Industrial standard part-oriented multi-agent anomaly detection system and method

Abstract

The application discloses a multi-agent abnormality detection system and method for industrial standard parts. The system comprises a data acquisition agent, an image detection agent, a parameter analysis agent, a document recall agent and a master control calling agent. The master control calling agent continuously executes one of four actions including executing the image detection agent, executing the parameter analysis agent, executing the document recall agent and stopping at each time step to obtain corresponding results until the master control calling agent outputs the action of stopping. Then, the final result obtained by the document recall agent is output as the final abnormality detection result, and the system ends. The application applies the multi-agent technology to the field of industrial standard part quality detection, jointly analyzes and diagnoses multi-modal data such as images, texts and parameters, and dynamically schedules and chain inferences by the agent, thereby breaking through the limitation of traditional detection systems relying on single models or single inputs.

Description

Technical Field

[0001] This invention belongs to the field of anomaly detection of industrial standard parts, and specifically relates to a multi-agent anomaly detection system and method for industrial standard parts. Background Technology

[0002] Currently, in the field of quality inspection and anomaly identification of industrial standard parts, a combination of traditional image processing and manual inspection techniques is commonly used. Industrial cameras are frequently used on production lines in conjunction with algorithms such as shape matching, edge detection, and grayscale thresholding to detect surface defects in standard parts. However, this method relies on manually setting thresholds, has limited generalization ability of the detection model, struggles to cope with interference from factors such as changes in lighting, pose variations, and distance, and has a weak ability to judge minor damage, easily leading to false positives and false negatives. Furthermore, manual review is costly and inefficient, failing to meet the needs of large-scale manufacturing.

[0003] With the development of computer technology, some factories have adopted visual inspection solutions based on deep learning models such as CNNs and Transformers, which automatically extract image features and perform classification and recognition through neural networks. However, these methods typically rely on a large number of similarly distributed samples for training, and their robustness becomes difficult to guarantee when the number of inspection objects increases or the working conditions become more complex. Finally, these methods only focus on image-dimensional information and cannot effectively combine structured information such as historical inspection records of standard parts, parameter specifications, and assembly conditions, resulting in a lack of knowledge accumulation and a lack of interpretability in diagnostic results.

[0004] In summary, existing methods for detecting anomalies in industrial standard parts generally suffer from the following shortcomings: single-model or single-modal information cannot support comprehensive and accurate anomaly identification; they lack the ability to jointly reason based on images, parameters, document specifications, and historical case knowledge; and they have poor robustness in complex industrial environments, making it difficult to meet the needs of real-time online detection. Therefore, there is a need for an anomaly detection method that can integrate multi-source data, including visual information, structured parameters, industry document standards, and historical case data, to achieve multi-role collaborative reasoning, thereby improving the accuracy, stability, and interpretability of detection. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention provides a multi-agent anomaly detection system and method for industrial standard parts. It solves the technical problems of existing technologies, such as the inability of single-model or single-modal information to support comprehensive and accurate anomaly identification; the lack of joint reasoning ability based on images, parameters, document specifications, and historical case knowledge; and poor robustness in complex industrial environments, making it difficult to meet real-time online detection requirements. This invention is a system and method for anomaly detection that combines intelligent agent interaction, retrieval enhancement generation, and image feature fusion algorithms, mainly involving automated quality inspection and safety assessment analysis of industrial standard parts (industrial components, such as bearings, gears, bolts, and nuts).

[0006] The technical solution adopted in this invention is: I. A multi-agent anomaly detection system for industrial standard parts: The data acquisition agent acquires images of the component under test, a set of parameter measurements including dimensions and mass, and a set of files including industry standard specifications and historical component anomaly detection data. The image detection agent processes the images of the component under test to obtain a structured image defect list. The parameter analysis agent processes the set of parameter measurements of the component under test to obtain corresponding structured parameter results. The document retrieval agent obtains anomaly detection results based on the structured image defect list, structured parameter results, and file set. The master control agent continuously executes one of four actions at each time step: executing the image detection agent, executing the parameter analysis agent, executing the document retrieval agent, and stopping. The process continues until the master control agent outputs "stop," at which point the final result obtained by the document retrieval agent is output as the final anomaly detection result, and the system terminates.

[0007] The execution steps of the image detection agent are as follows: First, the image of the component to be tested is augmented using a Gabor filter. Then, the YOLOv8 detection algorithm is used to perform preliminary defect identification on the augmented image, outputting candidate boxes, defect categories, and confidence scores to obtain preliminary detection results. Then, the pre-trained first language model is called to analyze and summarize the preliminary detection results of the image, generating a structured image defect list of the component to be tested.

[0008] The execution steps of the parameter analysis agent are as follows: normalize each parameter measurement value in the parameter measurement value set to obtain the corresponding normalized deviation value; calculate the corresponding Z-score statistical value based on each normalized deviation value; obtain the parameter health index set of the component under test based on the Z-score statistical value of all parameter measurement values; and call the pre-trained second language model to perform comprehensive analysis on the parameter health index set of the component under test to generate the structured parameter results of the component under test.

[0009] The structured parameter results of the component under test are obtained by processing the following formula: in, This indicates the structured parameter results of the component under test; This represents the second largest pre-trained language model; This represents the set of health indices of the parameters of the component under test. Indicates an index; Represents the first parameter in the set of parameter measurements. Each parameter measurement value; Indicates the measured value of the parameter The corresponding parameter weights; This indicates the total number of parameter measurements; Indicates the measured value of the parameter The corresponding Z-score statistics; Indicates the measured value of the parameter The corresponding normalized deviation value; Indicates the reference mean; Indicates the reference standard deviation; and This indicates the upper and lower limits of the set standard range; It is a constant used to avoid the denominator being zero.

[0010] The execution steps of the document retrieval agent are as follows: F1. A text vector representation model is used to encode the structured image defect list and structured parameter results of the component under test into semantic query text. Then, the text vector representation model maps the semantic query text into embedded vectors. The same text vector representation model is used to encode and map each document in the document collection into a corresponding text vector.

[0011] F2. Obtain coarse recall results for each document based on the semantic query text, embedding vector, and text vector of each document.

[0012] F3. Use a cross-encoder to process each document to obtain the consistency relevance of the corresponding document, and then weight the coarse recall result and the consistency relevance to obtain the final recall result of a single document.

[0013] F4. Then, based on the final recall results of all documents, sort each document in descending order, and select the Top K relevant documents from the descending sorted set to construct a structured knowledge set.

[0014] F5. The pre-trained third language model is invoked to summarize and synthesize the structured knowledge set, resulting in complete industry standards and solution suggestions, which are then used as anomaly detection results.

[0015] The coarse recall result for each document is set according to the following formula: in, For indexing; This represents the i-th document in the file set; Indicates the first Coarse recall results for one document; Indicates the weighting coefficient; and They represent the first BM25 similarity scores and vector similarity values ​​for each document; An algorithm for calculating the cosine similarity between two vectors; Represents an embedding vector; Indicates the first The text vector of a document; This represents the semantic query text.

[0016] The final recall result for each document is set according to the following formula: in, Indicates the first The final recall result for each document; Indicates the weighting coefficient; Indicates the first Consistency relevance of individual documents; Indicates the first Coarse recall results for one document; This refers to the Cross-Encoder module; Represents semantic query text; For indexing; This represents the i-th document in the file set.

[0017] The execution steps of the master control calling agent are as follows: The pre-trained fourth language model is called to process the state set of the current time step and output one of four actions, including executing the image detection agent, executing the parameter analysis agent, executing the document recall agent, and stopping execution. The master control calling agent executes the corresponding action to obtain the result of the current time step. Then, the master control calling agent updates the state set and uses the updated state set as the state set of the next time step. The master control calling agent is continuously executed in consecutive time steps until the action output by the master control calling agent is to stop execution. At this time, the system stops running and the final result generated by the document recall agent is the final anomaly detection result output.

[0018] The initial state set includes the image of the component under test, the set of parameter measurements, and user requirements; the non-initial state set includes the image of the component under test, the set of parameter measurements, user requirements, and all results obtained by the master controller calling the agent to perform various actions in all time steps before the current time step.

[0019] II. A multi-agent anomaly detection method for industrial standard parts: S1, the data acquisition agent acquires images of the parts to be tested, a set of parameter measurements including dimensions and mass, and a set of files including industry standard specifications and historical case data.

[0020] S2. The master controller calls the agent to call the pre-trained fourth language model to process its own state set at the current time step and output one of four actions, including executing the image detection agent, executing the parameter analysis agent, executing the document recall agent, and stopping. The master controller calls the agent to execute the corresponding action to obtain the result of the current time step. Then, the master controller calls the agent to update the state set and uses the updated state set as the state set for the next time step.

[0021] S3. Repeat step S2 continuously until the output action is "stop". The final result generated by the document recall agent is then used as the final anomaly detection result.

[0022] The beneficial effects of this invention are: 1. This invention is the first to apply multi-agent technology to the field of quality inspection of industrial standard parts. It performs joint analysis and diagnosis of multimodal data such as images, text, and parameters, and the agents perform dynamic scheduling and chain reasoning. This breaks through the limitations of traditional inspection systems that rely on a single model or a single input, thus possessing an advanced technical system.

[0023] 2. Compared with existing multimodal large language model detection schemes that forcibly unify the modeling of multimodal information, the multimodal processing of this invention is processed into standard results through different modules. New detection tools or industry standards can be added without retraining the entire model. The solution is more open and user-friendly and can be continuously updated and iterated according to actual production needs, thus achieving strong robustness.

[0024] 3. This invention integrates multi-source information for intelligent detection, making comprehensive use of image features, structured parameter data, industry document standards, and historical component anomaly detection data to achieve accurate identification and judgment of abnormal states of the components under test.

[0025] 4. This invention enhances the robustness and generalization ability of the system. Through multi-agent collaborative reasoning and knowledge-enhanced retrieval mechanisms, it improves the adaptability and stability of the detection model under complex working conditions and different types of standard parts.

[0026] 5. This invention improves the interpretability of the detection process, outputting the reasoning chain and evidence while giving the conclusion, reducing the intensity of human intervention, and improving the efficiency and reliability of quality detection results. Attached Figure Description

[0027] Figure 1 This is a system module diagram of the present invention.

[0028] Figure 2 This is a flowchart of the method of the present invention.

[0029] Figure 3 This is a diagram of the component to be tested in this embodiment.

[0030] Figure 4 This is the result of processing the image of the component under test in this embodiment using a Gabor filter.

[0031] Figure 5 This is a preliminary inspection result diagram of the component to be tested in this embodiment. Detailed Implementation

[0032] The present invention will now be described in more detail with reference to the accompanying drawings and embodiments. However, the present invention is not limited thereto. For those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

[0033] like Figure 1 As shown in this embodiment, a multi-agent anomaly detection system for industrial standard parts includes: The data acquisition agent acquires images of the parts under test, a set of parameter measurements including dimensions and mass, and a set of files including industry standard specifications and historical parts anomaly detection data.

[0034] Historical component anomaly detection data includes images of each component, a set of parameter measurements, user requirements, and anomaly detection results.

[0035] Images of the parts to be tested (bolts in this example) are acquired using an industrial camera on the production line, such as... Figure 3 As shown, a set of measured values, including dimensions and mass, of the component under test are obtained through a distance-measuring and weighing device. A set of documents, including industry standard specifications and historical case data, is acquired through big data collection methods.

[0036] In practice, the set of parameter measurement values ​​acquired by the ranging and weighing equipment is as follows: , This indicates the measured values ​​of various parameters (such as length, diameter, chamfer, etc.). Indicates weight.

[0037] The image detection agent processes images of the parts under test to obtain a corresponding structured image defect list: The images of the parts under test are first augmented using a Gabor filter to improve the visibility of subtle features such as surface cracks, dents, and scratches. Figure 4 As shown, the enhanced image is then preliminarily defect-identified using a finely tuned YOLOv8 detection algorithm. The algorithm outputs candidate bounding boxes, defect categories, and confidence scores, yielding preliminary detection results, as shown below. Figure 5 As shown; then, the pre-trained first language model is called and the preliminary detection results of the image are analyzed and summarized using preset prompts to generate a structured image defect list of the parts to be tested. The pre-trained first language model adopts Qwen3-VL-32B-Thinking.

[0038] Furthermore, the pre-trained first language model outputs the logical chain and basis for reasoning along with the structured image defect list.

[0039] The specific processing steps of the image detection agent are as follows: The input image is first augmented using a Gabor filter, and then the YOLOv8 detection algorithm is used to perform preliminary defect identification on the augmented image to obtain the corresponding preliminary detection results. in, This indicates the preliminary test results for the component under test; Represents the YOLOv8 detection algorithm. Represents an enhanced image. , and These represent the components under test. , and The confidence level of the defect category.

[0040] Next, the image detection agent analyzes and summarizes the preliminary detection results of the image to generate a list of structured defects. This process is defined as follows: in, The first pre-trained language model, This indicates the preliminary test results for the component under test. A list of structured image defects representing the component under test.

[0041] The parameter analysis agent processes the set of parameter measurements for the component under test to obtain corresponding structured parameter results. It normalizes each parameter measurement to obtain a normalized deviation value, calculates the corresponding Z-score based on each deviation, and obtains a set of parameter health indices for the component based on the Z-scores of all parameter measurements. Finally, it calls a pre-trained second language model to comprehensively analyze the parameter health indices and generate the structured parameter results for the component. The pre-trained second language model uses Qwen3-32B-Thinking.

[0042] The specific processing procedure of the parameter analysis agent is as follows: The normalized deviation value is obtained by processing the measured value of each parameter according to the following formula: in, Indicates an index; Represents the first parameter in the set of parameter measurements. Each parameter measurement value; Indicates the measured value of the parameter The corresponding normalized deviation value; and Indicates the upper and lower limits of the set standard range; It is a very small constant used to avoid the denominator being zero.

[0043] The Z-score statistical value for each parameter measurement is obtained by processing it according to the following formula: in, Indicates the measured value of the parameter The corresponding Z-score statistics; Indicates the reference mean; This represents the reference standard deviation. Both the reference mean and reference standard deviation are determined based on industry standards and specifications, historical case data, etc.

[0044] The Z-score statistics of all parameter measurements are processed using the following formula to obtain the set of parameter health indices for the component under test: in, This represents the set of health indices of the parameters of the component under test. Indicates the measured value of the parameter The parameter weights corresponding to the Z-score statistics can be set according to the importance of each dimension parameter of the component; Indicates the measured value of the parameter The corresponding Z-score statistics; Indicates an index; This indicates the total number of parameter measurements.

[0045] Finally, the pre-trained second language model is invoked, and preset prompts are used to perform a comprehensive analysis of the parameter health index set, generating structured parameter results: in, This indicates the structured parameter results of the component under test; This represents the second largest pre-trained language model; This represents the set of health indices of the parameters of the component to be tested.

[0046] Furthermore, the pre-trained second language model outputs the logical chain and basis of reasoning along with the structured parameter results.

[0047] The document recall agent obtains anomaly detection results based on a structured image defect list, structured parameter results, and a set of documents.

[0048] The specific processing steps of the document retrieval agent are as follows: F1. A text vector representation model (Embedding) is used to encode the structured image defect list and structured parameter results of the component under test into semantic query text. Then, the text vector representation model maps the semantic query text into embedding vectors. The same text vector representation model is used to encode and map each document in the document set into a corresponding text vector. Specifically, the text vector representation model (Embedding) used is Qwen3-Embedding-8B.

[0049] Embedded vector The formula obtained after processing is: ,in Represents an embedding vector. This represents a text vector representation model; This represents the semantic query text.

[0050] F2. Obtain coarse recall results for each document based on the semantic query text, embedding vector, and text vector of each document.

[0051] The coarse recall result for each document is set according to the following formula: in, For indexing; This represents the i-th document in the file set; Indicates the first Coarse recall results for one document; Indicates the weighting coefficient; and They represent the first BM25 similarity scores and vector similarity values ​​for each document; An algorithm for calculating the cosine similarity between two vectors; Represents an embedding vector; Indicates the first The text vector of a document; This represents the semantic query text.

[0052] F3. Use a cross-encoder to process each document to obtain the consistency relevance of the corresponding document, and then weight the coarse recall result and the consistency relevance to obtain the final recall result of a single document.

[0053] The final recall result for each document is set according to the following formula: in, Indicates the first The final recall result for each document; Indicates the weighting coefficient; Indicates the first Consistency relevance of individual documents; Indicates the first Coarse recall results for one document; This refers to the Cross-Encoder module; Represents semantic query text; For indexing; This represents the i-th document in the file set.

[0054] F4. Then, based on the final recall results of all documents, sort each document in descending order, and select the Top K (the first K, where K is a preset constant) most relevant documents from the descending sorted set to construct a structured knowledge set.

[0055] F5. The pre-trained third language model is invoked, and preset prompts are used to summarize and synthesize the structured knowledge set, resulting in comprehensive industry standards and solution suggestions, which are then used as anomaly detection results. The pre-trained third language model uses Qwen3-32B-Thinking.

[0056] The final, comprehensive industry standards and solutions are obtained using the following formula: in, This indicates the final, comprehensive industry standards and proposed solutions; It is the third largest pre-trained language model; This represents the structured knowledge set that has been constructed.

[0057] Furthermore, the pre-trained third language model outputs not only complete industry standards and solutions, but also the logical chain and basis for reasoning.

[0058] The master control agent continuously executes one of four actions at each time step: image detection, parameter analysis, document retrieval, and stopping. It obtains the corresponding result until the master control agent outputs "stop," at which point the final result from the document retrieval agent is output as the final anomaly detection result, and the system terminates. The master control agent, acting as the system's top-level decision-making and scheduling center, is built on a ReAct+CoT architecture and possesses the ability to autonomously invoke multiple downstream agents.

[0059] The master control agent's processing procedure is as follows: It calls the pre-trained fourth language model and uses preset prompts to process the state set at the current time step, outputting one of four actions: executing the image detection agent, executing the parameter analysis agent, executing the document retrieval agent, or stopping execution. The master control agent executes the corresponding action to obtain the result at the current time step, then updates the state set and uses the updated state set as the state set for the next time step. The master control agent is continuously executed at consecutive time steps until its output action is "stop execution," at which point the system stops running, and the final result generated by the document retrieval agent becomes the final anomaly detection result. The pre-trained fourth language model uses Qwen3-32B-Thinking.

[0060] The initial state set includes the image of the component under test, the set of parameter measurements, and user requirements; the non-initial state set includes the image of the component under test, the set of parameter measurements, user requirements, and all results obtained by the master controller calling the agent to perform various actions in all time steps before the current time step.

[0061] like Figure 2 As shown, this embodiment is implemented according to the following steps: S1, the data acquisition agent acquires images of the parts to be tested, a set of parameter measurement values ​​including dimensions and mass, and a set of files including industry standard specifications and historical case data.

[0062] S2. The master controller calls the agent to call the pre-trained fourth language model and uses preset prompts to process its own state set at the current time step. The output includes one of four actions: executing the image detection agent, executing the parameter analysis agent, executing the document recall agent, and stopping. The master controller calls the agent to execute the corresponding action to obtain the result of the current time step. Then, the master controller calls the agent to update the state set and uses the updated state set as the state set for the next time step.

[0063] S3. Repeat step S2 continuously until the output action is stopped. The final result generated by the document recall agent is used as the final anomaly detection result. This realizes the multi-source joint judgment of image defects, parameter deviations, process specifications and historical similar defects of the parts under test, and completes the final quality inspection of industrial standard parts.

[0064] This invention is the first to apply multi-agent technology to the field of quality inspection of industrial standard parts. It performs joint analysis and diagnosis of multimodal data such as images, text, and parameters, and the agents perform dynamic scheduling and chain reasoning. This breaks through the limitations of traditional inspection systems that rely on a single model or a single input, thus possessing an advanced technical system.

[0065] Compared to existing multimodal large language model detection schemes that forcibly unify multimodal information into a single model, this invention processes multimodal information into standard results through different modules. This eliminates the need to retrain the entire model to add new detection tools or industry standards. The solution is more open and user-friendly, and can be continuously updated and iterated according to actual production needs, thus achieving strong robustness.

[0066] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A multi-agent anomaly detection system for industrial standard parts, characterized in that, include: The data acquisition agent collects images of the parts under test, a set of parameter measurements including dimensions and mass, and a set of files including industry standard specifications and historical parts anomaly detection data. The image detection agent processes images of the parts to be tested to obtain a corresponding list of structured image defects. The parameter analysis agent processes the set of parameter measurement values ​​of the component under test to obtain the corresponding structured parameter results; The document recall agent obtains anomaly detection results based on a structured image defect list, structured parameter results, and a set of documents. The master controller calls the intelligent agent, which continuously executes one of four actions at each time step, including executing the image detection intelligent agent, executing the parameter analysis intelligent agent, executing the document recall intelligent agent, and stopping, and obtains the corresponding result. This continues until the action output by the master controller's intelligent agent is "stop". Then, the final result obtained by the document recall intelligent agent is output as the final anomaly detection result, and the system operation ends.

2. The multi-agent anomaly detection system for industrial standard parts according to claim 1, characterized in that, The execution steps of the image detection agent are as follows: The image of the component under test is first augmented using a Gabor filter, and then the YOLOv8 detection algorithm is used to perform preliminary defect identification on the augmented image, outputting candidate boxes, defect categories and confidence scores to obtain preliminary detection results. Then, the pre-trained first language model is called to analyze and summarize the preliminary detection results of the image to generate a structured image defect list of the component under test.

3. The multi-agent anomaly detection system for industrial standard parts according to claim 1, characterized in that, The execution steps of the parameter analysis agent are as follows: Each parameter measurement value in the parameter measurement set is normalized to obtain the corresponding normalized deviation value. The corresponding Z-score statistical value is calculated based on each normalized deviation value. The parameter health index set of the component under test is obtained based on the Z-score statistical value of all parameter measurements. The pre-trained second language model is called to comprehensively analyze the parameter health index set of the component under test to generate the structured parameter results of the component under test.

4. The multi-agent anomaly detection system for industrial standard parts according to claim 3, characterized in that: The structured parameter results of the component under test are obtained by processing the following formula: in, This represents the structured parameter results of the component under test; This represents the second largest pre-trained language model; This represents the set of health indices of the parameters of the component under test. Indicates an index; Represents the first parameter in the set of measured values. Each parameter measurement value; Indicates the measured value of the parameter The corresponding parameter weights; This indicates the total number of parameter measurements; Indicates the measured value of the parameter The corresponding Z-score statistics; Indicates the measured value of the parameter The corresponding normalized deviation value; Indicates the reference mean; Indicates the reference standard deviation; and This indicates the upper and lower limits of the set standard range; It is a constant used to avoid the denominator being zero.

5. The multi-agent anomaly detection system for industrial standard parts according to claim 1, characterized in that, The execution steps of the document retrieval agent are as follows: F1. A text vector representation model is used to encode the structured image defect list and structured parameter results of the component under test into semantic query text. Then, the text vector representation model maps the semantic query text into embedded vectors. The same text vector representation model is used to encode and map each document in the document collection to a corresponding text vector. F2. Obtain coarse recall results for each document based on the semantic query text, embedding vector, and text vector of each document; F3. Use a cross-encoder to process each document to obtain the consistency relevance of the corresponding document, and then weight the coarse recall result and the consistency relevance to obtain the final recall result of a single document. F4. Then, based on the final recall results of all documents, sort each document in descending order, and select the Top K relevant documents from the descending sorted set to construct a structured knowledge set. F5. The pre-trained third language model is invoked to summarize and synthesize the structured knowledge set, resulting in complete industry standards and solution suggestions, which are then used as anomaly detection results.

6. The multi-agent anomaly detection system for industrial standard parts according to claim 5, characterized in that: The coarse recall result for each document is set according to the following formula: in, For indexing; This represents the i-th document in the file set; Indicates the first Coarse recall results for one document; Indicates the weighting coefficient; and They represent the first BM25 similarity scores and vector similarity values ​​for each document; An algorithm for calculating the cosine similarity between two vectors; Represents an embedding vector; Indicates the first The text vector of a document; This represents the semantic query text.

7. A multi-agent anomaly detection system for industrial standard parts according to claim 5, characterized in that: The final recall result for each document is set according to the following formula: in, Indicates the first The final recall result for each document; Indicates the weighting coefficient; Indicates the first Consistency relevance of individual documents; Indicates the first Coarse recall results for one document; This refers to the Cross-Encoder module; Represents semantic query text; For indexing; This represents the i-th document in the file set.

8. A multi-agent anomaly detection system for industrial standard parts according to claim 1, characterized in that, The execution steps of the intelligent agent invoked by the master controller are as follows: The pre-trained fourth language model is invoked to process the state set of the current time step and output one of four actions: executing an image detection agent, executing a parameter analysis agent, executing a document retrieval agent, or stopping execution. The master controller invokes the agent to execute the corresponding action to obtain the result of the current time step. Then, the master controller invokes the agent to update the state set and uses the updated state set as the state set for the next time step. The master controller invokes the agent continuously at consecutive time steps until the action output by the master controller invokes the agent is to stop execution. At this point, the system stops running, and the final result generated by the document retrieval agent is the final anomaly detection result output.

9. A multi-agent anomaly detection system for industrial standard parts according to claim 8, characterized in that: The initial state set includes the image of the component under test, the set of parameter measurements, and user requirements; the non-initial state set includes the image of the component under test, the set of parameter measurements, user requirements, and all results obtained by the master controller calling the agent to perform various actions in all time steps before the current time step.

10. A multi-agent anomaly detection method for industrial standard parts using the system described in any one of claims 1-9, characterized in that, Includes the following steps: S1. The data acquisition agent acquires images of the parts to be tested, a set of parameter measurement values ​​including size and mass, and a set of files including industry standard specifications and historical case data. S2. The master controller calls the agent to call the pre-trained fourth language model to process its own state set at the current time step and output one of four actions, including executing the image detection agent, executing the parameter analysis agent, executing the document recall agent, and stopping. The master controller calls the agent to execute the corresponding action to obtain the result of the current time step. Then the master controller calls the agent to update the state set and uses the updated state set as the state set for the next time step. S3. Repeat step S2 continuously until the output action is "stop". The final result generated by the document recall agent is then used as the final anomaly detection result.

Images