Flexible production line real-time deviation correction method and system based on AI fool-proof digital large model

Abstract

The invention provides a flexible production line real-time deviation correction method and system based on an AI fool-proof digital large model. Belongs to the technical field of intelligent manufacturing. The method comprises the steps of performing multi-dimensional data source access configuration on a new energy automobile flexible production line, and generating a flexible production line multi-source data set; performing data preprocessing and feature extraction by using an AI fool-proof digital large model, and constructing an initial data model fused with multi-modal features; performing dynamic boundary condition learning and adaptive adjustment on the initial data model by using an AI fool-proof digital large model to generate a dynamic boundary condition model; and the multi-source data set of the flexible production line is monitored in real time. By means of deep access and fusion of an AI fool-proof digital large model to a multi-dimensional data source of a new energy automobile flexible production line, the barrier of data splitting in a traditional mode is broken, and association analysis of multi-modal data such as equipment, vision and materials is achieved.

Description

Technical Field

[0001] This invention proposes a real-time correction method and system for flexible production lines based on an AI-based error-proof digital model, belonging to the field of intelligent manufacturing technology. Background Technology

[0002] In the current booming development of the new energy vehicle industry, flexible production lines have become key to improving production flexibility and efficiency, covering core processes such as battery pack assembly, electric drive assembly, and vehicle body welding. However, the PLC logic control and MES system monitoring mode commonly used in current flexible production lines have many drawbacks.

[0003] In terms of rules, the over-reliance on preset thresholds and finite state machines proves inadequate in handling dynamic boundary condition changes caused by frequent vehicle model switching and continuous process modifications, easily leading to false alarms or missed alarms and disrupting normal production. At the perception level, data from equipment, vision, and materials are processed independently by different subsystems, lacking cross-modal correlation analysis, making it difficult to detect complex deviations such as incorrect tightening sequence of correct parts. In terms of response, anomalies are often only discovered at the final inspection station or after the product has finished production, at which point rework costs are high, and defective products have already entered subsequent processes. Corrective decisions rely on engineer experience, resulting in slow response and poor consistency, making it difficult to meet the demands of high-paced production. Summary of the Invention

[0004] This invention provides a method and system for real-time deviation correction in flexible production lines based on an AI-based error-proofing digital model, to solve the problems mentioned in the background section above:

[0005] The present invention proposes a real-time deviation correction method for flexible production lines based on an AI-based error-proofing digital model, the method comprising:

[0006] S1. Configure multi-dimensional data source access for the flexible production line of new energy vehicles to generate a multi-source data set for the flexible production line; use AI error-proof digital big data model for data preprocessing and feature extraction to build an initial data model that integrates multi-modal features;

[0007] S2. Utilize an AI-powered error-proof digital model to dynamically learn and adaptively adjust the initial data model to generate a dynamic boundary condition model; and conduct real-time monitoring of the multi-source data set of the flexible production line.

[0008] S3. Use the AI-based error-proofing digital big data model to perform cross-modal correlation analysis on potential abnormal data to generate composite deviation identification results data; based on the composite deviation identification results data, use the AI-based error-proofing digital big data model to classify deviation types and assess their severity to obtain deviation classification and severity data.

[0009] S4. By combining the AI ​​error-proof digital model with the historical correction case library and expert knowledge base, the deviation classification and severity data are inferred in real time to automatically generate correction decision instructions; the correction decision instructions are sent to the relevant execution equipment of the flexible production line to trigger real-time correction actions and generate correction execution feedback data.

[0010] S5. Use the AI ​​error prevention digital model to evaluate the effect of the error correction execution feedback data and generate error correction effect evaluation data; and optimize and adjust the parameters of the AI ​​error prevention digital model to generate optimized AI error prevention digital model parameter data.

[0011] S6. Based on the optimized AI error-proofing digital model parameter data, the entire process of real-time correction of the flexible production line is monitored and analyzed to generate production efficiency improvement data and product quality improvement data; and the overall benefits of the flexible production line real-time correction method based on the AI ​​error-proofing digital model are further evaluated to generate comprehensive benefit evaluation data.

[0012] The present invention proposes a system for implementing the real-time deviation correction method for flexible production lines based on the AI-based error-proofing digital model as described above, the system comprising:

[0013] Model building module: Configure multi-dimensional data source access for the flexible production line of new energy vehicles to generate a multi-source data set of the flexible production line; perform data preprocessing and feature extraction to build an initial data model that integrates multi-modal features;

[0014] Real-time monitoring module: performs dynamic boundary condition learning and adaptive adjustment on the initial data model to generate a dynamic boundary condition model; and performs real-time monitoring of the multi-source data set of the flexible production line;

[0015] Deviation identification module: Performs cross-modal correlation analysis on potential abnormal data to generate composite deviation identification results; performs deviation type classification and severity assessment to obtain deviation classification and severity data;

[0016] Real-time correction module: Combining historical correction case library and expert knowledge base, it performs real-time reasoning on deviation classification and severity data, automatically generates correction decision instructions; sends correction decision instructions to relevant execution equipment on flexible production line, triggers real-time correction actions, and generates correction execution feedback data;

[0017] Data optimization module: Utilizes the AI ​​error prevention digital model to evaluate the effectiveness of the error correction execution feedback data, generating error correction effectiveness evaluation data; and optimizes and adjusts the parameters of the AI ​​error prevention digital model, generating optimized AI error prevention digital model parameter data;

[0018] Benefit assessment module: Based on the optimized AI error-proof digital model parameter data, the module performs full-process monitoring and data analysis on the real-time correction process of the entire flexible production line, generating data on production efficiency improvement and product quality improvement; and evaluates the overall benefits to generate comprehensive benefit assessment data.

[0019] The beneficial effects of this invention are as follows: By leveraging an AI-powered error-proofing digital model for deep integration and fusion of multi-dimensional data sources in flexible production lines for new energy vehicles, the barriers of data fragmentation in traditional models are broken down, enabling the correlation analysis of multi-modal data such as equipment, vision, and materials. This not only accurately identifies complex deviations such as incorrect tightening sequence of correct parts, but also effectively reduces misjudgments and omissions caused by isolated data, significantly improving the accuracy of production deviation identification. Utilizing the model's adaptive learning and adjustment of dynamic boundary conditions, it breaks free from the constraints of preset thresholds and finite state machine rules, flexibly responding to complex changes brought about by vehicle model switching and process modifications. During real-time monitoring, potential anomalies can be quickly captured and corrective commands can be automatically triggered, achieving millisecond-level response, successfully preventing defective products from flowing into subsequent processes, reducing rework costs, and significantly improving production efficiency. By combining historical correction case libraries and expert knowledge bases for real-time reasoning to generate correction decisions, it eliminates excessive reliance on human experience, making correction decisions more scientific, consistent, and efficient. Attached Figure Description

[0020] Figure 1 This is a diagram illustrating the steps of the method described in this invention;

[0021] Figure 2 This is a system module diagram of the present invention. Detailed Implementation

[0022] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0023] One embodiment of the present invention, such as Figure 1 As shown, a real-time deviation correction method for flexible production lines based on an AI-based error-proofing digital model includes:

[0024] S1. Configure multi-dimensional data source access for the flexible production line of new energy vehicles, covering equipment data, visual image data, and material information data; the equipment data includes torque and displacement; the visual image data includes welding points and adhesive application; the material information data includes part batches, generating a multi-source data set for the flexible production line; based on this multi-source data set for the flexible production line, use an AI-based error-proof digital model for data preprocessing and feature extraction, and construct an initial data model that integrates multi-modal features;

[0025] S2. Utilize AI-based error-proof digital big data models to dynamically learn and adaptively adjust the initial data model, breaking the rigid constraints of preset thresholds and finite state machine (FSM) rules, and generating dynamic boundary condition models that adapt to vehicle model switching and process changes; based on the dynamic boundary condition model, monitor the multi-source data set of the flexible production line in real time, and mark the data as potential abnormal data when the data exceeds the dynamic boundary conditions.

[0026] S3. Utilize the AI-powered error-proofing digital model to perform cross-modal correlation analysis on potential abnormal data, uncover the intrinsic connections between equipment data, visual image data, and material information data, accurately identify compound deviations, such as correct parts + incorrect tightening sequence, and generate compound deviation identification result data; based on the compound deviation identification result data, use the AI-powered error-proofing digital model to classify deviation types and assess their severity, obtaining deviation classification and severity data;

[0027] S4. By combining the AI ​​error-proof digital model with the historical correction case library and expert knowledge base, the deviation classification and severity data are inferred in real time to automatically generate correction decision instructions. The correction decision instructions are sent to the relevant execution equipment of the flexible production line to trigger real-time correction actions, prevent defective products from flowing into subsequent processes, and generate correction execution feedback data.

[0028] S5. Utilize the AI ​​error-proofing digital model to evaluate the effectiveness of the correction execution feedback data, analyze whether the correction actions effectively eliminate deviations, and generate correction effect evaluation data; based on the correction effect evaluation data, optimize and adjust the parameters of the AI ​​error-proofing digital model to continuously improve its accuracy in identifying abnormalities on flexible production lines and the rationality of correction decisions, and generate optimized AI error-proofing digital model parameter data.

[0029] S6. Based on the optimized AI error-proofing digital model parameter data, the entire process of real-time deviation correction on the flexible production line is monitored and analyzed to generate production efficiency improvement data and product quality improvement data. Based on the production efficiency improvement data and product quality improvement data, the overall benefits of the flexible production line real-time deviation correction method based on the AI ​​error-proofing digital model are further evaluated, and comprehensive benefit evaluation data is generated to provide a decision-making basis for the continuous optimization of the flexible production line.

[0030] The working principle and effects of the above technical solution are as follows: Through multi-dimensional data access and preprocessing of the AI ​​error-proof model, multi-modal features such as equipment, vision, and materials can be accurately extracted, significantly improving the recognition accuracy of complex deviations. Hidden problems such as correct parts and incorrect tightening sequences can also be identified in a timely manner. The dynamic boundary model breaks the rigid limitations of traditional fixed thresholds, eliminating the need for repeated parameter adjustments when switching production line models or adjusting processes, thus enhancing flexible adaptability. The model combines case libraries and knowledge bases for real-time reasoning, generating correction instructions in milliseconds, reducing reliance on manual decision-making and minimizing the time spent on anomaly handling. Instructions directly trigger corrections on the executing equipment, directly preventing defective products from flowing into subsequent processes, reducing rework costs and significantly lowering the product defect rate. Subsequently, by evaluating the correction effect and optimizing model parameters, the recognition and correction accuracy will steadily improve, making the production line operation more stable. The efficiency and quality data generated by the full-process monitoring can also provide solid evidence for production line optimization, further improving production efficiency and providing direction for continuous production line improvement.

[0031] In one embodiment of the present invention, S1 includes:

[0032] S11. Analyze the core production processes of the flexible production line for new energy vehicles. These core processes include body welding, battery assembly, chassis assembly, and adhesive sealing. Determine the multi-dimensional data types to be collected for each process: equipment operation data, visual inspection data, and material traceability data. The equipment operation data includes torque, displacement, speed, and current. The visual inspection data includes weld point morphology, adhesive width / thickness, and part assembly gaps. The material traceability data includes part batches, supplier information, warehousing time, and expiration date. Simultaneously, adapt to the communication interfaces of different industrial equipment, such as OPCUA, Profinet, and Ethernet / IP, formulate differentiated data source access protocols, and generate a flexible production line data source access solution.

[0033] S12. Based on the data source access scheme, deploy data acquisition terminals at key workstations on the production line: install sensors (torque sensors, displacement sensors) on the equipment side, set up industrial cameras (2D vision cameras for size detection, 3D vision cameras for 3D assembly posture recognition) on the vision side, and affix RFID tags and configure readers on the material side; perform preliminary aggregation and format unification of the collected real-time data through the edge computing gateway (e.g., convert equipment data into JSON format, and vision data into standardized image tensors) to generate a flexible production line multi-source data set.

[0034] S13. Input the multi-source data set of the flexible production line into the preprocessing module of the AI ​​error-proof digital model, and perform data preprocessing in sequence. The preprocessing includes cleaning (removing sensor outliers and repairing image noise), data normalization (mapping physical quantities such as torque and displacement to the [0,1] interval), and data alignment (synchronizing equipment, vision, and material data according to timestamps). Then, extract spatial features through the model's multimodal feature extraction unit (for example, using CNN (convolutional neural network) to extract spatial features such as weld defects and glue coating abnormalities in visual images), and use LSTM (long short-term memory network) to extract equipment temporal features (for example, the temporal features include torque fluctuations and speed changes, and use an attention mechanism to weight and fuse multi-dimensional features to generate a multimodal fusion feature set).

[0035] S14. Combining the basic process parameters of the flexible production line for new energy vehicles (for example, the process parameters include the production cycle time of different models, standard assembly tolerances, and equipment rated parameters), the multimodal fusion feature set is input into the initial modeling module of the AI ​​error-proof digital big model. The mapping relationship between features and production line operation status is constructed through a fully connected layer to generate an initial data model that integrates multimodal features.

[0036] The working principle and effects of the above technical solution are as follows: First, the core processes are identified and the data types of each stage are clarified. Dedicated access protocols are also developed to adapt to different equipment interfaces, avoiding confusion and incompatibility issues during data access and improving the targeting and stability of data source access. Dedicated acquisition terminals are deployed at key workstations, and data is aggregated and formatted uniformly using edge computing gateways, reducing the time spent integrating different types of data and enhancing data processing efficiency. During data preprocessing, outliers are removed, noise is repaired, and synchronization is performed, effectively improving data quality and preventing interference from inferior data in subsequent analysis. Multiple algorithms are used to extract multi-dimensional features from visual and equipment perspectives, and then an attention mechanism is used to accurately allocate weights for fusion, enhancing the effectiveness of the features and enabling the model to capture key information about production line operation. An initial model is built based on actual process parameters, ensuring the model fits the actual operating conditions of the production line from the outset, laying a solid foundation for subsequent anomaly identification and real-time correction, guaranteeing both the reliability of the data link and the practical value of the model.

[0037] In one embodiment of the present invention, S13 includes:

[0038] S131. Input the multi-source data set of the flexible production line into the preprocessing module of the AI ​​error-proof digital model to perform data cleaning: use the 3σ principle to remove sensor outliers in the equipment operation data (torque, displacement), use the median filtering algorithm to repair noise in the visual inspection images, and verify and delete invalid records (such as duplicated warehousing information) in the material traceability data according to the part batch rules to generate cleaned multi-source data; based on the cleaned multi-source data, perform normalization for different types of data: map the equipment operation data (torque 0-500N·m, displacement 0-100mm) to [Min-Max normalization] In the range of 0,1, the pixel values ​​(0-255) of the visual image are normalized to floating-point data, while maintaining the text format of the material traceability data and converting it into an encoded vector to eliminate differences in data units and generate normalized multimodal data. Based on the normalized multimodal data, the timestamps of the production line's central clock system are used as a reference to synchronize three types of data: equipment, vision, and materials. Equipment torque acquisition data (100Hz), visual image data (10fps), and material RFID reading data (1 time / part) are grouped according to the same timestamp to ensure that multidimensional data at the same time point correspond to the same production action, generating time-aligned multimodal data.

[0039] S132. Input the visual detection image from the time-aligned multimodal data into the multimodal feature extraction unit of the model. Adopt a CNN (e.g., ResNet-50) architecture, extract spatial features (e.g., solder joint edges, glue coating contours) from the image through convolutional layers, compress the feature dimension through pooling layers, and finally output a 128-dimensional visual feature vector. The visual feature vector includes key information such as solder joint defects and assembly gaps, and generates a visual spatial feature set.

[0040] S133. Input the equipment operation data in the time-aligned multimodal data into the feature extraction unit, use the LSTM network to capture the temporal patterns of torque fluctuation and speed change (e.g., the upward trend of torque during tightening), learn the temporal dependency of the data through the hidden layer, output a 64-dimensional temporal feature vector, and generate the equipment temporal feature set.

[0041] S134. Based on the visual spatial feature set and the equipment temporal feature set, an attention mechanism is introduced to assign dynamic weights to the two types of features (for example, increasing the weight of visual features when the solder joint is abnormal, and increasing the weight of temporal features when the torque fluctuates). The features are then fused into a unified 256-dimensional feature vector through a fully connected layer. At the same time, the encoded vector of the material traceability data is embedded to generate a multimodal fusion feature set.

[0042] The working principle and effects of the above technical solution are as follows: During data cleaning, the 3σ principle is used to remove equipment anomalies, median filtering is used to repair image noise, and invalid material records are deleted according to rules, effectively improving data quality and avoiding interference from inferior data in subsequent analysis. Normalization unifies the scale of different types of data, eliminating the dimensional differences of torque, pixel, and other values, and reducing errors during feature extraction. The three types of data are aligned according to the central clock timestamp to ensure that data at the same time point corresponds to the same production action, avoiding misjudgments caused by data misalignment; visual data uses CNN to extract features, which can accurately capture details such as solder joint edges and glue application contours, improving the accuracy of spatial features; equipment data uses LSTM to capture temporal patterns such as torque fluctuations, enhancing the effectiveness of temporal features. An attention mechanism dynamically allocates weights, for example, emphasizing visual features when solder joints are abnormal, and emphasizing equipment features when torque fluctuates, and then embedding material codes to make the fused features more in line with actual needs. This ensures both the accuracy and comprehensiveness of features and lays a solid foundation for subsequent modeling.

[0043] In one embodiment of the present invention, S134 includes:

[0044] Based on material traceability data in time-aligned multimodal data, which includes part batches, supplier information and warehousing time, the textual material information is converted into a 64-dimensional standardized vector (adapted to the equipment time-series feature dimension) using word embedding (Word2Vec) combined with one-hot encoding technology, eliminating the format differences between material data and other features, and generating a material coding feature set;

[0045] The visual spatial feature set and the device temporal feature set are input into the Self-Attention module. By calculating the similarity matrix and attention score of the two types of features, when a solder joint defect is detected in the visual features, the weight of the visual features is adjusted to 0.6 and the weight of the device temporal features is 0.4; when torque fluctuations occur in the device temporal features, the weight of the device temporal features is adjusted to 0.6 and the weight of the visual features is 0.4, generating a dynamic weight allocation result for the features.

[0046] Based on the dynamic weight allocation results of features, the 128-dimensional visual space feature vector and the 64-dimensional device time-series feature vector are weighted respectively (visual feature × corresponding weight + device time-series feature × corresponding weight). Then, the two types of features are merged into a 192-dimensional preliminary fusion vector by vector superposition to generate a weighted fusion feature set.

[0047] The material coding feature set (64 dimensions) and the weighted fusion feature set (192 dimensions) are concatenated in dimensional order. Any dimensional deviations are filled in with zero padding to form an initial fusion vector of 256 dimensions (192 dimensions + 64 dimensions), ensuring the complete integration of the three types of features and generating a feature concatenation vector set.

[0048] The concatenated feature vector set is input into a fully connected layer, and the nonlinear expression of the features is enhanced by the ReLU activation function. Then, the feature distribution bias is eliminated by batch normalization, and finally a standardized 256-dimensional feature vector is output to generate a multimodal fusion feature set.

[0049] The working principle and effects of the above technical solution are as follows: Word embedding combined with one-hot encoding transforms textual material data into a 64-dimensional standardized vector, perfectly matching the temporal feature dimension of the equipment. This completely solves the format incompatibility problem between material data and visual and equipment features, avoiding dimensional chaos during integration. The Self-Attention module dynamically adjusts weights according to actual working conditions; for example, it emphasizes visual features when detecting weld defects and emphasizes equipment temporal features when detecting torque fluctuations, making key features more prominent and avoiding the problem of important information being masked under fixed weights, thus improving the targeting of feature fusion. After weighted superposition of visual and equipment features, the material encoding vector is concatenated, and zero-padding is used to complete the bias, ensuring the complete integration of the three types of features and enhancing the comprehensiveness of the fused features. The fully connected layer, combined with ReLU activation and batch normalization, strengthens the non-linear expression of features and eliminates distribution bias, making the output 256-dimensional features more standardized. This entire process ensures the effective fusion of multiple types of features and ensures that the final features accurately meet the actual needs of the production line, providing high-quality feature support for subsequent model construction.

[0050] In one embodiment of the present invention, S2 includes:

[0051] S21. Import the initial data model into the dynamic learning module of the AI ​​error-proof digital model, and import the historical production data of the flexible production line for the past 3 years. The historical production data includes normal working condition data and typical abnormal working condition data, covering 10+ model switching scenarios and 20+ process change scenarios. Use transfer learning to reuse the boundary learning experience of similar production lines, and combine reinforcement learning to allow the model to autonomously adjust its boundary cognition in simulated working condition switching, explore the dynamic boundary rules of production line operation under different scenarios, and generate a dynamic boundary learning dataset.

[0052] S22. Based on the dynamic boundary learning dataset, the AI-powered error-proof digital model breaks through the fixed process thresholds of traditional finite state machines (FSMs): for example, when switching from SUV to sedan, it automatically lowers the upper limit threshold of the welding equipment torque; when switching from room temperature glue to high temperature glue, it adjusts the acceptable range of glue thickness; through iterative optimization of boundary parameters in multiple scenarios, it constructs a dynamic boundary condition model that adapts to changes in production line conditions in real time, and outputs the core boundary parameters of the model (such as equipment parameter fluctuation range, visual inspection pass threshold, and material matching rules).

[0053] S23. Deploy the dynamic boundary condition model to the flexible production line real-time monitoring system. The system samples and compares the real-time data in the multi-source data set at a frequency of 10ms / time. If the real-time torque of the equipment exceeds the fluctuation range in the dynamic boundary parameters, the diameter of the weld point in the visual image is less than the qualified threshold, or the material batch does not match the current production model, the data is automatically marked as potential abnormal data, and the time of occurrence of the abnormality, workstation number, and data dimension are recorded simultaneously.

[0054] The working principle and effects of the above technical solution are as follows: It imports historical data covering nearly 3 years of vehicle model switching and 20+ process changes, combines this with transfer learning to reuse experience from similar production lines, and then uses reinforcement learning to allow the model to autonomously adjust its cognition in simulated working conditions. The resulting dynamic boundary patterns are more realistic, completely avoiding the problem of traditional fixed thresholds failing to keep up with changes in production line conditions. Based on these patterns, the model can automatically adapt to different scenario requirements. For example, when the vehicle model changes from SUV to sedan, it will proactively lower the upper limit of welding torque; when the adhesive coating process changes to high-temperature adhesive, it will adjust the acceptable range of adhesive coating thickness, eliminating the need for repeated manual parameter adjustments. This enhances the flexibility of production line condition adaptation and reduces the cost and errors of manual intervention. The deployed real-time monitoring system samples and compares data at a frequency of 10ms / time. Once it detects torque exceeding the range, weld joint non-compliance, or material mismatch, it immediately marks the anomaly and records information such as time and workstation. This not only improves the timeliness of anomaly identification but also avoids the problem of ambiguous location after anomalies occur, laying a solid foundation for accurately identifying the root cause and quickly correcting deviations.

[0055] In one embodiment of the present invention, S3 includes:

[0056] S31. Import potential abnormal data into the cross-modal correlation analysis module of the AI ​​error-proof digital big data model, and construct the correlation relationship between equipment, vision, and material data through graph neural network (GNN): take the abnormal workstation as the core node, and associate the equipment operation data (e.g., torque change in the last 10 seconds), visual inspection data (e.g., 3 frames of assembly images of the same part), and material information data (e.g., the batch and supplier of the parts used in the workstation) of that workstation to form an abnormal correlation map, which intuitively presents the implicit connection between different data sources (e.g., the correlation path from part batch A to equipment 1 to weld point image B).

[0057] S32. Based on the abnormal correlation graph, the deviation identification unit of the AI ​​error-proof digital big data model performs in-depth analysis on potential abnormal data. The in-depth analysis includes distinguishing between single-dimensional abnormalities (e.g., only the equipment torque exceeds the standard, with no other data abnormalities) and compound deviations (e.g., torque fluctuations caused by correct part batch + incorrect tightening sequence, abnormal glue application image + material expiration date). By comparing the structural differences between the normal correlation graph and the abnormal correlation graph, the root dimension of the deviation is located (e.g., the deviation is caused by an incorrect tightening sequence, rather than a problem with the part itself), and compound deviation identification result data is generated.

[0058] S33. Based on the quality standards of the new energy vehicle industry and the process requirements of flexible production lines, construct a deviation classification system (e.g., assembly deviation, material deviation, equipment deviation, process deviation) and severity grading standards. The grading standards are as follows: Minor: Does not affect product function and can be corrected later; Moderate: Affects local performance and requires immediate adjustment; Severe: Affects core function and requires suspension of the process; Critical: Causes product scrap and requires comprehensive investigation.

[0059] S34. The AI ​​error-proof digital model is based on this system and standard. It classifies and labels composite deviation identification results data (such as assembly deviation and tightening sequence error) and quantifies them (such as fatal deviation score ≥90 points, minor deviation score ≤30 points), and obtains deviation classification and severity data.

[0060] The working principle and effects of the above technical solution are as follows: A graph neural network is used to construct an anomaly correlation map, linking equipment data, visual images, and material information at abnormal workstations. This intuitively presents implicit connections, such as those between part batches, equipment, and weld joints, avoiding the problem of missing complex deviations when only looking at single data points, thus improving the accuracy of root cause localization. Deep analysis can clearly distinguish between single anomalies and complex deviations, such as accurately identifying problems like correct parts with incorrect tightening sequences. It can also pinpoint the problem as a tightening sequence issue rather than a problem with the part itself, reducing the possibility of misjudging process problems as material problems and enhancing the accuracy of deviation identification. A classification system and severity standards are established based on industry standards, providing a unified basis for deviation classification (e.g., assembly, material deviations) and grading (minor to critical), avoiding subjective confusion during evaluation and improving the standardization of subsequent processing. Finally, the categories are labeled and quantified for scoring; for example, assembly deviations such as incorrect tightening sequences are also scored. This allows staff to quickly understand the type and severity of deviations, enabling them to prioritize critical deviations to avoid scrapping and develop targeted solutions based on the category, paving the way for efficient subsequent correction.

[0061] In one embodiment of the present invention, step S4 includes:

[0062] S41. Collect historical correction cases from the past 5 years of the flexible production line. The historical correction cases include deviation types, correction solutions, and execution effect data. Also, record the process experience of industry experts (e.g., when the glue application width is insufficient, prioritize checking the glue gun dispensing pressure). Use knowledge graph technology to structure the cases and expert knowledge, establish the mapping relationship between deviation types and correction solutions, and generate a structured knowledge base for the AI ​​error-proof digital model.

[0063] S42. Input the deviation classification and severity data into the inference module of the AI ​​error-proof digital model. The inference module uses case-based reasoning (CBR) to match historical cases with similarity ≥90% from the knowledge base, and uses rule-based reasoning (RBR) to call expert knowledge to verify the case suitability. For new types of deviations (without historical case matching), innovative correction schemes are generated through the model's autonomous decision-making algorithm (e.g., Q-Learning in reinforcement learning). Finally, correction decision instructions are automatically generated (e.g., at the chassis assembly station, the tightening sequence is adjusted to left front to right front to left rear to right rear, and the torque parameter is set to 85 N·m).

[0064] S43. Transmit the correction decision command to the execution equipment control system (e.g., PLC controller, industrial robot control cabinet) of the flexible production line via industrial Ethernet to trigger the real-time action of the execution equipment. The real-time action includes the robot adjusting the tightening sequence and torque, the glue gun adjusting the glue dispensing pressure, and the material conveying line switching part batches.

[0065] S44. Simultaneously deploy a feedback acquisition module to obtain the action status of the execution equipment in real time (e.g., successful command reception or 100% completion of action execution) and the production line data after correction (e.g., adjusted torque value, glue application width), and generate correction execution feedback data.

[0066] The working principle and effects of the above technical solution are as follows: Five years of historical error correction cases and expert process experience are collected and structured using a knowledge graph to build a knowledge base. This transforms scattered experiences such as adjusting parameters for excessive torque and checking pressure for insufficient glue application into precise mapping relationships, avoiding the inefficiency of finding solutions from memory and improving the efficiency of subsequent reasoning. The reasoning module uses case-based reasoning to match historical cases with over 90% similarity, and then uses rule-based reasoning to verify suitability. When encountering new deviations, it can autonomously generate solutions, ensuring the accuracy of correcting old problems while solving the problem of lacking evidence for new deviations, thus enhancing the ability to handle complex deviations. Instructions are directly transmitted to the execution equipment via industrial Ethernet, allowing robots to immediately adjust tightening sequences and glue guns to promptly adjust glue pressure—much faster than manual communication, improving the error correction response speed and preventing defective products from flowing into the next process. The feedback module collects action status and post-correction data in real time, such as the completion rate of instruction execution and the adjusted torque value, confirming whether the correction has been implemented and accumulating real data for subsequent model optimization. This entire process can quickly and accurately resolve deviations while continuously accumulating experience for iterative optimization.

[0067] In one embodiment of the present invention, step S5 includes:

[0068] S51. Import the feedback data of the correction execution and the deviation data before correction into the effect evaluation module of the AI ​​error-proof digital model to construct a correction effect evaluation index system. The correction effect evaluation index system includes deviation elimination rate (the proportion of normal data after correction), process recovery time (the time taken from the issuance of the instruction to the normalization of the process), and defect rate change (the difference in product defect rate before and after correction). By comparing and analyzing the differences in indicators before and after correction (for example, the torque abnormality rate was 15% before correction and dropped to 0.5% after correction), the effectiveness of the correction action is quantitatively evaluated, and correction effect evaluation data is generated.

[0069] S52. Based on the error correction effect evaluation data, identify the shortcomings of the AI ​​error prevention digital model: if the error elimination rate is less than 90%, the model is judged to have insufficient error recognition accuracy; if the process recovery time exceeds 100ms, the model is judged to have slow decision-making and reasoning speed; if the error correction effect of new types of errors is poor, the model is judged to have incomplete knowledge base coverage; determine the model parameters to be optimized (e.g., feature extraction weight, inference threshold, learning rate) and knowledge base supplementation direction according to the shortcomings, and generate a list of model optimization requirements.

[0070] S53. Based on the model optimization requirements list, the gradient descent method is used to adjust the weights of the model's feature extraction layer (e.g., increase the weights of visual image features to enhance the accuracy of weld point defect recognition), and the particle swarm optimization algorithm (PSO) is used to optimize the inference threshold (e.g., reduce the decision trigger threshold for slight deviations to reduce missed judgments). At the same time, correction cases for new types of deviations are added to the knowledge base.

[0071] S54. Through offline simulation (using historical data to verify the optimization effect) and online testing (pilot operation at one workstation on the production line), the parameters are iteratively adjusted to the optimal state, and the optimized AI error-proof digital model parameter data is generated.

[0072] The working principle and effects of the above technical solution are as follows: The feedback from the correction execution and previous deviation data are imported into the evaluation module. Quantitative indicators such as deviation elimination rate and process recovery time are established. By comparing the data before and after correction, for example, if the torque anomaly rate decreases from 15% to 0.5%, it is clear whether the correction is effective, avoiding the problem of judging the effect based on subjective feelings and improving the scientific nature of the evaluation. Based on the evaluation data, model deficiencies can be accurately identified. For example, a low elimination rate indicates insufficient recognition accuracy, and a long recovery time indicates slow inference speed. This prevents blind parameter tuning during optimization and reduces wasted effort. Gradient descent is used to adjust feature extraction weights, such as increasing visual weights to enhance weld point recognition. Particle swarm optimization optimizes the inference threshold, reducing missed detections of minor deviations, and new cases are added to the knowledge base, making the model more targeted. Offline simulation uses historical data for verification, while online pilot testing at a single workstation repeatedly adjusts parameters to the optimal level, avoiding problems that might arise from direct full-production line deployment after optimization and improving parameter reliability.

[0073] In one embodiment of the present invention, S53 includes:

[0074] Based on the model optimization requirement list, three core optimization objects are identified. These three core optimization objects include feature extraction layer weights that need to be adjusted (e.g., visual image feature weights, device timing feature weights), inference thresholds that need to be optimized (e.g., minor deviation trigger thresholds, severe deviation judgment thresholds), and new types of deviation case categories that need to be added (e.g., composite deviation cases of mis-installed parts + path offset). The target value for each optimization is determined (e.g., the visual feature weight needs to be increased from 0.4 to 0.6), and parameter and case optimization details are generated.

[0075] Based on the parameter and case optimization details, the gradient descent method is initiated with the goal of improving the deviation recognition accuracy by 10%. The initial learning rate is set to 0.001. The current weights of the feature extraction layer (e.g., visual feature weight 0.4, device time sequence feature weight 0.6) are used as the initial values. The gradient of the weights with respect to the recognition error is calculated iteratively, and the weights are gradually adjusted until the error is minimized (e.g., the visual feature weight is increased to 0.6). The adjusted feature extraction layer weight data is then generated.

[0076] Based on the threshold optimization requirements in the parameter and case optimization details data, the particle swarm optimization algorithm is run; including setting the number of particles to 50 and the number of iterations to 30, with the goal of a false negative rate of ≤1% and a false negative rate of ≤0.5%, searching for the optimal solution of the inference threshold (e.g., reducing the threshold for triggering minor deviations from the original 0.3 to 0.25), verifying the effectiveness of the threshold through 1000 sets of historical deviation data, and generating optimized inference threshold data;

[0077] Based on the supplementary case requirements in the parameter and case optimization details data, complete case information for new types of deviations is collected. The complete case information includes the deviation phenomenon, the workstation where it occurred, the correction plan, and the execution effect. It is organized in a structured format according to the deviation type, workstation, plan, and effect (e.g., part misassembly + path deviation - battery pack assembly workstation - recalibrate positioning sensor + adjust conveyor path - deviation elimination rate 98%), and imported into the knowledge base of the AI ​​error-proof digital model to generate updated knowledge base data.

[0078] The adjusted feature extraction layer weight data, optimized inference threshold data, and updated knowledge base data are correlated and verified to ensure that the weight adjustment matches the threshold optimization (for example, after the visual feature weight is increased, the corresponding deviation recognition threshold is adapted synchronously), the case supplement is compatible with the model inference logic, and the model parameters and knowledge base optimization integration data are generated.

[0079] The working principle and effects of the above technical solution are as follows: First, the optimization requirements are broken down into three categories: weights, thresholds, and cases. Target values ​​are also clearly defined, such as increasing the visual weight from 0.4 to 0.6. This provides a clear direction for optimization, avoiding blindly adjusting parameters and improving the targeting of the optimization. Gradient descent is used to adjust the feature extraction weights, aiming to improve recognition accuracy by 10%. The gradient is iteratively calculated until the error is minimized; for example, the visual weight is precisely adjusted to 0.6, effectively enhancing the recognition ability of visually related deviations such as weld defects. When optimizing the threshold using the particle swarm optimization algorithm, the number of particles and the number of iterations are set, and 1000 sets of historical data are used for verification, keeping the false negative rate within 1% and the false positive rate within 0.5%, reducing the chance of slight deviations being missed or misjudged under normal operating conditions. New cases are added in a structured format; for example, cases of incorrect part assembly + path offset clearly state the workstation, solution, and effect, enriching the knowledge base and allowing the model to handle more new types of deviations. Finally, the weights, thresholds, and cases are correlated and verified to ensure that the recognition thresholds are also adapted after the visual weights are increased. This avoids incompatibility issues between different parts after optimization, making the optimization effect more reliable. It can accurately optimize model parameters and improve the knowledge base, allowing the model performance to steadily improve.

[0080] In one embodiment of the present invention, step S6 includes:

[0081] S61. Deploy the optimized AI error-proof digital model parameter data to the flexible production line full-process monitoring platform. The platform collects the full-process data of the correction process at a frequency of 5ms / time. The full-process data of the correction process includes the time consumed for anomaly identification, the response time of correction command, the accuracy of the execution equipment action, the product pass rate, and the equipment downtime. The production line operation status (e.g., the ranking of the anomaly occurrence rate of each workstation) and the correction effect (e.g., the trend of the number of corrections per day) are displayed in real time through data visualization tools (e.g., Dashboard, heat map), and real-time monitoring data reports are generated.

[0082] S62. Based on real-time monitoring data reports, the data analysis module of the AI ​​error-proof digital big data model calculates core benefit indicators, including production efficiency improvement indicators (output per unit time increase rate = (daily output after optimization - daily output before optimization) / daily output before optimization × 100%, process waiting time reduction rate = (waiting time before optimization - waiting time after optimization) / waiting time before optimization × 100%); product quality improvement indicators (defect rate reduction rate = (defect rate before optimization - defect rate after optimization) / defect rate before optimization × 100%, rework rate reduction rate = (rework rate before optimization - rework rate after optimization) / rework rate before optimization × 100%), generating production efficiency improvement data and product quality improvement data;

[0083] S63. Construct a comprehensive benefit assessment model to transform production efficiency improvement data into economic benefits (e.g., unit product cost reduction = (unit cost before optimization - unit cost after optimization) × annual output), and product quality improvement data into quality benefits (e.g., customer complaint reduction = number of complaints before optimization - number of complaints after optimization; brand reputation improvement value is calculated according to industry benchmarks), while deducting model deployment costs (hardware investment, software maintenance, personnel training costs); generate comprehensive benefit assessment data through corresponding dimensions, including return on investment (ROI = (total benefits - total costs) / total costs × 100%) and long-term benefit forecast (benefit growth trend over the next 3 years); and form a flexible production line continuous optimization report.

[0084] The working principle and effects of the above technical solution are as follows: The optimized model parameters are deployed to a full-process monitoring platform, collecting data such as anomaly identification time and equipment movement accuracy at a frequency of 5ms / time. Visual tools are also used to display workstation anomaly rankings and daily correction trends, allowing staff to monitor production line dynamics in real time, avoiding delays in detecting anomalies and improving monitoring timeliness. The data analysis module calculates quantitative indicators such as output increase rate and defect rate reduction rate. For example, formulas are used to calculate how much output increases per unit time and how much rework rate decreases, providing clear data support for efficiency and quality improvements, avoiding the ambiguity of subjective evaluations, and enhancing the accuracy of assessments. The comprehensive benefit evaluation model translates efficiency improvements into unit cost reductions, quality improvements into reduced complaints and enhanced brand value, and deducts hardware and maintenance costs. It then comprehensively measures these improvements through ROI and three-year revenue forecasts, avoiding one-sided evaluations that only consider returns without considering inputs, making the benefit situation more realistic. The resulting continuous optimization report clarifies the direction for future improvements, enabling managers to understand the practical value of the methods and providing solid evidence for further production line optimization, thus forming a closed loop of monitoring, evaluation, and optimization.

[0085] One embodiment of the present invention, such as Figure 2 As shown, a system for implementing the real-time deviation correction method for flexible production lines based on the AI-based error-proofing digital model as described above is provided, the system comprising:

[0086] Model building module: Configure multi-dimensional data source access for the flexible production line of new energy vehicles to generate a multi-source data set of the flexible production line; perform data preprocessing and feature extraction to build an initial data model that integrates multi-modal features;

[0087] Real-time monitoring module: performs dynamic boundary condition learning and adaptive adjustment on the initial data model to generate a dynamic boundary condition model; and performs real-time monitoring of the multi-source data set of the flexible production line;

[0088] Deviation identification module: Performs cross-modal correlation analysis on potential abnormal data to generate composite deviation identification results; performs deviation type classification and severity assessment to obtain deviation classification and severity data;

[0089] Real-time correction module: Combining historical correction case library and expert knowledge base, it performs real-time reasoning on deviation classification and severity data, automatically generates correction decision instructions; sends correction decision instructions to relevant execution equipment on flexible production line, triggers real-time correction actions, and generates correction execution feedback data;

[0090] Data optimization module: Utilizes the AI ​​error prevention digital model to evaluate the effectiveness of the error correction execution feedback data, generating error correction effectiveness evaluation data; and optimizes and adjusts the parameters of the AI ​​error prevention digital model, generating optimized AI error prevention digital model parameter data;

[0091] Benefit assessment module: Based on the optimized AI error-proof digital model parameter data, the module performs full-process monitoring and data analysis on the real-time correction process of the entire flexible production line, generating data on production efficiency improvement and product quality improvement; and evaluates the overall benefits to generate comprehensive benefit assessment data.

[0092] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A real-time deviation correction method for flexible production lines based on an AI-based error-proofing digital model, characterized in that: The method includes: S1. Configure multi-dimensional data source access for the flexible production line of new energy vehicles to generate a multi-source data set for the flexible production line; perform data preprocessing and feature extraction to construct an initial data model that integrates multi-modal features; S2. Perform dynamic boundary condition learning and adaptive adjustment on the initial data model to generate a dynamic boundary condition model; and monitor the multi-source data set of the flexible production line in real time. S3. Perform cross-modal correlation analysis on potential abnormal data to generate composite deviation identification results data; perform deviation type classification and severity assessment to obtain deviation classification and severity data. S4. Combining historical correction case database and expert knowledge base, perform real-time reasoning on deviation classification and severity data to automatically generate correction decision instructions; send the correction decision instructions to relevant execution equipment on the flexible production line to trigger real-time correction actions and generate correction execution feedback data. S5. Use the AI ​​error prevention digital model to evaluate the effect of the error correction execution feedback data and generate error correction effect evaluation data; and optimize and adjust the parameters of the AI ​​error prevention digital model to generate optimized AI error prevention digital model parameter data. S6. Based on the optimized AI error-proof digital model parameter data, the entire flexible production line real-time correction process is monitored and analyzed to generate production efficiency improvement data and product quality improvement data; and the overall benefits are evaluated to generate comprehensive benefit assessment data.

2. The real-time deviation correction method for flexible production lines based on an AI-based error-proofing digital model as described in claim 1, characterized in that, S1 includes: S11. Analyze the core production processes of the flexible production line for new energy vehicles, determine the multi-dimensional data types that need to be collected for each process, formulate differentiated data source access protocols, and generate a flexible production line data source access plan. S12. Based on the data source access scheme, deploy data acquisition terminals at key workstations on the production line; use the edge computing gateway to perform preliminary aggregation and format unification of the collected real-time data, and generate a flexible production line multi-source data set; S13. Input the multi-source data set of the flexible production line into the preprocessing module of the AI ​​error-proof digital model, perform data preprocessing in sequence, extract spatial features through the multimodal feature extraction unit of the model, extract equipment temporal features using LSTM, and use attention mechanism to weight and fuse multi-dimensional features to generate a multimodal fusion feature set. S14. Combining the basic process parameters of the flexible production line for new energy vehicles, the multimodal fusion feature set is input into the initial modeling module of the AI ​​error-proof digital big model. The mapping relationship between features and production line operation status is constructed through a fully connected layer to generate an initial data model that integrates multimodal features.

3. The real-time deviation correction method for flexible production lines based on an AI-based error-proofing digital model as described in claim 2, characterized in that, S13 includes: S131. Input the multi-source data set of the flexible production line into the preprocessing module of the AI ​​error-proof digital big data model, perform data cleaning, and generate cleaned multi-source data; based on the cleaned multi-source data, perform normalization for different types of data to generate normalized multimodal data; based on the normalized multimodal data, use the timestamp of the production line central clock system as the benchmark to synchronize the three types of data: equipment, vision, and materials to generate time-aligned multimodal data. S132. Input the visual detection images from the time-aligned multimodal data into the multimodal feature extraction unit of the model. Using a CNN architecture, the spatial features in the images are extracted through convolutional layers and the feature dimensions are compressed through pooling layers. Finally, a 128-dimensional visual feature vector is output to generate a visual spatial feature set. S133. Input the equipment operation data in the time-aligned multimodal data into the feature extraction unit, use the LSTM network to capture the temporal patterns of torque fluctuation and speed change, learn the temporal dependency of the data through the hidden layer, output a 64-dimensional temporal feature vector, and generate the equipment temporal feature set. S134. Based on the visual spatial feature set and the device temporal feature set, an attention mechanism is introduced to assign dynamic weights to the two types of features. The features are then fused into a unified 256-dimensional feature vector through a fully connected layer. At the same time, the encoded vector of the material traceability data is embedded to generate a multimodal fusion feature set.

4. The real-time deviation correction method for flexible production lines based on an AI-based error-proofing digital model according to claim 3, characterized in that, S134 includes: Based on material traceability data in time-aligned multimodal data, word embedding combined with one-hot coding technology is used to convert textual material information into 64-dimensional standardized vectors to generate material coding feature sets; The visual spatial feature set and the device temporal feature set are input into the Self-Attention module. By calculating the similarity matrix and attention score of the two types of features, the dynamic weight allocation result of the features is generated. Based on the dynamic weight allocation results of features, weighted calculations are performed on the 128-dimensional visual space feature vector and the 64-dimensional device temporal feature vector respectively. Then, the two types of features are merged into a 192-dimensional preliminary fusion vector by vector superposition to generate a weighted fusion feature set. The material coding feature set and the weighted fusion feature set are concatenated in dimensional order, and any dimensional deviations are filled with zero padding to form an initial fusion vector of 256 dimensions, which then generates a feature concatenation vector set. The concatenated feature vector set is input into a fully connected layer, and the nonlinear expression of the features is enhanced by the ReLU activation function. Then, the feature distribution bias is eliminated by batch normalization, and finally a standardized 256-dimensional feature vector is output to generate a multimodal fusion feature set.

5. The real-time deviation correction method for flexible production lines based on an AI-based error-proofing digital model according to claim 1, characterized in that, S2 includes: S21. Import the initial data model into the dynamic learning module of the AI ​​error-proof digital model, and import the historical production data of the flexible production line for the past 3 years; use transfer learning to reuse the boundary learning experience of similar production lines, and combine reinforcement learning to allow the model to autonomously adjust its boundary cognition in simulated working condition switching, explore the dynamic boundary rules of production line operation under different scenarios, and generate a dynamic boundary learning dataset. S22. Based on the dynamic boundary learning dataset, the AI ​​error-proof digital big model breaks through the fixed process threshold of the traditional finite state machine; through iterative optimization of boundary parameters in multiple scenarios, a dynamic boundary condition model that adapts to changes in production line conditions in real time is constructed, and the core boundary parameters of the model are output. S23. Deploy the dynamic boundary condition model to the flexible production line real-time monitoring system. The system samples and compares the real-time data in the multi-source data set at a frequency of 10ms / time. If the real-time torque of the equipment exceeds the fluctuation range in the dynamic boundary parameters, the diameter of the weld point in the visual image is less than the qualified threshold, or the material batch does not match the current production model, the data is automatically marked as potential abnormal data, and the time of occurrence of the abnormality, workstation number, and data dimension are recorded simultaneously.

6. The real-time deviation correction method for flexible production lines based on an AI-based error-proofing digital model according to claim 1, characterized in that, The S3 includes: S31. Import potential abnormal data into the cross-modal correlation analysis module of the AI ​​error-proof digital big data model, and construct the correlation relationship between equipment, vision, and material data through graph neural network to form an abnormal correlation map. S32. Based on the abnormal correlation graph, the deviation identification unit of the AI ​​error-proof digital big data model performs in-depth analysis of potential abnormal data. By comparing the structural differences between the normal correlation graph and the abnormal correlation graph, it locates the root dimension of the deviation and generates composite deviation identification result data. S33. Based on the quality standards of the new energy vehicle industry and the process requirements of flexible production lines, construct a deviation classification system and severity grading standards; S34, the AI ​​error-proof digital model, based on this system and standard, classifies, labels, and quantifies the composite deviation identification results data to obtain deviation classification and severity data.

7. The real-time deviation correction method for flexible production lines based on an AI-based error-proofing digital model according to claim 1, characterized in that, The S4 includes: S41. Collect historical correction cases from the past 5 years of flexible production lines and input process experience; use knowledge graph technology to structure the cases and expert knowledge, establish the mapping relationship between deviation types and correction schemes, and generate a structured knowledge base for the AI ​​error-proof digital model. S42. Input the deviation classification and severity data into the inference module of the AI ​​error prevention digital model, and finally automatically generate the correction decision instruction; S43. Transmit the correction decision command to the execution equipment control system of the flexible production line via industrial Ethernet to trigger the real-time action of the execution equipment. S44. Simultaneously deploy a feedback acquisition module to obtain the action status of the execution equipment and the production line data after correction in real time, and generate correction execution feedback data.

8. The real-time deviation correction method for flexible production lines based on an AI-based error-proofing digital model according to claim 1, characterized in that, The S5 includes: S51. Import the feedback data of the correction execution and the deviation data before correction into the effect evaluation module of the AI ​​error prevention digital model to construct a correction effect evaluation index system; by comparing and analyzing the differences in indicators before and after correction, quantify the effectiveness of the correction action and generate correction effect evaluation data. S52. Based on the error correction effect evaluation data, identify the shortcomings of the AI ​​error prevention digital model; determine the model parameters to be optimized and the direction of knowledge base supplementation according to the shortcomings, and generate a list of model optimization requirements. S53. Based on the model optimization requirements list, the gradient descent method is used to adjust the weights of the model's feature extraction layer, the particle swarm optimization algorithm is used to optimize the inference threshold, and new types of bias correction cases are added to the knowledge base. S54. Through offline simulation and online testing, the parameters are iteratively adjusted to the optimal state to generate optimized AI error-proof digital model parameter data.

9. The real-time deviation correction method for flexible production lines based on an AI-based error-proofing digital model according to claim 1, characterized in that, S53 includes: Based on the model optimization requirement list, three core optimization objects are identified, and the target value for each optimization is determined, generating parameter and case optimization details data. Based on the parameter and case optimization details, the gradient descent method is launched with the goal of improving the deviation recognition accuracy by 10%. The initial learning rate is set to 0.

001. The current weights of the feature extraction layer are used as the initial values. The gradient of the weights with respect to the recognition error is calculated iteratively. The weights are gradually adjusted until the error is minimized, and the adjusted feature extraction layer weight data is generated. Based on the threshold optimization requirements in the parameter and case optimization details data, the particle swarm optimization algorithm is run; including setting the number of particles to 50 and the number of iterations to 30, with the goal of a false negative rate of ≤1% and a false negative rate of ≤0.5%, searching for the optimal solution of the inference threshold, verifying the effectiveness of the threshold through 1000 sets of historical deviation data, and generating optimized inference threshold data; Based on the parameter and case optimization details, we collect complete case information of new types of deviations, import it into the knowledge base of the AI ​​error-proof digital model, and generate updated knowledge base data. The adjusted feature extraction layer weight data, optimized inference threshold data, and updated knowledge base data are correlated and verified. Case supplementation is compatible with model inference logic, and model parameters and knowledge base optimization integration data are generated.

10. A system for implementing the real-time deviation correction method for flexible production lines based on an AI-based error-proofing digital model as described in claim 1, characterized in that, The system includes: Model building module: Configure multi-dimensional data source access for the flexible production line of new energy vehicles to generate a multi-source data set of the flexible production line; perform data preprocessing and feature extraction to build an initial data model that integrates multi-modal features; Real-time monitoring module: performs dynamic boundary condition learning and adaptive adjustment on the initial data model to generate a dynamic boundary condition model; and performs real-time monitoring of the multi-source data set of the flexible production line; Deviation identification module: Performs cross-modal correlation analysis on potential abnormal data to generate composite deviation identification results; performs deviation type classification and severity assessment to obtain deviation classification and severity data; Real-time correction module: Combining historical correction case library and expert knowledge base, it performs real-time reasoning on deviation classification and severity data, automatically generates correction decision instructions; sends correction decision instructions to relevant execution equipment on flexible production line, triggers real-time correction actions, and generates correction execution feedback data; Data optimization module: Utilizes the AI ​​error prevention digital model to evaluate the effectiveness of the error correction execution feedback data, generating error correction effectiveness evaluation data; and optimizes and adjusts the parameters of the AI ​​error prevention digital model, generating optimized AI error prevention digital model parameter data; Benefit assessment module: Based on the optimized AI error-proof digital model parameter data, the module performs full-process monitoring and data analysis on the real-time correction process of the entire flexible production line, generating data on production efficiency improvement and product quality improvement; and evaluates the overall benefits to generate comprehensive benefit assessment data.

Images