Wheat germ production abnormity root cause tracing method and system based on NLP

Abstract

The invention provides a wheat germ production abnormity root cause tracing method and system based on NLP, and the method comprises the steps: firstly obtaining a multi-batch production record data set of a target production line, covering a production process text, an environment monitoring time sequence and raw material quality index data, then carrying out the analysis of the production process text data to obtain a process semantic feature vector, and carrying out the analysis of the process semantic feature vector; the method comprises the following steps: extracting environment monitoring time sequence data to obtain an environment fluctuation feature vector, fusing the environment fluctuation feature vector and the environment fluctuation feature vector to generate a fused production feature set, distributing an abnormal root cause weight to the fused feature set by using a pre-trained multi-layer sensing network model, generating an abnormal association degree score set, and performing combined tracing on raw material and process data to obtain an abnormal root cause tracing result. And finally, generating a dynamic optimization strategy according to a tracing result, and feeding back the dynamic optimization strategy to a production control system to calibrate parameters, thereby realizing wheat germ production abnormity root cause tracing and production line optimization.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology. Specifically, it relates to a method and system for tracing the root cause of abnormal wheat germ production based on NLP. Background Art

[0002] In the field of wheat germ production, the existing production anomaly diagnosis and root cause tracing technologies face many challenges, which seriously restrict the improvement of production efficiency and product quality. Currently, the production anomaly analysis methods commonly used in the industry mainly rely on single data sources or simple data statistical analysis, lacking effective integration and in-depth mining of multi-source heterogeneous data. Specifically, when processing production process text data, most of the existing technologies only perform keyword matching or simple text classification, unable to accurately parse the process semantics and extract its deep features, resulting in a superficial understanding of the production process. At the same time, for environmental monitoring time series data, the existing methods usually only perform basic statistical descriptions or trend analysis, making it difficult to capture the dynamic fluctuation characteristics of environmental factors and their potential impacts on the production process. Summary of the Invention

[0003] In view of the problems mentioned above, in combination with the first aspect of this application, embodiments of this application provide a method for tracing the root cause of abnormal wheat germ production based on NLP. The method includes: Obtain a multi-batch wheat germ production record data set corresponding to the target production line. The production record data set includes production process text data, environmental monitoring time series data, and raw material quality index data for each production batch; Perform semantic feature parsing processing on the production process text data to obtain process semantic feature vectors, perform dynamic fluctuation feature extraction processing on the environmental monitoring time series data to obtain environmental fluctuation feature vectors, and perform cross-modal feature fusion processing on the process semantic feature vectors and the environmental fluctuation feature vectors to generate a fused production feature set; Call a pre-trained multi-layer perceptron network model to perform abnormal root cause weight assignment processing on the fused production feature set to generate an abnormal correlation score set corresponding to the production batch, where the abnormal correlation score set includes the abnormal correlation distribution between the raw material quality index data and each production process; Perform joint root cause tracing processing on the raw material quality index data and the production process text data based on the abnormal correlation distribution to generate an abnormal root cause tracing result for the production batch. The abnormal root cause tracing result is used to indicate the production link and raw material defect type of the abnormal source; Generate a dynamic optimization strategy for production line parameters according to the abnormal root cause tracing result, and feedback the dynamic optimization strategy to the production control system to trigger parameter calibration operations.

[0004] In another aspect, an embodiment of the present application further provides a production service system, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions, or codes, and the processor is used to run the programs, instructions, or codes in the machine-readable storage medium to implement the above method.

[0005] Based on the above aspects, the embodiment of the present application deeply integrates the semantic parsing of production process text data, the dynamic feature extraction of environmental monitoring time-series data, and the quantitative analysis of raw material quality index data. Through cross-modal feature fusion processing, a set of fusion production features that comprehensively reflects the internal relationship of the production process is generated. On this basis, the pre-trained multi-layer perceptron network model not only reveals the complex interaction mechanism between raw material quality indicators and each production process through accurate abnormal root cause weight allocation, but also quantifies the contribution degree of each factor to production anomalies. Further, based on the joint root cause tracing process of abnormal correlation distribution, the accurate mapping from data association to actual production links and raw material defect types is realized, significantly improving the accuracy and efficiency of root cause tracing. Finally, by generating a dynamic optimization strategy for production line parameters and feeding it back to the production control system, a closed-loop control mechanism is formed, which can not only effectively solve the current production anomaly problems, but also prevent potential risks. Description of the Drawings

[0006] Figure 1 It is a schematic execution flowchart of the method for tracing the root cause of production anomalies in wheat germ production based on NLP provided by an embodiment of the present application.

[0007] Figure 2 It is a schematic hardware architecture diagram of the production service system provided by an embodiment of the present application. Detailed Embodiments

[0008] The present application will be specifically described below with reference to the accompanying drawings of the specification. Figure 1 It is a schematic flowchart of the method for tracing the root cause of production anomalies in wheat germ production based on NLP provided by an embodiment of the present application. The method for tracing the root cause of production anomalies in wheat germ production based on NLP will be introduced in detail below.

[0009] Step S110, obtain a set of production record data corresponding to the target production line for wheat germ, where the set of production record data includes production process text data, environmental monitoring time-series data, and raw material quality index data for each production batch.

[0010] For example, in a wheat germ production factory, there is a target production line for producing wheat germ. In order to comprehensively monitor and optimize the production process, it is necessary to obtain a set of production record data corresponding to this production line for multiple batches.

[0011] Each production batch has detailed production process text data records. For example, in a certain batch, the production process text data details multiple steps starting from the screening of raw wheat, followed by cleaning, crushing, separation, drying, etc. In the screening step, it is recorded that "a vibrating screen is used to screen the wheat, with a screen aperture of 5 mm to remove impurities and stones larger than 5 mm"; the cleaning step records "the screened wheat is put into the cleaning tank and rinsed for 15 minutes at a water flow rate of 10 liters per minute", etc., with detailed operation descriptions.

[0012] In terms of environmental monitoring time-series data, the factory has installed multiple sensors in the production workshop to monitor environmental parameters such as temperature, humidity, and air pressure in real time. Taking a day of production as an example, starting production at 8 am, the temperature sensor records data every 10 minutes. For example, the temperature is 22 degrees Celsius at 8 am and 22.5 degrees Celsius at 8:10 am; the humidity sensor also records every 10 minutes, with a humidity of 50% at 8 am and 51% at 8:10 am; the air pressure sensor records at the same frequency, with an air pressure of 101.3 kPa at 8 am and 101.2 kPa at 8:10 am. These data form a time-varying environmental monitoring time-series data sequence.

[0013] The raw material quality index data covers multiple aspects of wheat. For example, the moisture content of wheat is 12%, the protein content is 15%, and the impurity content is 1%, etc. For each batch of raw wheat, a comprehensive quality inspection is carried out and these index data are recorded. By collecting these data from multiple batches, a multi-batch wheat germ production record data set corresponding to the target production line is formed.

[0014] Step S120, perform semantic feature parsing processing on the production process text data to obtain a process semantic feature vector, perform dynamic fluctuation feature extraction processing on the environmental monitoring time-series data to obtain an environmental fluctuation feature vector, and perform cross-modal feature fusion processing on the process semantic feature vector and the environmental fluctuation feature vector to generate a fused production feature set.

[0015] In this embodiment, for the production process text data, first, process step segmentation processing is performed. Taking the production process text data of a certain batch mentioned above as an example, it is segmented into multiple process step description text fragments according to different operation steps. For example, "using a vibrating screen to screen the wheat, with a screen aperture of 5 mm to remove impurities and stones larger than 5 mm" becomes an independent fragment, and "putting the screened wheat into the cleaning tank and rinsing for 15 minutes at a water flow rate of 10 liters per minute" also becomes a fragment, etc.

[0016] Next, call the pre-trained language representation model to perform context semantic encoding processing on each process step description text segment. For example, for the segment "Use a vibrating screen to screen wheat, with a screen aperture of 5 mm, to remove impurities and stones larger than 5 mm", the language representation model can analyze the meaning of each word in the entire text context and encode it into an initial semantic feature vector. Suppose the initial semantic feature vector is represented by a set of numbers as [0.2, 0.3, …, 0.5], which contains the semantic information of this process step description.

[0017] Then, perform process domain feature enhancement processing. For example, match the set of domain entities associated with the current process step description text segment from the preset wheat germ production knowledge base. For the screening process, the set of domain entities matched in the knowledge base may include "vibrating screen", "screen specification", "impurity type", etc. Input these sets of domain entities into the language representation model for entity semantic encoding processing to generate a set of entity feature vectors. For example, "vibrating screen" is encoded into a vector [0.1, 0.4, …, 0.6], "screen specification" is encoded into a vector [0.3, 0.5, …, 0.7], etc. Perform attention mechanism weighted fusion on this set of entity feature vectors and the initial semantic feature vector. For example, according to the attention mechanism, the weights of entities such as "vibrating screen", "screen specification", and "impurity type" are calculated as 0.4, 0.3, and 0.3 respectively, and an enhanced semantic feature vector is generated through weighted calculation.

[0018] Finally, perform temporal position encoding splicing processing on the enhanced semantic feature vectors of each process step description text segment to generate a process semantic feature vector. Suppose after the above processing, enhanced semantic feature vectors of 5 process steps are obtained, namely vector A, vector B, vector C, vector D, and vector E. According to the chronological order of the production process, splice them together in sequence to form a complete process semantic feature vector.

[0019] For environmental monitoring time series data, first perform abnormal fluctuation interval detection processing. By analyzing the temperature, humidity, and air pressure monitoring data over a period of time, identify the abnormal fluctuation time window. For example, during a continuous week of production, it is found that between 2 pm and 4 pm on a certain day, the temperature rapidly rises from the normal 22 degrees Celsius to 28 degrees Celsius, the humidity also drops from 50% to 40%, and the air pressure fluctuates from 101.3 kPa to 101.8 kPa. This time period is identified as an abnormal fluctuation time window.

[0020] Perform multi-scale sliding window sampling processing on the original monitoring data within this abnormal fluctuation time window. For example, taking 10 minutes as a small window, starting from 2 pm, sequentially collect multiple local time series data segments such as from 2 pm to 2:10 pm, from 2:10 pm to 2:20 pm, etc. Each segment contains the change data of temperature, humidity, and air pressure within these 10 minutes.

[0021] Call the pre-trained temporal convolutional network model to perform local fluctuation feature extraction processing on each local time series data segment, generating local fluctuation feature vectors. For example, for the local time series data segment from 2 pm to 2:10 pm, after being processed by the temporal convolutional network model, a local fluctuation feature vector [0.3, 0.4, …, 0.6] is generated, which reflects the fluctuation characteristics of environmental parameters during this time period.

[0022] Perform global temporal attention aggregation processing on the local fluctuation feature vectors of each local time series data segment. First, calculate the similarity score between each local fluctuation feature vector and a preset global fluctuation pattern template. Assume the preset global fluctuation pattern template vector is [0.2, 0.5, …, 0.7]. For a certain local fluctuation feature vector, by calculating the similarity algorithm between them, the similarity score is obtained as 0.6. Based on these similarity scores, perform dynamic weighted summation on each local fluctuation feature vector. For example, there are 5 local fluctuation feature vectors, and the similarity scores are 0.6, 0.5, 0.7, 0.4, 0.8 respectively. Calculate the dynamic weights according to these scores, and then perform weighted summation on these local fluctuation feature vectors to obtain a weighted fluctuation feature vector. Finally, perform feature difference processing on the weighted fluctuation feature vector and the global fluctuation pattern template to generate an environmental fluctuation feature vector.

[0023] Perform cross-modal feature fusion processing on the process semantic feature vector and the environmental fluctuation feature vector. First, perform time dimension alignment processing on the process semantic feature vector, mapping it to the same time granularity as the environmental fluctuation feature vector. For example, the environmental fluctuation feature vector has a time granularity of 10 minutes, and the process semantic feature vector was originally in units of each process step. Through appropriate interpolation or aggregation operations, the process semantic feature vector is also converted to a time granularity of 10 minutes.

[0024] Construct a process-environment cross-attention mechanism to calculate the cross-modal correlation matrix between the semantic features of each time step in the process semantic feature vector and the fluctuation features of the corresponding time step in the environmental fluctuation feature vector. For example, within a certain 10-minute time step, the process semantic feature vector represents the relevant semantics of the cleaning process, and the environmental fluctuation feature vector represents the temperature, humidity, and air pressure fluctuations during this time period. By calculating the correlation value between them, a correlation matrix is formed.

[0025] Based on this cross-modal correlation matrix, two-way feature interaction processing is performed on the process semantic feature vector and the environmental fluctuation feature vector to generate an interaction semantic feature vector and an interaction fluctuation feature vector. For example, according to the correlation matrix, some information in the process semantic feature vector that is strongly correlated with the environmental fluctuation feature is transmitted to the environmental fluctuation feature vector, and at the same time, relevant information in the environmental fluctuation feature vector is transmitted to the process semantic feature vector, thereby generating a new interaction semantic feature vector and an interaction fluctuation feature vector.

[0026] Finally, the interaction semantic feature vector and the interaction fluctuation feature vector are processed by a gating mechanism fusion. First, calculate the feature complementarity score between the interaction semantic feature vector and the interaction fluctuation feature vector. For example, through a certain calculation method, their feature complementarity score is obtained as 0.7. Based on this score, a dynamic fusion weight vector is generated. Suppose the weight of the interaction semantic feature vector is calculated as 0.4 and the weight of the interaction fluctuation feature vector is 0.6 according to the score. Use this dynamic fusion weight vector to perform weighted splicing on the interaction semantic feature vector and the interaction fluctuation feature vector to generate a fused production feature set.

[0027] Step S130, call a pre-trained multi-layer perceptron network model to perform abnormal root cause weight assignment processing on the fused production feature set, and generate an abnormal correlation score set corresponding to the production batch, where the abnormal correlation score set includes the abnormal correlation distribution between the raw material quality index data and each production process.

[0028] In this embodiment, the fused production feature set can be input into the sparse feature screening layer of the multi-layer perceptron network model for redundant feature filtering processing. The fused production feature set contains feature information after fusion of multiple aspects such as processes and environments. For example, there are 100 feature dimensions. The sparse feature screening layer will analyze the correlation and redundancy between these features. For example, it is found that there are two feature dimensions, one is the operation speed of a certain process at a specific time, and the other is the operation strength of the same process at a similar time. After analysis, the correlation between them is as high as 0.9. Then, one of the more representative features will be retained and the redundant feature will be removed. After this processing, a key production feature set after screening is obtained. Suppose it is reduced from the original 100 feature dimensions to 60.

[0029] Invoke the cross - feature generation layer of the multi - layer perceptron network model to perform high - order feature combination processing on the set of key production features. For example, for the 60 key production features after screening, the cross - feature generation layer combines different features. For instance, it combines feature A (operation time of a certain process) and feature B (ambient temperature) to form a new high - order feature "the correlation feature between the operation time of a certain process and the ambient temperature". Through such a combination method, a cross - production feature matrix is generated. Assume the size of this cross - production feature matrix is 60×60, and each element represents a different high - order feature combination.

[0030] Invoke the attention allocation layer of the multi - layer perceptron network model to calculate the anomaly sensitivity weights for the cross - production feature matrix. The attention allocation layer analyzes the sensitivity of each high - order feature combination to anomalies based on historical data and model training. For example, after analysis, it is found that "the correlation feature between the operation time of a certain process and the ambient temperature" often shows significant changes in past abnormal production situations. Then, a relatively high anomaly sensitivity weight, assume the weight is 0.8, will be assigned to this cross - production feature. For some feature combinations that do not change significantly in abnormal situations, a lower weight, such as 0.2, will be assigned. In this way, the anomaly sensitivity distribution for each production feature dimension is generated.

[0031] Perform feature - weighted aggregation processing on the cross - production feature matrix based on the anomaly sensitivity distribution. For example, for each row or column in the cross - production feature matrix, weighted summation is performed according to the corresponding anomaly sensitivity weights. Assume a row has 60 elements (representing different high - order feature combinations), and the corresponding anomaly sensitivity weights are 0.2, 0.3, …, 0.8, etc. Through weighted calculation, the elements in this row are aggregated into a value, and finally an anomaly correlation score set is generated. In this anomaly correlation score set, each anomaly correlation score corresponds to the correlation strength between the raw material quality index and a specific production process. For example, a relatively high - scoring item may indicate a strong abnormal correlation between the moisture content of raw material wheat and the drying process.

[0032] Step S140, perform joint root - cause tracing processing on the raw material quality index data and the production process text data based on the anomaly correlation distribution, and generate the anomaly root - cause tracing result for the production batch. The anomaly root - cause tracing result is used to indicate the production link where the anomaly source is located and the type of raw material defect.

[0033] In this embodiment, abnormal correlation items exceeding a preset threshold can be screened out according to the abnormal correlation degree distribution to generate a candidate root cause feature set. For example, the preset threshold is 0.6. In the abnormal correlation score set, it is found that 10 score items exceed 0.6. One item indicates that the abnormal correlation degree between the protein content of raw wheat and the crushing process is 0.7, and another item indicates that the abnormal correlation degree between the environmental humidity and the separation process is 0.8, etc. These items exceeding the threshold constitute the candidate root cause feature set.

[0034] Perform reverse semantic parsing on each abnormal correlation item in the candidate root cause feature set to determine its corresponding raw material quality defect description and production process abnormality description. For the item "the abnormal correlation degree between the protein content of raw wheat and the crushing process is 0.7", through reverse semantic parsing, it is found that the protein content of raw wheat may be too low, resulting in uneven particles in the crushing process; for the item "the abnormal correlation degree between the environmental humidity and the separation process is 0.8", descriptions such as too high environmental humidity, which reduces the efficiency of the separation process and results in poor separation effect are parsed.

[0035] Call the pre-trained root cause inference model to perform joint causal reasoning on the raw material quality defect description and the production process abnormality description. Input the raw material quality defect description into the raw material defect encoder of the root cause inference model for defect type encoding. For example, for "the protein content of raw wheat is too low", a defect type feature vector [0.3, 0.5, …, 0.7] is encoded, which represents the characteristics of this raw material quality defect. Input the production process abnormality description into the process abnormality encoder of the root cause inference model for abnormal mode encoding. For example, for "uneven particles in the crushing process", an abnormal mode feature vector [0.4, 0.6, …, 0.8] is encoded.

[0036] Construct a defect-abnormality causal graph network, and input the defect type feature vector and the abnormal mode feature vector as node features into the causal graph network for multi-hop causal reasoning. In the causal graph network, multi-hop reasoning is performed by analyzing the connection relationships and weights between different nodes. For example, starting from the node of too low protein content of raw wheat, through the connection relationships in the network, it is found that it may affect the subsequent grinding process, and further affect the extraction quality of wheat germ. Output the causal association path between the defect type feature vector and each abnormal mode feature vector through the path generation layer of the causal graph network to generate a root cause inference path set.

[0037] Perform confidence evaluation processing on the set of root cause reasoning paths, and select the top k reasoning paths with the highest confidence as the abnormal root cause tracing results. For example, there are 20 reasoning paths in the set of root cause reasoning paths. Through a certain confidence evaluation algorithm, the confidence of each path is calculated. Suppose the top 3 reasoning paths with the highest confidence are as follows: the first path indicates that the too low protein content of the raw wheat leads to uneven particles in the crushing process, which in turn affects the subsequent processes; the second path points out that the too high environmental humidity affects the efficiency of the separation process, resulting in a decrease in product purity; the third path shows that the too high impurity content of the raw material is not completely removed in the screening process, affecting the subsequent processing quality. These top 3 reasoning paths are used as the abnormal root cause tracing results to indicate the production links and raw material defect types of the abnormal sources.

[0038] Step S150, generate a dynamic optimization strategy for the production line parameters according to the abnormal root cause tracing results, and feedback the dynamic optimization strategy to the production control system to trigger parameter calibration operations.

[0039] In this embodiment, the production links and raw material defect types of the abnormal sources in the abnormal root cause tracing results can be parsed. For example, from the above abnormal root cause tracing results, it is clear that the production links of the abnormal sources include the crushing process, the separation process, the screening process, etc., and the raw material defect types are too low protein content, too high humidity, too high impurity content, etc.

[0040] Match the set of historical parameter adjustment records associated with the production links of the abnormal sources from the historical optimization strategy library. The historical optimization strategy library stores the parameter adjustment records for different production links and problems in the past. For the crushing process, some historical records are found, such as once adjusting the rotation speed of the crushing equipment from 1000 revolutions per minute to 1200 revolutions per minute to improve particle uniformity; for the separation process, there is a record of extending the separation time from 30 minutes to 40 minutes to improve the separation effect; for the screening process, there is a record of changing the screen aperture from 5 mm to 4 mm to better remove impurities, etc.

[0041] Perform validity filtering processing on the set of historical parameter adjustment records based on the raw material defect types. For example, for the situation of too low protein content of the raw material, those historical records for adjusting the moisture content of the raw material are filtered out. After filtering, an effective adjustment strategy set is obtained. Suppose for too low protein content, the strategies of adjusting the rotation speed of the crushing equipment and adjusting the raw material ratio are retained; for too high humidity, the strategies of increasing the power of the dehumidification equipment and adjusting the temperature of the separation process are retained; for too high impurity content, the strategies of changing the screen aperture and strengthening the screening process time are retained, etc.

[0042] Perform multi-objective optimization on the set of effective adjustment strategies to generate dynamic optimization strategies that meet the current production constraint conditions. First, construct a three-dimensional optimization objective space that includes production efficiency, raw material loss rate, and abnormal recurrence rate. Map each effective adjustment strategy to the corresponding coordinate point in the three-dimensional optimization objective space. For example, for the strategy of "adjusting the rotation speed of the crushing equipment", after evaluation, the production efficiency has increased by 10%, the raw material loss rate has decreased by 5%, and the abnormal recurrence rate is expected to decrease by 30%. Map it to a coordinate point (0.1, -0.05, -0.3) in the three-dimensional space.

[0043] Use the Pareto front analysis algorithm to screen out the set of candidate optimization strategies located on the Pareto optimal front. The Pareto front analysis algorithm will analyze the performance of each strategy on the three objectives and find those strategies that cannot further improve a certain objective without reducing other objectives. For example, after analysis, 5 strategies are located on the Pareto optimal front, namely Strategy A, Strategy B, Strategy C, Strategy D, and Strategy E.

[0044] Call the strategy recommendation model to perform an adaptability scoring process on the set of candidate optimization strategies based on the real-time load status of the current production line. Assume that the real-time load status of the current production line shows that the equipment is running relatively stably, but the raw material supply is slightly tight. The strategy recommendation model can score the 5 candidate strategies according to this status. For example, although Strategy A can significantly improve production efficiency, it has a large demand for raw materials, and its adaptability score is 0.4 under the current tight raw material supply situation; Strategy B has a relatively small demand for raw materials while improving production efficiency, and its adaptability score is 0.7.

[0045] Finally, select the candidate optimization strategy with the highest adaptability score as the dynamic optimization strategy. In the above example, Strategy B has the highest adaptability score, so Strategy B is used as the dynamic optimization strategy and fed back to the production control system to trigger the parameter calibration operation. For example, after receiving Strategy B, the production control system will automatically adjust the parameters of relevant equipment, such as adjusting the rotation speed and temperature of the crushing equipment, to achieve the optimized operation of the production line, improve production quality and efficiency, and reduce the probability of raw material loss and abnormal occurrences.

[0046] Based on the above steps, the embodiments of the present application deeply integrate the semantic parsing of production process text data, the dynamic feature extraction of environmental monitoring time-series data, and the quantitative analysis of raw material quality index data. Through cross-modal feature fusion processing, a set of fusion production features that comprehensively reflect the internal relationships of the production process is generated. On this basis, the pre-trained multi-layer perceptron network model not only reveals the complex interaction mechanism between raw material quality indicators and each production process through accurate abnormal root cause weight allocation, but also quantifies the contribution degree of each factor to production anomalies. Further, based on the joint root cause tracing process of the abnormal correlation distribution, the accurate mapping from data correlation to actual production links and raw material defect types is realized, significantly improving the accuracy and efficiency of root cause tracing. Finally, by generating a dynamic optimization strategy for production line parameters and feeding it back to the production control system, a closed-loop control mechanism is formed, which can not only effectively solve the current production anomaly problems, but also prevent potential risks.

[0047] In a possible implementation manner, the semantic feature parsing process performed on the production process text data to obtain a process semantic feature vector includes: Step S121, perform process step segmentation processing on the production process text data to obtain multiple process step description text segments.

[0048] Taking the production process text data of a certain batch as an example, this production process text data details the entire process from the input of raw material wheat to the output of the final wheat germ product. The entire production process includes large stages such as raw material preparation, pretreatment, core processing, and post-treatment. In the raw material preparation stage, it can be further divided into process steps such as wheat procurement and acceptance, and wheat storage; the pretreatment stage can be divided into wheat screening, wheat cleaning, etc.; the core processing stage includes crushing, separation, extraction, etc.; the post-treatment stage includes drying, packaging, etc. Each subdivided process step forms an independent process step description text segment. For example, the process step description text segment of the wheat screening process is "Use a vibrating screen to screen the purchased wheat. The model of the vibrating screen is XYZ-50, the screen aperture is set to 4 mm, and the screening time lasts for 30 minutes. The purpose is to remove large particle impurities such as stones and straw in the wheat."

[0049] Step S122, call the pre-trained language representation model to perform context semantic encoding processing on each of the process step description text segments to generate an initial semantic feature vector.

[0050] Taking the text fragment describing the wheat screening process steps just now as an example, the language representation model will deeply analyze each word and sentence structure in the text. It will understand that the "vibrating screen" is a tool for screening, "XYZ-50" is the specific model of this tool, "4 mm" specifies the key parameter of the sieve mesh, "30 minutes" stipulates the duration of the screening operation, and "removing large particle impurities" elaborates on the purpose of the operation. Based on the understanding of the entire text context, these semantic information are encoded into an initial semantic feature vector. This initial semantic feature vector can be understood as an information packet containing various semantic information. For example, a certain dimension represents tool-related information, and a certain dimension represents operation duration information, etc. Suppose the generated initial semantic feature vector is [0.1, 0.3, 0.4, 0.2, 0.5, 0.1, 0.3], and each value represents the feature intensity of different semantic dimensions.

[0051] Step S123, perform process domain feature enhancement processing on the initial semantic feature vector to obtain an enhanced semantic feature vector. The process domain feature enhancement processing includes the following steps: Match the set of domain entities associated with the current process step description text fragment from the preset wheat germ production knowledge base. Input the set of domain entities into the language representation model for entity semantic encoding processing to generate a set of entity feature vectors. Perform attention mechanism weighted fusion on the set of entity feature vectors and the initial semantic feature vector to generate the enhanced semantic feature vector.

[0052] In this embodiment, a large amount of professional knowledge related to wheat germ production is stored in the preset wheat germ production knowledge base. After matching, the obtained set of domain entities includes "the working principle of the vibrating screen", "the influence of different sieve mesh apertures on the screening effect", "common impurity types and removal methods", etc. Input these sets of domain entities into the language representation model for entity semantic encoding processing. For example, "the working principle of the vibrating screen" is encoded into an entity feature vector [0.2, 0.4, 0.3, 0.1, 0.5], "the influence of different sieve mesh apertures on the screening effect" is encoded as [0.3, 0.5, 0.2, 0.4, 0.1], and "common impurity types and removal methods" is encoded as [0.4, 0.3, 0.2, 0.5, 0.1]. These vectors constitute the set of entity feature vectors.

[0053] The attention mechanism assigns weights according to the relevance of each entity to the current process step description. For example, after analysis and calculation, the weight of "the working principle of the vibrating screen" is 0.3, the weight of "the influence of different screen aperture sizes on the screening effect" is 0.4, and the weight of "common impurity types and removal methods" is 0.3. Then, according to the weighted calculation rule, the initial semantic feature vector is fused with these entity feature vectors. Taking the first dimension as an example, calculate the value of the first dimension of the enhanced semantic feature vector: the value of the first dimension of the initial semantic feature vector is 0.1, the value of the first dimension of the entity feature vector of "the working principle of the vibrating screen" is 0.2, the value of the first dimension of the entity feature vector of "the influence of different screen aperture sizes on the screening effect" is 0.3, the value of the first dimension of the entity feature vector of "common impurity types and removal methods" is 0.4, and the fused value of the first dimension is 0.1×0.3 + 0.2×0.4 + 0.3×0.3 = 0.2. By analogy, calculate each dimension, and finally generate the enhanced semantic feature vector.

[0054] Step S124: Perform a temporal position encoding splicing process on the enhanced semantic feature vectors of the text fragments of each process step description to generate the process semantic feature vector.

[0055] For example, according to the sequence of production processes, such as wheat screening first, then wheat cleaning, and then crushing, etc. Assume that the enhanced semantic feature vector of the wheat screening process is vector A, the enhanced semantic feature vector of the wheat cleaning process is vector B, the enhanced semantic feature vector of the crushing process is vector C, etc. Concatenate these vectors in sequence according to the temporal position. For example, first arrange all the dimension values of vector A in the front, then arrange all the dimension values of vector B, and then arrange all the dimension values of vector C. Finally, generate the process semantic feature vector.

[0056] In a possible implementation manner, the performing dynamic fluctuation feature extraction processing on the environmental monitoring time series data to obtain an environmental fluctuation feature vector includes: Step S125: Perform an abnormal fluctuation interval detection process on the environmental monitoring time series data to identify the abnormal fluctuation time windows in the temperature, humidity, and air pressure monitoring data.

[0057] In this embodiment, sensors installed in the production workshop continuously collect temperature, humidity, and air pressure monitoring data, and the above data is recorded at regular time intervals, for example, once per minute. By analyzing the data over a period of time, such as analyzing the temperature data within a week, it is found that between 10 am and 11 am on a certain day, the temperature rapidly rises from the normal 23 degrees Celsius to 28 degrees Celsius, and then rapidly drops to 22 degrees Celsius between 11 am and 12 pm. This time period is identified as an abnormal fluctuation time window in the temperature monitoring data; at the same time, within this time period, the humidity drops from 55% to 45%, and the air pressure fluctuates from 101.2 kPa to 101.8 kPa, which also constitutes an abnormal fluctuation time window for humidity and air pressure monitoring data.

[0058] Step S126, perform multi-scale sliding window sampling processing on the original monitoring data within the abnormal fluctuation time window to obtain multiple local time series data segments.

[0059] For example, taking the temperature data as an example, set the small-scale sliding window to 5 minutes, the medium-scale sliding window to 10 minutes, and the large-scale sliding window to 15 minutes. Starting from 10 am, the small-scale sliding window collects the temperature data from 10 am to 10:05 am as a local time series data segment, and then collects the data from 10:05 am to 10:10 am as the next segment; the medium-scale sliding window collects the temperature data from 10 am to 10:10 am as a local time series data segment, and then collects the data from 10:10 am to 10:20 am as the next segment; the large-scale sliding window collects the temperature data from 10 am to 10:15 am as a local time series data segment, and then collects the data from 10:15 am to 10:30 am as the next segment. The same multi-scale sliding window sampling processing is also performed on the humidity and air pressure data, so as to obtain multiple local time series data segments.

[0060] Step S127, call the pre-trained temporal convolutional network model to perform local fluctuation feature extraction processing on each local time series data segment to generate local fluctuation feature vectors.

[0061] Taking a 5 - minute local time - series data segment of temperature as an example, this segment records the temperature values per minute from 10:00 to 10:05, which are 23 degrees Celsius, 24 degrees Celsius, 25 degrees Celsius, 26 degrees Celsius, and 27 degrees Celsius respectively. The temporal convolutional network model will analyze features such as the change trend and change amplitude of these data. For example, the temperature difference between adjacent time points can be calculated. The temperature increased by 1 degree Celsius from 10:00 to 10:01, and by 1 degree Celsius from 10:01 to 10:02, etc. Through a series of convolutional calculations and feature extraction operations, a local fluctuation feature vector is generated. Suppose the generated local fluctuation feature vector is [0.3, 0.4, 0.5, 0.2, 0.1], and this local fluctuation feature vector reflects the fluctuation characteristics of the temperature data within these 5 minutes. Such processing is performed on all local time - series data segments, including temperature, humidity, and air - pressure data segments collected by different - scale sliding windows, to generate their respective local fluctuation feature vectors.

[0062] Step S128: Perform global temporal attention aggregation processing on the local fluctuation feature vectors of each of the local time - series data segments to generate the environmental fluctuation feature vector, where the global temporal attention aggregation processing includes: Step S1281: Calculate the similarity score between each local fluctuation feature vector and a preset global fluctuation pattern template.

[0063] In this embodiment, the preset global fluctuation pattern template is determined based on a large amount of historical environmental data and experience, and represents normal and common environmental fluctuation patterns. Taking a local temperature fluctuation feature vector as an example, assume the global fluctuation pattern template vector is [0.2, 0.5, 0.4, 0.3, 0.1]. When calculating the similarity score, a certain similarity calculation method is adopted, such as calculating the reciprocal of the sum of the squared differences of the corresponding dimension values of the two vectors. For the first dimension, the difference is 0.3 - 0.2 = 0.1, and the square is 0.01; for the second dimension, the difference is 0.4 - 0.5=-0.1, and the square is 0.01; for the third dimension, the difference is 0.5 - 0.4 = 0.1, and the square is 0.01; for the fourth dimension, the difference is 0.2 - 0.3=-0.1, and the square is 0.01; for the fifth dimension, the difference is 0.1 - 0.1 = 0, and the square is 0. The sum of the squared differences is 0.01 + 0.01 + 0.01 + 0.01+0 = 0.04, and the reciprocal is 25, that is, the similarity score between this local fluctuation feature vector and the global fluctuation pattern template is 25. Such similarity score calculations are performed on all local fluctuation feature vectors.

[0064] Step S1282: Perform dynamic weighted summation on each local fluctuation feature vector based on the similarity score to obtain a weighted fluctuation feature vector.

[0065] Suppose that through similarity score calculation, the similarity scores of 10 local fluctuation feature vectors are 20, 30, 25, 15, 22, 28, 35, 18, 24, and 26 respectively. Calculate the dynamic weights based on these scores. The weight calculation method can be dividing each score by the sum of all scores. The sum of all scores is 20 + 30 + 25 + 15 + 22 + 28 + 35 + 18 + 24 + 26 = 243. The weight of the first local fluctuation feature vector is 20÷243≈0.082, the second weight is 30÷243≈0.123, and so on. Then, perform weighted summation on each local fluctuation feature vector according to the weights. Taking the first dimension as an example, suppose the first dimension values of 10 local fluctuation feature vectors are 0.1, 0.2, 0.3, 0.1, 0.2, 0.3, 0.4, 0.1, 0.2, 0.3 respectively. The weighted sum is 0.1×0.082 + 0.2×0.123 + 0.3×0.103 + 0.1×0.062 + 0.2×0.090 + 0.3×0.115 + 0.4×0.144 + 0.1×0.074 + 0.2×0.099 + 0.3×0.107≈0.23. Perform such calculations for each dimension to obtain the weighted fluctuation feature vector.

[0066] Step S1283, perform feature difference processing on the weighted fluctuation feature vector and the global fluctuation pattern template to generate the environmental fluctuation feature vector.

[0067] In this embodiment, still taking the first dimension as an example, the first dimension value of the weighted fluctuation feature vector is 0.23, and the first dimension value of the global fluctuation pattern template is 0.2. The difference is 0.23 - 0.2 = 0.03. Calculate such differences for each dimension, and finally generate the environmental fluctuation feature vector. This vector comprehensively reflects the fluctuation characteristics of environmental monitoring data within the abnormal fluctuation time window and provides key information for subsequent production analysis and optimization.

[0068] In a possible implementation manner, the cross-modal feature fusion processing of the process semantic feature vector and the environmental fluctuation feature vector to generate the fused production feature set includes: Step S129, perform time dimension alignment processing on the process semantic feature vector and map it to the same time granularity as the environmental fluctuation feature vector.

[0069] In this embodiment, the process semantic feature vector is generated through a series of processes based on production process text data, while the environmental fluctuation feature vector is extracted from environmental monitoring time series data. In this factory, the process semantic feature vector is based on each production process step and has a relatively coarse time granularity, while the environmental fluctuation feature vector is based on data collected at fixed short time intervals (such as every 10 minutes).

[0070] Taking a certain batch of production as an example, the process semantic feature vectors were originally divided according to process steps such as raw material screening, cleaning, and crushing, and the time spans of each process step were different. For example, the raw material screening process lasted for 30 minutes, the cleaning process lasted for 20 minutes, etc. In order to align with the environmental fluctuation feature vectors with a time step of every 10 minutes, the raw material screening process was divided according to a time granularity of 10 minutes, and semantic feature representations of 3 time steps were formed within these 30 minutes. Through the re - sorting and integration of the process operation content and time, each time step of the process semantic feature vector corresponded to the time step of the environmental fluctuation feature vector, completing the alignment in the time dimension.

[0071] Step S1210: Construct a process - environment cross - attention mechanism, and calculate the cross - modal correlation matrix between the semantic features of each time step in the process semantic feature vector and the fluctuation features of the corresponding time step in the environmental fluctuation feature vector.

[0072] For example, within a certain 10 - minute time step, the process semantic feature vector represents the operation semantic information such as the equipment rotation speed and pressure in the crushing process, and the environmental fluctuation feature vector represents the fluctuation information of temperature, humidity, and air pressure during this period. Through a specific calculation method, the degree of association between the operation semantics of the crushing process and the environmental factor fluctuations is measured. When calculating the correlation between the equipment rotation speed and temperature, considering that the increase in temperature may affect the performance of the equipment and thus the crushing effect, the correlation value between them is determined by analyzing historical data and production knowledge. Such calculations are performed for the process semantic features and environmental fluctuation features of each time step, and finally a cross - modal correlation matrix is formed. Each element of this matrix represents the association strength between the process and environmental features at a specific time step.

[0073] Step S1211: Based on the cross - modal correlation matrix, perform two - way feature interaction processing on the process semantic feature vector and the environmental fluctuation feature vector to generate an interaction semantic feature vector and an interaction fluctuation feature vector.

[0074] Taking the above-mentioned crushing process and the environmental characteristics of the corresponding time steps as an example, according to the correlation matrix, part of the information in the process semantic feature vector that is strongly correlated with environmental fluctuations is transmitted to the environmental fluctuation feature vector. For example, if the correlation degree shows a high correlation between the equipment rotation speed and the temperature, then the relevant semantic information of the equipment rotation speed is incorporated into the part about temperature in the environmental fluctuation feature vector, so that the environmental fluctuation feature vector can better reflect this correlation. At the same time, the relevant information in the environmental fluctuation feature vector is transmitted to the process semantic feature vector. For example, the potential impact information of temperature on the crushing effect is fed back to the process semantic feature vector to make it more comprehensive when representing the crushing process. Through this two-way transmission, an interactive semantic feature vector and an interactive fluctuation feature vector are generated.

[0075] Step S1212, perform gated mechanism fusion processing on the interactive semantic feature vector and the interactive fluctuation feature vector to generate the fused production feature set, where the gated mechanism fusion processing includes: Step S1212-1, calculate the feature complementarity score between the interactive semantic feature vector and the interactive fluctuation feature vector.

[0076] In this embodiment, by analyzing the information contained in the interactive semantic feature vector and the interactive fluctuation feature vector. For example, the interactive semantic feature vector contains detailed process operation procedures and technical parameter information, and the interactive fluctuation feature vector contains information about the dynamic impact of environmental factors on the production process. Calculate their complementarity in different dimensions. For a dimension, if the interactive semantic feature vector represents the operation intensity of the process in this dimension, and the interactive fluctuation feature vector represents the potential impact of environmental factors on the operation intensity in this dimension, then through a certain calculation method, measure the complementarity of the two in this dimension. Integrate the complementarity of all dimensions to obtain a feature complementarity score. For example, after calculation, the feature complementarity score is 0.7, indicating that they are complementary to a large extent.

[0077] Step S1212-2, generate a dynamic fusion weight vector based on the feature complementarity score.

[0078] In this embodiment, the weights of the interactive semantic feature vector and the interactive fluctuation feature vector in the fusion process are determined according to the feature complementarity score. Assume a simple calculation method is adopted. The weight of the interactive semantic feature vector is the score value divided by 2 plus 0.1, and the weight of the interactive fluctuation feature vector is 1 minus the weight of the interactive semantic feature vector. For a score of 0.7, the weight of the interactive semantic feature vector is 0.7÷2 + 0.1 = 0.45, and the weight of the interactive fluctuation feature vector is 1 - 0.45 = 0.55, thus generating a dynamic fusion weight vector.

[0079] Step S1212-3: Use the dynamic fusion weight vector to perform weighted splicing on the interactive semantic feature vector and the interactive fluctuation feature vector to generate the fusion production feature set.

[0080] In this embodiment, the interactive semantic feature vector and the interactive fluctuation feature vector are spliced according to their respective weights. For example, the interactive semantic feature vector is [0.2, 0.3, 0.4, 0.5], the interactive fluctuation feature vector is [0.6, 0.7, 0.8, 0.9], the weight of the interactive semantic feature vector is 0.45, and the weight of the interactive fluctuation feature vector is 0.55. When performing weighted splicing, the value of the first dimension is 0.2×0.45 + 0.6×0.55 = 0.42, the value of the second dimension is 0.3×0.45 + 0.7×0.55 = 0.53, the value of the third dimension is 0.4×0.45 + 0.8×0.55 = 0.62, and the value of the fourth dimension is 0.5×0.45 + 0.9×0.55 = 0.72. Finally, the fusion production feature set [0.42, 0.53, 0.62, 0.72] is obtained. This fusion production feature set combines the feature information of both the process and the environment.

[0081] In a possible implementation manner, step S130 includes: Step S131: Input the fusion production feature set into the sparse feature screening layer of the multi-layer perceptron network model for redundant feature filtering processing to obtain the screened key production feature set.

[0082] In this embodiment, the fusion production feature set contains rich information, but there may be some redundant features or features that contribute little to anomaly analysis. For example, in the fusion production feature set, there may be two features related to a certain temperature-related information in the production process. One is the real-time temperature monitored by the environment, and the other is the equivalent temperature based on the device feedback. After analysis, it is found that their correlation is extremely high and they almost provide the same information. The sparse feature screening layer will identify and remove one of the redundant features. Through the correlation analysis and importance evaluation of all features, the finally screened key production feature set is obtained. This key production feature set retains the features that are more valuable for anomaly analysis.

[0083] Step S132: Invoke the cross-feature generation layer of the multi-layer perceptron network model to perform high-order feature combination processing on the key production feature set to generate a cross-production feature matrix.

[0084] In this embodiment, for each feature in the set of key production features after screening, such as features including raw material protein content, pressure in the crushing process, environmental humidity, etc. The cross-feature generation layer will combine these different features to form high-order features. For example, the raw material protein content and the pressure in the crushing process are combined into a new feature, indicating the influence of the raw material protein content on the pressure in the crushing process; the environmental humidity and the efficiency of the separation process are combined into another feature, indicating the effect of the environmental humidity on the efficiency of the separation process. Through various combination methods, a cross-production feature matrix is generated, and each element in the matrix is a different high-order feature combination.

[0085] Step S133, perform an abnormal sensitivity weight calculation process on the cross-production feature matrix through the attention allocation layer of the multi-layer perceptron network model to generate the abnormal sensitivity distribution of each production feature dimension.

[0086] In this embodiment, the attention allocation layer will analyze the performance of each high-order feature combination in past abnormal production situations based on historical data and model training. For example, in past production records, when the combined feature of the raw material protein content and the pressure in the crushing process appears a specific value, product quality problems often occur, indicating that this feature combination is more sensitive to abnormal situations. Through the analysis of a large amount of historical data, an abnormal sensitivity weight is assigned to each high-order feature combination. For the combined feature of the raw material protein content and the pressure in the crushing process, a weight of 0.8 may be assigned, indicating a high sensitivity to abnormal situations; while for some feature combinations that do not change significantly in abnormal situations, a weight of 0.2 may be assigned. In this way, the abnormal sensitivity distribution of each production feature dimension is generated.

[0087] Step S134, perform a feature weighted aggregation process on the cross-production feature matrix based on the abnormal sensitivity distribution to generate the abnormal correlation score set, where each abnormal correlation score corresponds to the correlation strength between the raw material quality index and a specific production process.

[0088] For example, in the cross-production feature matrix, a row vector represents multiple high-order feature combinations related to raw material quality indicators and production processes, such as the combination feature of raw material protein content and crushing process pressure, the combination feature of environmental humidity and separation process efficiency, etc. The corresponding abnormal sensitivity weights are 0.8, 0.6, etc. Perform feature weighted aggregation processing on this row vector, and add the values of each feature combination multiplied by their corresponding weights. Suppose the value of the combination feature of raw material protein content and crushing process pressure is 0.5, and the value of the combination feature of environmental humidity and separation process efficiency is 0.4. The value after weighted aggregation is 0.5×0.8 + 0.4×0.6 = 0.64. Perform such calculations on each row of the cross-production feature matrix, and finally generate a set of abnormal correlation degree scores. Each abnormal correlation degree score in the set of abnormal correlation degree scores corresponds to the correlation strength between the raw material quality indicator and a specific production process. For example, the calculated 0.64 above may indicate a strong abnormal correlation strength between the raw material protein content and the crushing process, providing an important basis for subsequent abnormal root cause tracing and production optimization.

[0089] In a possible implementation manner, step S140 includes: Step S141, screening out abnormal correlation degree items that exceed a preset threshold according to the abnormal correlation degree distribution, and generating a candidate root cause feature set.

[0090] During the production process of a certain batch of wheat germ, the abnormal correlation degree distribution has been obtained. The preset threshold is set to 0.6, and the abnormal correlation degree distribution is screened according to this preset threshold. For example, in the set of abnormal correlation degree scores, there are the correlation degree scores of raw material wheat moisture content and drying process 0.7, the correlation degree score of raw material wheat impurity content and screening process 0.8, and the correlation degree score of environmental temperature and separation process 0.65, etc. These abnormal correlation degree items that exceed the preset threshold of 0.6 are selected to form a candidate root cause feature set. This candidate root cause feature set contains key factor combinations that may cause production abnormalities, such as the correlation between raw material wheat moisture content and drying process, the correlation between raw material wheat impurity content and screening process, and the correlation between environmental temperature and separation process, etc.

[0091] Step S142, performing reverse semantic parsing processing on each abnormal correlation degree item in the candidate root cause feature set to determine its corresponding raw material quality defect description and production process abnormality description.

[0092] For example, for the item where the correlation degree between the moisture content of raw wheat and the drying process is 0.7, through in-depth analysis of the production process text data and relevant knowledge, it is determined that the description of the raw material quality defect is that the moisture content of raw wheat is too high. Because the excessive moisture content exceeds the normal production requirement range, it may affect the drying effect. The description of the production process anomaly is that the drying time is too long and the moisture content of the product after drying still does not meet the standard. This is because the high moisture content of the raw material causes the drying equipment to take longer to remove the moisture, but the moisture content of the final product still does not meet the quality standard. For the item where the correlation degree between the impurity content of raw wheat and the screening process is 0.8, it is parsed that the description of the raw material quality defect is that the impurity content of raw wheat exceeds the standard, and the description of the production process anomaly is that the screening process fails to effectively remove impurities, which may be caused by improper selection of the sieve mesh or insufficient screening time resulting in impurity residue. For the item where the correlation degree between the environmental temperature and the separation process is 0.65, it is determined that the description of the raw material quality defect is empty (because this is mainly a problem caused by environmental factors), and the description of the production process anomaly is that the too high environmental temperature affects the separation effect, resulting in incomplete separation of wheat germ from other components.

[0093] Step S143, call the pre-trained root cause reasoning model to perform joint causal reasoning processing on the description of the raw material quality defect and the description of the production process anomaly, and generate a set of root cause reasoning paths.

[0094] In a possible implementation manner, step S143 includes: Step S1431, input the description of the raw material quality defect into the raw material defect encoder of the root cause reasoning model for defect type coding processing to generate a defect type feature vector.

[0095] Taking the excessive moisture content of raw wheat, too long drying time and the moisture content of the product after drying still not meeting the standard as an example, input the description of the raw material quality defect "the moisture content of raw wheat is too high" into the raw material defect encoder of the root cause reasoning model for defect type coding processing. The encoder can analyze this description and generate a defect type feature vector according to its influence and characteristics on the production process. For example, it can consider multiple dimensions such as the physical properties of the raw material and its influence on subsequent processes. Assuming that the dimension of the influence of raw material moisture on drying efficiency is assigned a value of 0.8, and the dimension of the potential influence on product quality is assigned a value of 0.7, etc., a defect type feature vector [0.8, 0.7,..., 0.5] is comprehensively generated.

[0096] Step S1432, input the description of the production process anomaly into the process anomaly encoder of the root cause reasoning model for anomaly mode coding processing to generate an anomaly mode feature vector.

[0097] For example, the production process anomaly description "the drying time is too long and the moisture content of the product after drying still does not meet the standard" is input into the process anomaly encoder of the root cause reasoning model for anomaly pattern encoding. The encoder will analyze factors such as the manifestation form and occurrence frequency of this anomaly. For example, a value of 0.9 is assigned from the dimension that the drying time exceeds the normal range, and a value of 0.8 is assigned from the dimension of the impact of unqualified product moisture on the overall quality, etc., to generate an anomaly pattern feature vector [0.9, 0.8, …, 0.6].

[0098] Step S1433, construct a causal graph network of defect - anomaly, and input the defect type feature vector and the anomaly pattern feature vector as node features into the causal graph network for multi - hop causal reasoning processing.

[0099] In this causal graph network, the nodes are connected by various causal relationships. Taking the node of "the moisture content of raw wheat is too high" and the node of "the drying time is too long and the moisture content of the product after drying still does not meet the standard" as an example, the network will analyze the possible direct and indirect causal relationships between them. For example, the too high moisture content of raw wheat directly leads to an increase in the heat required for drying, thus extending the drying time; although the drying time is extended, due to equipment performance limitations or process parameter setting problems, the moisture content of the final product still does not meet the standard. Through such multi - hop causal reasoning, the causal relationship chain is deeply explored.

[0100] Step S1434, output the causal association paths between the defect type feature vector and each anomaly pattern feature vector through the path generation layer of the causal graph network, and generate the root cause reasoning path set.

[0101] For example, for the group of "the moisture content of raw wheat is too high" and "the drying time is too long and the moisture content of the product after drying still does not meet the standard", the path generation layer will generate a causal association path, describing how the abnormal situation in the drying process is caused step by step from the too high raw material moisture. This path may be: the moisture content of raw wheat is too high → the heat required for drying increases → the continuous working time of the drying equipment extends → the moisture content of the product after drying still does not meet the standard. For the group of "the impurity content of raw wheat exceeds the standard" and "the screening process fails to effectively remove impurities", a corresponding causal association path will also be generated, such as the impurity content of raw wheat exceeds the standard → the screening difficulty increases → the probability of sieve blockage increases → the screening time is insufficient → the screening process fails to effectively remove impurities. For the group of "the environmental temperature is too high" and "the separation effect is not complete", the causal association path may be the environmental temperature is too high → the physical properties of substances change during the separation process → the working efficiency of the separation equipment decreases → the separation effect is not complete. These causal association paths together constitute the root cause reasoning path set.

[0102] Step S144, perform a confidence evaluation process on the root cause reasoning path set, and select the top k reasoning paths with the highest confidence as the abnormal root cause tracing result.

[0103] In this embodiment, various factors can be comprehensively considered in the evaluation process, such as the frequency of similar causal relationships in historical data, the similarity between current production conditions and past situations, etc. For each causal association path, there is a set of evaluation indicators and calculation methods. Taking the path that the excessive moisture content of raw wheat leads to drying problems as an example, by analyzing historical production records, it is found that when the moisture content of raw wheat exceeds a certain standard, the problem of extended drying time and unqualified product moisture occurs in 80% of the cases. At the same time, the similarity between the raw material procurement source and the drying equipment status of the current production batch and the historical situation is relatively high. Considering these factors comprehensively, through a specific calculation method (such as weighted calculation of factors such as historical occurrence frequency and current similarity), the confidence level of this path is obtained as 0.8. Confidence evaluations are performed on all paths in the root cause reasoning path set.

[0104] Suppose k is set to 3, and the top 3 reasoning paths with the highest confidence levels are selected from all the paths that have undergone confidence evaluation as the abnormal root cause tracing results. After comparison, the top 3 paths with the highest confidence levels are as follows: The first one is that the excessive impurity content of raw wheat causes the screening process to fail to effectively remove impurities, thereby affecting subsequent processes; the second one is that the excessive moisture content of raw wheat results in an overly long drying time and the product moisture content still not meeting the standard after drying; the third one is that the excessively high environmental temperature causes incomplete separation. These 3 paths clearly point out the possible roots of abnormal situations in the production process, providing a strong basis for subsequent targeted measures to solve production anomalies.

[0105] For example, in a possible implementation manner, step S1433 includes: Step S1433-1, taking the defect type feature vector and the abnormal pattern feature vector as the initial defect type node feature and the initial abnormal pattern node feature in the causal graph network respectively.

[0106] In this embodiment, taking the defect type feature vector [0.8, 0.7, …, 0.5] of the excessive moisture content of raw wheat and the abnormal pattern feature vector [0.9, 0.8, …, 0.6] of the overly long drying time and the product moisture content still not meeting the standard after drying obtained from the previous analysis as examples, these two vectors are taken as the initial defect type node feature and the initial abnormal pattern node feature in the causal graph network respectively. These two nodes become the starting points of the causal graph network reasoning, representing two key factors in the production process where problems occur.

[0107] Step S1433-2: Calculate the node association weight matrix according to the semantic similarity between the initial defect type node features and the initial abnormal pattern node features. The node association weight matrix includes the association strength values between each pair of defect type nodes and abnormal pattern nodes.

[0108] In this embodiment, the calculation of semantic similarity comprehensively considers factors from multiple dimensions. For example, in the scenario where the moisture content of raw wheat is too high and the drying is abnormal, from the dimension of the impact on drying efficiency, the corresponding dimension value in the defect type node feature vector is 0.8, and the corresponding dimension value in the abnormal pattern node feature vector is 0.9, indicating a high similarity between the two in terms of the impact on drying efficiency; from the dimension of the potential impact on product quality, the defect type node feature vector value is 0.7, and the abnormal pattern node feature vector value is 0.8, also showing a certain similarity. Calculate the absolute value of the difference between each pair of dimension values. The absolute value of the difference in the dimension of the impact on drying efficiency is |0.8 - 0.9| = 0.1, and the absolute value of the difference in the dimension of the potential impact on product quality is |0.7 - 0.8| = 0.1, etc. Add up the absolute values of the differences in all dimensions. Assuming there are a total of 5 dimensions, and the absolute values of the differences in other dimensions are 0.2, 0.1, 0.1 respectively, the sum is 0.1 + 0.1 + 0.2 + 0.1 + 0.1 = 0.6. Then subtract this sum from 1 to get the semantic similarity of 1 - 0.6 = 0.4. This semantic similarity is the association strength value between this pair of defect type nodes and abnormal pattern nodes. By analogy, calculate the association strength values between all possible defect type nodes and abnormal pattern nodes to form the node association weight matrix. This matrix comprehensively records the degree of association tightness between different nodes.

[0109] Step S1433-3: Generate the node relationship topological structure of the causal graph network based on the node association weight matrix. The node relationship topological structure includes the directed connection edges between defect type nodes and abnormal pattern nodes and their corresponding association strength values.

[0110] In the node relationship topological structure, taking the defect type node of the excessive moisture content of raw wheat and the abnormal pattern node of the too long drying time and the moisture content of the product still not meeting the standard after drying as an example, according to the association strength value of 0.4 between them, a directed connection edge pointing from the node of the excessive moisture content of raw wheat to the drying abnormal node will be generated, and the association strength value of 0.4 is marked on the edge. The relationships between other defect type nodes and abnormal pattern nodes are also constructed in a similar manner in the topological structure. Finally, a node relationship topological structure including all the directed connection edges between defect type nodes and abnormal pattern nodes and their corresponding association strength values is formed. This node relationship topological structure intuitively shows the potential connections between different problem factors in the production process.

[0111] Step S1433-4: Perform edge weight enhancement processing on each directed connection edge in the node relationship topology structure. The edge weight enhancement processing includes: extracting the defect type node features and abnormal pattern node features corresponding to the directed connection edge. Input the defect type node features and abnormal pattern node features into the edge weight enhancement layer of the causal graph network for feature interaction processing to generate an enhanced edge weight value.

[0112] Taking the directed connection edge from the node of excessive moisture content of raw wheat to the node of excessive drying time and still unqualified moisture content of the product after drying as an example, first extract the defect type node features [0.8, 0.7, …, 0.5] and abnormal pattern node features [0.9, 0.8, …, 0.6] corresponding to this directed connection edge. Input these two feature vectors into the edge weight enhancement layer of the causal graph network for feature interaction processing. In the edge weight enhancement layer, perform weighted multiplication and summation on the corresponding dimensions of the two vectors. For example, for the first dimension, the weighted multiplication is 0.8×0.9×0.3 (assuming the weighting coefficient of the first dimension is 0.3) = 0.216; for the second dimension, the weighted multiplication is 0.7×0.8×0.2 (assuming the weighting coefficient of the second dimension is 0.2) = 0.112, etc. Add up the calculation results of all dimensions. Assuming there are a total of 5 dimensions, and the calculation results of other dimensions are 0.15, 0.08, and 0.06 respectively, the sum is 0.216 + 0.112 + 0.15 + 0.08 + 0.06 = 0.618, and this value is the enhanced edge weight value.

[0113] Step S1433-5: Update the weights of the directed connection edges in the node relationship topology structure based on the enhanced edge weight values to generate an updated node relationship topology structure.

[0114] For example, for the directed connection edge from the node of excessive moisture content of raw wheat to the node of excessive drying time and still unqualified moisture content of the product after drying, update the original association strength value of 0.4 to the enhanced edge weight value of 0.618 to generate an updated node relationship topology structure, and this updated topology structure more accurately reflects the causal association degree between nodes.

[0115] Step S1433-6: Perform multi-hop path traversal processing in the updated node relationship topology structure. The multi-hop path traversal processing includes: starting from each defect type node, expanding the path along the directed connection edge with the maximum edge weight first, recording the passed node sequence and the cumulative weight value, and generating a set of candidate multi-hop causal paths.

[0116] For example, starting from the defect type node of excessive moisture content in raw wheat, the path expansion is performed along the directed connection edges with the maximum edge weight prioritized. Since the edge weight between the node of excessive moisture content in raw wheat and the node of excessive drying time and still unqualified moisture content of the product after drying is 0.618, which is the largest among the currently connected edges to this node, the drying anomaly node is first expanded, and the recorded node sequence passed through is the node of excessive moisture content in raw wheat and the node of excessive drying time and still unqualified moisture content of the product after drying, with the cumulative weight value being 0.618. Then, continue to expand from the drying anomaly node. Assume that there is a directed connection edge with an edge weight of 0.5 between the drying anomaly node and the node of unqualified product quality. This is the largest among the connected edges to the drying anomaly node, so continue to expand to the node of unqualified product quality. At this time, the passed node sequence is updated to the node of excessive moisture content in raw wheat, the node of excessive drying time and still unqualified moisture content of the product after drying, and the node of unqualified product quality, with the cumulative weight value being 0.618 + 0.5 = 1.118. In this way, starting from each defect type node for expansion, record all the passed node sequences and cumulative weight values to generate a candidate multi-hop causal path set. This candidate multi-hop causal path set contains various paths that may lead to production anomalies through multi-hop connections starting from different defect type nodes.

[0117] Step S1433-7, perform path validity verification processing on the candidate multi-hop causal path set. The path validity verification processing includes: extracting the node sequence features and cumulative weight values of each candidate multi-hop causal path. Input the node sequence features into the path verification layer of the causal graph network for logical consistency detection to generate a path confidence score.

[0118] Taking the candidate path of the node of excessive moisture content in raw wheat, the node of excessive drying time and still unqualified moisture content of the product after drying, and the node of unqualified product quality as an example, first extract the node sequence features of this candidate multi-hop causal path, that is, the characteristic information represented by the three nodes of excessive moisture content in raw wheat, drying anomaly, and unqualified product quality, as well as the cumulative weight value of 1.118. Input the node sequence features into the path verification layer of the causal graph network for logical consistency detection. The path verification layer will make a judgment based on the knowledge and logical rules of the production process. For example, in this production process, excessive moisture content in raw wheat leads to drying anomaly, and drying anomaly in turn affects product quality. This logic conforms to the actual production situation. Through a series of rule matching and reasoning calculations, a path confidence score is generated. Assume that after complex verification calculations, the path confidence score of this path is obtained as 0.8.

[0119] Step S1433-8: Weight and sort the candidate multi-hop causal path set according to the path confidence score and the cumulative weight value, and filter out the target multi-hop causal path set that meets the preset confidence threshold.

[0120] For each candidate path, set a weighted calculation method. For example, the comprehensive score of the path = path confidence score × 0.6 + cumulative weight value × 0.4. For the above path, the comprehensive score = 0.8 × 0.6 + 1.118 × 0.4 = 0.48 + 0.4472 = 0.9272. Perform such calculations on all paths in the candidate multi-hop causal path set, and sort them in descending order according to the comprehensive score. Assume that the preset confidence threshold is 0.7, and filter out the paths whose comprehensive scores meet this threshold from the sorted paths to form the target multi-hop causal path set.

[0121] Step S1433-9: Convert each target multi-hop causal path in the target multi-hop causal path set into a causal association node sequence to generate the root cause inference path set.

[0122] For example, there is a path in the target multi-hop causal path set that starts from the node of excessive impurity content in raw material wheat, passes through the node of ineffective impurity removal in the screening process, and then reaches the node of unqualified product purity. Convert this path into a causal association node sequence, that is, excessive impurity content in raw material wheat → ineffective impurity removal in the screening process → unqualified product purity. By analogy, perform such conversions on all target multi-hop causal paths, and finally generate the root cause inference path set. This root cause inference path set clearly shows the causal relationship chain from raw material quality defects to production process abnormalities and then to final product problems during the production process, providing a detailed and reliable basis for finding the root cause of production anomalies.

[0123] In a possible implementation manner, step S144 includes: Step S1441: Extract the causal association node sequence in each root cause inference path.

[0124] For example, in the previously generated root cause inference path set, there is a path that starts from the node of excessive moisture content in raw material wheat, passes through the nodes of too long drying time and still unqualified moisture content of the product after drying in sequence, and finally reaches the node of unqualified product quality. The causal association node sequence in this root cause inference path is excessive moisture content in raw material wheat, too long drying time and still unqualified moisture content of the product after drying, unqualified product quality. Another path is excessive impurity content in raw material wheat, ineffective impurity removal in the screening process, and subsequent processing processes are affected, and its causal association node sequence is the description of these three sequentially connected nodes. Perform such extraction operations on each path in the root cause inference path set to clearly identify the causal association nodes involved in each path.

[0125] Step S1442: Call the pre-trained path evaluation model to perform semantic coherence scoring and logical rationality scoring on each of the causal association node sequences.

[0126] Taking the causal association node sequence of "excessive moisture content of raw wheat, too long drying time, still unqualified moisture content of the product after drying, and unqualified product quality" as an example, when the path evaluation model performs semantic coherence scoring, it will analyze whether the semantic connection between nodes is natural and smooth. Semantically, if the moisture content of raw wheat is too high, according to normal production knowledge and logic, it is easy to understand that the drying time needs to be extended, and it is possible that the moisture content of the product after drying is still unqualified. And poor drying effect will inevitably affect the product quality. The whole semantic chain is closely connected. The model will score according to the tightness of this semantics. Assuming a full score of 10 points, after detailed analysis and judgment by the model, it is considered that the semantic coherence of this sequence is very strong and a score of 8 points is given.

[0127] When performing logical rationality scoring, the path evaluation model will combine the actual logic and historical data in the production process to judge. In the production of wheat germ, an increase in the moisture content of raw materials will indeed increase the drying difficulty, and more time and energy are required to remove the moisture, which is in line with physical principles and actual production experience; incomplete drying will directly affect the quality indicators of the product. For example, too high moisture content may cause the product to deteriorate easily and affect the taste, etc. By analyzing these logical relationships and comparing with similar situations in historical data, the model judges that this logic is reasonable. Assuming that the logical rationality score is also full score of 10 points, according to the evaluation criteria of the model, a score of 9 points is given to this sequence.

[0128] For the causal association node sequence of "excessive impurity content of raw wheat, the screening process fails to effectively remove impurities, and the subsequent processing process is affected", in terms of semantic coherence, excessive raw material impurities will pose greater challenges to the screening process, resulting in the failure to effectively remove impurities, and incomplete screening will naturally have a negative impact on the subsequent processing process. The semantic coherence degree is relatively high, and the model gives 7 points. In terms of logical rationality, from the actual production, excessive impurities will block the sieve mesh, reduce the screening efficiency, etc., and then affect the subsequent processing, which is consistent with the situation in historical production. The model gives 8 points. Such semantic coherence scoring and logical rationality scoring are performed on all causal association node sequences in the root cause reasoning path set.

[0129] Step S1443: Based on the preset evaluation weights, perform weighted summation on the semantic coherence score and the logical rationality score to obtain the confidence score of each root cause reasoning path.

[0130] Suppose the preset semantic coherence scoring weight is 0.4 and the logical rationality scoring weight is 0.6. For the path where the moisture content of the raw wheat is too high, the drying time is too long, the moisture content of the product after drying still does not meet the standard, and the product quality does not meet the standard, the confidence scoring is calculated as follows: The semantic coherence score of 8 points is multiplied by the weight of 0.4, that is, 8×0.4 = 3.2; the logical rationality score of 9 points is multiplied by the weight of 0.6, that is, 9×0.6 = 5.4. Adding the two together, 3.2 + 5.4 = 8.6 points, which is the confidence score of this root cause reasoning path. For the path where the impurity content of the raw wheat exceeds the standard, the screening process fails to effectively remove impurities, and the subsequent processing process is affected, the semantic coherence score of 7 points is multiplied by the weight of 0.4, that is, 7×0.4 = 2.8; the logical rationality score of 8 points is multiplied by the weight of 0.6, that is, 8×0.6 = 4.8. Adding the two together, 2.8 + 4.8 = 7.6 points, obtaining the confidence score of this path. Each path in the set of root cause reasoning paths is calculated in this way to obtain its respective confidence score.

[0131] Step S1444, sort the set of root cause reasoning paths according to the confidence score, and filter out the target reasoning paths whose confidence meets the preset conditions.

[0132] Suppose the preset condition is that the confidence score is greater than or equal to 7 points. Sort all the root cause reasoning paths from high to low according to the confidence score, and in the sorted path set, filter out the paths with a score greater than or equal to 7 points. For example, in addition to the above two paths, there are other paths whose calculated confidence scores are 6.5 points, 7.2 points, 8.1 points, etc. Then, the paths with confidence scores of 7.2 points, 8.1 points, and the previously calculated 8.6 points, 7.6 points meet the preset conditions and are filtered out as the target reasoning paths. These target reasoning paths have higher reliability and credibility in tracing the root causes of production anomalies, and can provide a strong basis for the factory to take targeted measures to solve problems in the production process, helping the factory better optimize the production process and improve product quality.

[0133] In a possible implementation manner, step S150 includes: Step S151, analyze the production link of the abnormal source and the type of raw material defect in the abnormal root cause tracing result.

[0134] For example, the previous abnormal root cause tracing results showed that the abnormalities in a certain batch of production mainly originated from the excessive impurity content of the raw material wheat and the abnormalities in the drying process. The raw material defect type was clearly defined as the impurity content of the raw material wheat being higher than the standard range, which might be due to the lax control of raw materials in the procurement link, resulting in the mixing of more impurities. The source of the abnormality in the production link was the drying process. The drying time was too long and the moisture content of the product after drying still did not meet the standard. This might be caused by factors such as unstable performance of the drying equipment and unreasonable setting of drying parameters.

[0135] Step S152, match the historical parameter adjustment record set associated with the abnormal source production link from the historical optimization strategy library.

[0136] In this embodiment, the historical optimization strategy library stores the parameter adjustment strategies and related records tried and applied by the factory for various production link problems in past production. For the drying process, there are various adjustment methods in the historical records. For example, once to solve the problem of too long drying time, the temperature of the drying equipment was increased from 80 degrees Celsius to 90 degrees Celsius, and at the same time, the ventilation rate was adjusted, increasing from 50 cubic meters per minute to 60 cubic meters per minute. Finally, the drying time was shortened from the original 60 minutes to 50 minutes. Another time, for the situation where the moisture content of the product after drying was too high, the drying time was extended by 10 minutes, and the humidity sensor parameters of the drying equipment were adjusted to make its detection of humidity more accurate, so as to better control the drying degree. In addition, in terms of raw material processing, when the impurity content of the raw material was too high, different aperture sieves were tried, changing from the original 4-mm aperture sieve to a 3-mm aperture sieve to improve the screening effect, and at the same time, the screening times were increased, from one screening to two screenings, effectively reducing the impurity content in the raw material. These historical records constitute the historical parameter adjustment record set related to the drying process and raw material impurity treatment.

[0137] Step S153, perform an effectiveness filtering process on the historical parameter adjustment record set based on the raw material defect type to obtain an effective adjustment strategy set.

[0138] For example, for the defect type of the excessive impurity content in the raw wheat this time, analyze each strategy in the historical parameter adjustment record set. Exclude those records that have nothing to do with the removal of raw material impurities, such as parameter adjustment records solely for equipment maintenance. Among the strategies related to raw material impurity screening, evaluate their feasibility and effectiveness under the current production conditions. For example, the strategy of replacing the sieve mesh with a 3-mm aperture and increasing the screening times is considered to have high feasibility and effectiveness considering the current impurity types and distribution of the raw materials, as well as the performance of the existing screening equipment in the factory; while some overly complex or costly impurity removal strategies, such as using high-precision laser screening equipment, although theoretically can effectively remove impurities, are not yet implementable considering the factory's budget and actual production scale and are filtered out. After such screening, an effective adjustment strategy set is obtained, which includes strategies such as replacing the sieve mesh with a 3-mm aperture, increasing the screening times, optimizing the temperature and ventilation rate of the drying equipment, etc.

[0139] Step S154, perform multi-objective optimization processing on the effective adjustment strategy set to generate a dynamic optimization strategy that meets the current production constraint conditions, where the multi-objective optimization processing includes production efficiency optimization, raw material loss rate reduction, and abnormal recurrence rate control.

[0140] Among them, step S154 includes: Step S1541, construct a three-dimensional optimization objective space including production efficiency, raw material loss rate, and abnormal recurrence rate.

[0141] For example, in terms of production efficiency, it is measured by the number of qualified wheat germ products produced per hour; the raw material loss rate refers to the proportion of raw material losses caused by various reasons during the production process to the total input raw materials; the abnormal recurrence rate refers to the probability that the same abnormal situation will occur again after optimization measures are taken.

[0142] Step S1542, map each effective adjustment strategy to the corresponding coordinate point in the three-dimensional optimization objective space.

[0143] For example, for the strategy of replacing the 3-mm aperture sieve and increasing the screening frequency, after evaluation and simulation calculations, it is expected that the production efficiency will increase from 100 kg per hour to 110 kg, the raw material loss rate is expected to decrease from the current 8% to 6%, and according to historical experience and data analysis, the abnormal recurrence rate is expected to decrease from 20% to 10%. Taking these data as coordinate values, the coordinate point (110, 6, 10) corresponding to this strategy is determined in the three-dimensional optimization objective space. For the strategy of optimizing the temperature and ventilation rate of the drying equipment, the same evaluation and calculation are carried out. Assuming that the production efficiency is expected to increase to 105 kg per hour, the raw material loss rate is reduced to 7%, and the abnormal recurrence rate is reduced to 15%, then the coordinate point of this strategy in the three-dimensional optimization objective space is (105, 7, 15). All the strategies in the effective adjustment strategy set are mapped to the three-dimensional optimization objective space in this way.

[0144] Step S1543, screen out the candidate optimization strategy set located on the Pareto optimal front surface through the Pareto front analysis algorithm.

[0145] In this embodiment, the Pareto front analysis algorithm will analyze and compare each coordinate point in the three-dimensional optimization objective space. For example, for the coordinate points (110, 6, 10) and (105, 7, 15), in terms of production efficiency, the strategy corresponding to (110, 6, 10) is higher; in terms of the raw material loss rate, the strategy corresponding to (110, 6, 10) is also lower; and in terms of the abnormal recurrence rate, the strategy corresponding to (110, 6, 10) is also better. Then the point (105, 7, 15) is considered a relatively inferior point. Through such comparison and screening, find out those points that cannot further improve a certain objective without reducing other objectives, and the strategies corresponding to these points constitute the Pareto optimal front surface. For example, after algorithm analysis, it is finally determined that the strategies corresponding to (110, 6, 10) and several other similar strategy points with better comprehensive performance in each objective form the candidate optimization strategy set.

[0146] Step S1544, call the strategy recommendation model to perform an adaptability scoring process on the candidate optimization strategy set based on the real-time load status of the current production line.

[0147] In this embodiment, the real-time load status of the current production line includes factors such as the operating conditions of equipment, the supply of raw materials, and the urgency of orders. Suppose the equipment on the current production line is running relatively stably, but the raw material supply is slightly tight and the order urgency is relatively high. The policy recommendation model will evaluate each policy in the candidate optimization policy set according to these real-time situations. For the policy of replacing the sieve mesh with a 3-mm aperture and increasing the screening frequency, although it can effectively reduce the raw material loss rate and the recurrence rate of abnormalities and improve production efficiency, it may cause a short-term shortage of raw material supply due to the increased screening frequency, affecting the order delivery speed. According to the evaluation rules of the model, the adaptability score of this policy is 70 points. For another policy of optimizing the temperature and ventilation rate of the drying equipment, although it is relatively weak in terms of improving production efficiency, it has less impact on the raw material supply and can better control the recurrence rate of abnormalities. Under the current order urgency, it can better ensure the continuity of production. The model gives the adaptability score of this policy as 80 points. Such adaptability scoring is performed on all policies in the candidate optimization policy set.

[0148] Step S1545, select the candidate optimization policy with the highest adaptability score as the dynamic optimization policy.

[0149] In the above example, the adaptability score of the policy of optimizing the temperature and ventilation rate of the drying equipment is 80 points, which is higher than the scores of other policies. Therefore, this policy is determined as the dynamic optimization policy. The factory will adjust the temperature and ventilation rate parameters of the drying equipment according to this dynamic optimization policy, and at the same time closely monitor the changes of various indicators in the production process to ensure that the production can optimize production efficiency, reduce the raw material loss rate, and effectively control the recurrence rate of abnormalities while meeting the current production constraint conditions, and ensure the stable and efficient operation of wheat germ production.

[0150] Figure 2 The hardware structure diagram of the production service system 100 for implementing the above NLP-based wheat germ production anomaly root cause tracing method provided by the embodiment of the present application is shown, as Figure 2 shown, the production service system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0151] In one possible design, the production service system 100 can be a single server or a group of servers. The group of servers can be centralized or distributed (for example, the production service system 100 can be a distributed system). In some embodiments, the production service system 100 can be local or remote. For example, the production service system 100 can access information and / or data stored in a machine-readable storage medium 120 via a network. As another example, the production service system 100 can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the production service system 100 can be implemented on the production service system. By way of example only, the production service system can include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any aggregation thereof.

[0152] The machine-readable storage medium 120 can store data and / or instructions. In some embodiments, the machine-readable storage medium 120 can store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 can store the data and / or instructions that the production service system 100 uses to execute or use to complete the exemplary methods described in this application.

[0153] In a specific implementation process, one or more processors 110 execute the computer-executable instructions stored in the machine-readable storage medium 120, so that the processors 110 can execute the NLP-based method for tracing the root cause of abnormal wheat germ production in the above method embodiments. The processors 110, the machine-readable storage medium 120, and the communication unit 140 are connected through a bus 130, and the processors 110 can be used to control the transceiver actions of the communication unit 140.

[0154] For the specific implementation process of the processors 110, reference can be made to the respective method embodiments executed by the above production service system 100. Their implementation principles and technical effects are similar, and will not be elaborated here in this embodiment.

[0155] In addition, an embodiment of the present application also provides a readable storage medium, in which computer-executable instructions are set. When a processor runs the computer-executable instructions, the above NLP-based method for tracing the root cause of abnormal wheat germ production is implemented.

[0156] It should be noted that, in order to simplify the presentation of the disclosure of the present application and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, sometimes multiple features are merged into one embodiment, drawing, or description thereof. Similarly, it should be noted that, in order to simplify the presentation of the disclosure of the present application and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. A method for tracing the root cause of abnormal wheat germ production based on NLP, characterized in that: The method comprises: Acquire a set of production record data of multiple batches of wheat germ corresponding to the target production line, wherein the production record data set includes production process text data, environmental monitoring time series data and raw material quality index data of each production batch; Performing semantic feature parsing processing on the production process text data to obtain a process semantic feature vector, performing dynamic fluctuation feature extraction processing on the environmental monitoring time series data to obtain an environmental fluctuation feature vector, performing cross-modal feature fusion processing on the process semantic feature vector and the environmental fluctuation feature vector to generate a fused production feature set; Calling a pre-trained multi-layer perception network model to perform abnormal root cause weight distribution processing on the fused production feature set, and generating an abnormal correlation score set corresponding to the production batch, wherein the abnormal correlation score set includes the abnormal correlation distribution between the raw material quality index data and each production process; Based on the abnormal correlation distribution, the raw material quality index data and the production process text data are jointly traced for root cause, and an abnormal root cause tracing result of the production batch is generated, where the abnormal root cause tracing result is used to indicate the production link and raw material defect type of the abnormal source; A dynamic optimization strategy for production line parameters is generated according to the abnormal root cause tracing result, and the dynamic optimization strategy is fed back to the production control system to trigger a parameter calibration operation.

2. The method for tracing the root cause of abnormal wheat germ production based on NLP according to claim 1, characterized in that: The performing semantic feature parsing processing on the production process text data to obtain a process semantic feature vector includes: Performing process step segmentation processing on the production process text data to obtain a plurality of process step description text segments; Calling a pre-trained language representation model to perform contextual semantic encoding processing on each process step description text segment to generate an initial semantic feature vector; The initial semantic feature vector is subjected to process domain feature enhancement processing to obtain an enhanced semantic feature vector, wherein the process domain feature enhancement processing comprises the following steps: matching a domain entity set associated with a current process step description text fragment from a preset wheat germ production knowledge base; inputting the domain entity set into the language representation model for entity semantic encoding processing to generate an entity feature vector set; performing attention mechanism weighted fusion on the entity feature vector set and the initial semantic feature vector to generate the enhanced semantic feature vector; The enhanced semantic feature vectors of the text segments describing each process step are subjected to temporal position coding and splicing processing to generate the process semantic feature vectors.

3. The method for tracing the root cause of abnormal wheat germ production based on NLP according to claim 1, characterized in that: The step of performing dynamic fluctuation feature extraction processing on the environmental monitoring time series data to obtain an environmental fluctuation feature vector includes: Performing abnormal fluctuation interval detection processing on the environmental monitoring time series data to identify abnormal fluctuation time windows in the temperature, humidity and air pressure monitoring data; Performing multi-scale sliding window sampling processing on the original monitoring data within the abnormal fluctuation time window to obtain multiple local time series data fragments; Calling a pre-trained time series convolutional network model to perform local fluctuation feature extraction processing on each of the local time series data segments to generate a local fluctuation feature vector; Performing global temporal attention aggregation processing on the local fluctuation feature vectors of each of the local temporal data segments to generate the environmental fluctuation feature vector, wherein the global temporal attention aggregation processing includes: Calculate the similarity score between each local fluctuation feature vector and the preset global fluctuation pattern template; Performing dynamic weighted summation on each local fluctuation feature vector based on the similarity score to obtain a weighted fluctuation feature vector; The weighted fluctuation feature vector and the global fluctuation pattern template are subjected to feature difference processing to generate the environmental fluctuation feature vector.

4. The method for tracing the root cause of abnormal wheat germ production based on NLP according to claim 1, characterized in that: The step of performing cross-modal feature fusion processing on the process semantic feature vector and the environment fluctuation feature vector to generate a fused production feature set includes: Performing time dimension alignment processing on the process semantic feature vector, mapping it to the same time granularity as the environment fluctuation feature vector; Constructing a process-environment cross attention mechanism, calculating a cross-modal correlation matrix between the semantic features of each time step in the process semantic feature vector and the fluctuation features of the corresponding time step in the environment fluctuation feature vector; Based on the cross-modal correlation matrix, bidirectional feature interaction processing is performed on the process semantic feature vector and the environment fluctuation feature vector to generate an interactive semantic feature vector and an interactive fluctuation feature vector; The interactive semantic feature vector and the interactive fluctuation feature vector are subjected to a gating mechanism fusion process to generate the fused production feature set, wherein the gating mechanism fusion process includes: Calculating a feature complementarity score between the interactive semantic feature vector and the interactive fluctuation feature vector; generating a dynamic fusion weight vector based on the feature complementarity score; The interactive semantic feature vector and the interactive fluctuation feature vector are weightedly concatenated using the dynamic fusion weight vector to generate the fused production feature set.

5. The method for tracing the root cause of abnormal wheat germ production based on NLP according to claim 1, characterized in that: The calling of the pre-trained multi-layer perception network model performs abnormal root cause weight distribution processing on the fused production feature set to generate an abnormal correlation score set corresponding to the production batch, including: Inputting the fused production feature set into the sparse feature screening layer of the multi-layer perception network model to perform redundant feature filtering processing to obtain a screened key production feature set; Calling the cross-feature generation layer of the multi-layer perception network model to perform high-order feature combination processing on the key production feature set to generate a cross-production feature matrix; Performing abnormal sensitivity weight calculation processing on the cross-production feature matrix through the attention allocation layer of the multi-layer perception network model to generate abnormal sensitivity distribution of each production feature dimension; The cross-production feature matrix is ​​subjected to feature weighted aggregation processing based on the abnormal sensitivity distribution to generate the abnormal correlation score set, wherein each abnormal correlation score corresponds to the strength of correlation between the raw material quality index and a specific production process.

6. The method for tracing the root cause of abnormal wheat germ production based on NLP according to claim 1, characterized in that: The performing joint root cause tracing processing on the raw material quality index data and the production process text data based on the abnormal correlation distribution to generate the abnormal root cause tracing result of the production batch includes: According to the abnormal correlation distribution, abnormal correlation items exceeding a preset threshold are screened out to generate a candidate root cause feature set; Performing reverse semantic parsing processing on each abnormal correlation item in the candidate root cause feature set to determine its corresponding raw material quality defect description and production process abnormality description; Calling a pre-trained root cause reasoning model to perform joint causal reasoning processing on the raw material quality defect description and the production process abnormality description to generate a root cause reasoning path set; A confidence evaluation process is performed on the root cause reasoning path set, and top k reasoning paths with the highest confidence are screened out as the abnormal root cause tracing results.

7. The method for tracing the root cause of abnormal wheat germ production based on NLP according to claim 6, characterized in that: The calling of the pre-trained root cause reasoning model performs joint causal reasoning processing on the raw material quality defect description and the production process abnormality description to generate a root cause reasoning path set, including: Inputting the raw material quality defect description into the raw material defect encoder of the root cause reasoning model to perform defect type encoding processing to generate a defect type feature vector; Inputting the production process abnormality description into the process abnormality encoder of the root cause reasoning model to perform abnormal pattern encoding processing to generate an abnormal pattern feature vector; Constructing a defect-anomaly causal graph network, and inputting the defect type feature vector and the anomaly pattern feature vector as node features into the causal graph network for multi-hop causal reasoning processing; The causal association path between the defect type feature vector and each abnormal pattern feature vector is outputted through the path generation layer of the causal graph network to generate the root cause reasoning path set.

8. The method for tracing the root cause of abnormal wheat germ production based on NLP according to claim 6, characterized in that: The confidence evaluation process of the root cause reasoning path set includes: Extract the causal association node sequence in each root cause reasoning path; Calling a pre-trained path evaluation model to perform semantic coherence scoring and logical rationality scoring on each of the causal association node sequences; Performing weighted summation of the semantic coherence score and the logical rationality score based on preset evaluation weights to obtain a confidence score for each root cause reasoning path; The root cause reasoning path set is sorted according to the confidence score, and the target reasoning path whose confidence meets the preset condition is screened out.

9. The method for tracing the root cause of abnormal wheat germ production based on NLP according to claim 1, characterized in that: The generating of a dynamic optimization strategy for production line parameters according to the abnormal root cause tracing result includes: Analyze the abnormal source production links and raw material defect types in the abnormal root cause tracing results; Matching a set of historical parameter adjustment records associated with the abnormal source production link from a historical optimization strategy library; Performing validity filtering on the historical parameter adjustment record set based on the raw material defect type to obtain a valid adjustment strategy set; Performing multi-objective optimization processing on the effective adjustment strategy set to generate a dynamic optimization strategy that meets current production constraints, wherein the multi-objective optimization processing includes optimizing production efficiency, reducing raw material loss rate, and controlling abnormal recurrence rate; The multi-objective optimization process is performed on the effective adjustment strategy set to generate a dynamic optimization strategy that meets the current production constraints, including: Construct a three-dimensional optimization target space including production efficiency, raw material loss rate and abnormal recurrence rate; Mapping each effective adjustment strategy to a corresponding coordinate point in the three-dimensional optimization target space; The set of candidate optimization strategies located on the Pareto optimal frontier is screened out through the Pareto frontier analysis algorithm; The strategy recommendation model is called to perform adaptive scoring processing on the candidate optimization strategy set based on the real-time load status of the current production line; The candidate optimization strategy with the highest adaptability score is selected as the dynamic optimization strategy.

10. A production service system, characterized in that: The production service system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to run the programs, instructions or codes in the memory to implement the NLP-based root cause tracing method for abnormal wheat germ production as described in any one of claims 1 to 9.

Images