Automobile manufacturing parameter anomaly detection rule base construction method and device and automobile manufacturing parameter anomaly detection method and device

Abstract

The invention provides an automobile manufacturing parameter anomaly detection rule base construction and anomaly detection method and device, and relates to the technical field of automobile manufacturing. Multi-source manufacturing process knowledge reference data is matched with a fixed format rule based on a predefined regular expression mode; the method comprises the following steps: carrying out entity recognition based on a natural language processing model to obtain an unstructured text rule, extracting a multi-source heterogeneous data rule based on search enhancement, carrying out classification and similarity comparison and combination, sending into a preset large language model for reasoning to extract a standard detection rule, and constructing an automobile manufacturing parameter anomaly detection rule library. Parameter detection automation is realized based on the automobile manufacturing parameter anomaly detection rule base, and the detection accuracy is improved. According to the method, correctness of parameters and logic can be identified efficiently, and rapid and accurate anomaly detection can be carried out on large-scale multi-mode automobile manufacturing parameters.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automobile manufacturing, and in particular to an automobile manufacturing parameter abnormality detection rule library construction method, an abnormality detection method and device. BACKGROUND

[0002] The automobile manufacturing production process highly depends on the precision of the equipment parameters and control logic of each link of the production line. With the development of the automobile industry towards intelligence and flexibility, modern production lines integrate a large number of heterogeneous devices (such as welding robots, conveying devices, visual inspection instruments, etc.) and multi-source sensing systems, involving tens of thousands of process parameters, and the control logic complexity grows exponentially. Traditional parameter abnormality detection techniques mainly rely on the following methods: (1) manual sampling mode: technical experts check key equipment parameters based on experience, and identify abnormalities through manual recording and comparison. For example, the B company production line uses manual sampling, and the detection cycle of a single vehicle type is as long as 72 hours, and the parameter coverage rate is less than 15%; (2) script-based detection tool: automatic detection of specific equipment by writing fixed scripts, but due to differences in communication protocols of heterogeneous devices, script compatibility is poor, and cross-system collaborative detection cannot be achieved; (3) statistical process control (SPC) method: setting threshold range based on historical data, but lacking adaptability to dynamically changing control logic and device coupling effects.

[0003] However, the prior art has significant defects: first, manual detection is inefficient and subjective, and it is difficult to cover the correlation abnormalities between a large number of parameters; second, script tools are only suitable for single device scenarios and cannot solve the data fusion problem of multi-source heterogeneous systems; third, traditional statistical methods cannot real-time analyze the chain abnormality of complex control logic, resulting in a high potential problem detection rate of up to 32% (according to industry research data). In addition, existing solutions generally lack global modeling capability for process parameters, device states and control logic, making it impossible to build a complete abnormality traceability system during the production line debugging stage, directly affecting the stability of new vehicle production. Therefore, it is urgent to develop an intelligent abnormality detection scheme integrating multi-modal data analysis capability. SUMMARY

[0004] In view of this, the embodiments of the present application provide an automobile manufacturing parameter abnormality detection rule library construction method, an abnormality detection method and device to eliminate or improve one or more defects in the prior art, and solve the problem that the prior art cannot efficiently and accurately detect abnormalities of large-scale multi-modal automobile manufacturing parameters.

[0005] The present application provides an automobile manufacturing parameter abnormality detection rule library construction method, which comprises the following steps: acquire a plurality of manufacturing process knowledge benchmark data for generating detection rules based on a plurality of preset first type data sources; the first type data sources include production logs, sensor data, process parameters, operation manuals, technical documents, and historical abnormality reports; match fixed format rules based on a predefined regular expression pattern, acquire unstructured text rules based on an entity recognition of a natural language processing model, and extract multi-source heterogeneous data rules based on search enhancement; After classifying and similarity-based rule merging of the fixed format rules, the unstructured text rules, and the multi-source heterogeneous data rules, calculate a comprehensive confidence for each rule obtained after merging, and deliver rules with a comprehensive confidence higher than a set value to a preset large language model for reasoning and extraction to generate standard detection rules for automobile manufacturing parameters, forming an automobile manufacturing parameter abnormality detection rule library.

[0006] In some embodiments, the manufacturing process knowledge benchmark data includes: design data for recording automobile design simulation and process requirements; manufacturing standards for recording production standards of each vehicle project; project data for recording progress, experience summary, and technical acceptance of each vehicle project; and downtime data for recording downtime reports, downtime technical analysis, and equipment failure mode analysis during production and mass production; After acquiring a plurality of manufacturing process knowledge benchmark data for generating detection rules based on a plurality of preset first type data sources, the method further includes: parsing and data cleaning of documents of each type of manufacturing process knowledge benchmark data, uniformly extracting as pure text format, and removing invalid information including headers, footers, tables of contents, and references.

[0007] In some embodiments, matching fixed format rules based on a predefined regular expression pattern includes: extracting parameter names and numerical value ranges, and extracting logical conditions between parameters to construct a plurality of regular expressions, and extracting the fixed format rules based on the regular expressions traversing the manufacturing process knowledge benchmark data; Acquiring unstructured text rules based on an entity recognition of a natural language processing model includes: using a multilingual BERT base model to extract unstructured text rules in unstructured text through entity recognition; Extracting multi-source heterogeneous data rules based on search enhancement includes: dividing long texts in the manufacturing process knowledge benchmark data into text blocks; performing word embedding processing on the text blocks using a BGE-m3 model to obtain vectorized representations, storing the vectorized representations to a vector database and establishing indexes; combining vector similarity retrieval and keyword retrieval, fusing results of hybrid retrieval, and reordering using a BGE-reranker-v2-m3 model; and extracting the multi-source heterogeneous data rules from the retrieved text blocks based on a natural language processing model.

[0008] In some embodiments, the classification of the fixed format rules, the unstructured text rules and the multi-source heterogeneous data rules comprises: dividing the fixed format rules, the unstructured text rules and the multi-source heterogeneous data rules into fixed value classes, threshold range classes, logical quantity classes, mandatory field classes and format, unit requirement classes using a pre-trained multilingual BERT base model. The rule merging based on similarity comprises: using Jaccard similarity internal deduplication to merge rules with similarity greater than a set value.

[0009] In some embodiments, the comprehensive confidence is calculated based on three parts of rule source, rule itself characteristics and text consistency.

[0010] The preset large language model is obtained by Lora fine-tuning of two models DeepSeek R1:32b and / or Qwen2.5:32b.

[0011] In some embodiments, the method further comprises: Before writing the standard detection rule into the automobile manufacturing parameter abnormality detection rule library, the rule is further submitted to a third-party audit end for auditing through a preset link, and in the case of passing the auditing, the rule is written into the labeling rule library.

[0012] In another aspect, the present application also provides an automobile manufacturing parameter abnormality detection method, comprising the following steps: Based on a plurality of preset second-type data sources, a plurality of automobile manufacturing parameters to be detected are obtained and converted into unified structured data; the plurality of automobile manufacturing parameters include device configuration parameters, device control logic, process parameters, material parameters and process quality control parameters in the production process. The structured plurality of automobile manufacturing parameters are subjected to data cleaning to supplement missing values, remove duplicate data and abnormal values. Different data in the plurality of automobile manufacturing parameters are subjected to routing planning to form a routing table for data routing and storage. The plurality of automobile manufacturing parameters are classified based on data sources and combined according to preset rules to obtain functional groups. An automobile manufacturing parameter abnormality detection rule library obtained by the automobile manufacturing parameter abnormality detection rule library construction method is obtained, each functional group is detected and analyzed according to the standard rule in the automobile manufacturing parameter abnormality detection rule library, and a corresponding detection result is output.

[0013] On the other hand, the present invention also provides an abnormal automobile manufacturing parameter detection device, including a processor, a memory, and a computer program / instructions stored in the memory. The processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the device / equipment / system implements the steps of the above method.

[0014] On the other hand, the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0015] On the other hand, the present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0016] The present invention relates to a rule base for anomaly detection in automotive manufacturing parameters, a method for anomaly detection, and an apparatus. This method uses predefined regular expression pattern matching rules to match fixed-format rules from multi-source manufacturing process knowledge benchmark data. It then uses a natural language processing model for entity recognition to obtain unstructured text rules. After extracting rules from multi-source heterogeneous data using search enhancement, these rules are merged through classification and similarity comparison, and then fed into a pre-set large language model for reasoning to extract standard detection rules and construct an anomaly detection rule base for automotive manufacturing parameters. Based on this rule base, parameter detection is automated, improving detection accuracy. This method can efficiently identify the correctness of parameters and logic, enabling rapid and accurate anomaly detection for large-scale, multimodal automotive manufacturing parameters.

[0017] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0018] Those skilled in the art will understand that the objectives and advantages achievable with this invention are not limited to those specifically described above, and that the above and other objectives achievable with this invention will become clearer from the following detailed description. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings: Figure 1 This is a flowchart illustrating the method for constructing an abnormal detection rule base for automobile manufacturing parameters according to an embodiment of the present invention.

[0020] Figure 2This is a schematic diagram of the structure of the automobile manufacturing parameters required in the manufacturing parameter detection method based on multi-source heterogeneous data fusion according to an embodiment of the present invention.

[0021] Figure 3 This is a logical diagram illustrating the construction of a rule base for a manufacturing parameter detection method based on multi-source heterogeneous data fusion, as described in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the structure for rule extraction in the manufacturing parameter detection method based on multi-source heterogeneous data fusion according to an embodiment of the present invention.

[0023] Figure 5 This is a logical schematic diagram of a manufacturing parameter detection method based on multi-source heterogeneous data fusion according to an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0025] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0026] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0027] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0028] Automobile manufacturing is controlled by parameters and logic on the production line. High-quality new models require a sound foundation of parameters and logic for production. However, due to the large number of heterogeneous equipment and sensing systems involved in automobile manufacturing, the vast amount of equipment and process parameters and complex control logic are difficult to comprehensively test. This can easily lead to some potential problems going unidentified during the production phase, affecting the stability of subsequent mass-produced products. According to a survey of Company B and its parent company's other factories worldwide, the current solution mainly relies on technical experts for random checks, using manual methods and scripts. This approach is inefficient and lacks comprehensive functionality.

[0029] The present invention provides a method for constructing a rule base for detecting anomalies in automobile manufacturing parameters, the method comprising the following steps S101~S103: Step S101: Obtain various manufacturing process knowledge benchmark data for generating detection rules based on multiple preset first-type data sources. The first-type data sources include production logs, sensor data, process parameters, operation manuals, technical documents, and historical anomaly reports.

[0030] Step S102: Based on predefined regular expression pattern matching fixed format rules, the manufacturing process knowledge benchmark data is used to obtain unstructured text rules through entity recognition based on a natural language processing model, and multi-source heterogeneous data rules are extracted based on search enhancement.

[0031] Step S103: After classifying fixed format rules, unstructured text rules, and multi-source heterogeneous data rules, and merging similarity-based rules, calculate the comprehensive confidence score of each merged rule. Rules with a comprehensive confidence score higher than the set value are given to the preset large language model for inference and extraction to generate standard detection rules for automobile manufacturing parameters, forming an automobile manufacturing parameter anomaly detection rule library.

[0032] In step S101, the manufacturing process knowledge baseline data includes: design data for recording automotive design simulation and process requirements; manufacturing standards for recording the production standards of each vehicle model project; project data for recording the progress, experience summary and technical acceptance of each vehicle model project; and downtime data for recording downtime reports, downtime technical analysis and equipment failure mode analysis during production and mass production. After acquiring various manufacturing process knowledge benchmark data for generating detection rules based on multiple preset first-type data sources, the method further includes: parsing and cleaning the documents of each type of manufacturing process knowledge benchmark data, uniformly extracting them into plain text format, and removing invalid information including headers, footers, tables of contents, and references.

[0033] In the automotive manufacturing process, acquiring comprehensive and accurate benchmark data is fundamental to building an anomaly detection rule base. This step ensures data diversity and coverage by extracting data from multiple pre-defined primary data sources. These data sources include, but are not limited to, production logs, sensor data, process parameters, operation manuals, technical documents, and historical anomaly reports. Structured data (such as time-series data collected by sensors and numerical parameters in production plans) provides clear numerical ranges and logical relationships, while unstructured data (such as text descriptions in operation manuals and diagrams in technical documents) contains rich contextual information and tacit knowledge. The integration of multi-source heterogeneous data further enhances the comprehensiveness of the data, for example, by extracting information from production records from different workshops, material specifications provided by external suppliers, and historical fault databases. The data acquisition process is typically implemented through automated interfaces or data crawlers to ensure the real-time nature and dynamic updates of the data. In addition, data preprocessing is a key step in this process, including data cleaning (removing noise and erroneous data), format unification (converting data of different formats into a unified structure), and data alignment (ensuring consistency of timestamps and parameters). Through these operations, the baseline data not only reflects the normal state of the manufacturing process but also captures potential abnormal patterns, providing high-quality input for subsequent rule generation. For example, temperature, pressure, and speed parameters extracted from sensor data can serve as numerical benchmarks for subsequent rule generation, while operational specifications in unstructured text can be transformed into logical conditional rules. This integration and preprocessing of multi-source data ensures the comprehensiveness and adaptability of the rule base, enabling it to cope with complex and ever-changing manufacturing environments.

[0034] In step S102, after acquiring the manufacturing process knowledge baseline data, detection rules are extracted from the data using three core methods to adapt to different data types and complexities. First, regular expression pattern matching is an efficient method for extracting fixed-format rules from structured data. Through predefined patterns (such as numerical ranges, time formats, status codes, etc.), the system can quickly identify rules conforming to specific formats. For example, extracting the rule "trigger an alarm when the temperature exceeds 100°C" from sensor logs, regular expressions can accurately match the numerical range and logical conditions, ensuring the accuracy and consistency of the rules. Second, Natural Language Processing (NLP) models extract key rules from unstructured text using entity recognition technology. This method is particularly suitable for textual data such as operation manuals, technical documents, and historical reports. For example, through Named Entity Recognition (NER), the system can identify rules such as "stop production when the pressure is below 20 kPa" and convert them into structured logical conditions. NLP models can also understand implicit relationships in text through semantic analysis, such as complex rules like "maintenance is required when the equipment has been running for more than 10 hours." Finally, the search-enhanced extraction method targets multi-source heterogeneous data, extracting rules from dispersed data sources through intelligent search algorithms. This method combines keyword matching, association rule mining, and cross-data source integration techniques, enabling the extraction of valuable rules from historical anomaly reports, external databases, and production records from different workshops. For example, the search algorithm discovers that "a certain model of engine frequently experiences overheating problems under specific process conditions," and transforms this pattern into a detection rule. The combination of these three methods ensures the comprehensiveness and flexibility of rule extraction, adapting to various data types from simple to complex, and providing rich input for subsequent rule optimization.

[0035] Specifically, it uses predefined regular expression pattern matching to match fixed format rules, including: extracting parameter names and numerical ranges, extracting logical conditions between parameters to construct multiple regular expressions, and using regular expressions to traverse manufacturing process knowledge benchmark data to extract fixed format rules.

[0036] Entity recognition based on natural language processing models to obtain unstructured text rules includes: using a multilingual BERT base model to extract unstructured text rules from unstructured text.

[0037] The algorithm for extracting multi-source heterogeneous data rules based on search enhancement includes: segmenting long texts in manufacturing process knowledge benchmark data into text blocks; using the BGE-m3 model to perform word embedding processing on the text blocks to obtain vectorized representations, storing them in a vector database and creating an index; combining vector similarity retrieval and keyword retrieval, fusing the results of the mixed retrieval and re-ranking them using the BGE-reranker-v2-m3 model; and extracting multi-source heterogeneous data rules from the retrieved text blocks based on a natural language processing model.

[0038] In step S103, after extracting various rules, this step transforms the original rules into a high-quality standard detection rule base through classification, merging, and optimization operations. First, rule classification is the process of organizing rules according to their type (e.g., numerical range, text description, logical conditions) and application scenario (e.g., temperature monitoring, pressure detection, equipment maintenance). The purpose of classification is to improve the manageability and retrieval of rules, ensuring that relevant rules can be quickly located in practical applications. For example, "temperature exceeds 100°C" is classified as a numerical range rule, while "stop production when pressure is below 20 kPa" is classified as a logical condition rule. Second, similarity rule merging integrates duplicate or similar rules by calculating textual similarity, logical similarity, or semantic similarity between rules. This method not only reduces redundancy in the rule base but also improves the universality and maintainability of the rules. For example, "alarm when temperature exceeds 100°C" and "trigger alarm when temperature is above 100°C" are merged into a single rule, avoiding duplicate detection and resource waste. The comprehensive confidence assessment calculates the confidence level of each rule using statistical analysis and machine learning methods. The confidence level is calculated based on the rule's frequency of occurrence, data support, and consistency with other rules. Only rules with a confidence level higher than a set threshold (e.g., 0.8) are retained to ensure accuracy and reliability. Finally, a pre-defined large language model (such as a Transformer-based GPT model) infers and optimizes the retained rules. Through contextual understanding and logical reasoning, the large language model can identify and correct potential problems in the rules. For example, "alarm when temperature exceeds 100°C" can be optimized to "alarm when temperature exceeds 100°C for more than 5 seconds" to avoid false alarms caused by instantaneous fluctuations. Through these steps, the final standard detection rule base not only possesses high confidence and logical consistency but also dynamically adapts to changes in the manufacturing process, providing robust support for anomaly detection in automotive manufacturing parameters.

[0039] In some embodiments, fixed-format rules, unstructured text rules, and multi-source heterogeneous data rules are classified, including: using a pre-trained multilingual BERT base model to classify fixed-format rules, unstructured text rules, and multi-source heterogeneous data rules into fixed-value classes, threshold range classes, logical quantity classes, required field classes, format or unit requirement classes, and other classes.

[0040] Similarity-based rule merging includes: using Jaccard similarity internal deduplication to merge rules with similarity greater than a set value.

[0041] In some embodiments, after classifying fixed-format rules, unstructured text rules, and multi-source heterogeneous data rules, and merging rules based on similarity, the method further includes: calculating a comprehensive confidence score for each merged rule, wherein the comprehensive confidence score is calculated based on three parts: rule source, rule characteristics, and text consistency.

[0042] First, the source of the rules is a crucial factor influencing the overall confidence level. Rules may originate from various sources, such as production logs, operation manuals, and historical anomaly reports. Rules from different sources have varying degrees of confidence. For example, regarding the rule source, regular expression weights are 0.3, NLP entity extraction weights are 0.2, and RAG retrieval generation weights are 0.5, thus assessing the reliability and authority of the rule source.

[0043] Secondly, the characteristics of the rule itself are also key to assessing its confidence level. These characteristics include the clarity of its expression, the consistency of its logic, and its relevance to the manufacturing process. For example, a rule that is clearly expressed, logically rigorous, and closely related to specific manufacturing parameters will generally have a higher confidence level. Furthermore, the complexity of the rule also affects its confidence level. Simple and clear rules are often easier to verify and apply, while complex rules may require more data support and verification.

[0044] Finally, textual consistency is also crucial when calculating overall confidence. For rules extracted from text, their consistency across different texts or data sources needs to be checked. A rule that appears in a similar form across multiple independent texts or data sources has a higher confidence level. Textual consistency can be assessed by comparing the rule's wording, frequency of occurrence, and context across different sources. For example, if the rule "trigger an alarm when the temperature exceeds 100°C" is mentioned in multiple operation manuals and historical anomaly reports, then this rule has high textual consistency, thus increasing its overall confidence.

[0045] In some embodiments, the preset large language model is obtained by fine-tuning two models, DeepSeek R1:32b and / or Qwen2.5:32b, using Lora.

[0046] In some embodiments, the method further includes: before writing the standard detection rules into the automobile manufacturing parameter anomaly detection rule library, submitting them to a third-party reviewer for review through a preset link, and writing them into the annotation rule library if the review is passed.

[0047] On the other hand, the present invention also provides a method for detecting abnormal automobile manufacturing parameters, comprising the following steps S201~S205: Step S201: Obtain the multi-source automobile manufacturing parameters to be detected based on multiple preset second-type data sources, and convert them into unified structured data; the multi-source automobile manufacturing parameters include equipment configuration parameters, equipment control logic, process parameters, material parameters and process quality control parameters in the production process.

[0048] Step S202: Perform data cleaning on the structured multi-source automobile manufacturing parameters to fill in missing values ​​and remove duplicate data and outliers.

[0049] Step S203: Perform routing planning for different data in the multi-source automobile manufacturing parameters to form a routing table for data routing and storage.

[0050] Step S204: Classify the multi-source automobile manufacturing parameters based on the data source, and combine them according to preset rules to obtain functional groups.

[0051] Step S205: Obtain the automobile manufacturing parameter anomaly detection rule library as described in the automobile manufacturing parameter anomaly detection rule library construction method above, and perform detection and analysis on each functional group according to the standard rules in the automobile manufacturing parameter anomaly detection rule library, and output the corresponding detection results.

[0052] On the other hand, the present invention also provides an abnormal automobile manufacturing parameter detection device, including a processor, a memory, and a computer program / instructions stored in the memory. The processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the device / equipment / system implements the steps of the above method.

[0053] On the other hand, the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0054] On the other hand, the present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0055] The present invention will now be described with reference to a specific embodiment: This embodiment presents a manufacturing parameter detection method based on multi-source heterogeneous data fusion. By fusing design data, manufacturing standard data, project data, downtime data, etc., and generating rules through a large model, it enables large-scale and automated detection of automotive manufacturing parameters. Specifically, it includes three parts: the data source required for detection, the construction of detection rules, and the parameter detection method.

[0056] In the automobile manufacturing process, manufacturing parameters are divided into two categories: The first category is equipment parameters. Equipment parameters refer to parameters related to the operation of the equipment. These can be further divided into equipment configuration parameters and equipment control logic. Equipment configuration parameters are modifiable configuration parameters within the equipment, such as the robot's global operating speed and the model of the sensors used for online measurement. Equipment control logic is the logic program that controls the operation of the equipment. Different devices use different programming languages; for example, Siemens robots use Ladder Diagrams (LAD), Function Blocks (FBD), and Structured Control Language (SCL), while KUKA robots use the KUKA programming language (KRL).

[0057] The second category is process parameters. Process parameters refer to a series of key parameters used to control product quality and production efficiency during automobile manufacturing. We further divide them into process parameters, material parameters, and process quality control parameters. Process parameters refer to the process and tooling configuration parameters in the manufacturing process, such as welding parameters like welding voltage, current, and time, and assembly parameters like tightening torque and assembly clearance. Material parameters include the material parameters of automotive parts and process-related equipment, such as the material and thickness of body panels, and the material and adhesion of coatings. Process quality control parameters refer to parameters that control process quality, such as the spot welding process stability (PSF), the welding process quality (UIP), and the maximum current threshold.

[0058] In addition to the automotive manufacturing parameters themselves, testing automotive manufacturing parameters also requires multi-source manufacturing process knowledge benchmark data, including design data, manufacturing standards, project data, and downtime data, such as... Figure 2 As shown.

[0059] Data is primarily acquired via Industrial Internet of Things (IIoT), based on Company B's IIoT platform, using a Manufacturing Service Bus (MSB) to ensure low latency and data integrity. The data format is JSON. Some equipment control logic, such as PLC programs and robot programs, is generally not transmitted via IIoT; instead, source files such as AAWL and GDC are processed directly.

[0060] The design data includes production line design simulation data and process requirement data. This article only requires equipment asset data, drawing data, and process design data. The data can be in a structured database or Excel format.

[0061] The manufacturing standards are the production standard documents for each vehicle model project, including production line design requirements, equipment requirements, and process standards. The data format is PDF.

[0062] Project data includes project documents for each vehicle model, project experience summaries, and technical acceptance documents. Data formats include PDF, WORD, and EXCEL.

[0063] The downtime data includes downtime reports, downtime technical analysis reports, and equipment failure mode analysis (FMEA) reports following equipment downtime during the commissioning and mass production periods of each vehicle model project. Data formats include PDF, PPT, and EXCEL.

[0064] Detection rule construction: In rule-based systems, rule generation is a core component, directly impacting system performance and accuracy. The goal of rule generation is to extract rules from data or knowledge that can be used for reasoning or decision-making. Common methods include rule generation based on expert knowledge and rule generation based on machine learning. To achieve high efficiency and 100% accuracy in rule generation, rules are first generated using a large model, and then manually reviewed and adjusted by experts to ensure their correctness. Figure 3 As shown.

[0065] The large-scale model generates rules using a three-layer rule extraction strategy: the first layer uses predefined regular expression pattern matching for structured rules; the second layer utilizes entity recognition for NLP rule extraction; and the third layer uses the RAG system combined with similarity and keyword hybrid retrieval and re-ranking for rule extraction. The system also integrates a BERT-based multilingual rule classifier for type identification of the extracted rules. This method effectively handles documents of different languages ​​and formats, extracting accurate parametric rules. The three extraction methods operate in parallel, each generating rules, classifying the generated rules, and then merging all classified rules in a merging phase. This phase merges rules from different sources, handles rule conflicts, removes duplicates, standardizes rule formats, and finally scores confidence and outputs the results. Each extraction method is independent but ultimately converges in the merging phase to ensure rule consistency and completeness. The rules extracted through this three-layer strategy are then used for large-scale model inference to generate the final rules. This embodiment combines the advantages of the RAG method and the large language model, integrating knowledge from external databases to improve the accuracy and reliability of the generated rules. Especially for knowledge-intensive tasks, RAG collaboratively combines the intrinsic knowledge of large language models with a vast, dynamic external database repository, supporting continuous knowledge updates and the integration of domain-specific information. Large model generation rules include... Figure 4 As shown, the specific implementation method is as follows: 1) Document Parsing and Cleaning. Due to the diverse sources and formats of documents, a unified format conversion is necessary. Content from files with different formats is extracted into plain text, and the document content is cleaned for subsequent processing. Here, we use PyMuPDFLoader to load PDF files, and Unstructured to load WORD, PPT, and TEXT files. EXCEL files are converted to text descriptions line by line. The document content is cleaned to remove irrelevant information, including headers, footers, tables of contents, and references.

[0066] 2) Regular expression-based rule extraction. Regular expressions are used to extract rules with a fixed format.

[0067] 3) Rule extraction based on NLP. Latent rules are extracted from unstructured text using the bert-base-multilingual-cased entity recognition model.

[0068] 4) RAG-based rule extraction. More accurate rules are extracted by combining vector indexing and hybrid retrieval with contextual information, as detailed below: Text segmentation. For long texts, segmenting the document into text blocks is a common and necessary operation. Common segmentation strategies include fixed-size segmentation, recursive segmentation, sliding window segmentation, paragraph-based segmentation, and semantic segmentation. This study selects appropriate granularity text block retrieval for text segmentation based on the document's characteristics. For long documents such as PDFs, the semantic segmentation strategy SpacyTextSplitter is used to segment long documents into smaller text blocks, segmenting based on natural boundaries such as paragraphs and chapters. Dynamic segmentation based on sliding windows introduces adaptive adjustment of overlap, increasing overlap for semantically dense parts to facilitate subsequent processing. For short documents such as HTML, the entire document is written to maintain high information integrity.

[0069] Word embedding. This converts text into vectors. Since the documents are primarily in English, Chinese, and German, the embedding model needs to support multiple languages. After comparing models such as Microsoft's multilingual-e5-large-instruct and Zhiyuan's BGE series, the BGE-m3 model was chosen.

[0070] The data is stored in a vector database. Due to the large number of documents, a batch writing strategy is adopted. After training and writing the FAISS vector database index, the FAISS index and metadata are saved. The index uses the IndexIVFFlat method, which is an indexing method used to accelerate vector search. It divides the dataset into multiple cluster spaces through an inverted index structure, thereby reducing the number of target vectors that need to be distance-calculated and improving search efficiency.

[0071] Hybrid retrieval. The retrieval process is typically based on semantic matching techniques, calculating the similarity between the query and document fragments to filter out the most relevant document fragments. To improve retrieval accuracy, we use a hybrid retrieval approach combining vector similarity retrieval and keyword retrieval (BM25). This retrieves Top-K relevant documents from a vector database, while simultaneously utilizing a sparse embedding model to cover full-text search capabilities. The results of keyword and vector retrieval are then fused to improve recall and precision. Furthermore, re-ranking is used to enhance semantic relevance and search quality; here, the BGE-reranker-v2-m3 model is employed for re-ranking.

[0072] 5) Rule Classification. Based on the source, rules extracted by the regular expression module are labeled with the category tag "regex", rules extracted by the NLP module are labeled with the category tag "nlp", and rules extracted by the RAG module are labeled with the category tag "rag". The rule types are then categorized into six types using bert-base-multilingual-cased: fixed value, threshold range, logical variable, required field, format / unit requirement, and others.

[0073] 6) Rule Merging. Each source uses Jaccard similarity internal deduplication; rules with a similarity greater than 0.9 are removed. A comprehensive confidence score is designed and calculated, derived from three parts: rule source, rule features, and text consistency. For rule sources, regularization weight is 0.3, NLP entity extraction weight is 0.2, and RAG retrieval generation weight is 0.5. Rule features include scores for reasonable rule length, complete rule structure, and keyword richness. Text consistency considers rules generated simultaneously from multiple sources; if two or more sources extract the same rule, the score is increased. Rules with a comprehensive confidence score higher than 0.7 are fed into the large model for inference.

[0074] 7) Large Language Model Inference. Large language models are developing rapidly and are constantly being updated. Here, we can choose the latest and better performing model. Since it is a local deployment, we currently use DeepSeek R1:32b and Qwen2.5:32b as the base models, and fine-tune them using LoRa.

[0075] After initial rule extraction and generation, the rules are submitted to domain experts for review, and adjustments and optimizations are made based on their feedback. This combination of manual and automated methods not only improves the accuracy of the rules but also enhances their interpretability, making them easier for engineers to understand and apply. Manual adjustments include three steps: rule review, rule refinement, and rule classification.

[0076] 1) Rule Review. This includes checking the accuracy and completeness of the rules. Accuracy checks ensure each rule meets the actual requirements. Completeness checks ensure the rules cover all key parameters and conditions.

[0077] 2) Rule Refinement. This includes specifying conditions and providing supplementary context. Specific conditions refer to making ambiguous conditions concrete. For example: Supplementing context adds contextual information to the rules to facilitate subsequent application. For example:

[0078] 3) Rule Classification. Parameters are further classified according to vehicle model, technology category, and function. Prioritization is then based on the importance of each function.

[0079] After the rules are adjusted, they are written to the rule base for storage. This article uses a MySQL database to store the rules. The fields of the rule base table include: rule ID, technology name, function category, parameter name, rule content, rule source, and priority.

[0080] Throughout the parameter testing process, rule formulation and optimization are crucial. Testing rules need to be regularly updated and adjusted based on the latest automotive manufacturing standards and technical requirements to ensure the timeliness and accuracy of the test results. Rule updates are incremental, updating only the changed parts of the rule base, including adding new rules, modifying existing rules, or deleting outdated rules. A version control mechanism is introduced to record historical versions of the rule base and to perform rollbacks or comparative analysis when necessary.

[0081] Parameter detection methods such as Figure 5 As shown: First, multi-source heterogeneous data is transformed into unified structured data. Through data preprocessing, data routing, data transformation, data association, and data retention strategies, automotive manufacturing parameters, design data, and rule bases generated from manufacturing standards, project knowledge, and downtime reports are transformed into unified structured data.

[0082] Data preprocessing involves cleaning the data to fill in missing values ​​and remove duplicate and outlier values.

[0083] Data routing plans routes for different data, forms a routing table, and executes data routing and storage. Due to the large volume of IoT data and the presence of data from multiple topics across different technologies, especially in new vehicle projects, IoT connections from different technologies are gradually activated during the commissioning phase. Early data is unstable. To facilitate data management, data routing provides unified management of data access. Routing control policies configure whether data from each topic of each technology is accessible. Once the control policy is configured and enabled for that topic, the data will be accessed and sent.

[0084] Data transformation standardizes and structures semi-structured data from the Industrial Internet of Things (IIoT), unstructured data from device logic, structured data from design, and rule information obtained through identification, making them conform to a unified structured data format and definition.

[0085] Data association links data elements in the data to be tested based on time, vehicle model, technology type, equipment name, and process point name.

[0086] The data retention strategy manages the data lifecycle. Different technologies and data types can have their data retention periods customized by modifying the default values, which are 3 years by default.

[0087] Next, parameter testing is performed. Functional groups are divided, and based on rules, automotive manufacturing parameters are tested, with the results output. The purpose is to ensure that the parameters of each functional group meet preset standards and rules, thereby guaranteeing the vehicle's performance and quality.

[0088] Based on the different data sources, the parameters are divided into different categories and inherit the subcategories of the data source to facilitate the identification of data sources and subsequent management.

[0089] After classification, the parameters are combined according to rules to form functional groups. The division of functional groups is based on the correlation and functional consistency between parameters, ensuring that the parameters within each functional group can work together to achieve a specific function. This functional grouping not only helps simplify the detection process but also improves the accuracy and targeting of the detection.

[0090] After the functional groups are divided, the system automatically calculates and checks the parameters in each functional group based on the detection rules. The system automatically calculates the parameters of each functional group to be tested according to the rules and outputs the detection results. The detection results are usually divided into two types: pass or fail. If all parameters in a functional group meet the detection rules, the system will output a "pass" result, indicating that the parameter settings of that functional group are reasonable and meet the design requirements. However, if any parameter in a functional group fails the detection, the system will output a "fail" result, further indicating which parameters do not meet the rules and the reasons for the non-compliance.

[0091] For parameter groups that fail the test, the system not only identifies the problem but also provides the correct parameter values ​​or threshold ranges to facilitate adjustments and corrections by engineers. This feedback mechanism not only helps to quickly locate problems but also provides guidance for subsequent parameter optimization. In this way, the parameter testing process can not only identify potential problems but also provide specific solutions, thereby ensuring the smooth progress of the automobile manufacturing process. Furthermore, automated parameter testing avoids human intervention and errors, significantly improving the accuracy and reliability of parameter testing in automobile manufacturing.

[0092] Corresponding to the above method, the present invention also provides an apparatus / system including a computer device, the computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the apparatus / system performs the steps of the method as described above.

[0093] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned edge computing server deployment method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0094] In summary, the automobile manufacturing parameter anomaly detection rule base construction, anomaly detection method, and apparatus described in this invention utilize predefined regular expression pattern matching fixed-format rules based on multi-source manufacturing process knowledge benchmark data. It then uses a natural language processing model for entity recognition to obtain unstructured text rules, extracts rules from multi-source heterogeneous data using search enhancement, merges these rules through classification and similarity comparison, and feeds them into a pre-set large language model for reasoning to extract standard detection rules and construct an automobile manufacturing parameter anomaly detection rule base. Based on this rule base, parameter detection is automated, improving detection accuracy. This method can efficiently identify the correctness of parameters and logic, enabling rapid and accurate anomaly detection for large-scale, multimodal automobile manufacturing parameters.

[0095] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0096] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0097] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0098] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a rule base for detecting abnormal automobile manufacturing parameters, characterized in that, The method includes the following steps: Based on multiple preset first-type data sources, various manufacturing process knowledge benchmark data are acquired to generate detection rules; the first-type data sources include production logs, sensor data, process parameters, operation manuals, technical documents, and historical anomaly reports; The manufacturing process knowledge benchmark data is used to obtain unstructured text rules by matching fixed format rules based on predefined regular expression patterns, and entity recognition is performed based on natural language processing models. Multi-source heterogeneous data rules are extracted based on search enhancement. After classifying the fixed format rules, the unstructured text rules, and the multi-source heterogeneous data rules, and merging the rules based on similarity, a comprehensive confidence score is calculated for each merged rule. Rules with a comprehensive confidence score higher than a set value are then used by a preset large language model for inference and extraction to generate standard detection rules for automobile manufacturing parameters, forming an automobile manufacturing parameter anomaly detection rule library.

2. The method for constructing a rule base for detecting abnormal automobile manufacturing parameters according to claim 1, characterized in that, The manufacturing process knowledge benchmark data includes: design data for recording automotive design simulation and process requirements; manufacturing standards for recording the production standards of each vehicle model project; project data for recording the progress, experience summary and technical acceptance of each vehicle model project; and downtime data for recording downtime reports, downtime technical analysis and equipment failure mode analysis during production and mass production. After acquiring various manufacturing process knowledge benchmark data for generating detection rules based on multiple preset first-type data sources, the method further includes: parsing and cleaning the documents of each type of manufacturing process knowledge benchmark data, uniformly extracting them into plain text format, and removing invalid information including headers, footers, tables of contents, and references.

3. The method for constructing a rule base for detecting abnormal automobile manufacturing parameters according to claim 2, characterized in that, Based on predefined regular expression pattern matching fixed format rules, the method includes: extracting parameter names and numerical ranges and extracting logical conditions between parameters to construct multiple regular expressions, and traversing the manufacturing process knowledge benchmark data based on the regular expressions to extract the fixed format rules. Entity recognition based on natural language processing models to obtain unstructured text rules, including: using a multilingual BERT base model to extract unstructured text rules from unstructured text; The method for extracting multi-source heterogeneous data rules based on search enhancement includes: segmenting long texts in the manufacturing process knowledge benchmark data into text blocks; using the BGE-m3 model to perform word embedding processing on the text blocks to obtain vectorized representations, storing them in a vector database and establishing an index; combining vector similarity retrieval and keyword retrieval, fusing the results of the mixed retrieval and re-ranking them using the BGE-reranker-v2-m3 model; and extracting the multi-source heterogeneous data rules from the retrieved text blocks based on a natural language processing model.

4. The method for constructing a rule base for detecting abnormal automobile manufacturing parameters according to claim 1, characterized in that, The fixed format rules, unstructured text rules, and multi-source heterogeneous data rules are classified, including: using a pre-trained multilingual BERT base model to divide the fixed format rules, unstructured text rules, and multi-source heterogeneous data rules into fixed value classes, threshold range classes, logical quantity classes, required field classes, and format or unit requirement classes; Similarity-based rule merging includes: using Jaccard similarity internal deduplication to merge rules with similarity greater than a set value.

5. The method for constructing a rule base for detecting abnormal automobile manufacturing parameters according to claim 1, characterized in that, The overall confidence score is calculated by assigning scores to three parts: the source of the rule, the characteristics of the rule itself, and the consistency of the text. Among them, the characteristics of the rule include the clarity of its expression, the consistency of its logic, and its relevance to the manufacturing process.

6. The method for constructing an abnormal automobile manufacturing parameter detection rule base according to claim 1, characterized in that, The method further includes: Before the standard detection rules are written into the automobile manufacturing parameter anomaly detection rule library, they are also submitted to a third-party reviewer for review through a preset link. If the review is passed, the rules are written into the annotation rule library.

7. A method for detecting abnormal automobile manufacturing parameters, characterized in that, Includes the following steps: The multi-source automotive manufacturing parameters to be detected are obtained based on multiple preset second-type data sources and converted into unified structured data; the multi-source automotive manufacturing parameters include equipment configuration parameters, equipment control logic, process parameters, material parameters and process quality control parameters in the production process; The structured multi-source automotive manufacturing parameters are cleaned to fill in missing values ​​and remove duplicate and outlier values. Routing planning is performed on the different data in the multi-source automobile manufacturing parameters to form a routing table for data routing and storage; The multi-source automobile manufacturing parameters are classified based on the data source and combined according to preset rules to obtain functional groups; Obtain the automobile manufacturing parameter anomaly detection rule library obtained in the automobile manufacturing parameter anomaly detection rule library construction method as described in any one of claims 1 to 6, and perform anomaly detection analysis on each functional group according to the standard rules in the automobile manufacturing parameter anomaly detection rule library, and output the corresponding anomaly detection results.

8. A device for detecting abnormal automobile manufacturing parameters, comprising a processor, a memory, and a computer program / instructions stored in the memory, characterized in that, The processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the device / equipment / system implements the steps of the method as described in claim 7.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 7.

Images