory testing, or independent testing equipment to belong to other process states, and form verification requirements for other process states according to the same verification fields; form a set of verification conditions for the target process state based on the verification requirements of all verification fields corresponding to the target process state, and form a set of verification conditions for other process states based on the verification requirements of all verification fields corresponding to other process states; if there exists a set of verification field values ​​that can simultaneously satisfy all verification requirements corresponding to the target process state and any other process state, it is determined that the verification conditions for the two process states overlap and cannot distinguish between process states; if there is no combination of verification field values ​​that can simultaneously satisfy all verification requirements for two process states, it is determined that the existing process state verification conditions can distinguish the corresponding process states.

[0100] When there is an overlap between the verification conditions of the target process state and other process states that cannot be distinguished, the process state verification conditions obtained based on historical samples are not directly adopted; when the state determination boundary can be obtained from the process specification or actual production test, the corresponding state determination boundary is adopted; when neither the process specification nor the actual production test can obtain the state determination boundary, the corresponding process state verification conditions are not established based on historical samples, the independent data support results of the corresponding process state are determined as unsupported, the reason for unsupported is recorded as the absence of process state verification conditions with state discrimination function, and the corresponding condition item is determined as unconfirmed in Example 3; when the process specification, actual production test results and historical samples cannot form process state verification conditions, they are handled in the same way.

[0101] Read an independent data record or set of independent data used for process status verification, along with the corresponding process status verification conditions; when the verification field exists directly in the independent data record, read the corresponding verification field value; when the verification field is calculated from multiple independent data, obtain the corresponding verification field value according to the calculation rules stored in the process status verification conditions; when the process status verification conditions include one verification field, compare the verification field value with the corresponding verification requirement; when the process status verification conditions include multiple verification fields, perform the comparison of each verification field separately, and merge multiple comparison results according to the judgment logic recorded in the process status verification conditions.

[0102] When all required verification requirements are met, a verification result indicating support for the corresponding process state is obtained, and the independent data support result for the corresponding process state is determined as supported. When any required verification requirement is not met, a verification result indicating non-support for the corresponding process state is obtained, and the independent data support result for the corresponding process state is determined as unsupported. The independent data support result is associated with and saved with the training samples, process knowledge, and corresponding process state.

[0103] When the independent data support result is "not supported," it is determined according to the condition confirmation result in Example 3 where it is not supported. If the reason for "not supported" is that the independent data does not meet the process status verification conditions, it means that the independent data currently used for verification does not meet the corresponding process status verification conditions. If the reason for "not supported" is that no associated independent data has been obtained, it means that no associated independent data that can be used to verify the corresponding process status has been obtained. If the reason for "not supported" is that the independent data field required to complete the process status verification is missing, it means that the current associated independent data cannot complete the process status verification. If the reason for "not supported" is that a process status verification condition with status discrimination function has not been formed, it means that there is currently no verification rule that can determine the corresponding process status based on the independent data. The aforementioned "not supported" results are not used to indicate that the corresponding process status is not valid. When new independent data is obtained in the future, verification will be performed again according to Example 6. When new process status verification conditions are formed in the future, verification will be performed again according to the process status verification process in this example.

[0104] Example 3: See Figure 3 Based on Examples 1 and 2, this embodiment further explains the process of determining whether process knowledge is used to constrain the training of the model to be trained when the applicable conditions of process knowledge include multiple conditions.

[0105] Read all the conditions in the applicable conditions of the process knowledge and the logical relationships between multiple conditions; for each condition, read the process state corresponding to the training sample and compare it according to the condition condition establishment rules; when the process state meets the condition condition establishment rules, record the corresponding condition as established; after completing the judgment of all conditions, determine the combination of conditions that can make the entire applicable conditions of the process knowledge established from the established conditions based on the logical relationships between multiple conditions.

[0106] When a logical relation requires multiple condition terms to be true simultaneously, all condition terms that need to be true simultaneously are placed into the same condition combination. When a logical relation allows any one of multiple condition terms to be true, each condition term that can individually make the corresponding logical branch true is formed into a condition combination. When a logical relation includes multiple layers of AND and OR logic, it is expanded layer by layer according to Boolean logic into condition combinations that can make the conditions applicable to the entire process knowledge true. Each condition combination includes a set of condition terms that need to be true simultaneously to make the conditions applicable to the entire process knowledge true. When two or more OR logic branches need to be true simultaneously, a condition term that has already been true is selected from each OR logic branch and combined to obtain all condition combinations that can make multiple OR logic branches true simultaneously.

[0107] When the logical relation indicates that condition A must be true, and the application of process knowledge is made possible by the truth of either condition B or condition C, the process state corresponding to the training sample makes condition A, condition B, and condition C all true, thus forming a condition combination consisting of condition A and condition B, and a condition combination consisting of condition A and condition C; the two condition combinations then enter the condition confirmation result determination process respectively.

[0108] For each combination of conditions, the data generation process and dependencies corresponding to each condition item in the combination are read. If the process state is directly determined by the data of the model processing results that do not depend on the model to be trained, and the condition confirmation result is determined as direct confirmation when the condition item is determined to be true according to the corresponding process state, the condition confirmation result is determined as direct confirmation. If the process state is determined by directly reading sensor detection data, detection equipment detection data, production control system recorded data, or manual detection data, the condition confirmation result is determined according to the aforementioned rules. If the process state is determined by processing the corresponding data according to the calculation rules of the model processing results that are fixed in advance and do not use the model to be trained, or by processing the corresponding data using other models with fixed parameters and input data that do not contain the model processing results of the model to be trained, the definition of directly determining the process state by the data of the model processing results that do not depend on the model to be trained is followed in the core terminology.

[0109] When the process status depends on the model processing result of the model to be trained, read the independent data support result obtained in Example 2; when the independent data support result is supported, determine the condition confirmation result of the corresponding condition item as supported and confirmed; when the independent data support result is not supported, determine the condition confirmation result of the corresponding condition item as unconfirmed; save the condition confirmation results corresponding to all condition items in each condition combination.

[0110] For the same process knowledge, first check if there is a condition combination where all condition items have a direct confirmation result; if there is a condition combination where all condition items have a direct confirmation result, determine that the process knowledge is used to constrain the training of the model to be trained, and form knowledge training constraints based on the process constraint relationship; after completing the above judgment, directly adopt the corresponding process knowledge usage result, and no longer change the process knowledge usage result based on the support confirmation or non-confirmation in other condition combinations.

[0111] If there is no condition combination in which all condition items are directly confirmed, check if there is a condition combination that does not include unconfirmed condition items and includes condition items that support confirmation; if there is a condition combination that does not include unconfirmed condition items and includes condition items that support confirmation, determine that process knowledge is used to constrain the training of the model to be trained, and adopt restricted knowledge training constraints; the restricted knowledge training constraints are formed according to Examples 4 and 5.

[0112] If the first two judgment results do not exist, and each condition combination includes a condition item whose condition confirmation result is unconfirmed, it is determined that the process knowledge is not used to constrain the training of the model to be trained; the training sample continues to participate in the training of the model to be trained according to the basic training process in Example 1; the corresponding process knowledge does not form the knowledge training loss of the current training sample; and a process knowledge usage result is obtained for the same process knowledge according to the aforementioned judgment order.

[0113] Example 4: This embodiment further explains the specific processing method of restricted knowledge training constraints based on embodiment 3; the different processing methods in embodiment 4 are applicable to the corresponding types of process constraint relationships; the first type of restricted knowledge training constraint, the second type of restricted knowledge training constraint, and the third type of restricted knowledge training constraint are described below.

[0114] In one implementation, the first type of restricted knowledge training constraint is applicable to situations where the process constraint relationship includes the target value or allowable range of the model prediction result, and the direction of change relationship between the set target process parameters and the model prediction result; the target value or allowable range and direction of change relationship of the model prediction result are read; when the first type of restricted knowledge training constraint is used, the knowledge training loss is not calculated based on the target value or allowable range of the model prediction result, but only based on the direction of change relationship; when performing scaling of the model prediction result in Example 1, the original prediction result and the changed prediction result are processed using the same scale before comparison and difference calculation are performed.

[0115] When the target process parameter increases, the model prediction result does not increase. The original prediction result and the change prediction result are formed according to Example 5. When the change prediction result is less than or equal to the original prediction result, the knowledge training loss corresponding to the change direction relationship is recorded as 0. When the change prediction result is greater than the original prediction result, the difference between the change prediction result and the original prediction result is calculated, and the square of the difference is used as the knowledge training loss corresponding to the change direction relationship.

[0116] When the target process parameter increases, the model prediction result should not decrease. If the change prediction result is greater than or equal to the original prediction result, the knowledge training loss corresponding to the change direction relationship is recorded as 0. If the change prediction result is less than the original prediction result, the difference between the original prediction result and the change prediction result is calculated, and the square of the difference is used as the knowledge training loss corresponding to the change direction relationship.

[0117] When the target process parameter decreases, the model prediction result does not increase. If the changed prediction result is less than or equal to the original prediction result, the knowledge training loss corresponding to the changed direction relationship is recorded as 0. If the changed prediction result is greater than the original prediction result, the square of the difference between the changed prediction result and the original prediction result is used as the knowledge training loss corresponding to the changed direction relationship.

[0118] In another implementation, the second type of restricted knowledge training constraint is applicable when the process constraint relationship includes the allowable range of the model prediction results; the lower limit and upper limit of the allowable range are read, and an existing data source is selected to determine the expanded allowable range; when the process specification has given the allowable deviation of the physical quantity corresponding to the model prediction results, the allowable deviation of the process specification is used first, the allowable deviation is subtracted from the lower limit of the allowable range and the allowable deviation is added to the upper limit of the allowable range; when determining the expanded allowable range at one time, only one expansion method is used, and the expansion amounts from different sources are not repeatedly superimposed.

[0119] When the process specification does not provide allowable deviations, and the test results of the testing equipment and the model prediction results correspond to the same physical quantity, use the same units, and the actual production results corresponding to the model prediction results in the training samples originate from the testing equipment, the upper limit of measurement error given in the technical data or calibration record of the testing equipment shall be adopted; the lower limit of the allowable range shall be subtracted from the upper limit of measurement error, and the upper limit of the allowable range shall be added to the upper limit of measurement error.

[0120] When neither the allowable deviation of the process specification nor the upper limit of the measurement error that meets the aforementioned conditions can be obtained, historical samples are selected from the historical samples that meet the applicable conditions of the current process knowledge and whose process status used to determine the applicable conditions of the current process knowledge has been confirmed by manual inspection, laboratory inspection, or independent testing equipment. Further, historical samples with actual production results corresponding to the model prediction results in the process constraint relationship, and whose actual production results have been confirmed as acceptable by the corresponding process specification, product inspection results, or production test results, are retained. The minimum value in the actual production results corresponding to the selected historical samples is compared with the original lower limit of the allowable range, and the maximum value in the actual production results is compared with the original upper limit of the allowable range. The expanded lower limit of the allowable range is the minimum of the original lower limit and the minimum value, and the expanded upper limit of the allowable range is the maximum of the original upper limit and the maximum value.

[0121] If there are no historical samples that meet the aforementioned screening conditions and have corresponding actual production results, the second type of restricted knowledge training constraint will not be used; if the current process knowledge also includes the process constraint relationship that can be used by the first type of restricted knowledge training constraint or the third type of restricted knowledge training constraint in Example 4, the corresponding restricted knowledge training constraint will be used; if there are no other process constraint relationships that can form knowledge training loss, the current training sample will only participate in the training of the model to be trained based on the basic training loss.

[0122] The expanded allowable range includes the original allowable range; when the model prediction result is within the expanded allowable range, the knowledge training loss corresponding to the allowable range is recorded as 0; when the model prediction result is lower than the lower limit of the expanded allowable range or higher than the upper limit of the expanded allowable range, the square of the difference between the model prediction result and the boundary of the expanded allowable range closest to the model prediction result is used as the knowledge training loss.

[0123] In another implementation, the third type of restricted knowledge training constraint is applicable when the process constraint relationship includes multiple constraint items that correspond to different process states. During process knowledge structuring, a correspondence is established between the process state corresponding to each constraint item and the process state corresponding to a condition item in the applicable conditions of the process knowledge. In Example 3, after determining the condition confirmation result for each condition item, the condition confirmation result corresponding to each constraint item is searched based on the aforementioned correspondence. Constraint items that cannot establish a correspondence with the process state corresponding to a condition item in the applicable conditions of the process knowledge are not screened using the third type of restricted knowledge training constraint. The condition confirmation result corresponding to the process state is retained as a constraint item that supports confirmation, and the retained constraint items are used to form restricted knowledge training constraints.

[0124] Constraints whose condition confirmation results for the corresponding process state do not support confirmation are not added to the current restricted knowledge training constraints. For each retained constraint, the corresponding knowledge training loss is calculated according to the target value, allowable range, or direction of change recorded in the process constraint relationship. When there are multiple retained constraint items, the knowledge training losses corresponding to the multiple retained constraint items are added together to obtain the knowledge training loss corresponding to the restricted knowledge training constraint of the current process knowledge. When there are no retainable constraint items, the third type of restricted knowledge training constraint is not used. When there are other process constraint relationships that can form knowledge training loss for the current process knowledge, the corresponding restricted knowledge training constraint is used. When there are no other process constraint relationships that can form knowledge training loss, only the basic training loss is used.

[0125] Example 5: See Figure 4 This embodiment, based on embodiment 4, further explains the specific processing method of applying knowledge training constraints to the model to be trained according to the relationship of changing direction.

[0126] First, check whether the applicable scope of process knowledge for the target process parameters is recorded according to the corresponding change direction relationship. If the applicable scope of process knowledge has been recorded, read the applicable scope of process knowledge, the allowable adjustment range of the equipment, and the allowable range of actual production corresponding to the target process parameters. The applicable scope of process knowledge is obtained according to the value limit of the target process parameters recorded in step S101. The allowable adjustment range of the equipment is read from the parameter limit of the equipment operation specifications or the production control system. The allowable range of actual production is read from the process specifications used in the current production batch or the parameter limit currently in effect of the production control system. The three ranges use the same unit to represent the target process parameters. If the applicable scope of process knowledge is not recorded, the processing method of determining knowledge training constraints based on change training samples is not adopted, and the processing method of determining knowledge training constraints based on two training samples is adopted instead.

[0127] When the scope of application of process knowledge has both a lower and upper limit, the intersection of the scope of application of process knowledge, the allowable adjustment range of equipment, and the allowable range of actual production is taken. The lower limit of the intersection is the maximum value among the three lower limits, and the upper limit of the intersection is the minimum value among the three upper limits. When the scope of application of process knowledge is limited only to the upper limit, the lower limit of the intersection is the maximum value among the lower limit of the allowable adjustment range of equipment and the lower limit of the allowable range of actual production, and the upper limit of the intersection is the minimum value among the upper limit of the scope of application of process knowledge, the upper limit of the allowable adjustment range of equipment, and the upper limit of the allowable range of actual production. When the scope of application of process knowledge is limited only to the lower limit, the lower limit of the intersection is the maximum value among the lower limit of the scope of application of process knowledge, the lower limit of the allowable adjustment range of equipment, and the lower limit of the allowable range of actual production, and the upper limit of the intersection is the minimum value among the upper limit of the allowable adjustment range of equipment and the upper limit of the allowable range of actual production.

[0128] When the lower limit of the intersection is less than or equal to the upper limit of the intersection, the intersection is determined as the range of variation of the target process parameters; when the lower limit of the intersection is greater than the upper limit of the intersection, no variation training samples are generated, and the processing method of determining knowledge training constraints based on two training samples in this embodiment is adopted.

[0129] In one possible implementation, when the change direction relationship requires the target process parameter to increase, the target process parameter in the training sample is compared with the upper limit of the change range; when the upper limit of the change range is greater than the target process parameter in the training sample, the target process parameter in the training sample is replaced with the upper limit of the change range to form a change training sample to be checked; when the change direction relationship requires the target process parameter to decrease, and the lower limit of the change range is less than the target process parameter in the training sample, the target process parameter in the training sample is replaced with the lower limit of the change range to form a change training sample to be checked; when a change in the target process parameter cannot be formed, the process switches to a method of determining knowledge training constraints based on two training samples.

[0130] When the target process parameter in the training sample has a clearly defined linkage relationship with other input parameters in the process specification, the corresponding input parameter in the changed training sample to be checked is modified synchronously according to the linkage relationship given in the process specification; when the process specification does not specify a linkage relationship, only the target process parameter is changed; all input parameters for which parameter limits have been set in the production control system are checked item by item; when the production control system or training sample data source has saved the allowable combination range of the target process parameter and other input parameters, the changed training sample to be checked is also checked to see if it is within the corresponding allowable combination range; when all input parameters meet the corresponding parameter limits and the combination of input parameters to be checked meets the allowable combination range, the changed training sample to be checked is determined as a changed training sample; when any check result does not meet the requirements, the corresponding modification result is not used as a changed training sample, and the process is switched to the method of determining knowledge training constraints based on two training samples.

[0131] After forming the variation training samples, the training samples and variation training samples are input into the same model to be trained to obtain the original prediction results and the variation prediction results. Example 1 requires that when performing scaling processing on the model prediction results, the original prediction results and the variation prediction results are processed using the same scale. The original prediction results and the variation prediction results after the same scaling processing are compared according to the change direction relationship in the process knowledge. When the model prediction results satisfy the change direction relationship, the knowledge training loss corresponding to the change direction relationship is recorded as 0. When the model prediction results violate the change direction relationship, the knowledge training loss is obtained according to the difference square calculation method of the first restricted knowledge training constraint in Example 4.

[0132] Multiple training samples and preset similar conditions for comparing the corresponding process states of the multiple training samples are obtained; the process states to be compared are determined to be all process states in the applicable conditions of process knowledge, excluding the target process parameter, which are used to limit the applicable environment of the change direction relationship; when the condition item in the applicable conditions of process knowledge directly uses the target process parameter as the judgment object, the target process parameter is not used as the process state to be compared; preset similar conditions are set for each process state to be compared.

[0133] When the process state being compared is a category-based process state, the same category is used as the preset similarity condition; when the process state being compared is a numerical process state, the absolute value of the difference between the two process state values ​​does not exceed the allowable fluctuation value of the corresponding process state in the process specification is used as the preset similarity condition; when the process specification does not provide an allowable fluctuation value and the measurement error range given by the corresponding detection equipment or sensor can be obtained, the measurement error range is used as the comparison range.

[0134] When the allowable fluctuation value in the process specification and the measurement error range of the corresponding detection equipment or sensor cannot be obtained, the corresponding numerical process state cannot form a preset similar condition, and the processing method of determining knowledge training constraints based on two training samples is not adopted; when all preset similar conditions can be formed, for each training sample to be compared, the dependency relationship corresponding to each process state to be compared is determined according to Examples 1 to 3; independent data support results are read only for process states whose dependency relationship indicates that the process state depends on the model processing result of the model to be trained.

[0135] When the process states involved in the comparison do not depend on the model processing results of the model to be trained, the corresponding process states are used for the judgment based on preset similarity conditions. When the process states involved in the comparison depend on the model processing results of the model to be trained, the corresponding process states are used for the judgment based on preset similarity conditions only when the corresponding independent data support results are supported. When any process state involved in the comparison depends on the model processing results of the model to be trained and the corresponding independent data support results are not supported, the corresponding training samples are not used to form training sample pairs for comparing the relationship of change direction.

[0136] If all process states involved in the comparison in two training samples can be used to judge the pre-set similarity conditions, the process states involved in the comparison in the two training samples are compared item by item; if all process states involved in the comparison meet the corresponding pre-set similarity conditions, and the target process parameters in the two training samples are different, then the two training samples are determined as the two training samples for comparing the relationship of change direction.

[0137] Two training samples are input into the same training model to obtain two model prediction results. Example 1 requires that the two model prediction results be processed using the same scale when performing scaling of the model prediction results. The magnitude relationship of the target process parameters of the two training samples is compared, and then the magnitude relationship that the two model prediction results need to satisfy is determined according to the change direction relationship in the process knowledge. When the two model prediction results satisfy the change direction relationship, the knowledge training loss is recorded as 0. When the two model prediction results violate the change direction relationship, the square of the difference between the two model prediction results that violate the change direction relationship is used as the knowledge training loss.

[0138] Determining knowledge training constraints based on changing training samples and determining knowledge training constraints based on two training samples are two separate processing methods. When changing training samples can be formed within the range of target process parameter variations, the changing training sample processing method is used. When changing training samples cannot be formed, it is checked whether there are two training samples that meet preset similar conditions and have different target process parameters; if two training samples that meet the conditions exist, the two training sample processing method is used. When changing training samples cannot be formed, and there are no two training samples among multiple training samples that meet preset similar conditions and have different target process parameters, no knowledge training loss is formed based on the current change direction relationship. When there are no other process constraint relationships that can form knowledge training loss for the current process knowledge, the current training sample only participates in the training of the model to be trained based on the basic training loss.

[0139] Example 6: This embodiment further explains the re-verification process after obtaining newly added independent data, based on embodiments 1 to 3.

[0140] When process knowledge is not used to constrain the training of the model to be trained, record the process knowledge, the sample identifier corresponding to the training sample, and the process status that needs to be independently verified by data; the process knowledge is associated with the complete process knowledge record stored in the process knowledge storage location; when the training sample has no sample identifier, record the production batch, equipment identifier, material identifier, collection location, or collection time that the training sample can obtain; and save the process knowledge, training sample, and process status together with the same record.

[0141] After obtaining the newly added independent data, read the sample identifier, production batch, equipment identifier, material identifier, collection location or collection time corresponding to the newly added independent data; according to the field semantic check, association relationship judgment, independent data combination and verification field formation rules in Example 2, find the training samples that are related to the newly added independent data, and determine a newly added independent data record or set of newly added independent data that can form all verification fields; after finding the related training samples, obtain the corresponding process status and process knowledge according to the related records, and read the process status verification conditions that have been established for the corresponding process status.

[0142] When a newly added independent data record or a newly added independent data set can form all the verification fields required for the process state verification conditions, all verification fields are compared with the process state verification conditions and an updated verification result is obtained. If the updated verification result indicates that the corresponding process state is supported, the independent data support result for the process state is updated to supported. If the updated verification result indicates that the corresponding process state is not supported, the independent data support result for the process state is updated to unsupported. If no associated newly added independent data is obtained or the associated newly added independent data cannot form all the verification fields, the original independent data support result is not updated, and it is recorded that no new verification result has been formed. Only when an updated verification result is obtained is it re-determined, according to Example 3, whether the corresponding process knowledge is used to constrain the subsequent training of the model to be trained.

[0143] When re-determining process knowledge to constrain the subsequent training of the model to be trained, knowledge training constraints are formed based on the results of the re-determined use of process knowledge, starting from the next use of the corresponding training sample or the training sample that matches the corresponding process state. When re-determining process knowledge not to constrain the subsequent training of the model to be trained, the corresponding training continues to be performed using only the basic training loss. New independent data only updates the process knowledge judgment results used in subsequent training, and does not require re-executing the already completed training process of the model to be trained from the first training batch.

[0144] Example 7: This embodiment, based on embodiment 1, further explains the processing method when the same training sample is matched with multiple process knowledge.

[0145] When matching training samples with multiple process knowledge, each process knowledge is used to constrain the training of the model to be trained, according to Examples 1 to 5 respectively, and knowledge training constraints are formed for each process knowledge that can be used to constrain the training of the model to be trained; process constraint relationship conflict checks are performed only for two or more process knowledge that have been determined to constrain the training of the model to be trained; the process constraint relationship corresponding to the process knowledge participating in the conflict check is read, and checked according to the model prediction result field and the target process parameter input field respectively; when the process constraint relationship includes multiple constraint items, each constraint item is first split into corresponding comparison content according to the target value, allowable range or change direction relationship recorded in the constraint item.

[0146] For target values ​​and allowable ranges acting on the same model prediction result field, the target value is converted into a range where both the upper and lower limits are equal to the target value, while the allowable range remains the original range. The intersection of the ranges corresponding to all target values ​​and allowable ranges is calculated. If the intersection has common values, it is determined that there are common satisfyable values ​​between the target value and the allowable range. If the intersection is empty, it is determined that the corresponding process constraints cannot be simultaneously satisfied. When there is only one target value or one allowable range in the same model prediction result field, the corresponding target value or allowable range is directly used as the result of the common satisfyable value check.

[0147] For the directional relationship, grouping is done according to the target process parameter input field, the direction of change of the target process parameter, and the model prediction result field. When all directional relationships in the same group require the model prediction result not to increase or not to decrease, it is determined that the directional relationships in the same group can be simultaneously satisfied. When the directional relationships in the same group require the model prediction result not to increase and not to decrease respectively, the commonly satisfied relationship is determined to be that the changed model prediction result is equal to the unchanged model prediction result, and it is determined that the two directional relationships can be simultaneously satisfied. Only when multiple directional relationships do not have a commonly satisfied relationship in the magnitude of the model prediction results is it determined that the corresponding process constraint relationship cannot be simultaneously satisfied. The directional relationships corresponding to different target process parameter input fields, different target process parameter change directions, or different model prediction result fields are checked separately.

[0148] When different types of process constraints apply to the same model prediction result field, first perform a check on the common satisfyable values ​​of the target value and the allowable range, and then perform a check on the common satisfyable magnitude relationship of the change direction relationship. If the target value and the allowable range have common satisfyable values, and multiple change direction relationships have common satisfyable model prediction result magnitude relationships, then using the common satisfyable values ​​as both the model prediction result before and after the change can satisfy the change direction relationship that allows the model prediction result to remain unchanged. Therefore, it is determined that the corresponding process constraints can be established simultaneously. If the target value and the allowable range do not have common satisfyable values, or multiple change direction relationships do not have common satisfyable model prediction result magnitude relationships, it is determined that the corresponding process constraints cannot be established simultaneously. The constraint item is first converted into the aforementioned comparison content according to the target value, allowable range, or change direction relationship recorded in the constraint item, and then the same check is performed. The process constraints applied to different model prediction result fields are checked separately. When the check results corresponding to all model prediction result fields are all simultaneously valid, it is determined that the process constraints corresponding to multiple process knowledge can be established simultaneously.

[0149] When the process constraints corresponding to multiple pieces of process knowledge can be simultaneously satisfied, the knowledge training loss corresponding to the multiple pieces of process knowledge is added to the basic training loss, and the model parameters of the model to be trained are updated based on the added training loss. When the process constraints corresponding to multiple pieces of process knowledge cannot be simultaneously satisfied, the multiple pieces of process knowledge for which the corresponding process constraints cannot be simultaneously satisfied are identified as a set of conflicting process knowledge. In the current training, the knowledge training loss corresponding to the set of conflicting process knowledge is not added to the training loss used in the current training. Only the basic training loss and the training losses of other available knowledge that do not belong to the set of conflicting process knowledge are added together to update the model parameters of the model to be trained. The model training execution device records the set of conflicting process knowledge and the corresponding training samples.

[0150] If the process constraint relationships corresponding to multiple process knowledges cannot be established simultaneously, it means that the knowledge training constraints corresponding to the conflicting process knowledge sets are not used together in the current training. After the current training is completed, the independent use judgment result of each process knowledge obtained according to Examples 1 to 5 is retained. When subsequent training samples only match a single process knowledge in the conflicting process knowledge set or when the conflict relationship disappears, the process knowledge usage result is processed according to the process knowledge that has been obtained or re-determined.

[0151] Example 8: This embodiment uses the drying production process as an application scenario to illustrate the processing relationships in Embodiments 1 to 7. The model to be trained is used to predict the remaining processing time based on the processing temperature, weight, equipment operation data, and other production data that have been used as inputs to the model to be trained in the training samples. The actual remaining processing time corresponding to the training samples is obtained based on the time difference between the collection time of the training samples and the completion time of the corresponding production batch, and is only used as the supervision result for calculating the basic training loss. The process state in the applicable conditions of process knowledge is determined according to the data generation process that can be obtained in the actual application stage of the model to be trained. The actual remaining processing time is not used as the data generation source for the current drying stage, so that the training stage and the actual application stage of the model to be trained adopt the same process state generation path.

[0152] The process knowledge stipulates that when the material is in the later stage of drying, the moisture content is lower than the moisture content specified in the process knowledge, and the equipment is in normal operation, the remaining processing time is within the allowable range specified in the process knowledge, and the remaining processing time does not increase when the processing temperature increases. The conditions of being in the later stage of drying, the material moisture content being lower than the moisture content specified in the process knowledge, and the equipment being in normal operation are respectively set as conditions in the applicable conditions of the process knowledge, and all three conditions are set to be true simultaneously. The relationship between the allowable range of the remaining processing time and the direction of change of the remaining processing time not increasing when the processing temperature increases is recorded as a process constraint relationship. The applicable range corresponding to the processing temperature is read from the process specification from which the process knowledge is derived, and the corresponding applicable range is recorded as the applicable range of the process knowledge for the processing temperature.

[0153] The material moisture content is obtained by an independent moisture content detection device, and the device's operating status is recorded by the production control system. The corresponding condition is determined directly based on the material moisture content and the device's operating status. The remaining processing time boundary for entering the drying stage is read from the process procedure executed by the current production batch. The remaining processing time boundary and the remaining processing time output by the model to be trained use the same time unit. The training sample is input into the model to be trained to obtain the remaining processing time. When the remaining processing time is less than or equal to the remaining processing time boundary, the current process state is determined to be the drying stage. After reading the data generation process corresponding to the current drying stage, the model processing result dependent on the model to be trained is determined.

[0154] According to the association rules in Example 2, independent moisture content detection data and weight detection data that are associated with the training samples are obtained; the weight detection data is obtained by the weight detection equipment, and the data generation process of the weight detection data does not use the output results, intermediate processing results, or model processing results of the model to be trained; the verification conditions of the current process state corresponding to the drying stage are established based on historical samples that have been confirmed to have entered the drying stage by manual inspection; the independent moisture content detection data corresponding to the historical samples are read, and the weight detection data associated with the historical samples are arranged in ascending order of collection time, and the weight detection data with the smallest time difference with the historical sample collection time is selected before and after the historical sample collection time.

[0155] When the acquisition time of a later weight measurement data point is greater than that of a previous weight measurement data point, the weight of the later weight measurement data point is subtracted from the weight of the previous weight measurement data point, and then divided by the time difference between the later and previous acquisition times to obtain the weight change rate per unit time. If there is a lack of weight measurement data before or after the historical sample acquisition time, or if the acquisition times of two weight measurement data points are the same, the weight change rate per unit time for the corresponding historical sample is not calculated. Only historical samples for which independent moisture content measurement data can be obtained and whose weight change rate per unit time can be calculated are retained.

[0156] For example, if the first weight measurement data is 100kg and the second weight measurement data is 98kg, and the time interval between the two weight measurement data is 10min, the weight change rate per unit time is -0.2kg / min. When the verification range of the weight change rate per unit time in the process status verification conditions is -0.25kg / min to -0.15kg / min, -0.2kg / min is within the corresponding verification range. Therefore, the weight change rate per unit time meets the corresponding verification requirements.

[0157] When historical samples are retained, all retained historical samples are processed in the same way, and the minimum and maximum values ​​of independent moisture content detection data and weight change rate per unit time are used as the range boundaries of the two verification fields, respectively. According to Example 2, the verification condition sets corresponding to the current drying stage and other process states are formed respectively, and it is determined whether there is a set of independent moisture content detection data and weight change rate per unit time that can simultaneously meet all the verification requirements corresponding to the two process states. When there is a combination of verification field values ​​that can simultaneously meet all the verification requirements of the two process states, the state determination boundary in the process specification or actual production test is adopted according to Example 2. If the state determination boundary cannot be obtained, the historical sample range is not used to establish the process state verification conditions corresponding to the current drying stage.

[0158] If there are no retained historical samples, or if the range of historical samples cannot form process state verification conditions with state discrimination function due to the aforementioned overlap, process state verification conditions shall be established sequentially from the state judgment criteria in the process specification or the actual production test results in accordance with Example 2; if neither the state judgment criteria in the process specification nor the actual production test results can form process state verification conditions, the current process knowledge shall not be used to form knowledge training loss, and the current training samples shall only participate in the training of the model to be trained based on the basic training loss.

[0159] Once the process state verification conditions corresponding to the current drying stage have been established, an independent data set is formed, consisting of the independent moisture content detection data associated with the current training sample and the weight detection data used to calculate the weight change rate per unit time before and after the current training sample collection time. When independent moisture content detection data and two weight detection data are available, and the collection time of the latter weight detection data is greater than that of the former weight detection data, the weight change rate per unit time corresponding to the current training sample is obtained according to the calculation rules saved in the process state verification conditions. Then, the independent moisture content detection data and the weight change rate per unit time are compared with the corresponding verification ranges. When both verification fields are within the corresponding verification ranges, the independent data support result for the current drying stage is determined to be supported. When either verification field is outside the corresponding verification range, the independent data support result for the current drying stage is determined to be unsupported.

[0160] If there is a lack of weight detection data before or after the current training sample collection time, or if two weight detection data collection times are the same, the independent data that can complete the process status verification is not obtained, and the independent data support result of the current drying stage is determined as unsupported; if the data generation process of material moisture content and equipment operating status does not depend on the model processing result of the model to be trained, and both conditions are met, the condition confirmation result of the two conditions is determined as direct confirmation.

[0161] The current processing result of the model to be trained depends on the drying stage, and the independent data support result of the current drying stage is supportive. Therefore, the condition confirmation result of the corresponding condition item in the current drying stage is determined as support confirmation. The condition combination consisting of the three condition items does not include unconfirmed condition items and includes support confirmation condition items. Therefore, process knowledge is determined to constrain the training of the model to be trained, and restricted knowledge training constraints are adopted. The process constraint relationship includes both the allowable range of the remaining processing time and the change direction relationship that the remaining processing time does not increase when the processing temperature increases. Therefore, according to the first restricted knowledge training constraint processing in Example 4, the knowledge training loss is not formed based on the allowable range of the remaining processing time, but only based on the change direction relationship that the remaining processing time does not increase when the processing temperature increases.

[0162] When the process constraint relationship includes the direction of change of the remaining processing time, the applicable range of process knowledge, the allowable adjustment range of the equipment, and the allowable range of actual production corresponding to the processing temperature are obtained according to Example 5, and the range of change of processing temperature is determined; within the range of change of processing temperature, the processing temperature in the training sample is increased to form a change training sample; the training sample and the change training sample are processed respectively to obtain the original prediction result and the change prediction result.

[0163] The change direction relationship stipulates that the remaining processing time does not increase when the processing temperature increases; when the change prediction result is less than or equal to the original prediction result, the knowledge training loss corresponding to the change direction relationship is recorded as 0; when the change prediction result is greater than the original prediction result, the square of the difference between the change prediction result and the original prediction result is used as the knowledge training loss; the knowledge training loss is added to the basic training loss corresponding to the training sample and the model parameters of the model to be trained are updated, thereby completing the process knowledge enhancement training corresponding to the current training sample.

[0164] Example 9: See Figure 5 This embodiment provides a process knowledge enhancement system for a model; the process knowledge enhancement system for a model includes a knowledge acquisition module, a process state determination module, a dependency determination module, a process state verification module, a knowledge constraint determination module, and a model training module.

[0165] The knowledge acquisition module is used to acquire training samples and process knowledge to assist in training the model to be trained, and provides the training samples and process knowledge to the process state determination module; the process knowledge includes the applicable conditions of the process knowledge and the process constraint relationship that is established when the applicable conditions of the process knowledge are met; the data reading method performed by the knowledge acquisition module is the same as the data reading method in step S101 of embodiment 1.

[0166] The process state determination module is used to determine the corresponding process state based on the training samples, and to determine the process knowledge that matches the training samples based on the process state and the applicable conditions of the process knowledge. The process state determination module provides the process knowledge that matches the training samples and the process state used to determine the applicable conditions of the process knowledge to the dependency determination module. The processing procedure performed by the process state determination module is the same as the processing procedure of step S102 in Example 1.

[0167] The dependency determination module is used to determine the dependency relationship between the process state used to determine the applicable conditions of process knowledge and the model processing result of the model to be trained; the dependency determination module reads the data generation process corresponding to the process state and determines the dependency relationship according to the data source tracing method in Example 2; the dependency determination module provides the dependency relationship to the process state verification module and the knowledge constraint determination module.

[0168] The process state verification module is used to verify the process state that the process state depends on the model processing result of the model to be trained, which represents the dependency relationship, to obtain the verification result, and to determine the independent data support result of the corresponding process state based on the verification result. The process state verification module obtains independent data that are related to the training samples according to the association method in Example 2, and determines an independent data record or independent data set for process state verification according to the source of the verification field, and then completes the verification according to the process state verification conditions of the corresponding process state.

[0169] The knowledge constraint determination module is used to determine whether process knowledge matching the training samples is used to constrain the training of the model to be trained, based on the dependency relationship and the independent data support results corresponding to the process state of the model processing results that depend on the model to be trained. When the process knowledge application conditions include multiple conditions, the knowledge constraint determination module determines the condition combination and condition confirmation results according to Example 3, and determines the process knowledge usage results according to the judgment order in Example 3.

[0170] When the knowledge constraint determination module determines to use restricted knowledge training constraints, it determines the restricted knowledge training constraints corresponding to the process constraint relationship according to Example 4, and forms the knowledge training loss according to Example 5 when the process constraint relationship includes the change direction relationship; when new independent data is obtained, the knowledge constraint determination module updates the process knowledge usage results according to Example 6; when the training sample matches multiple process knowledge, the knowledge constraint determination module determines whether the process constraint relationship corresponding to multiple process knowledge can be established at the same time according to Example 7.

[0171] The model training module is used to train the model to be trained using training samples, and when process knowledge is used to constrain the training of the model to be trained, knowledge training constraints are applied to the model to be trained according to the process constraint relationship; the model training module calculates the basic training loss and knowledge training loss according to step S106 in Example 1, and updates the model parameters of the model to be trained according to the training loss used in the current training.

[0172] The knowledge acquisition module, process state determination module, dependency determination module, process state verification module, knowledge constraint determination module, and model training module in the model's process knowledge enhancement system can be deployed in the same model training execution device, or they can be deployed in multiple computing devices capable of data communication. The training samples, process knowledge, process state, dependency, independent data support results, and knowledge training constraints transmitted between multiple computing devices use the same data field names.

[0173] The above embodiments are used to illustrate the technical solution of this application; while keeping the processing relationship between training samples, process knowledge, process status, dependency relationship, independent data support results, process knowledge usage results and knowledge training constraints unchanged, the execution entity can be set in a model training server, industrial computing device or computing device capable of executing model training program according to the model training deployment method; the data source and storage location can be configured according to the existing data interface of the industrial production system.

[0174] The method processing procedures in Examples 1 to 8 correspond to the module processing procedures in Example 9. The same processing procedure uses the exact same training samples, process knowledge, applicable conditions of process knowledge, process constraint relationships, process states, model to be trained, model processing results, independent data, independent data support results, condition combinations, condition confirmation results, and knowledge training constraints in different examples.

[0175] The above description is only a preferred embodiment of this application. Without departing from the essence of the technical solution of this application, any improvements and modifications made by those skilled in the art should fall within the protection scope of this application. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of this application should also be considered within the protection scope of this application.

Claims

1. A method for enhancing the process knowledge of a model, characterized in that, include: Acquire training samples and process knowledge used to assist in training the model to be trained. The process knowledge includes the applicable conditions of the process knowledge and the process constraint relationships that hold when the applicable conditions of the process knowledge are met. The corresponding process state is determined based on the training samples, and the process knowledge that matches the training samples is determined based on the process state and the applicable conditions of the process knowledge. Determine the dependency between the process state used to determine the applicability of process knowledge and the model processing results of the model to be trained; The process states that the process states depend on the model processing results of the model to be trained are validated, the validation results are obtained, and the independent data support results for the corresponding process states are determined based on the validation results. Based on the dependencies and the independent data support results corresponding to the process states of the model processing results that depend on the model to be trained, determine whether the process knowledge that matches the training samples is used to constrain the training of the model to be trained. The training samples are used to train the model to be trained, and when process knowledge is used to constrain the training of the model to be trained, knowledge training constraints are applied to the model to be trained according to the process constraint relationship.

2. The method according to claim 1, characterized in that, Determine the dependency between the process state used to determine the applicability of process knowledge and the model processing results of the model to be trained, including: The process of generating data corresponding to the process status; When the data generation process uses the output results, intermediate processing results, or data obtained by further processing the model processing results of the model to be trained, it is determined that there is a dependency between the process state and the model processing results of the model to be trained. Otherwise, there is no dependency between determining the process state and the model processing results of the model to be trained.

3. The method according to claim 1, characterized in that, The process state is validated based on the process state of the model processing results, which represent the dependence of the process state on the model being trained. This includes: Obtain independent data and the corresponding sample identifier, production batch, equipment identifier, material identifier, collection location or collection time for training samples and independent data; Determine the association between independent data and training samples based on sample identifier, production batch, equipment identifier, material identifier, collection location or collection time; The corresponding process state is validated based on independent data that are correlated with the training samples.

4. The method according to claim 3, characterized in that, Validation of the corresponding process state is performed based on independent data that are correlated with the training samples, including: Obtain the process status verification conditions used to verify the corresponding process status; When independent data meets the process status verification conditions, a verification result indicating support for the corresponding process status is obtained; When independent data does not meet the process status verification conditions, a verification result is obtained indicating that the corresponding process status is not supported.

5. The method according to claim 4, characterized in that, The conditions for verifying the process status are determined based on the status judgment criteria in the established process specifications, the results of actual production tests, or historical samples and the independent data corresponding to the historical samples. When determining process status verification conditions based on historical samples and corresponding independent data, the process status corresponding to the historical samples is confirmed by manual testing, laboratory testing, or independent testing equipment.

6. The method according to claim 1, characterized in that, The applicable conditions of process knowledge include multiple condition items that correspond to the process state, and process knowledge also includes the logical relationship between multiple condition items; Based on the process status, determine whether multiple conditions are true or false, and based on the logical relationship between multiple conditions, determine one or more combinations of conditions from the conditions that are true, which can make the process knowledge applicable.

7. The method according to claim 6, characterized in that, Also includes: Independent data supports the results, including both supported and unsupported data. Determine the condition confirmation result for each condition item in each condition combination; When the process state corresponding to a condition item is directly determined by data from the model processing results that do not depend on the model to be trained, and the condition item is determined to be valid based on the corresponding process state, the condition confirmation result is determined as direct confirmation. When the process state corresponding to the condition item depends on the model processing result of the model to be trained and the independent data support result is supported, the condition confirmation result will be determined as support confirmation. When the process state corresponding to a condition item depends on the model processing result of the model to be trained and the independent data support result is not supported, the condition confirmation result will be determined as unconfirmed.

8. The method according to claim 7, characterized in that, Determining whether the process knowledge matching the training samples is used to constrain the training of the model to be trained includes: When there exists a combination of conditions, and the confirmation results of all conditions in the combination are direct confirmations, process knowledge is determined to constrain the training of the model to be trained, and knowledge training constraints are applied to the model to be trained according to the process constraint relationship.

9. The method according to claim 7, characterized in that, Determining whether the process knowledge matching the training samples is used to constrain the training of the model to be trained includes: When there is no combination of conditions where all condition confirmation results are directly confirmed, but there is a combination of conditions that does not include unconfirmed condition items and includes condition items that support confirmation, process knowledge is determined to constrain the training of the model to be trained, and the constrained knowledge training constraint is adopted.

10. The method according to claim 7, characterized in that, Determining whether the process knowledge matching the training samples is used to constrain the training of the model to be trained includes: When all combinations of conditions include condition items whose confirmation result is unconfirmed, it is determined that process knowledge is not used to constrain the training of the model to be trained.

11. The method according to claim 9, characterized in that, The process constraints include the target value or allowable range of the model prediction results, as well as the relationship between the set target process parameters and the change direction of the model prediction results. The change direction relationship is used to characterize the change direction of the model prediction results when the target process parameters increase or decrease. The constrained knowledge training constraint does not impose constraints based on the target value or the allowable range, but imposes knowledge training constraints on the model to be trained based on the relationship of the direction of change.

12. The method according to claim 9, characterized in that, Process constraints include the allowable range of model prediction results; The constrained knowledge training constraint applies knowledge training constraints to the model to be trained using an expanded allowable range, where the expanded allowable range includes the allowable range in the process constraint relationship.

13. The method according to claim 9, characterized in that, The process constraints include multiple constraint items that correspond to different process states. The restricted knowledge training constraints include retaining the condition confirmation results of the corresponding process state from multiple constraint items as the constraint items that support confirmation, and applying knowledge training constraints to the model to be trained based on the retained constraint items.

14. The method according to claim 11, characterized in that, Based on the relationship of changing directions, knowledge training constraints are imposed on the model to be trained, including: Obtain the applicable scope of process knowledge, the allowable adjustment range of equipment, and the allowable range of actual production corresponding to the target process parameters, and determine the range of variation of the target process parameters based on the applicable scope of process knowledge, the allowable adjustment range of equipment, and the allowable range of actual production. By changing the target process parameters in the training samples within the range of variation, a variable training sample is obtained. The training samples and the changed training samples are processed by the model to be trained respectively to obtain the original prediction results and the changed prediction results. The knowledge training constraints are determined based on whether the original prediction results and the changed prediction results satisfy the relationship of the change direction.

15. The method according to claim 11, characterized in that, Based on the relationship of changing directions, knowledge training constraints are imposed on the model to be trained, including: Acquire multiple training samples and preset similarity conditions for comparing the corresponding process states of the multiple training samples; Identify two training samples from multiple training samples that meet preset similar conditions for process states but have different target process parameters; Two training samples are processed using the model to be trained to obtain two model prediction results. Knowledge training constraints are determined based on whether the two model prediction results satisfy the relationship of change direction.

16. The method according to claim 1, characterized in that, Also includes: Record the process knowledge that was not used to constrain the training of the model to be trained, as well as the corresponding training samples and process states; Acquire new independent data associated with the training samples, and re-validate the process status based on the new independent data to obtain updated validation results; The independent data support results for the process state are updated based on the updated validation results, and the process knowledge is re-determined whether it should be used to constrain the subsequent training of the model to be trained based on the updated independent data support results.

17. The method according to claim 1, characterized in that, When training samples are matched with multiple process knowledge sets, this also includes: Determine whether the process constraint relationships corresponding to multiple process knowledge can be valid simultaneously. When multiple process constraints corresponding to process knowledge can be simultaneously established, the knowledge training constraints corresponding to multiple process knowledge are used together to constrain the training of the model to be trained. When multiple process constraints corresponding to process knowledge cannot be simultaneously established, the knowledge training constraints corresponding to multiple process knowledge will not be used together to constrain the training of the model to be trained.

18. A process knowledge enhancement system for a model, characterized in that, include: The knowledge acquisition module is used to acquire training samples and process knowledge to assist in training the model to be trained. The process knowledge includes the applicable conditions of the process knowledge and the process constraint relationships that hold when the applicable conditions of the process knowledge are met. The process state determination module is used to determine the corresponding process state based on the training samples, and to determine the process knowledge that matches the training samples based on the process state and the applicable conditions of process knowledge. The dependency determination module is used to determine the dependency between the process state used to determine the applicability of process knowledge and the model processing results of the model to be trained. The process state verification module is used to verify the process state that represents the process state's dependence on the model's processing results, obtain the verification results, and determine the independent data support results for the corresponding process state based on the verification results. The knowledge constraint determination module is used to determine whether the process knowledge matching the training samples is used to constrain the training of the model, based on the dependency relationship and the independent data support results corresponding to the process state that depends on the model processing results of the model to be trained. The model training module is used to train the model to be trained using training samples, and to apply knowledge training constraints to the model to be trained based on the process constraint relationship when process knowledge is used to constrain the training of the model to be trained.

Images