High-precision instrument assembly fault backtracking method and system

Abstract

This invention discloses a method and system for backtracking faults in high-precision instrument assembly, belonging to the field of precision manufacturing. Addressing the problems of fragmented and unreliable assembly data, ambiguous fault root cause localization, and reliance on large amounts of data for adaptation to new scenarios in traditional backtracking methods, the method includes: collecting multi-source structured assembly data, video images, and environmental data; filtering out low-quality images and assigning unique identifiers to products; fusing multimodal data to generate a deep feature matrix; constructing an anomaly detection model to output risk scores and labels; hashing the data and storing it in a product-specific private blockchain; extracting on-chain data using the unique identifier during a fault, reconstructing the assembly process using a graph neural network, and comparing it with standards to locate the root cause; incrementally training the model based on fault reports, and optimizing parameters using meta-reinforcement learning. This invention ensures data authenticity, improves the accuracy and efficiency of fault backtracking, quickly adapts to new scenarios, reduces production rework rates, and maintains stable assembly quality.

Description

Technical Field

[0001] This invention relates to the field of precision manufacturing, and more specifically, to a method and system for tracing back faults in high-precision instrument assembly. Background Technology

[0002] High-precision instruments, as key equipment in industrial manufacturing and scientific research, directly determine the performance, accuracy, and reliability of the final product through their assembly quality. These instruments are typically characterized by complex structures, stringent tolerance requirements, and high functional integration. They are widely used in aerospace, precision measurement, medical equipment, and high-end scientific research instruments. Their manufacturing process involves the orderly assembly of hundreds or even thousands of precision components, adhering to strict technological processes, environmental control standards, and multi-level quality inspection mechanisms. In actual production, any minute assembly deviation, such as improper torque control, insufficient part cleanliness, misalignment of positioning references, or incorrect operating sequence, can lead to serious malfunctions such as performance drift, signal distortion, or even functional failure during instrument operation. Due to the high value and critical application scenarios of high-precision instruments, malfunctions can result in not only costly rework and delivery delays but also significant impacts on the user's production safety and research progress. Therefore, achieving refined control over the entire assembly process, early identification of potential defects, and accurate backtracking of past failures has become a requirement for improving the manufacturing quality and reliability of high-precision instruments.

[0003] Currently, in the field of precision manufacturing, the monitoring and fault analysis of the assembly process mainly rely on traditional methods. The commonly used approach is a post-incident troubleshooting mechanism based on manual experience. When an instrument malfunctions during factory testing or field use, senior engineers use the fault symptoms, test data, and paper or electronic assembly records, combined with their personal experience, to reverse-engineer the problem and attempt to pinpoint the root cause. However, this method has significant efficiency bottlenecks and reliability defects: First, the assembly process involves numerous operational details, and manual recording alone cannot comprehensively cover all key actions and environmental parameters; second, the causes of faults are often hidden and multifactorial, possibly the result of the accumulation of minor deviations in multiple processes, making systematic analysis difficult based solely on individual experience; third, with technological iterations and personnel turnover, there is a risk of knowledge gaps in transmission, leading to the recurrence of similar problems. Another technological approach is to utilize Manufacturing Execution Systems (MES) for the collection and management of production data. These systems can record structured data such as work order information, material batches, and process parameters, and to a certain extent, achieve process traceability. However, existing MES systems still have certain limitations in the breadth and depth of data collection: their data sources are mainly limited to equipment interfaces and manual input, making it difficult to effectively capture a large amount of key unstructured information during the assembly process, such as the specific action trajectory of operators, the order and method of tool use, the temporary placement of parts, environmental temperature and humidity fluctuations, and visual confirmation processes. Although this information is difficult to quantify, it is often an important cause of assembly deviations, resulting in only fragmented and superficial inferences when tracing faults, making it difficult to accurately locate the root cause and make preventive improvements. Summary of the Invention

[0004] To address the problems existing in the prior art, the present invention aims to provide a high-precision instrument assembly fault backtracking method and system, which can accurately identify assembly deviation points and locate the root cause of the fault when a fault occurs, effectively solving the problems of low data reliability and ambiguous root cause location in traditional backtracking methods.

[0005] To solve the above problems, the present invention adopts the following technical solution.

[0006] Firstly, a method for tracing back faults in the assembly of high-precision instruments includes the following steps:

[0007] Step 1: Collect structured data, video images and environmental data of the assembly process in real time through multi-source data acquisition devices, assign a unique identifier and process label to each instrument product, update the dynamic information label of the components using edge computing, filter low-quality images using image quality detection algorithms, and output a multimodal dataset with identifiers.

[0008] Step 2: The multimodal dataset is fused to extract the spatiotemporal and textual features of the video, which are then combined with the structured data to form a unified feature vector. The environmental data is then standardized, and the data correlation is enhanced through a cross-modal attention mechanism to generate a deep feature matrix.

[0009] Step 3: Construct an anomaly detection model, train neural networks and time series networks using historical normal data, calculate warning coefficients and use adaptive thresholds to identify anomalies, use graph neural networks to model process relationships to identify operational deviations, and combine interpretable AI output risk scores and binary labels.

[0010] Step 4: Process the abnormal results and the original data hash, create a private blockchain for each product, with blocks containing data hashes and timestamps, verify consistency through smart contracts and achieve tamper-proof storage;

[0011] Step 5: When a fault occurs, extract blockchain data based on the unique identifier, remodel the assembly process using a graph neural network, identify deviations by comparing with the standard process, locate the root cause of the fault by combining risk scoring and video playback, and generate an analysis report.

[0012] Step 6: Use the fault reports as labeled data to retrain the anomaly detection model, optimize the process parameters using multi-objective reinforcement learning and meta-reinforcement learning, and dynamically update the model and thresholds based on real-time performance data.

[0013] Furthermore, step 3 includes the following steps:

[0014] Step S31: Collect historical normal assembly data of the instrument, clean and label it to form a model training dataset, and output the historical normal assembly training dataset.

[0015] Step S32: Construct an anomaly detection model architecture, including a neural network module for learning normal assembly patterns and a time series network module for processing assembly time series data, and output the model architecture.

[0016] Step S33: Input the historical normal assembly training dataset into the model architecture, iteratively train and optimize the parameters to master the normal features, and output the trained anomaly detection model.

[0017] Step S34: Input the deep feature matrix from step 2 into the trained model. The reconstruction error and prediction error are calculated by the two modules, and the assembly warning coefficient is obtained by weighted summation. The warning coefficient is then output.

[0018] Step S35: Analyze the warning coefficients using an adaptive threshold algorithm, determine the anomaly judgment thresholds according to the coefficient distribution, and output the dynamic anomaly thresholds.

[0019] Step S36: Call the graph neural network to construct a process relationship model based on the process data of the deep feature matrix, identify operational deviations, and output the deviation identification results;

[0020] Step S37: Introduce interpretable AI technology to analyze the characteristic contribution of the early warning coefficient and the deviation results, clarify the impact, and output a risk score;

[0021] Step S38: Compare the warning coefficient with the dynamic threshold. If the threshold is met, output an abnormal label; otherwise, output a normal label. Associate the risk score to form an abnormal detection result.

[0022] Furthermore, step S36 also includes the following steps:

[0023] Step S361: Based on the deep feature matrix, process attributes and assembly features are fused to generate multi-dimensional node feature vectors;

[0024] Step S362: Based on node features, analyze the dependencies between processes and dynamically construct a process relationship graph model;

[0025] Step S363: A hierarchical identification mechanism is adopted to identify node feature deviations and dependency deviations respectively, and the anomaly degree is calculated in a comprehensive manner to locate the deviation type and location.

[0026] Furthermore, the implementation of the immutable storage also includes the following steps:

[0027] Step S41: Extract the anomaly detection results from step 3 and the multi-source raw data from step 1, unify the data format and remove redundant fields to form a dataset to be hashed, and output the dataset.

[0028] Step S42: Use a hash algorithm to calculate the dataset to be hashed, generate a unique hash value, synchronously record the data collection timestamp and device information, and output the hash value and associated timestamp and device information;

[0029] Step S43: Based on the unique identifier of the instrument product in Step 1, create a private blockchain for each product, define the block structure, and output the private blockchain framework.

[0030] Step S44: According to the block structure of step S43, encapsulate the hash value, timestamp, device information and assembly data of step S42 into an initial block and output the block.

[0031] Step S45: Write a smart contract containing data consistency verification rules, deploy it to the private blockchain of step S43, and output the blockchain of the deployed contract.

[0032] Step S46: Upload the initial block to the blockchain where the contract is deployed. The contract automatically verifies the data consistency. Once the verification is successful, the data is uploaded to the blockchain, achieving immutable data storage and outputting the updated private blockchain.

[0033] Furthermore, the generation of the fault root cause location and analysis report also includes the following steps:

[0034] Step S51: When the instrument fails later, the backtracking process is triggered. The system receives the basic information of the faulty instrument and outputs the fault backtracking start command.

[0035] Step S52: Based on the startup command and basic information, retrieve the unique identifier assigned to the instrument in step 1 and output the unique identifier of the faulty instrument.

[0036] Step S53: Input the identifier into the private blockchain of step 4, extract the full process assembly data through the retrieval module, and output the full process assembly dataset.

[0037] Step S54: Call the graph neural network to reconstruct the assembly process model based on the dataset and output the reconstructed model;

[0038] Step S55: Retrieve the standard assembly process, compare and reconstruct the model step by step, identify assembly deviation points, and output a list of deviation points.

[0039] Step S56: Combining the risk score from step 3 with the video playback of the dataset, analyze the deviation points to locate the root cause of the fault and output the root cause analysis results.

[0040] Step S57: Based on the root cause analysis results, compile a structured report and visualization charts containing the faulty process, cause, and improvement suggestions, and output the report and charts.

[0041] Furthermore, the model optimization and closed-loop system construction also include the following steps:

[0042] Step S61: Using the root cause analysis report of the failure in step 5 as labeled data, extract key information and classify and label it to form a labeled training dataset, and output the dataset.

[0043] Step S62: Input the labeled training dataset into the anomaly detection model of step 3, and update the parameters through incremental training to improve the fault feature recognition capability, and output the updated anomaly detection model.

[0044] Step S63: Determine the multi-objective optimization parameters based on the actual assembly requirements, clarify the optimization weight of each parameter, and output the multi-objective optimization parameters and weight configuration.

[0045] Step S64: A multi-objective reinforcement learning algorithm is used, with the parameters and weights of step S63 as the objectives, to optimize the assembly process parameters in combination with real-time production line data, and output the optimized process parameters.

[0046] Step S65: Introduce meta-reinforcement learning technology, collect a small amount of assembly data for the new scene, learn the assembly features of the new scene, and output model adjustment parameters adapted to the new scene.

[0047] Step S66: Based on the real-time performance data of the production line, combined with the updated model in step S62 and the adaptation parameters in step S65, dynamically adjust the anomaly judgment threshold and model parameters in step 3, and output the adjusted threshold and parameters.

[0048] Step S67: Apply the adjusted threshold and parameters to the production line to verify the optimization effect. If the preset target is not met, return to step S61 for iteration until the target is met.

[0049] Furthermore, by introducing meta-reinforcement learning techniques, collecting a small amount of assembly data for new scenarios, learning the assembly features of the new scenarios, and outputting model adjustment parameters adapted to the new scenarios, the process also includes the following steps:

[0050] Step S651: Preprocess the small amount of assembly data in the new scenario and anchor reusable general features through feature comparison.

[0051] Step S652: Based on the preprocessed data, the pre-trained meta-model is used to quickly optimize the decision-making logic specific to the new scenario through small sample iterative learning.

[0052] Step S653: Verify the initial adaptation parameters through trial assembly, fine-tune them based on performance feedback, and output the final model adjustment parameters.

[0053] Furthermore, the acquisition of the multimodal dataset also includes the following steps:

[0054] Step S11: Select high-definition cameras, IoT sensors and assembly equipment controllers according to the assembly process requirements, clarify the collection categories of structured data, video images and environmental data, and output the equipment list and data collection category table.

[0055] Step S12: Collect multi-source raw data in real time according to the aforementioned devices and categories, temporarily store it in the temporary storage module of the edge computing device, and output the multi-source raw dataset.

[0056] Step S13: Assign a unique identifier to each instrument product, apply a unique label according to the assembly process and associate it with the original data, and output multi-source data with preliminary identification.

[0057] Step S14: The edge computing device reads the component information with the preliminary identification data, generates dynamic tags, receives component status updates in real time and synchronizes tag content, and outputs multi-source data with real-time updated tags.

[0058] Step S15: Call a lightweight image quality detection algorithm to filter out low-quality image frames that are blurry, occluded, or have abnormal lighting, retain high-quality images, and output multi-source data with low-quality images removed.

[0059] Step S16: Integrate the multi-source data with low-quality images filtered out from step S15, the unique identifiers and exclusive labels output from step S13, and the dynamic labels in the multi-source data with real-time updated labels output from step S14 to form a multimodal dataset with unique identifiers.

[0060] Furthermore, the generation of the deep feature matrix also includes the following steps:

[0061] Step S21: Receive the multimodal dataset with a unique identifier, preprocess it, and output the preprocessed multimodal dataset;

[0062] Step S22: Call the pre-trained spatiotemporal feature extraction model to extract various spatiotemporal features from the preprocessed video image and output the video spatiotemporal feature vector;

[0063] Step S23: Using natural language processing technology, the key text tags in the encoded data are encoded and converted into standardized text feature vectors, and the text feature vectors are output.

[0064] Step S24: The video spatiotemporal feature vector, text feature vector and preprocessed structured data are fused together and concatenated into an initial unified feature vector according to the data association rules, and the vector is output.

[0065] Step S25: Extract environmental data, standardize and normalize the initial unified feature vector to eliminate temperature and humidity interference, and output the environmentally corrected unified feature vector.

[0066] Step S26 introduces a cross-modal attention mechanism to enhance the correlation of corrected vectors, strengthen the intrinsic correlation of multimodal data, and generate a deep feature matrix.

[0067] Secondly, a high-precision instrument assembly fault tracing system, applied to the aforementioned high-precision instrument assembly fault tracing method, includes:

[0068] The tag acquisition module is configured to acquire structured data, video images, and environmental data of the assembly process in real time through multi-source data acquisition devices, assign a unique identifier and process tag to each instrument product, update the dynamic information tags of components using edge computing, filter low-quality images using image quality detection algorithms, and output a multimodal dataset with identifiers.

[0069] The feature fusion module is used to fuse multimodal datasets, extract spatiotemporal and textual features from videos, combine them with structured data to form a unified feature vector, standardize environmental data, enhance data correlation through a cross-modal attention mechanism, and generate a deep feature matrix.

[0070] The anomaly detection module is used to build an anomaly detection model. It trains neural networks and time series networks using historical normal data, calculates warning coefficients and uses adaptive thresholds to determine anomalies, uses graph neural networks to model process relationships to identify operational deviations, and combines interpretable AI to output risk scores and binary labels.

[0071] The block storage module is used to process abnormal results and raw data hashes, creating a private blockchain for each product. The blocks contain data hashes and timestamps, and the consistency is verified through smart contracts to achieve tamper-proof storage.

[0072] The fault backtracking module is configured to extract blockchain data based on a unique identifier when a fault occurs, remodel the assembly process using a graph neural network, identify deviations by comparing with the standard process, locate the root cause of the fault by combining risk scoring and video playback, and generate an analysis report.

[0073] The adaptive optimization module is used to retrain the anomaly detection model using fault reports as labeled data. It optimizes process parameters using multi-objective reinforcement learning and meta-reinforcement learning, and dynamically updates the model and thresholds based on real-time performance data.

[0074] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0075] (1) This solution combines multi-source data acquisition with blockchain immutable storage technology, which not only ensures the authenticity, integrity and traceability of the data throughout the assembly process, but also relies on technologies such as multimodal data fusion and graph neural network reconstruction of the assembly process to accurately identify assembly deviation points and locate the root cause of the fault when a fault occurs. This effectively solves the problems of low data credibility and ambiguous root cause location in traditional backtracking methods, and greatly improves the efficiency and accuracy of fault backtracking.

[0076] (2) This solution constructs a closed-loop system from anomaly detection to fault backtracking and then to model optimization. On the one hand, it improves the accuracy of anomaly detection through adaptive thresholds and interpretable AI. On the other hand, it introduces meta-reinforcement learning technology, which can quickly adapt to changes such as new products and new processes with only a small amount of new scenario data, avoiding the dependence of traditional models on a large amount of labeled data. At the same time, it continuously optimizes assembly process parameters by incrementally training the model through fault reports, maintains the stability of high-precision instrument assembly quality in the long term, and reduces production rework rate and cost. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0078] Figure 1 This is a flowchart illustrating the method execution of the present invention;

[0079] Figure 2 This is a flowchart illustrating the interaction between the modules of this invention. Detailed Implementation

[0080] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0081] Please see Figures 1 to 2 A method for tracing back faults in high-precision instrument assembly, comprising:

[0082] Step 1: Collect structured data, video images and environmental data of the assembly process in real time through multi-source data acquisition devices, assign a unique identifier and process label to each instrument product, update the dynamic information label of the components using edge computing, filter low-quality images using image quality detection algorithms, and output a multimodal dataset with identifiers.

[0083] Step 2: The multimodal dataset is fused to extract the spatiotemporal and textual features of the video, which are then combined with the structured data to form a unified feature vector. The environmental data is then standardized, and the data correlation is enhanced through a cross-modal attention mechanism to generate a deep feature matrix.

[0084] Step 3: Construct an anomaly detection model, train neural networks and time series networks using historical normal data, calculate warning coefficients and use adaptive thresholds to identify anomalies, use graph neural networks to model process relationships to identify operational deviations, and combine interpretable AI output risk scores and binary labels.

[0085] Step 4: Process the abnormal results and the original data hash, create a private blockchain for each product, with blocks containing data hashes and timestamps, verify consistency through smart contracts and achieve tamper-proof storage;

[0086] Step 5: When a fault occurs, extract blockchain data based on the unique identifier, remodel the assembly process using a graph neural network, identify deviations by comparing with the standard process, locate the root cause of the fault by combining risk scoring and video playback, and generate an analysis report.

[0087] Step 6: Use the fault reports as labeled data to retrain the anomaly detection model, optimize the process parameters using multi-objective reinforcement learning and meta-reinforcement learning, and dynamically update the model and thresholds based on real-time performance data.

[0088] Step 1 includes the following steps:

[0089] Step S11: Select high-definition cameras, IoT sensors and assembly equipment controllers according to the assembly process requirements, clarify the collection categories of structured data, video images and environmental data, and output the equipment list and data collection category table.

[0090] Step S12: Collect multi-source raw data in real time according to the aforementioned devices and categories, temporarily store it in the temporary storage module of the edge computing device, and output the multi-source raw dataset.

[0091] Step S13: Assign a unique identifier to each instrument product, apply a unique label according to the assembly process and associate it with the original data, and output multi-source data with preliminary identification.

[0092] Step S14: The edge computing device reads the component information with the preliminary identification data, generates dynamic tags, receives component status updates in real time and synchronizes tag content, and outputs multi-source data with real-time updated tags.

[0093] Step S15: Call a lightweight image quality detection algorithm to filter out low-quality image frames that are blurry, occluded, or have abnormal lighting, retain high-quality images, and output multi-source data with low-quality images removed.

[0094] Step S16: Integrate the multi-source data with low-quality images filtered out from step S15, the unique identifiers and exclusive labels output from step S13, and the dynamic labels in the multi-source data with real-time updated labels output from step S14 to form a multimodal dataset with unique identifiers.

[0095] In this embodiment, firstly, a data acquisition system needs to be built based on the specific requirements of the high-precision instrument assembly process. Considering the need to acquire multi-dimensional information such as visual, physical parameters, and environmental data during the assembly process, suitable acquisition devices will be selected first: high-definition cameras are used to capture assembly action details and component status, such as component installation positions and connection gaps; IoT sensors are responsible for collecting environmental data, such as temperature, humidity, cleanliness, and vibration frequency of the assembly area; and assembly equipment controllers synchronously record structured data, such as assembly torque, equipment operating parameters, and process execution time. Simultaneously, the acquisition categories for various types of data need to be clearly defined. For example, structured data must include process number, equipment ID, operating parameter thresholds, and actual operating values; video image data must specify a frame rate ≥30fps, resolution ≥1080P, and shooting angle covering key areas of the assembly station; and environmental data must specify a temperature and humidity acquisition interval ≤1 minute and a cleanliness detection accuracy of Class 1000. The final output includes an equipment list containing equipment model, quantity, and installation location, as well as a collection category table with corresponding data types, acquisition frequencies, and accuracy requirements.

[0096] Based on the identified equipment and data acquisition categories, the system initiates a real-time data acquisition process. High-definition cameras continuously capture the assembly process at a set frame rate, generating a continuous video image stream. IoT sensors convert real-time monitored environmental parameters into electrical signals, which are then converted into digital signals by a data conversion module before transmission. The assembly equipment controller extracts equipment operating status and operation command data in real time, forming a structured data sequence. Because high-precision instrument assembly requires high real-time data processing, all multi-source raw data is temporarily stored in the temporary storage module of an edge computing device. The edge computing device's proximity to the data acquisition source significantly reduces data transmission latency and avoids bandwidth consumption caused by directly uploading large amounts of raw data to the cloud. This ensures that the data acquisition process does not affect the normal progress of the assembly process, ultimately forming a multi-source raw dataset containing video images, structured data, and environmental data.

[0097] To achieve precise data binding to specific instrument products, a unique identifier must be assigned to each instrument product entering the assembly process. This identifier will serve as a data identity card throughout the entire assembly and traceability process, ensuring that the data of each product is independently traceable. Simultaneously, based on the assembly process steps, such as housing assembly, sensor calibration, and circuit soldering, a unique step label is assigned to the raw data generated by the corresponding step. A data association algorithm then binds the unique identifier, step label, and raw data, ensuring that each piece of data clearly corresponds to a specific assembly step of a particular instrument product. The final output is multi-source data with preliminary identification.

[0098] Edge computing devices further dynamically process the multi-source data with initial labels: First, they read the component information contained in the data, such as component model, installation status, and test results. Using preset label generation rules, such as component model, installation status, and test time, they generate dynamic labels for the components. For example, a pressure sensor Model X-100 might be labeled "Installed, Initial Inspection Passed, 2025XXXX". During subsequent assembly, if the component status changes, such as a deviation detected during secondary testing or an adjustment in installation position, the edge computing device receives real-time status update signals from the testing equipment or operating terminal and synchronously modifies the corresponding dynamic label content. This ensures the labels accurately reflect the real-time status of the components, avoiding information discrepancies caused by data lag. Ultimately, it outputs multi-source data with real-time updated labels, guaranteeing data timeliness.

[0099] Considering that low-quality frames in video image data may be generated due to factors such as shooting angle deviation, lighting changes, and component occlusion, these images can interfere with the accuracy of subsequent spatiotemporal feature extraction. Therefore, a lightweight image quality detection algorithm is needed for filtering. This algorithm analyzes indicators such as image sharpness, occlusion rate, and illumination uniformity. Sharpness can be calculated based on edge gradient values, occlusion rate can be based on pixel region segmentation statistics, and illumination uniformity can be based on grayscale value distribution analysis. Then, corresponding thresholds are set, such as sharpness below 80 points, occlusion rate exceeding 30%, and illumination uniformity deviating from the normal range by ±20%, automatically filtering out low-quality image frames that do not meet the requirements, retaining only image data that meets the quality standards. A lightweight algorithm is chosen to adapt to the computing power requirements of edge computing devices, ensuring filtering effectiveness while avoiding excessive consumption of computing resources. The final output is multi-source data with low-quality images filtered out, ensuring the validity of the image data.

[0100] Finally, the high-quality image data, structured data, environmental data, unique identifiers, process tags, and component dynamic information tags obtained from the aforementioned processing are integrated. During the integration process, data association algorithms are used to map and bind various types of data corresponding to the same identifier, ensuring the consistency of multiple data types in the temporal and spatial dimensions. For example, video images of a certain process are accurately matched with environmental data and equipment operation data for the corresponding time period. This ultimately forms a multimodal dataset with unique identifiers that contains multimodal data and complete identification information. This dataset not only covers all dimensions of the assembly process but also achieves data traceability through the identification system.

[0101] In some embodiments, step S2 includes the following steps:

[0102] Step S21: Receive the multimodal dataset with a unique identifier, preprocess it, and output the preprocessed multimodal dataset;

[0103] Step S22: Call the pre-trained spatiotemporal feature extraction model to extract various spatiotemporal features from the preprocessed video image and output the video spatiotemporal feature vector;

[0104] Step S23: Using natural language processing technology, the key text tags in the encoded data are encoded and converted into standardized text feature vectors, and the text feature vectors are output.

[0105] Step S24: The video spatiotemporal feature vector, text feature vector and preprocessed structured data are fused together and concatenated into an initial unified feature vector according to the data association rules, and the vector is output.

[0106] Step S25: Extract environmental data, standardize and normalize the initial unified feature vector to eliminate temperature and humidity interference, and output the environmentally corrected unified feature vector.

[0107] Step S26 introduces a cross-modal attention mechanism to enhance the correlation of corrected vectors, strengthen the intrinsic correlation of multimodal data, and generate a deep feature matrix.

[0108] In this embodiment, a multimodal dataset with a unique identifier is received. This dataset includes video images of the instrument assembly process, such as workstation operation screens and component status images, as well as structured data, such as assembly torque, equipment operating parameters, and process execution time, and text tags, such as component dynamic information tags (e.g., pressure sensor Model X-100 has been installed and passed initial inspection), process tags (e.g., housing assembly, process 2), and environmental data, such as temperature, humidity, and cleanliness of the assembly area. All data is bound to a specific instrument product through a unique identifier. To eliminate data noise and standardize the format, preprocessing is required: For video images, timestamp alignment technology is used to match consecutive frames with corresponding process time periods, removing duplicate or broken frames caused by equipment lag, and grayscale normalization is performed to reduce feature interference caused by lighting differences; for structured data, mean imputation is used to fill in a small number of missing values. The principle is to identify and remove outliers that exceed the normal range, such as assembly torque values ​​far exceeding the equipment's range. Then, all parameters are converted to a unified unit, such as converting process durations from seconds to minutes. For text labels, special characters are cleaned using regular expressions, and non-standardized expressions, such as "sensor calibration is complete," are standardized to a preset format, such as "sensor calibration is qualified." Environmental data requires verification of the consistency between the acquisition time and the process time, and data with misaligned times is deleted. After the above processing, a preprocessed multimodal dataset with a unified format, noise removal, and logical correlation is output.

[0109] Based on the preprocessed multimodal dataset, the system calls pre-trained spatiotemporal feature extraction models, such as the I3D model and SlowFast model adapted for industrial scenarios, to extract spatiotemporal features from video images. These pre-trained models have learned spatiotemporal feature extraction capabilities on a large amount of general video data. For instrument assembly scenarios, in the spatial dimension, the model captures detailed features of assembly actions through convolutional layers, such as the relative positions of screwdrivers and screws, the alignment accuracy of component installation, and the trajectory contours of the operator's hand movements. In the temporal dimension, the model analyzes the logical relationships of action sequences through temporal convolution or recurrent structures, such as the sequence of actions from picking up components to positioning, tightening screws, and finally detection, as well as the logical order of tool usage, such as soldering circuits with a soldering iron first, and then using a multimeter to check continuity. During the extraction process, the model uses process labels as constraints to ensure that each video segment corresponds to only a single process's spatiotemporal features, avoiding cross-process information confusion. Finally, the extracted high-dimensional spatiotemporal information is compressed into a fixed-dimensional vector, outputting a video spatiotemporal feature vector that accurately represents the interaction patterns between assembly actions and tools.

[0110] For textual information in the data, such as component dynamic information tags and process tags, the system uses natural language processing (NLP) technology for encoding and conversion. Considering the semantic uniformity and standardization of industrial texts, lightweight and efficient word embedding models, such as Word2Vec and pre-trained BERT mini-models for industrial applications, are prioritized. First, the text is segmented, breaking down long labels into semantically independent lexical units. For example, "Pressure sensor Model X-100 installed and initially qualified" is broken down into "pressure sensor," "Model X-100," "installed," and "initially qualified." Then, each word is converted into a low-dimensional dense vector using the word embedding model. This vector reflects the semantic relationships between words; for example, the vectors for "installed" and "not installed" are far apart, while the vectors for "initially qualified" and "re-qualified" are close. Finally, all lexical vectors for the same text label are averaged or attention-weighted to generate standardized text feature vectors with uniform dimensions. For example, process labels, circuit welding, and process 5 are encoded as 128-dimensional vectors; component dynamic labels, resistor R101, welding completed, and no false welds are encoded as 128-dimensional vectors, ensuring that text information can participate in subsequent feature fusion in numerical form, outputting text feature vectors.

[0111] After completing the single-modal feature extraction, the video spatiotemporal feature vector, text feature vector, and preprocessed structured data need to be fused into an initial unified feature vector. Unique identifiers and timestamps serve as dual constraints to ensure that the fused data comprises three types of data from the same instrument product and the same assembly process time period. First, the structured data, such as the assembly torque value of 5 N·m, the process execution time of 2.5 min, and the equipment speed of 1500 r / min, is converted into a one-dimensional numerical vector, such as [5, 2.5, 1500]. Then, the vectors are concatenated in the order of video spatiotemporal feature vector to text feature vector, and then to structured data vector. If the video spatiotemporal feature vector is 512-dimensional, the text feature vector is 128-dimensional, and the structured data vector is 8-dimensional, then the concatenated vector forms an initial unified feature vector with dimensions of 512 + 128 + 8 = 648. This splicing method can not only completely preserve the original information of each modality data, but also ensure the spatiotemporal consistency of data within the vector through association rules. For example, the welding action features, welding text labels, and welding torque data of a certain process are accurately matched to avoid feature misalignment and finally output an initial unified feature vector.

[0112] Because high-precision instrument assembly is sensitive to the environment—for example, excessively high temperatures may cause plastic components to deform, and excessive humidity may affect circuit insulation—environmental data can interfere with the effectiveness of the initial unified feature vector. Therefore, it is necessary to extract environmental data for the corresponding process period, such as the average temperature T and average humidity H during the process, and then standardize and normalize the initial vector. The standardization method uses Z-score standardization, and the formula is as follows: ;

[0113] in These are the standardized eigenvalues; These are the original feature values ​​in the initial unified feature vector, such as the torque value of a certain process. This is the mean of this feature in historical normal assembly data; The standard deviation of this feature in historical normal assembly data is used. By converting the original feature values ​​to a standard normal distribution with a mean of 0 and a standard deviation of 1, the dimensional differences between different features are eliminated, such as the different units of torque and time, while also reducing the impact of extreme environmental data on the features. After standardization, Min-Max normalization is used to map the feature values ​​to the [0,1] interval to adapt to the input requirements of subsequent models. The final output is a unified feature vector after environmental correction, which eliminates environmental interference such as temperature and humidity and allows for direct comparison of features from different processes.

[0114] To strengthen the intrinsic correlation between features of different modalities, such as the correlation between calibration actions in videos and calibration procedures in text, and the correlation between calibration accuracy values ​​in structured data, the system introduces a cross-modal attention mechanism to optimize the unified feature vector after environmental correction. First, the corrected vector is split into three modal branches: video sub-vectors, text sub-vectors, and structured data sub-vectors. Then, the attention weights of features within each modal branch are calculated. For example, in the sensor calibration procedure, features such as the contact action between the calibration tool and the sensor in the video sub-vector, calibration keywords in the text sub-vector, and calibration error values ​​in the structured sub-vector receive higher weights, while irrelevant features, such as the operator's irrelevant actions in the video, receive lower weights. Simultaneously, the correlation weights between different modal branches are calculated to strengthen the synergistic relationship between actions, text, and parameters. Finally, the three sub-vectors are weighted and fused to generate a deep feature matrix with the same dimensions as the corrected vector but significantly enhanced feature correlation. Each row of the matrix corresponds to a multimodal feature at a time step, and each column corresponds to the feature dimension of a modality. This preserves the temporal information of the assembly process while highlighting key correlated features through the attention mechanism.

[0115] In some embodiments, step S3 includes the following steps:

[0116] Step S31: Collect historical normal assembly data of the instrument, clean and label it to form a model training dataset, and output the historical normal assembly training dataset.

[0117] Step S32: Construct an anomaly detection model architecture, including a neural network module for learning normal assembly patterns and a time series network module for processing assembly time series data, and output the model architecture.

[0118] Step S33: Input the historical normal assembly training dataset into the model architecture, iteratively train and optimize the parameters to master the normal features, and output the trained anomaly detection model.

[0119] Step S34: Input the deep feature matrix from step 2 into the trained model. The reconstruction error and prediction error are calculated by the two modules, and the assembly warning coefficient is obtained by weighted summation. The warning coefficient is then output.

[0120] Step S35: Analyze the warning coefficients using an adaptive threshold algorithm, determine the anomaly judgment thresholds according to the coefficient distribution, and output the dynamic anomaly thresholds.

[0121] Step S36: Call the graph neural network to construct a process relationship model based on the process data of the deep feature matrix, identify operational deviations, and output the deviation identification results;

[0122] Step S37: Introduce interpretable AI technology to analyze the characteristic contribution of the early warning coefficient and the deviation results, clarify the impact, and output a risk score;

[0123] Step S38: Compare the warning coefficient with the dynamic threshold. If the threshold is met, output an abnormal label; otherwise, output a normal label. Associate the risk score to form an abnormal detection result.

[0124] In this embodiment, comprehensive data from inspected and qualified products during past high-precision instrument assembly processes are collected. This includes multimodal raw data, such as video images of normal assembly, equipment operating parameters, and environmental data, as well as processed data, such as structured features and text tags. Because historical data may contain noise, missing values, or formatting issues, it needs to be cleaned first: for numerical data, such as assembly torque and process duration, a mean-filling method is used to fill in any missing values. The principles are as follows: Extreme values ​​exceeding the normal range are removed, such as abnormal torque caused by temporary equipment malfunctions; for video image data, blurry or obscured invalid frames are deleted, retaining only clear, normal assembly images; for text labels, a standardized format is used, such as unifying "sensor calibrated" and "sensor calibration completed" as "sensor calibration qualified." The cleaned data must be labeled with a "normal" tag, clearly specifying the assembly process, product model, and qualification status for each data point. This ultimately forms a historical normal assembly training dataset covering different instrument models and assembly processes, ensuring the model can learn diverse normal assembly characteristics.

[0125] Based on the characteristic attributes of historical normal assembly data, an adapted anomaly detection model architecture needs to be constructed. Among these, the neural network module for learning normal assembly patterns preferentially adopts an autoencoder architecture. The autoencoder compresses the high-dimensional multimodal features of the input into low-dimensional latent vectors through an encoder, and then reconstructs the original features through a decoder. Its advantage is that it can learn the inherent patterns of data in unsupervised scenarios, making it particularly suitable for anomaly detection tasks that rely solely on normal data for training. It can effectively capture the normal spatial features of component states and operational actions during the assembly process, such as the standard angle of screw tightening and the alignment accuracy of component installation. The time-series network module for processing assembly time-series data adopts a Long Short-Term Memory (LSTM) network or a Temporal Convolutional Network (TCN). The assembly process has obvious time dependencies, such as the fixed time sequence from component retrieval to positioning, installation, and finally detection. LSTM / TCN can capture the long-term dependencies of time-series data through gating mechanisms or convolutional kernels, avoiding the gradient vanishing problem of conventional recurrent neural networks. It can accurately learn the normal patterns of process execution order and operational parameter changes over time, such as the standard execution time of a certain process and the torque change curve over time. The two modules are integrated in parallel within the model architecture, covering both the assembly details in the spatial dimension and the process dependency features in the temporal dimension, ultimately outputting a structurally complete and functionally adaptable anomaly detection model architecture.

[0126] During the model training phase, the historical normal assembly training dataset needs to be proportionally divided into training and validation sets, which are then input into the constructed model architecture. The training process aims to minimize reconstruction and prediction errors: For the autoencoder module, after the training data is compressed by the encoder and reconstructed by the decoder, the mean square error between the reconstructed output and the original input is calculated. The Adam optimizer iteratively adjusts the network weights to gradually reduce the reconstruction error until the model can accurately reconstruct the spatial features of normal assemblies. For the LSTM / TCN module, the temporal data in the training set is input step by step. The model predicts the features of the next time step based on the features of the previous time step, and the error between the predicted and actual values ​​is calculated. Similarly, the optimizer adjusts the parameters to enable the model to master the temporal patterns of normal assemblies. An early stopping mechanism is introduced during training. Training stops when the validation set error no longer decreases after several consecutive rounds to avoid overfitting, i.e., the model can only identify normal patterns in the training data and cannot generalize to new data. The final output is a trained anomaly detection model that can stably learn normal assembly features and has generalization ability.

[0127] When the model is put into practical application, the deep feature matrix generated earlier needs to be input into the trained model. The two main modules calculate the error separately and then sum them with weights to obtain the assembly warning coefficient. The autoencoder module receives the spatial feature portion of the deep feature matrix, such as component installation features and operation action features at a certain moment. After reconstruction by the decoder, the difference between the reconstructed output and the input features is calculated, i.e., the reconstruction error. If the actual assembly features deviate from the normal pattern, the reconstruction error will increase significantly. The LSTM / TCN module receives the temporal feature portion of the deep feature matrix, such as the process execution parameters and action sequences of continuous time steps, predicts the temporal features of the next moment, and calculates the difference between the predicted value and the actual features of the next moment, i.e., the prediction error. If the execution sequence of processes is disordered or the parameters change abnormally over time, the prediction error will increase. To comprehensively reflect the degree of anomalies in both the spatial and temporal dimensions, the two types of errors need to be weighted and summed, as shown in the formula:

[0128]

[0129] Since reconstruction error and prediction error reflect assembly anomalies from spatial and temporal dimensions, respectively, weights need to be assigned based on the contribution of each type of error to historical anomaly cases. For example, if historical data shows that spatial anomalies have a greater impact on failures, then... And the sum of the weights is Ensure the early warning coefficient It can reflect both types of abnormal information in a balanced way, avoiding misjudgments caused by a single error; among which For assembly early warning coefficient; The weights for the reconstruction error range from 0 to 1; The reconstruction error output by the autoencoder module; The weight for the prediction error, with a value ranging from 0 to 1; This represents the prediction error output by the LSTM / TCN module. The warning coefficient is calculated using this formula. It can quantify the abnormal risks in the current assembly process. The higher the coefficient, the greater the probability of abnormality, and finally outputs the assembly early warning coefficient.

[0130] To avoid fixed thresholds failing to adapt to fluctuations in the assembly process, such as subtle differences in assembly parameters between different batches of instruments, an adaptive threshold algorithm is needed to analyze the warning coefficients and determine dynamic anomaly thresholds. The system first collects warning coefficients from normal assembly processes over a period of time, statistically analyzing their distribution characteristics, such as the mean μ and standard deviation σ. An adaptive method based on statistical distribution is prioritized. Principle: If the warning coefficient follows a normal distribution, the anomaly threshold is set to μ+3σ. This threshold covers 99.73% of normal assembly scenarios, avoiding misjudging normal fluctuations as anomalies while effectively capturing true anomalies exceeding the normal range. If the warning coefficient does not follow a normal distribution, kernel density estimation is used. By fitting the coefficient distribution curve, a critical point with extremely low probability density is found as the threshold, such as a coefficient value with a probability below 0.1%. Furthermore, the threshold is dynamically updated with new normal assembly data, such as monthly re-statistical distribution of coefficients and adjustment of the threshold, ensuring it always adapts to the current assembly conditions. The final output is a dynamic anomaly threshold that flexibly responds to process fluctuations.

[0131] Each assembly process is treated as a node in a Generative Neural Network (GNN). Node features are the multimodal features of the corresponding process in the deep feature matrix, such as process execution time, operating parameters, and component status labels. Based on the assembly process logic (e.g., housing assembly is a prerequisite for sensor installation), directed edges are constructed between nodes. The edge weights represent the strength of the dependency relationship between processes; edges between critical prerequisite processes and subsequent processes have higher weights, forming an initial process relationship graph. Subsequently, the GNN uses a message passing mechanism to allow each node to exchange feature information with its neighbors, such as between prerequisite and subsequent processes. It calculates the deviation between node features and normal process features, as well as the deviation between edge weights and the strength of normal dependencies. If a node's assembly parameters, such as torque, are significantly outside the normal range, it is considered a node feature deviation (e.g., abnormal screw tightening process parameters). If an edge weight is significantly lower than normal (e.g., a subsequent process is executed before a prerequisite process is completed), it is considered a dependency deviation (e.g., disordered process sequence). Finally, by combining the two types of deviation results, a deviation identification result containing the deviation location and type is output, achieving a deeper understanding from overall anomaly warning to local deviation localization.

[0132] To clarify the key causes of anomalies and improve the interpretability of detection results, interpretable AI (XAI) technology needs to be introduced to analyze the feature contribution of warning coefficients and deviation results. SHAP or LIME algorithms are preferred: For warning coefficients, the SHAP value quantifies the contribution of each feature to the coefficient. If the SHAP value of a feature is positive and has a large absolute value, it indicates that this feature is the main factor causing the warning coefficient to rise. For example, if the torque exceeds the standard value by 50%, its SHAP value will be significantly positive. For deviation identification results, LIME constructs a locally interpretable model to analyze which features caused specific deviations. For example, sensor installation deviations are mainly caused by two features: installation angle deviation and insufficient torque of the fixing screws. Based on the magnitude of feature contribution, it is converted into a risk score: the score range is set to 0-100 points. The higher the contribution, that is, the greater the impact on the anomaly, the higher the risk score. For example, if the contribution of a key feature reaches 80%, the corresponding risk score is 90 points; if the contribution of a secondary feature is 20%, the corresponding risk score is 30 points. At the same time, the impact feature and the degree of impact corresponding to each score are clearly defined. For example, a risk score of 85 points is mainly caused by the screw torque exceeding the standard value by 40%. Finally, a quantifiable and interpretable risk score is output.

[0133] Finally, by comparing the assembly warning coefficient with the dynamic anomaly threshold and combining it with the risk score, the final anomaly detection result is formed. If the warning coefficient is greater than the dynamic anomaly threshold, it indicates that there is a significant deviation in the current assembly process. The system outputs an anomaly with a binary label, such as a label of 1, and associates it with the corresponding risk score and deviation identification result. For example, anomaly label 1, risk score 92 points, deviation location: sensor calibration process, deviation type: calibration accuracy out of tolerance. If the warning coefficient is less than or equal to the dynamic anomaly threshold, it indicates that the assembly process conforms to the normal mode. The system outputs a normal result with a binary label, such as a label of 0, a risk score of 0 points, and a deviation identification result of no deviation. The final anomaly detection result must include four pieces of information: label, warning coefficient, risk score, and deviation details. This clearly defines the current assembly status and provides complete data support for subsequent blockchain storage and fault backtracking.

[0134] In some embodiments, step S36 includes the following steps:

[0135] Step S361: Based on the deep feature matrix, process attributes and assembly features are fused to generate multi-dimensional node feature vectors;

[0136] Step S362: Based on node features, analyze the dependencies between processes and dynamically construct a process relationship graph model;

[0137] Step S363: A hierarchical identification mechanism is adopted to identify node feature deviations and dependency deviations respectively, and the anomaly degree is calculated in a comprehensive manner to locate the deviation type and location.

[0138] In this embodiment, based on the deep feature matrix generated above, process attributes and assembly features are fused to generate multi-dimensional node feature vectors. The deep feature matrix integrates video spatiotemporal features, text features, and structured data, and strengthens the inherent correlation through a cross-modal attention mechanism, comprehensively reflecting the state of the assembly process. Process attributes include process specification information such as the process number, type, standard parameter range, estimated execution time, and required tool model. The fusion process uses unique identifiers and process labels as constraints, concatenating the process attributes of the corresponding process with the assembly feature vector in the deep feature matrix to form a multi-dimensional node feature vector for each process. For example, if the assembly feature vector of a certain process in the deep feature matrix is ​​512-dimensional, and the attribute information of that process, such as standard torque, execution time, and tool model code, is finally converted into a numerical vector of 8 dimensions, then the fusion results in a 520-dimensional node feature vector. This fusion method not only retains real-time assembly status information but also incorporates process specification information, enabling subsequent deviation identification to compare actual execution with historical normal modes and directly compare with process standards, improving the accuracy and interpretability of deviation location.

[0139] Next, based on node features, the dependencies between processes are analyzed, and a process relationship graph model is dynamically constructed. The assembly process has clear sequential dependencies; for example, the housing assembly must be completed before sensor installation. These dependencies constitute the edge structure of the graph model. The construction process consists of two steps: First, the basic dependencies between processes are determined according to the process specifications, forming an initial directed graph. Second, the similarity or correlation between processes is calculated using node feature vectors, and the edge weights are dynamically adjusted. For example, the matching degree of feature vectors of adjacent process nodes is calculated using cosine similarity. If the similarity is high, it indicates that the actual execution process conforms to the standard dependencies, and the edge weight remains high; if the similarity is low, there may be process execution anomalies or sequence errors, and the edge weight is reduced accordingly. Furthermore, the system dynamically updates the graph structure based on real-time assembly data. If a process is detected to be executed earlier or later than expected, the graph model will adjust the node connections in real time to ensure that the model remains consistent with the actual production process. The dynamically constructed process relationship graph not only reflects the static dependencies of the process design but also reflects the dynamic characteristics of actual execution.

[0140] Subsequently, a hierarchical identification mechanism is employed to identify node feature deviations and dependency deviations separately. Node feature deviation identification focuses on whether the execution status of a single process conforms to a normal pattern: by comparing the current node feature vector with the feature distribution of the corresponding process in historical normal assembly data, the degree of difference between the two is calculated, such as using Euclidean distance or cosine distance. If the difference exceeds a preset threshold, the process is determined to have a feature deviation, which may manifest as abnormal operating parameters, such as excessive or insufficient torque, incorrect tool use, or improper component installation. Dependency deviation identification focuses on whether the connections between processes conform to specifications: by analyzing the changes in edge weights in the graph model, it is determined whether the dependency strength between processes deviates from the normal range. For example, if the edge weight from preceding process A to subsequent process B is lower than the historical average, it may indicate that process A was not completed according to standards before entering process B, suggesting a risk of incorrect sequence or missing processes.

[0141] Finally, the anomaly score is calculated comprehensively to pinpoint the type and location of deviations. The anomaly score calculation considers the severity of both node characteristic deviations and dependency relationship deviations, weighted and summed according to preset weights to form a comprehensive anomaly score for each process node. Weight allocation can be determined based on the contribution of these two types of deviations to the final failure in historical failure data; if a node characteristic deviation has a higher proportion of failures in historical cases, its weight is increased accordingly. Based on the comprehensive anomaly score, the system can rank and locate the process nodes with the most severe anomalies and generate a detailed deviation report, including the specific process number, deviation type, impact characteristics, and recommended measures. This hierarchical identification and comprehensive evaluation method not only improves the accuracy of deviation location but also...

[0142] In some embodiments, step S4 includes the following steps:

[0143] Step S41: Extract the anomaly detection results from step 3 and the multi-source raw data from step 1, unify the data format and remove redundant fields to form a dataset to be hashed, and output the dataset.

[0144] Step S42: Use a hash algorithm to calculate the dataset to be hashed, generate a unique hash value, synchronously record the data collection timestamp and device information, and output the hash value and associated timestamp and device information;

[0145] Step S43: Based on the unique identifier of the instrument product in Step 1, create a private blockchain for each product, define the block structure, and output the private blockchain framework.

[0146] Step S44: According to the block structure of step S43, encapsulate the hash value, timestamp, device information and assembly data of step S42 into an initial block and output the block.

[0147] Step S45: Write a smart contract containing data consistency verification rules, deploy it to the private blockchain of step S43, and output the blockchain of the deployed contract.

[0148] Step S46: Upload the initial block to the blockchain where the contract is deployed. The contract automatically verifies the data consistency. Once the verification is successful, the data is uploaded to the blockchain, achieving immutable data storage and outputting the updated private blockchain.

[0149] In this embodiment, the generated anomaly detection results and the collected multi-source raw data need to be extracted first and stored as immutable content. The anomaly detection results include key information such as warning coefficients, risk scores, anomaly labels, and deviation identification details, while the multi-source raw data covers video images, structured data, environmental data, and various labels. Since these data come from different sources and have different formats, they need to be processed uniformly: converting data of different formats to standard JSON or Protocol Buffers formats, unifying field naming and data types to ensure consistency in subsequent hash calculations and on-chain storage; and removing redundant fields, such as duplicate records, intermediate calculation process data, and auxiliary information unrelated to assembly quality, to reduce storage capacity and improve processing efficiency. The processed dataset is called the hashable dataset, which contains complete key information about the instrument product during the assembly process.

[0150] Next, a secure hash algorithm is used to calculate a unique hash value on the prepared dataset. While calculating the hash value, the system simultaneously records the precise timestamp of data collection and detailed information about the collection device, such as device ID, firmware version, and collection port. This information is stored along with the hash value for subsequent verification of the data's generation time and origin. The generated hash value serves as a digital fingerprint of the data, which can be used not only for data integrity verification but also as a unique index in the blockchain, facilitating quick location and retrieval of the corresponding data.

[0151] Based on a unique identifier assigned to each instrument product, the system creates a dedicated private blockchain for each product. The advantages of a private blockchain lie in its high controllability, fast transaction speed, and strong security, making it suitable for industrial production environments with strict requirements for data access permissions and processing efficiency. When creating a private blockchain, the block structure needs to be defined, including a block header and a block body. The block header contains metadata such as the version number, the hash value of the previous block, the Merkle Root, timestamp, difficulty target, and nonce. The block body contains the specific data records, which in this case are the hash value to be stored, timestamp, equipment information, and assembly data. Through this structural design, each block contains a summary of its own data and forms an inseparable chain structure with the entire chain through the hash value of the previous block, thus ensuring the immutability of the data.

[0152] After defining the block structure, the system encapsulates the generated hash value, timestamp, device information, and assembly data according to the preset block structure to form the initial block, also known as the genesis block. The genesis block is the first block of the blockchain and has no preceding block; therefore, the hash value field of the previous block in its block header is usually set to zero or a specific placeholder. The genesis block contains the initial key data for the assembly process of this instrument product.

[0153] To ensure data consistency and reliability in a blockchain, smart contracts containing data consistency verification rules need to be written and deployed to a private blockchain. Smart contracts are automated scripts running on the blockchain that can automatically execute corresponding operations when preset conditions are met. In this scenario, the smart contract mainly includes the following verification rules: data field integrity verification, ensuring that the uploaded data contains all required fields; timestamp validity verification, ensuring that the timestamp is no earlier than the previous block and no later than the current time; device information legitimacy verification, verifying whether the data collection device is in the authorized list; and hash value correctness verification, ensuring that the data matches the hash value. After deploying the smart contract, all data operations on the blockchain will be subject to the contract rules. Any data that does not conform to the rules will be automatically rejected, thus ensuring the legality and consistency of the data from a technical perspective.

[0154] Finally, the encapsulated initial block is uploaded to the private blockchain with the deployed smart contract. Upon receiving a new block, the blockchain network invokes the smart contract to automatically execute the data consistency verification process. If the verification passes, the block is added to the end of the blockchain, and the blockchain state is updated; if the verification fails, the block is discarded, ensuring that only legitimate data can enter the blockchain. Due to the blockchain's chain structure and hash encryption mechanism, any tampering with the data already on the chain requires recalculating the hash value of that block and all subsequent blocks, which is computationally infeasible, thus achieving immutable data storage. The final output, the updated private blockchain, contains a complete, verifiable, and immutable record of the instrument product's assembly process data.

[0155] In some embodiments, step S5 includes the following steps:

[0156] Step S51: When the instrument fails later, the backtracking process is triggered. The system receives the basic information of the faulty instrument and outputs the fault backtracking start command.

[0157] Step S52: Based on the startup command and basic information, retrieve the unique identifier assigned to the instrument in step 1 and output the unique identifier of the faulty instrument.

[0158] Step S53: Input the identifier into the private blockchain of step 4, extract the full process assembly data through the retrieval module, and output the full process assembly dataset.

[0159] Step S54: Call the graph neural network to reconstruct the assembly process model based on the dataset and output the reconstructed model;

[0160] Step S55: Retrieve the standard assembly process, compare and reconstruct the model step by step, identify assembly deviation points, and output a list of deviation points.

[0161] Step S56: Combining the risk score from step 3 with the video playback of the dataset, analyze the deviation points to locate the root cause of the fault and output the root cause analysis results.

[0162] Step S57: Based on the root cause analysis results, compile a structured report and visualization charts containing the faulty process, cause, and improvement suggestions, and output the report and charts.

[0163] In this embodiment, when an instrument malfunctions during later use or testing, the system will trigger a backtracking process. The triggering method can be automatic reporting by the quality inspection system or fault information entered by after-sales maintenance personnel through a human-machine interface. The basic information received by the system includes the faulty instrument's model, serial number, production batch, description of the fault phenomenon, and the time of occurrence. Based on this information, the system generates a fault backtracking initiation command, which includes a unique task ID and the traceability scope, such as the involved processes, time intervals, and priority parameters, ensuring the orderly execution of subsequent steps.

[0164] Next, based on the startup command and the basic information of the faulty instrument, the system retrieves the unique identifier assigned to the instrument at the initial stage of assembly from the metadata database of the production execution system or data acquisition system. This unique identifier serves as a data identity card throughout the entire assembly and traceability process, allowing for precise association with the entire process data stored in the blockchain. The retrieval process employs a combination of index matching and fuzzy search to ensure that even with incomplete input information, the unique identifier of the target instrument can be efficiently located and output for subsequent data extraction.

[0165] Subsequently, a unique identifier is input into the private blockchain constructed in step 4, and the on-chain retrieval module extracts the complete assembly data of the instrument. Each block in the blockchain contains a hash value, timestamp, equipment information, and corresponding multi-source raw data and anomaly detection results. The retrieval module traverses relevant blocks based on the identifier, extracting the complete data chain from online assembly to final inspection, and uses smart contract data consistency verification to ensure that the extracted data has not been tampered with and that the timeline is continuous. The output complete assembly dataset includes video image clips, structured process parameters, environmental data, component dynamic labels, process labels, and anomaly detection results.

[0166] After acquiring the full-process assembly dataset, the system invokes a Graph Neural Network (GNN) to reconstruct the assembly process model. The reconstruction process uses process steps as nodes and dependencies between steps as edges, employing multi-dimensional features extracted from the blockchain data, such as process execution time, operational parameters, and component states, as node feature vectors. The GNN propagates feature information between nodes through a message passing mechanism, learning the dynamic dependencies between processes to generate a graph structure model that reflects the actual assembly execution process. This reconstructed model not only preserves the temporal information of the processes but also reflects the correlation and influence strength between them, providing a structured analytical framework for deviation comparison.

[0167] Next, the system retrieves the standard assembly process, i.e., the ideal process model of the process design, and compares it step-by-step with the reconstructed model. The comparison includes two aspects: node feature deviation and dependency deviation. Node feature deviation is calculated by distributing the distance between the feature vectors of each process node in the reconstructed model and the feature vectors of the corresponding nodes in the standard model, using Euclidean distance as a measure. If the difference exceeds a preset threshold, it is considered that the process has problems such as abnormal execution parameters, incorrect tool usage, or component installation deviation. Dependency deviation is identified by comparing the weights of edges with those in the standard process. If the difference is significant, it may indicate an incorrect process sequence, unmet preconditions, or abnormal inter-process coordination. The comparison results will generate a detailed list of deviation points, including process number, deviation type, deviation degree, and associated characteristics.

[0168] Based on the identified deviation points, the system combines the risk score generated in step 3 with full-process video playback to conduct root cause analysis. The risk score reflects the contribution of each feature to the anomaly warning; features with high risk scores are often highly correlated with the root cause of the fault. Video playback provides intuitive visual evidence, verifying whether the deviation actually occurred and the specific scenario in which it occurred. By combining the two, the system can cross-validate the causal relationship between deviation points and fault phenomena, eliminating deviations with weak correlations and focusing on the key processes and factors most likely to cause the fault. For example, if the risk score shows that sensor calibration error is the main contributing factor, and the video playback shows that the operator did not use the calibration tool according to specifications during the calibration process, then the root cause of the fault can be determined to be improper use of the calibration tool leading to excessive calibration accuracy.

[0169] Finally, based on the root cause analysis results, the system generates a structured fault analysis report and visualization charts. The report includes fault process location, root cause description, impact assessment on product performance, improvement suggestions such as optimizing process parameters, strengthening personnel training, and adding automated testing steps, as well as preventative measures. The visualization charts visually display the fault evolution process and key influencing factors through time-series trend charts, deviation heatmaps, and process dependency diagrams, facilitating production managers to quickly understand the essence of the problem and formulate improvement plans. The report and charts will be stored in the system's knowledge base as knowledge accumulation and can also be used as labeled data for subsequent retraining of the anomaly detection model, achieving closed-loop management of fault prevention and process optimization.

[0170] In some embodiments, step S6 includes the following steps:

[0171] Step S61: Using the root cause analysis report of the failure in step 5 as labeled data, extract key information and classify and label it to form a labeled training dataset, and output the dataset.

[0172] Step S62: Input the labeled training dataset into the anomaly detection model of step 3, and update the parameters through incremental training to improve the fault feature recognition capability, and output the updated anomaly detection model.

[0173] Step S63: Determine the multi-objective optimization parameters based on the actual assembly requirements, clarify the optimization weight of each parameter, and output the multi-objective optimization parameters and weight configuration.

[0174] Step S64: A multi-objective reinforcement learning algorithm is used, with the parameters and weights of step S63 as the objectives, to optimize the assembly process parameters in combination with real-time production line data, and output the optimized process parameters.

[0175] Step S65: Introduce meta-reinforcement learning technology, collect a small amount of assembly data for the new scene, learn the assembly features of the new scene, and output model adjustment parameters adapted to the new scene.

[0176] Step S66: Based on the real-time performance data of the production line, combined with the updated model in step S62 and the adaptation parameters in step S65, dynamically adjust the anomaly judgment threshold and model parameters in step 3, and output the adjusted threshold and parameters.

[0177] Step S67: Apply the adjusted threshold and parameters to the production line to verify the optimization effect. If the preset target is not met, return to step S61 for iteration until the target is met.

[0178] In this embodiment, a labeled training dataset is constructed based on the generated root cause analysis report of the failure. The report contains rich information such as the location of the failure process, the type of deviation, key impact features, and risk scores. The system extracts this key information using natural language processing technology, including failure features such as abnormal torque and installation angle deviation; deviation types such as abnormal parameters and incorrect sequence; degree of impact and related processes. Subsequently, this information is associated and labeled with the corresponding assembly data to form training samples with clear labels. For example, a failure instance of torque deviation in the sensor calibration process is bound to the video features, structured parameters, and environmental data of that process and marked as an abnormal sample; at the same time, sufficient normal samples are retained to maintain class balance. After data cleaning and format standardization, a labeled training dataset that can be directly used for model training is output.

[0179] Next, the labeled training dataset is input into the anomaly detection model constructed in step 3 for incremental training. The advantage of incremental training is that it can quickly absorb new fault features without losing existing knowledge. During training, the system freezes some of the model's lower-level weights and the layers responsible for extracting general features, updating only the upper-level classifier weights to avoid catastrophic forgetting. Through multiple rounds of iterative optimization, the model gradually masters new fault patterns, improving its ability to identify rare or novel faults. After training, the updated anomaly detection model is output, which maintains the original detection accuracy while exhibiting higher sensitivity to the latest fault types.

[0180] Based on actual assembly requirements, the system determines multi-objective optimization parameters and weight configurations. Optimization objectives typically include: improving anomaly detection accuracy, reducing false alarm rate, shortening detection response time, and reducing assembly rework rate. Each objective corresponds to a set of adjustable process parameters, such as sensor sampling frequency, image analysis frame rate, anomaly judgment threshold, and process execution time window. The priority weights of each objective are determined through the analytic hierarchy process (AHP) or expert evaluation, forming a mathematical expression of the multi-objective optimization problem. The weight configuration is dynamically adjusted according to production strategies; for example, increasing the weight of detection accuracy during quality-sensitive periods and appropriately reducing detection intensity to improve production efficiency during periods of tight capacity.

[0181] A multi-objective reinforcement learning algorithm is employed, using defined optimization parameters and weights as objectives, and continuously optimizing assembly process parameters in conjunction with real-time production line data. The reinforcement learning agent treats the assembly system as its environment, process parameter settings as actions, and the weighted sum of the multi-objective results as a reward signal. Through continuous interaction with the environment, the agent learns the optimal parameter strategy under different operating conditions. For example, in environments with large temperature and humidity fluctuations, the agent automatically increases the sampling frequency of environmental data and adjusts the anomaly detection threshold to maintain detection accuracy. During the optimization process, the system periodically evaluates the strategy performance and employs an elite retention strategy to ensure that high-quality parameter combinations are not discarded. The final output is an optimized process parameter configuration that can be directly applied to the production line.

[0182] To address new scenarios such as new product models, process changes, or equipment upgrades, the system introduces meta-reinforcement learning technology. This technology rapidly learns adaptation strategies using a small amount of assembly data from these new scenarios. Meta-reinforcement learning, through pre-training on multiple tasks, acquires general strategies capable of quickly adapting to new environments. When a new scenario emerges, the system only needs to collect a small amount of trial assembly data and fine-tune the model parameters using the meta-learning algorithm to generate model adjustment parameters adapted to the new scenario. This process significantly shortens the model adaptation cycle under new scenarios, ensuring that the system maintains stable detection performance even during production changes.

[0183] The system dynamically adjusts the parameters and thresholds of the anomaly detection model based on real-time performance data from the production line. Real-time performance data includes key indicators such as detection accuracy, false alarm rate, and detection latency. Combining the updated anomaly detection model with parameters adapted to new scenarios, the system constructs a performance feedback mechanism: when detection accuracy decreases or the false alarm rate increases, the model weights and thresholds are automatically adjusted; when the system is running stably, the current configuration is maintained to ensure production continuity. This dynamic adjustment mechanism enables the system to adapt to slow changes in the production environment, such as equipment aging and fluctuations in personnel skills.

[0184] Finally, the adjusted thresholds and parameters are applied to the production line for verification. During the verification process, the system continuously monitors key performance indicators and compares them with preset targets. If the target requirements are met or exceeded, the current configuration is officially adopted; if not, the system returns to step S61 to regenerate the target training dataset and initiates a new round of model updates and parameter optimization. Through this continuous iterative approach, the system continuously accumulates fault knowledge, optimizes detection strategies, and ultimately forms a self-evolving and continuously improving intelligent assembly quality control closed-loop system.

[0185] In multi-objective optimization, a weighted summation method can be used to transform multiple objectives into a single-objective optimization problem, as shown in the following formula:

[0186]

[0187] in The value of the comprehensive objective function; To optimize the target quantity; Let be the weight of the i-th objective, and satisfy . ; Let be the function value of the i-th objective; To optimize the parameter vector, multiple optimization objectives are linearly combined into a single objective function by assigning corresponding weights to each objective. This facilitates the solution using optimization algorithms such as reinforcement learning. The magnitude of the weights reflects the relative importance of the corresponding objectives and can be dynamically adjusted according to production needs.

[0188] In meta-reinforcement learning, the parameter tuning rule can be expressed as: ;

[0189] in To adapt the model parameters to the new scenario; Basic model parameters; The learning rate; Based on a small amount of new scenario data Calculated gradient; This is the reward function for the new scenario. The performance gradient is calculated using a small amount of new scenario data. Starting with the basic model parameters, the parameters are updated according to a certain learning rate to achieve rapid adaptation. This method allows the model to quickly adjust to the new environment even with limited data.

[0190] In some embodiments, step S65 includes the following steps:

[0191] Step S651: Preprocess the small amount of assembly data in the new scenario and anchor reusable general features through feature comparison.

[0192] Step S652: Based on the preprocessed data, the pre-trained meta-model is used to quickly optimize the decision-making logic specific to the new scenario through small sample iterative learning.

[0193] Step S653: Verify the initial adaptation parameters through trial assembly, fine-tune them based on performance feedback, and output the final model adjustment parameters.

[0194] In this embodiment, the small amount of assembly data collected in the new scenario undergoes comprehensive preprocessing. This includes data cleaning, format standardization, time alignment, and feature extraction, such as extracting representative features from videos, structured data, and text tags.

[0195] After preprocessing, the system anchors reusable general features using feature comparison methods. Specifically, it compares the feature vectors of the new scene data with the learned feature space from historical scenes, such as calculating cosine similarity, to identify stable feature patterns across scenes. These general features typically include basic assembly action patterns, standard process conversion logic, and common equipment status characteristics. In this way, the system can quickly locate directly reusable parts of the existing knowledge system in new scenes, laying the foundation for subsequent few-shot learning and reducing reliance on entirely new data.

[0196] Based on anchored general features, the system calls a pre-trained meta-model to quickly optimize the decision logic specific to new scenarios. The meta-model has been trained on a large number of different scenarios, possesses the ability to recognize various patterns in the assembly process, and has learned learning strategies to quickly adapt to new environments.

[0197] For new scenarios, the system inputs preprocessed data into the meta-model. The meta-model uses its existing knowledge framework to analyze the new data, focusing on scenario-specific features beyond general characteristics. Through iterative learning with few samples, the meta-model can quickly adjust its internal parameters to form a dedicated decision-making logic for the new scenario.

[0198] After the initial adaptation parameters are generated, the system needs to be verified through a trial assembly process. These parameters are then applied to the actual production environment, and performance data of the system in the new scenario is collected, including key indicators such as anomaly detection accuracy, false alarm rate, and detection latency.

[0199] Based on the performance feedback from the trial assembly, the system fine-tunes the initial adaptation parameters. The fine-tuning process typically employs an incremental learning strategy, adjusting only the model parameters strongly correlated with the new scene while maintaining the stability of the general feature parts, in order to avoid the model forgetting knowledge of the old scene.

[0200] This validation and fine-tuning loop may run multiple times until the model's performance in the new scenario reaches the preset target. The final output model adjustment parameters enable the system to achieve high-precision anomaly detection and quality control in the new scenario, while maintaining compatibility with the old scenario. In this way, the system can maintain continuous adaptability and high performance in the face of constantly changing production environments, achieving a true intelligent manufacturing closed loop.

[0201] A high-precision instrument assembly fault tracing system, applied to the aforementioned printed circuit board resin via positioning full-process management method, includes:

[0202] The tag acquisition module is configured to acquire structured data, video images, and environmental data of the assembly process in real time through multi-source data acquisition devices, assign a unique identifier and process tag to each instrument product, update the dynamic information tags of components using edge computing, filter low-quality images using image quality detection algorithms, and output a multimodal dataset with identifiers.

[0203] The feature fusion module is used to fuse multimodal datasets, extract spatiotemporal and textual features from videos, combine them with structured data to form a unified feature vector, standardize environmental data, enhance data correlation through a cross-modal attention mechanism, and generate a deep feature matrix.

[0204] The anomaly detection module is used to build an anomaly detection model. It trains neural networks and time series networks using historical normal data, calculates warning coefficients and uses adaptive thresholds to determine anomalies, uses graph neural networks to model process relationships to identify operational deviations, and combines interpretable AI to output risk scores and binary labels.

[0205] The block storage module is used to process abnormal results and raw data hashes, creating a private blockchain for each product. The blocks contain data hashes and timestamps, and the consistency is verified through smart contracts to achieve tamper-proof storage.

[0206] The fault backtracking module is configured to extract blockchain data based on a unique identifier when a fault occurs, remodel the assembly process using a graph neural network, identify deviations by comparing with the standard process, locate the root cause of the fault by combining risk scoring and video playback, and generate an analysis report.

[0207] The adaptive optimization module is used to retrain the anomaly detection model using fault reports as labeled data. It optimizes process parameters using multi-objective reinforcement learning and meta-reinforcement learning, and dynamically updates the model and thresholds based on real-time performance data.

[0208] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A method for tracing back faults in the assembly of high-precision instruments, characterized in that, Includes the following steps: Step 1: Collect structured data, video images and environmental data of the assembly process in real time through multi-source data acquisition equipment, and assign a unique identifier and process label to each instrument product; Utilize edge computing devices to read component information, generate and update component dynamic information tags in real time; call a lightweight image quality detection algorithm to filter low-quality image frames, and output a multimodal dataset with unique identifiers; Step 2: After preprocessing the multimodal dataset, extract the spatiotemporal features and text features of the video, and concatenate them with the structured data to form an initial unified feature vector; combine the environmental data to standardize and normalize the initial unified feature vector to eliminate environmental interference; introduce a cross-modal attention mechanism to generate a deep feature matrix; Step 3: Construct and train an anomaly detection model that includes an autoencoder module and a temporal network module; input the deep feature matrix into the trained model, calculate the reconstruction error and prediction error respectively, and sum them by weight to obtain the assembly warning coefficient; An adaptive threshold algorithm is used to determine the dynamic anomaly threshold; a graph neural network is invoked to construct a process relationship graph model based on a deep feature matrix, with process as nodes and dynamic logical dependencies between processes as edges. A hierarchical identification mechanism is adopted to simultaneously calculate the node feature deviation of a single process and the dependency relationship deviation between processes in the process relationship graph model; interpretable AI technology is introduced to analyze the feature contribution of assembly early warning coefficient, node feature deviation and dependency relationship deviation, and output a quantifiable risk score; the assembly early warning coefficient is compared with the dynamic abnormal threshold to output abnormal / normal binary labels. Step 4: Extract the anomaly detection results output in Step 3 and the multi-source raw data from Step 1, unify the data format and remove redundant fields to generate the dataset to be hashed. A hash algorithm is used to calculate a unique hash value for the dataset to be hashed, and the data collection timestamp and device information are recorded synchronously. Each instrument product is given a unique identifier to create its own private blockchain. The block structure is defined and the hash value, timestamp, equipment information and assembly data are encapsulated as data blocks. A smart contract containing data consistency verification rules is written and deployed to the private blockchain. The data blocks are uploaded to the blockchain and automatically verified by the smart contract to complete the on-chain process, thus achieving tamper-proof data storage. Step 5: When an instrument malfunctions, the backtracking process is triggered to retrieve its unique identifier based on the basic information of the malfunctioning instrument. Extract the entire assembly process data from the instrument's private blockchain; The graph neural network is invoked to reconstruct the actual assembly process model of the faulty instrument based on the full-process assembly data; the standard assembly process is retrieved and compared with the reconstructed actual assembly process model step by step to identify the process nodes with deviations and the problematic inter-process dependencies, generating a list of deviation points; combined with the risk score output in step 3 and the video playback extracted from the blockchain, the list of deviation points is cross-validated and causal analysis is performed to accurately locate the root cause process and root cause of the failure, and a structured analysis report containing the faulty process, root cause description and improvement suggestions is generated; Step 6: Use the structured analysis report generated in Step 5 as labeled data, extract key information to form a retraining dataset, and incrementally train the anomaly detection model in Step 3 to update the model parameters. Based on the assembly quality objectives, multi-objective optimization parameters and weight configurations are determined. A multi-objective reinforcement learning algorithm is adopted, combined with real-time performance data of the production line, to optimize the assembly process parameters online. Meta-reinforcement learning technology is introduced to collect a small amount of assembly data under new scenarios, quickly learn the assembly characteristics of new scenarios, and generate model adjustment parameters adapted to new scenarios. Based on the real-time performance data of the production line, the parameters of the anomaly detection model and the anomaly judgment threshold are dynamically updated.

2. The method for tracing back faults in high-precision instrument assembly according to claim 1, characterized in that, Step 3 includes the following steps: Step S31: Collect historical normal assembly data of the instrument, clean and label it to form a model training dataset, and output the historical normal assembly training dataset. Step S32: Construct an anomaly detection model architecture, including a neural network module for learning normal assembly patterns and a time series network module for processing assembly time series data, and output the model architecture. Step S33: Input the historical normal assembly training dataset into the model architecture, iteratively train and optimize the parameters to master the normal features, and output the trained anomaly detection model. Step S34: Input the deep feature matrix from step 2 into the trained model. The reconstruction error and prediction error are calculated by the two modules, and the assembly warning coefficient is obtained by weighted summation. The warning coefficient is then output. Step S35: Analyze the warning coefficients using an adaptive threshold algorithm, determine the anomaly judgment thresholds according to the coefficient distribution, and output the dynamic anomaly thresholds. Step S36: Call the graph neural network to construct a process relationship model based on the process data of the deep feature matrix, identify operational deviations, and output the deviation identification results; Step S37: Introduce interpretable AI technology to analyze the characteristic contribution of the early warning coefficient and the deviation results, clarify the impact, and output a risk score; Step S38: Compare the warning coefficient with the dynamic threshold. If the threshold is met, output an abnormal label; otherwise, output a normal label. Associate the risk score to form an abnormal detection result.

3. The method for tracing back faults in high-precision instrument assembly according to claim 2, characterized in that, Step S36 further includes the following steps: Step S361: Based on the deep feature matrix, process attributes and assembly features are fused to generate multi-dimensional node feature vectors; Step S362: Based on node characteristics, analyze the dependencies between processes and dynamically construct a process relationship graph model; Step S363: A hierarchical identification mechanism is adopted to identify node feature deviations and dependency deviations respectively, and the anomaly degree is calculated in a comprehensive manner to locate the deviation type and location.

4. The method for tracing back faults in high-precision instrument assembly according to claim 3, characterized in that, Step 4 includes the following steps: Step S41: Extract the anomaly detection results from step 3 and the multi-source raw data from step 1, unify the data format and remove redundant fields to form a dataset to be hashed, and output the dataset. Step S42: Calculate the dataset to be hashed using a hash algorithm, generate a unique hash value, synchronously record the data collection timestamp and device information, and output the hash value and associated timestamp and device information; Step S43: Based on the unique identifier of the instrument product in Step 1, create a private blockchain for each product, define the block structure, and output the private blockchain framework. Step S44: According to the block structure of step S43, encapsulate the hash value, timestamp, device information and assembly data of step S42 into an initial block and output the block. Step S45: Write a smart contract containing data consistency verification rules, deploy it to the private blockchain of step S43, and output the blockchain of the deployed contract. Step S46: Upload the initial block to the blockchain where the contract is deployed. The contract automatically verifies the data consistency. Once the verification is successful, the data is uploaded to the blockchain, achieving immutable data storage and outputting the updated private blockchain.

5. The method for tracing back faults in high-precision instrument assembly according to claim 4, characterized in that, Step 5 includes the following steps: Step S51: When the instrument fails later, the backtracking process is triggered. The system receives the basic information of the faulty instrument and outputs the fault backtracking start command. Step S52: Based on the startup command and basic information, retrieve the unique identifier assigned to the instrument in step 1 and output the unique identifier of the faulty instrument. Step S53: Input the identifier into the private blockchain of step 4, extract the full process assembly data through the retrieval module, and output the full process assembly dataset. Step S54: Call the graph neural network to reconstruct the assembly process model based on the dataset and output the reconstructed model; Step S55: Retrieve the standard assembly process, compare and reconstruct the model step by step, identify assembly deviation points, and output a list of deviation points. Step S56: Combining the risk score from step 3 with the video playback of the dataset, analyze the deviation points to locate the root cause of the fault and output the root cause analysis results. Step S57: Based on the root cause analysis results, compile a structured report and visualization charts containing the faulty process, cause, and improvement suggestions, and output the report and charts.

6. The method for tracing back faults in high-precision instrument assembly according to claim 5, characterized in that, Step 6 includes the following steps: Step S61: Using the root cause analysis report of the failure in step 5 as labeled data, extract key information and classify and label it to form a labeled training dataset, and output the dataset. Step S62: Input the labeled training dataset into the anomaly detection model of step 3, update the parameters through incremental training to improve the fault feature recognition capability, and output the updated anomaly detection model. Step S63: Determine the multi-objective optimization parameters based on the actual assembly requirements, clarify the optimization weight of each parameter, and output the multi-objective optimization parameters and weight configuration. Step S64: A multi-objective reinforcement learning algorithm is used, with the parameters and weights of step S63 as the objectives, to optimize the assembly process parameters in combination with real-time production line data, and output the optimized process parameters. Step S65: Introduce meta-reinforcement learning technology, collect a small amount of assembly data for the new scene, learn the assembly features of the new scene, and output model adjustment parameters adapted to the new scene. Step S66: Based on the real-time performance data of the production line, combined with the updated model in step S62 and the adaptation parameters in step S65, dynamically adjust the anomaly judgment threshold and model parameters in step 3, and output the adjusted threshold and parameters. Step S67: Apply the adjusted threshold and parameters to the production line to verify the optimization effect. If the preset target is not met, return to step S61 for iteration until the target is met.

7. The method for tracing back faults in high-precision instrument assembly according to claim 6, characterized in that, Step S65 includes the following steps: Step S651: Preprocess the small amount of assembly data in the new scenario and anchor reusable general features through feature comparison. Step S652: Based on the preprocessed data, the pre-trained meta-model is used to quickly optimize the decision-making logic specific to the new scenario through small sample iterative learning. Step S653: Verify the initial adaptation parameters through trial assembly, fine-tune them based on performance feedback, and output the final model adjustment parameters.

8. The method for tracing back faults in high-precision instrument assembly according to claim 1, characterized in that, Step 1 includes the following steps: Step S11: Select high-definition cameras, IoT sensors and assembly equipment controllers according to the assembly process requirements, clarify the collection categories of structured data, video images and environmental data, and output the equipment list and data collection category table. Step S12: Collect multi-source raw data in real time according to the aforementioned devices and categories, temporarily store it in the temporary storage module of the edge computing device, and output the multi-source raw dataset. Step S13: Assign a unique identifier to each instrument product, apply a unique label according to the assembly process and associate it with the original data, and output multi-source data with preliminary identification. Step S14: The edge computing device reads the aforementioned multi-source data component information with preliminary identification, generates dynamic tags, receives component status updates in real time and synchronizes tag content, and outputs multi-source data with real-time updated tags. Step S15: Call a lightweight image quality detection algorithm to filter out low-quality image frames that are blurry, occluded, or have abnormal lighting, retain high-quality images, and output multi-source data with low-quality images removed. Step S16: Integrate the multi-source data with low-quality images filtered out from step S15, the unique identifiers and exclusive labels output from step S13, and the dynamic labels in the multi-source data with real-time updated labels output from step S14 to form a multimodal dataset with unique identifiers.

9. A method for tracing back faults in high-precision instrument assembly according to claim 8, characterized in that, Step 2 includes the following steps: Step S21: Receive the multimodal dataset with a unique identifier, preprocess it, and output the preprocessed multimodal dataset; Step S22: Call the pre-trained spatiotemporal feature extraction model to extract various spatiotemporal features from the preprocessed video image and output the video spatiotemporal feature vector; Step S23: Using natural language processing technology, the key text tags in the encoded data are encoded and converted into standardized text feature vectors, and the text feature vectors are output. Step S24: The video spatiotemporal feature vector, text feature vector and preprocessed structured data are fused together and concatenated into an initial unified feature vector according to the data association rules, and the vector is output. Step S25: Extract environmental data, standardize and normalize the initial unified feature vector to eliminate temperature and humidity interference, and output the environmentally corrected unified feature vector. Step S26 introduces a cross-modal attention mechanism to enhance the correlation of corrected vectors, strengthen the intrinsic correlation of multimodal data, and generate a deep feature matrix.

10. A high-precision instrument assembly fault tracing system, applied to the high-precision instrument assembly fault tracing method described in any one of claims 1-9, characterized in that, include: The tag acquisition module is configured to acquire structured data, video images, and environmental data of the assembly process in real time through multi-source data acquisition devices, assign a unique identifier and process tag to each instrument product, update the dynamic information tags of components using edge computing, filter low-quality images using image quality detection algorithms, and output a multimodal dataset with identifiers. The feature fusion module is used to fuse multimodal datasets, extract spatiotemporal and textual features from videos, combine them with structured data to form a unified feature vector, standardize environmental data, enhance data correlation through a cross-modal attention mechanism, and generate a deep feature matrix. The anomaly detection module is used to build an anomaly detection model. It trains neural networks and time series networks using historical normal data, calculates warning coefficients and uses adaptive thresholds to determine anomalies, uses graph neural networks to model process relationships to identify operational deviations, and combines interpretable AI to output risk scores and binary labels. The block storage module is used to process abnormal results and raw data hashes, creating a private blockchain for each product. The blocks contain data hashes and timestamps, and the consistency is verified through smart contracts to achieve tamper-proof storage. The fault backtracking module is configured to extract blockchain data based on a unique identifier when a fault occurs, remodel the assembly process using a graph neural network, identify deviations by comparing with the standard process, locate the root cause of the fault by combining risk scoring and video playback, and generate an analysis report. The adaptive optimization module is used to retrain the anomaly detection model using fault reports as labeled data. It employs multi-objective reinforcement learning and meta-reinforcement learning to optimize process parameters and dynamically updates the model and thresholds based on real-time performance data.

Images