A vehicle electric control assembly plant monitoring management method

Abstract

The present application relates to a kind of vehicle electric control assembly workshop monitoring management method, belong to workshop monitoring management technical field.The method includes: through the tightening experiment of covering all height screw, collection qualified and failure torque-angle curve, calculate slope safety threshold, slope attenuation coefficient and stroke coefficient as physical parameter, through operating condition parameter vectorization and with physical parameter as label establish parameter library;Based on the parameter library constructed, establish identification model and compensation model, and optimize model by transfer learning;By reading bolt specification, real-time generation static feature vector using identification model, using compensation model to fuse static feature and real-time torque sequence updates physical parameter, and through double-strategy monitoring tightening state, classify abnormal type when failure.The present application constructs screw physical parameter library, combines static operating condition identification and dynamic friction compensation model, utilizes real-time matching bolt specification and adopts double-strategy monitoring torque curve, realizes height self-adapting tightening control.

Description

Technical Field

[0001] This invention belongs to the field of workshop monitoring and management technology, and specifically relates to a monitoring and management method for a vehicle electronic control assembly workshop. Background Technology

[0002] Traditional electronic assembly technology still largely relies on manual tightening guns for the circuit board tightening process, depending on operators visually confirming that the screws on the circuit board are properly tightened. Some semi-automated equipment uses screw counting to control the tightening and reduce the risk of missing screws. However, manual visual inspection and counting methods suffer from low accuracy and susceptibility to human error, making it difficult to effectively guarantee assembly quality. To improve the stability and reliability of the tightening process, current methods introduce fixed torque thresholds to monitor the tightening process, thereby compensating to some extent for the shortcomings of traditional technologies in quality control.

[0003] However, circuit boards often contain connectors and other structures, leading to variations in screw tightening depths at different locations. Although some tightening guns can identify torque values, situations may still arise where the torque is sufficient but the screw is not fully tightened. Under the same procedure, monitoring the tightening angle relies on static thresholds, which cannot effectively distinguish between screws at different depths. Furthermore, actual workshop environments are complex, with various screw sizes used at different heights on circuit boards. In addition, factors such as oil contamination and temperature changes can cause fluctuations in the coefficient of friction, and transient faults such as loose screws and thread damage frequently occur, making it difficult for a single threshold monitoring method to handle these diverse operational scenarios. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a monitoring and management method for vehicle electronic control assembly workshops. By constructing a screw physical parameter library, combining static working condition identification and dynamic friction compensation models, and utilizing real-time matching of bolt specifications and employing a dual-strategy monitoring torque curve, highly adaptive tightening control is achieved.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] This disclosure provides a method for monitoring and managing a vehicle electronic control assembly workshop, including the following steps:

[0007] S1. Parameter library construction: By collecting qualified and faulty torque-angle curves through tightening experiments covering all height screws, the slope safety threshold, slope attenuation coefficient and stroke coefficient are calculated as physical parameters. The parameter library is established by vectorizing the working condition parameters and using the physical parameters as labels.

[0008] S2. Machine learning model training: Based on the constructed parameter library, establish recognition and compensation models, and optimize the models through transfer learning;

[0009] S3. Tightening Status Monitoring: By reading the bolt specifications, a static feature vector is generated in real time using an identification model. A compensation model is used to fuse the static features with the real-time torque sequence to update the physical parameters. The tightening status is monitored through a dual strategy, and the abnormality type is classified when a fault occurs.

[0010] Furthermore, the construction of the parameter library includes the following steps:

[0011] S11. Data Acquisition: Tightening tests were conducted using screws covering all heights, and torque-angle curves for both qualified and faulty screws were collected for each category; faults included floating screws, thread damage, foreign metal objects, misalignment, and incorrect model.

[0012] S12. Parameter initialization: Based on the qualified torque-angle curve, the bolt deformation stage is divided into the screw-in stage, the tension stage and the final tightening stage. The slope of each stage of the qualified curve is calculated. The safe threshold range is obtained by eliminating 5% of the extreme values. At the same time, strain gauges are attached to the bolt shank. During the process of tightening the bolt to fracture, the point of sudden increase in strain is defined as the yield point. The slope attenuation coefficient and the stroke coefficient are calculated.

[0013] S13. Dataset Integration: By vectorizing operating parameters and generating datasets labeled with physical parameters, a parameter library is constructed.

[0014] Furthermore, the formula for calculating the slope attenuation coefficient β is as follows:

[0015] ;

[0016] In the formula, K e K represents the average slope of the stretching segment. y Where is the instantaneous slope at the yield point, and H is the bolt height;

[0017] The formula for calculating the travel coefficient γ is as follows:

[0018] ;

[0019] In the formula, θ y Let θ be the yield point angle. e The angle at the end of the stretching segment.

[0020] Furthermore, the training of the machine learning model includes the following steps:

[0021] S21. Working condition identification: Working condition identification is completed by establishing an identification model, and the ratio threshold and slope attenuation coefficient are obtained by static feature mapping of the working condition.

[0022] S22. Dynamic compensation: By establishing a compensation model to obtain friction fluctuations and instantaneous faults, dynamic compensation of the ratio threshold and slope attenuation coefficient is achieved.

[0023] S23. Performance Optimization: Enhance fault feature extraction through transfer learning, input the output features into the recognition model and compensation model, and optimize the model.

[0024] Furthermore, the operating condition identification includes the following steps:

[0025] S211. Feature Engineering: Convert the original features, such as bolt height, thread specification, material hardness, mounting surface roughness, and circuit board temperature, into feature vectors through encoding.

[0026] S212. Establish a CNN-Transformer model: Establish a model architecture that includes convolutional layers and a Transformer encoder, where convolutional layers are used to extract local feature combinations, and the Transformer encoder is used to calculate the global dependencies between features, with the output being a set of parameters including a ratio threshold and a slope decay coefficient.

[0027] S213. Model Training: Input the feature vector into the CNN-Transformer model to obtain the output prediction, calculate the mean square error between the predicted value and the working condition label in the parameter library, and use the optimizer to update the weights.

[0028] Furthermore, the dynamic compensation includes the following steps:

[0029] S221. Establish an LSTM-GRU model: With the torque sequence of the spiral segment as input and the correction amount of the ratio threshold and slope attenuation coefficient as output, construct a compensation model for parallel processing of LSTM and GRU, and splice the outputs of LSTM and GRU through a feature fusion layer.

[0030] S222, Model Training: By inputting torque sequence data and deviation label data, applying time-series Dropout forward propagation, calculating loss, and then injecting gradient noise for reverse update, a dynamic correction model that can predict operating condition deviations is finally obtained.

[0031] Furthermore, the performance optimization includes the following steps:

[0032] The screw-in torque sequence is converted into a time-spectrum image using STFT transformation; a ResNet architecture is adopted, and the first layer convolutional kernel is modified to adapt to the long strip structure of the spectrum image; the number of input channels is adjusted from RGB three channels to single channel to match the grayscale torque spectrum input; the ImageNet pre-trained weights of the first 3 ResBlocks are retained to maintain the general feature extraction capability.

[0033] Furthermore, the performance optimization also includes the steps of: unfreezing and training the last ResBlock and adaptive pooling layer to learn screw condition-specific features; compressing features through a global average pooling layer to output a 512-dimensional high semantic fault feature vector.

[0034] Furthermore, the tightening status monitoring includes the following steps:

[0035] S31. Parameter preloading: When the device starts up, it preloads the machine learning model and parameter library to establish a sensor communication link. The parameter library data includes ratio threshold, slope attenuation coefficient and height parameter mapping table.

[0036] S32. Screw identification: RFID is used to read the bolt specifications, and a static feature vector is generated in real time through the identification model to obtain the initial ratio threshold and slope attenuation coefficient.

[0037] S33. Real-time dynamic monitoring: Torque angle data is collected at 0.5° intervals, and moving average filtering and outlier removal are performed; a compensation model is used to fuse static features with real-time torque sequences, outputting the correction amount of ratio threshold and slope attenuation coefficient, and updating the actual ratio threshold and slope attenuation coefficient.

[0038] S34. Fault Diagnosis: When the alarm mechanism alarms, the data is frozen, the fault classifier is called to identify the anomaly type, and the anomaly segment curve and static feature vector are input into the fault classifier to obtain the fault type.

[0039] S35. Data closed-loop collection: Store all tightening curves and model decision data, and label low-confidence samples for self-evolution.

[0040] Furthermore, the real-time dynamic monitoring also includes the following steps:

[0041] Anomaly alerts are generated through a dual-strategy monitoring approach, employing both ratio monitoring and gradient-based monitoring.

[0042] Ratio monitoring: Verify whether the slope of each segment is within the actual ratio threshold range;

[0043] Gradient method monitoring: Real-time detection of whether slope decay triggers yield point or floating nail conditions.

[0044] The beneficial effects of this invention are as follows:

[0045] This invention optimizes bolt quality control during the tightening process through a systematic parameter library construction, machine learning model training, and real-time monitoring. First, tightening experiments covering all bolt heights are conducted to collect torque-angle curves for both faults and successes, and to calculate the slope safety threshold, slope decay coefficient, and stroke coefficient, establishing a database containing labels for operating conditions and physical parameters. Then, a CNN-Transformer model is used for static feature recognition under different operating conditions, while an LSTM-GRU model dynamically compensates for torque fluctuations and instantaneous faults, optimizing control accuracy during bolt tightening. Next, RFID technology is used to identify bolt specifications, torque data is collected in real time, and anomalies are detected through dual-strategy monitoring, ensuring timely alarms and classification when faults occur. Finally, the entire system collects tightening curves and model decision data through a data closed-loop mechanism for self-optimization and fault diagnosis. This method, through machine learning optimization and height-adaptive process parameter adjustment, effectively overcomes problems such as height differences and oversensitivity of small-stroke bolts, improving the precision of bolt tightening and the accuracy of fault diagnosis, ensuring the stability and reliability of the assembly process.

[0046] This invention quantifies the essential parameters of bolt deformation through physical experiments and constructs a static-dynamic AI collaborative monitoring architecture: CNN-Transformer identifies static working conditions and generates adaptive thresholds, while LSTM-GRU integrates real-time torque sequences to compensate for dynamic friction fluctuations. Furthermore, time-spectrum graph transfer learning is introduced to extract high-dimensional fault features, enhancing the generalization ability of the dual models and achieving a dual closed-loop drive of physical laws and data intelligence. By fusing material mechanical properties, dynamic response to working conditions, and fault spectrum fingerprints, a complete closed loop from physical essence to data intelligence is formed, fundamentally solving the industrial monitoring challenges of highly differentiated and complex fault scenarios. Attached Figure Description

[0047] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0048] Figure 1 This is a schematic diagram illustrating the steps of a vehicle electronic control assembly workshop monitoring and management method provided in an embodiment of the present invention. Detailed Implementation

[0049] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0050] This invention addresses the issue of inconsistent screw tightening depths at different locations on circuit boards due to connectors and other structures on the circuit board, which causes tightening guns to fail in determining the tightening process based on torque. It proposes a method for judging the tightening status based on the joint monitoring of torque and angle. Specifically, the following details this method for monitoring and managing vehicle electronic control assembly workshops:

[0051] This embodiment provides a monitoring and management method for a vehicle electronic control assembly workshop, such as... Figure 1 As shown, it includes the following steps:

[0052] S1. Parameter Library Construction: Pass and fail torque-angle curves are collected through tightening experiments covering all screw heights. The slope safety threshold, slope attenuation coefficient, and stroke coefficient are calculated as physical parameters. A parameter library is built by vectorizing the operating parameters and labeling them with the physical parameters, including the following steps:

[0053] S11. Data Acquisition: Tightening tests were conducted using screws covering all heights, and torque-angle curves for both qualified and faulty screws were collected for each category; faults included floating screws, thread damage, foreign metal objects, misalignment, and incorrect model.

[0054] It should be noted that in the fault curve experiment, a 0.05mm polyimide film was pre-placed at the bottom of the threaded hole to simulate non-adhesion; burrs were generated by machining the threaded hole with a tap with a wear of 0.2mm to simulate thread damage; chips were injected to simulate foreign metal objects; the tightening gun was tilted at 5° to simulate misalignment; and screws with mixed length deviations were used to simulate incorrect screw type. The torque-angle curve is the core data carrier characterizing the bolt tightening process. Fault simulations ensure that the collected data covers scenarios with abnormal curve shapes and yield point deviations. The data acquisition equipment includes a torque sensor and an optical encoder, which record changes of 0.1 N·m and the screw rotation trajectory, respectively.

[0055] S12. Parameter initialization: Based on the qualified torque-angle curve, the bolt deformation stage is divided into the screw-in stage, the tension stage, and the final tightening stage. The slope of each stage of the qualified curve is calculated. The safe threshold range is obtained by eliminating 5% of the extreme values. At the same time, strain gauges are attached to the bolt shank. During the bolt tightening process until fracture, the point of sudden increase in strain is defined as the yield point. The slope attenuation coefficient and the stroke coefficient are calculated.

[0056] It should be noted that the end point of the screw-in section is determined by the inflection point of the first derivative of the torque curve, the end point of the tension section is determined by the zero point of the second derivative of the torque curve, and the end point of the final tightening section is determined by the bolt entering the plastic deformation zone (irreversible damage). The slope attenuation coefficient is determined by the ratio of the yield point slope reduction rate to the bolt height; the slope reduction rate is the ratio of the difference between the average slope of the tension section and the instantaneous slope of the yield point to the average slope of the tension section; and the stroke coefficient is the ratio of the difference between the yield point angle and the angle at the end of the tension section to the bolt height.

[0057] Specifically, the formula for calculating the slope attenuation coefficient β is as follows:

[0058] ;

[0059] In the formula, K e K represents the average slope of the stretching segment. y Here, denoted by , represents the instantaneous slope at the yield point, and H represents the bolt height.

[0060] The formula for calculating the travel coefficient γ is:

[0061] ;

[0062] In the formula, θ y Let θ be the yield point angle. e The angle at the end of the stretching segment.

[0063] S13. Dataset Integration: By vectorizing operating parameters and generating datasets labeled with physical parameters, a parameter library is constructed.

[0064] It should be noted that the ratio monitoring threshold is established based on the statistical slope of the qualified curve through parameter initialization, and the gradient method β and γ parameters are calculated by determining the yield point through destructive experiments, thereby digitizing the physical properties of the bolt. The operating parameters include bolt height, thread specification, material hardness, mounting surface roughness, and circuit board temperature.

[0065] S2. Machine Learning Model Training: Based on the constructed parameter library, establish recognition and compensation models, and optimize the models through transfer learning, including the following steps:

[0066] S21. Operating Condition Identification: Operating condition identification is completed by establishing an identification model. The ratio threshold and slope attenuation coefficient are obtained through static feature mapping of the operating condition, including the following steps:

[0067] S211. Feature Engineering: Convert the original features, such as bolt height, thread specifications, material hardness, mounting surface roughness, and circuit board temperature, into feature vectors through encoding.

[0068] Understandably, the original feature numerical vectors are transformed, and the original features are then structured and encoded before being input into the subsequent machine learning model.

[0069] S212. Establish a CNN-Transformer model: Establish a model architecture that includes convolutional layers and a Transformer encoder. The convolutional layers are used to extract local feature combinations (such as the combined effect of height and hardness), and the Transformer encoder is used to calculate the global dependencies between features (such as the amplification effect of temperature on roughness). The output is a set of parameters including the ratio threshold and the slope decay coefficient.

[0070] S213. Model Training: Input the feature vector into the CNN-Transformer model to obtain the output prediction, calculate the mean square error between the predicted value and the working condition label in the parameter library, and use the optimizer to update the weights.

[0071] The CNN-Transformer model is a hybrid model that combines a convolutional neural network (CNN) and a transformer (Transformer) structure. It uses CNN for feature extraction and Transformer for sequence capture.

[0072] S22. Dynamic Compensation: By establishing a compensation model to obtain friction fluctuations and instantaneous faults, dynamic compensation of the ratio threshold and slope attenuation coefficient is achieved, including the following steps:

[0073] S221. Establish an LSTM-GRU model: With the torque sequence of the spiral segment as input and the correction amount of the ratio threshold and slope decay coefficient as output, construct a compensation model with LSTM (Long Short-Term Memory Network) and GRU (Gated Recurrent Unit) in parallel processing, and splice the outputs of LSTM and GRU through a feature fusion layer.

[0074] S222, Model Training: By inputting torque sequence data and deviation label data, applying time-series Dropout forward propagation, calculating loss, and then injecting gradient noise for reverse update, a dynamic correction model that can predict operating condition deviations is finally obtained.

[0075] It should be noted that the training data also includes ratio threshold and slope decay bias label data during the training process. The bias label data is obtained through test experiments on fault curves during the construction of the parameter library. Golden labels are obtained for different fault conditions. During the model training phase, the bias is mapped by learning fault features, thereby solving the problem of generalization.

[0076] S23. Performance Optimization: Enhance fault feature extraction through transfer learning, input the output features into the recognition model and compensation model, and optimize the model, including the following steps:

[0077] The screw-in torque sequence was converted into a time-spectrum image using STFT (Short Time Fourier Transform) (window length = 16 points, overlap rate 75%). A ResNet (Residual Network) architecture was adopted, and the first layer convolutional kernel was modified to 3×15 to adapt to the long strip structure of the spectrum image. The number of input channels was adjusted from RGB three channels to single channel to match the grayscale torque spectrum input. The ImageNet pre-trained weights of the first 3 ResBlocks were retained to maintain the general feature extraction capability. The last ResBlock and the adaptive pooling layer were unfrozen and trained to learn the screw condition-specific features. The features were compressed through a global average pooling layer to output a 512-dimensional high semantic fault feature vector.

[0078] It should be noted that the time-spectrum graph contains more fault mode information (such as high-frequency transients) than the original sequence. Secondly, the pre-trained ResNet can extract more robust features than manually designed ones. Finally, the 512-dimensional vector, as a shared feature, simultaneously optimizes the structural defects of both models. For CNN-Transformer, static features might have missed dynamic features related to the working condition (such as torque fluctuation patterns caused by oil contamination), but now the 512-dimensional vector supplements this information through the concat operation. This is even more crucial for LSTM-GRU. The original time series has a low signal-to-noise ratio for short-term abrupt changes, while the spectrogram features significantly enhance the ability to identify periodic fault modes. For example, screw cracks can cause energy anomalies in specific frequency bands, which are submerged in noise in the original torque data. The essence of the 512-dimensional feature is to encode physical fault phenomena into machine-readable high-dimensional fingerprints, providing downstream models with fault discrimination capabilities that surpass the original data, thereby achieving a breakthrough in prediction accuracy.

[0079] S3. Tightening Status Monitoring: By reading the bolt specifications, a static feature vector is generated in real time using an identification model. A compensation model is used to fuse the static features with the real-time torque sequence to update the physical parameters. A dual-strategy approach is used to monitor the tightening status, and in case of a fault, the abnormality type is classified, including the following steps:

[0080] S31. Parameter preloading: When the device starts up, it preloads the machine learning model and parameter library to establish a sensor communication link. The parameter library data includes ratio threshold, slope attenuation coefficient and height parameter mapping table.

[0081] It should be noted that in this embodiment, a millisecond-level communication link is established between the horse head controller and the torque / angle sensor, and an RFID reader is activated at the same time to identify the bolt identification tag.

[0082] S32. Screw identification: RFID is used to read the bolt specifications (height / diameter / material), and a static feature vector is generated in real time through the identification model to obtain the initial ratio threshold and slope attenuation coefficient.

[0083] S33. Real-time dynamic monitoring: Torque angle data is collected at 0.5° intervals, and moving average filtering and outlier removal are performed; a compensation model is used to fuse static features with real-time torque sequences, outputting the correction amount of ratio threshold and slope attenuation coefficient, and updating the actual ratio threshold and slope attenuation coefficient.

[0084] Anomaly alerts are generated through a dual-strategy monitoring approach, employing both ratio monitoring and gradient-based monitoring.

[0085] Ratio monitoring: Verify whether the slope of each segment is within the actual ratio threshold range;

[0086] Gradient method monitoring: Real-time detection of whether slope decay triggers yield point or floating nail conditions.

[0087] S34. Fault Diagnosis: When the alarm mechanism is triggered, the data is frozen, the fault classifier is called to identify the anomaly type, and the anomaly segment curve and static feature vector are input into the fault classifier to obtain the fault type.

[0088] It should be noted that the fault classifier is trained using the fault torque-angle curve in the parameter library and the static feature vector obtained from the recognition model, which will not be elaborated here.

[0089] S35. Data closed-loop collection: Store all tightening curves and model decision data, and label low-confidence samples for self-evolution.

[0090] It should be noted that ratio monitoring divides the complete torque-angle curve into multiple monitoring intervals. By calculating the torque change rate of each interval, if the actual ratio of any interval exceeds the threshold range during real-time tightening, an alarm is triggered. Real-time analysis of curve shape changes captures process anomalies (such as sudden changes in resistance during the screwing-in stage). Gradient method monitoring, on the other hand, detects the bolt's elastic deformation, yield point, and plastic deformation stages during tightening. During the elastic stage, torque and angle have a linear relationship (constant slope). At the yield point, the slope decreases significantly (the material begins plastic deformation). Gradient method monitoring accurately captures the yield point and terminates tightening by calculating the curve slope change rate, preventing bolt over-tightening failure. This embodiment addresses the issues of ratio monitoring's inability to resolve endpoint determination deviations caused by height differences and gradient method monitoring's oversensitivity to small-stroke screws by using height-adaptive dynamic matching of process parameters, thus solving the tightening quality control problem in highly differentiated scenarios.

[0091] This invention optimizes bolt quality control during the tightening process through a systematic parameter library construction, machine learning model training, and real-time monitoring. First, tightening experiments covering all bolt heights are conducted to collect torque-angle curves for both faulty and successful bolts, and the slope safety threshold, slope decay coefficient, and stroke coefficient are calculated to establish a database containing labels for operating conditions and physical parameters. Then, a CNN-Transformer model is used for static feature recognition under different operating conditions, while an LSTM-GRU model dynamically compensates for torque fluctuations and instantaneous faults, optimizing control accuracy during bolt tightening. Next, RFID technology is used to identify bolt specifications, torque data is collected in real time, and anomalies are detected through dual-strategy monitoring (ratio monitoring and gradient method monitoring) to ensure timely alarms and classification when faults occur. Finally, the entire system collects tightening curves and model decision data through a data closed-loop mechanism for self-optimization and fault diagnosis. This method, through machine learning optimization and height-adaptive process parameter adjustment, effectively overcomes problems such as height differences and oversensitivity of small-stroke bolts, improving the accuracy of bolt tightening and fault diagnosis, and ensuring the stability and reliability of the assembly process.

[0092] This invention quantifies the essential parameters of bolt deformation (slope attenuation coefficient β, stroke coefficient γ) through physical experiments, and constructs a static-dynamic AI collaborative monitoring architecture: CNN-Transformer identifies static working conditions and generates adaptive thresholds, LSTM-GRU integrates real-time torque sequences to compensate for dynamic friction fluctuations, and introduces time-spectrum graph transfer learning to extract high-dimensional fault features to enhance the generalization ability of the dual models, achieving a dual closed-loop drive of physical laws and data intelligence. By fusing material mechanical properties (β / γ), dynamic response to working conditions (LSTM-GRU compensation), and fault spectral fingerprint (ResNet features), a complete closed loop from physical essence to data intelligence is formed, fundamentally solving the industrial monitoring challenges of highly differentiated and complex fault scenarios.

[0093] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A monitoring and management method for a vehicle electronic control assembly workshop, characterized in that: Includes the following steps: Parameter library construction: Pass and failure torque-angle curves were collected through tightening experiments covering screws of all heights. The slope safety threshold, slope attenuation coefficient and stroke coefficient were calculated as physical parameters. The parameter library was built by vectorizing the working condition parameters and labeling them with physical parameters. Machine learning model training: Based on the constructed parameter library, a recognition model and a compensation model are built, and the model is optimized through transfer learning; Tightening status monitoring: By reading the bolt specifications, a static feature vector is generated in real time using an identification model. A compensation model is used to fuse the static features with the real-time torque sequence to update the physical parameters. The tightening status is monitored through a dual strategy, and the abnormality type is classified when a fault occurs. The parameter library construction includes the following steps: Data acquisition: Tightening tests were conducted using screws covering all heights, and torque-angle curves were collected for each category, including those for successful and failed tests; failures included floating screws, thread damage, foreign metal objects, misalignment, and incorrect model. Parameter initialization: Based on the qualified torque-angle curve, the bolt deformation stage is divided into the screw-in stage, the tension stage, and the final tightening stage. The slope of each stage of the qualified curve is calculated. The safe threshold range is obtained by eliminating 5% of the extreme values. At the same time, strain gauges are attached to the bolt shank. During the bolt tightening process until fracture, the point of sudden increase in strain is defined as the yield point. The slope attenuation coefficient and the stroke coefficient are calculated. Dataset integration: By vectorizing operating parameters and generating datasets labeled with physical parameters, a parameter library is constructed; The formula for calculating the slope attenuation coefficient β is as follows: ； In the formula, K e K represents the average slope of the stretching segment. y Where is the instantaneous slope at the yield point, and H is the bolt height; The formula for calculating the travel coefficient γ is as follows: ； In the formula, θ y Let θ be the yield point angle. e The angle at the end of the stretching segment.

2. The monitoring and management method for a vehicle electronic control assembly workshop according to claim 1, characterized in that: The training of the machine learning model includes the following steps: Operating condition identification: Operating condition identification is completed by establishing an identification model, and the ratio threshold and slope attenuation coefficient are obtained through static feature mapping of the operating condition; Dynamic compensation: By establishing a compensation model to obtain friction fluctuations and instantaneous faults, dynamic compensation of the ratio threshold and slope attenuation coefficient is achieved; Performance optimization: Fault feature extraction is enhanced through transfer learning, and the output features are input into the recognition model and compensation model to optimize the model.

3. The monitoring and management method for a vehicle electronic control assembly workshop according to claim 2, characterized in that: The operating condition identification Includes the following steps: Feature engineering involves converting the original features, such as bolt height, thread specifications, material hardness, mounting surface roughness, and circuit board temperature, into feature vectors through encoding. Establish a CNN-Transformer model: Establish a model architecture that includes convolutional layers and a Transformer encoder, where convolutional layers are used to extract local feature combinations, and the Transformer encoder is used to calculate the global dependencies between features, with the output being a set of parameters including a ratio threshold and a slope decay coefficient. Model training: Input the feature vector into the CNN-Transformer model to obtain the output prediction, calculate the mean square error between the predicted value and the working condition label in the parameter library, and use the optimizer to update the weights.

4. The monitoring and management method for a vehicle electronic control assembly workshop according to claim 3, characterized in that: The dynamic compensation includes the following steps: Establish an LSTM-GRU model: take the torque sequence of the spiral segment as input, and the correction amount of the ratio threshold and slope decay coefficient as output, construct a compensation model for parallel processing of LSTM and GRU, and concatenate the outputs of LSTM and GRU through a feature fusion layer; Model training: By inputting torque sequence data and deviation label data, time-series Dropout forward propagation is applied, and after calculating the loss, gradient noise is injected for reverse updating, finally obtaining a dynamic correction model that can predict operating condition deviations.

5. The monitoring and management method for a vehicle electronic control assembly workshop according to claim 4, characterized in that: The performance optimization includes the following steps: The screw-in torque sequence is converted into a time-spectrum image using STFT transformation; a ResNet architecture is adopted, and the first layer convolutional kernel is modified to adapt to the long strip structure of the spectrum image; the number of input channels is adjusted from RGB three channels to single channel to match the grayscale torque spectrum input; the ImageNet pre-trained weights of the first 3 ResBlocks are retained to maintain the general feature extraction capability.

6. The monitoring and management method for a vehicle electronic control assembly workshop according to claim 5, characterized in that: The performance optimization also includes the following steps: unfreezing and training the last ResBlock and adaptive pooling layer to learn screw condition-specific features; compressing features through a global average pooling layer to output a 512-dimensional high semantic fault feature vector.

7. The monitoring and management method for a vehicle electronic control assembly workshop according to claim 1, characterized in that: The tightening status monitoring includes the following steps: Parameter preloading: When the device starts up, it preloads the machine learning model and parameter library to establish a sensor communication link. The parameter library data includes ratio threshold, slope attenuation coefficient and height parameter mapping table. Screw identification: RFID is used to read bolt specifications, and a static feature vector is generated in real time through the identification model to obtain the initial ratio threshold and slope attenuation coefficient; Real-time dynamic monitoring: Torque angle data is collected at 0.5° intervals, and moving average filtering and outlier removal are performed; a compensation model is used to fuse static features with real-time torque sequences, outputting the correction amount of ratio threshold and slope attenuation coefficient, and updating the actual ratio threshold and slope attenuation coefficient. Fault diagnosis: When the alarm mechanism is triggered, the data is frozen, the fault classifier is called to identify the anomaly type, and the anomaly segment curve and static feature vector are input into the fault classifier to obtain the fault type; Data closed-loop collection: Store all tightening curves and model decision data, and label low-confidence samples for self-evolution.

8. The monitoring and management method for a vehicle electronic control assembly workshop according to claim 7, characterized in that: The real-time dynamic monitoring also includes the following steps: Anomaly alerts are generated through a dual-strategy monitoring approach, employing both ratio monitoring and gradient-based monitoring. Ratio monitoring: Verify whether the slope of each segment is within the actual ratio threshold range; Gradient method monitoring: Real-time detection of whether slope decay triggers yield point or floating nail conditions.

Images