Intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer

Abstract

The application discloses a bolt tightening process early warning system based on an industrial brain, and belongs to the technical field of intelligent manufacturing and quality control. The system adopts a three-level collaborative architecture of end, edge and cloud: intelligent tightening tools collect high-frequency original data; edge computing nodes perform millisecond-level feature extraction and real-time threshold alarm; the industrial brain platform fuses multi-source data, runs a torque attenuation prediction model based on physical constraints, and realizes long-term performance early warning; and a warning feedback terminal is used for receiving and executing the warning instruction. The core lies in that a material yield strength constraint term is introduced into a model training loss function, so that the physical rationality of prediction is ensured; meanwhile, edge side defines stick-slip oscillation characteristic values to realize instantaneous abnormality interception. The application realizes closed-loop management and control from millisecond-level real-time defect blocking to long-term service performance prediction, effectively improves the reliability of assembly quality, and reduces the whole life cycle cost.

Description

Technical Field

[0001] This invention relates to assembly process quality control technology in the field of industrial manufacturing technology, specifically an intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer. It focuses on building an online monitoring system covering the entire life cycle of key bolts through multi-dimensional data fusion analysis, including torque-angle curves, tightening speed, thread friction coefficient, etc. Background Technology

[0002] In discrete manufacturing industries such as automotive and machinery, bolted connections are crucial for ensuring the structural strength, safety, and reliability of products. Especially in the assembly of critical components like the automotive body-in-white, door hinges, and powertrain, the quality of bolt tightening directly impacts the vehicle's NVH (noise, vibration, and harshness) performance, long-term safety, and after-sales maintenance costs. Traditional tightening quality control relies primarily on tightening tools reaching preset torque or angle thresholds, supplemented by very low-frequency manual sampling and offline auditing. This approach cannot capture dynamic anomalies during the tightening process, nor can it predict the long-term performance of the connection points, leading to a large number of potential defects entering the market and resulting in high after-sales repair costs and damage to brand reputation.

[0003] With the development of Industrial Internet of Things (IIoT), edge computing, and artificial intelligence (AI) technologies, bolt tightening monitoring technology is gradually evolving from static threshold judgment to dynamic data analysis. Existing technologies have been explored at different levels, but significant limitations still exist.

[0004] Regarding real-time monitoring and parameter optimization during the tightening process, Chinese patent CN119830744B proposes a deep learning-based tightening tool parameter optimization system. This system uses an edge computing module and an RNN model to predict the optimal tightening parameters for a single tool in real time, and uses a cloud-based Transformer model to analyze load balancing among multiple tools to optimize collaborative operations. However, the core objective of this solution is dynamic adjustment of process parameters and multi-tool scheduling under real-time operating conditions. Its data processing closed loop focuses on short-term optimization of "tightening-feedback-parameter adjustment," and does not construct a predictive model for the long-term performance of bolt connection points (such as preload decay). Furthermore, its edge computing focuses on data cleaning and feature extraction for cloud analysis, lacking the ability to perform real-time, localized judgment and immediate production intervention for transient anomalies (such as microsecond-level stick-slip oscillations) during tightening, and cannot freeze the production line the instant a defect occurs.

[0005] In the area of ​​online assessment and defect diagnosis of tightening quality, Chinese patent CN120387249B utilizes an LSTM network to model the nonlinear fluctuations of the friction factor during repeated tightening processes to correct the torque-preload conversion relationship and provides visual monitoring based on digital twins. However, its core focus is on improving the quality assessment accuracy and process adaptability of a single tightening process. Its model output is used to adjust the tightening parameters in real time, rather than predicting the performance degradation of the connection point over subsequent weeks or months. Chinese patent application CN121117826A, on the other hand, constructs a MOE+Transformer hybrid model to extract multi-dimensional features from the torque-angle curve, achieving high-precision classification and diagnosis of more than ten tightening defects (such as stick-slip, over-tightening, and stripped threads). Although the defect classification accuracy is high, its complex model structure is more suitable for post-analysis or near-real-time diagnosis, making it difficult to meet the stringent requirements of production lines for millisecond-level real-time identification and interception of specific high-risk defects (such as stick-slip caused by thread jamming). Both of the above belong to "in-process" quality control and fail to extend to "service life" risk warning.

[0006] In the area of ​​long-term health monitoring of bolted connections, Chinese patent CN120293512B proposes an online loosening monitoring method based on multimodal sensors and graph neural networks (GNNs) for bolts already in service in wind turbine towers, bridges, and other applications. This solution achieves good service status assessment and early warning by fusing vibration, temperature, and stress data and utilizing a physically-enhanced LSTM network to predict loosening risk. However, this approach relies on the bolts being in service and exposed to complex environmental loads. The sensor deployment methods (such as externally attached vibration plates and fiber optic gratings) fundamentally conflict with the sealing and embedded assembly structures in automotive manufacturing, especially door hinges, making it unsuitable for automotive assembly lines. More importantly, this solution aims to monitor whether loosening has occurred or is about to occur, representing a "post-event" or "in-event" fault diagnosis. It cannot make prior predictions about long-term performance based on tightening process data at the moment of assembly tightening, thus failing to achieve true preventative quality control.

[0007] Furthermore, existing technologies generally lack expertise in knowledge reuse and rapid process adaptation across product platforms. Currently, the automotive industry is experiencing rapid model iteration, and the development of tightening processes for hinges in new models heavily relies on engineers' experience and extensive physical trial and error. The massive amounts of tightening process data accumulated over time have become "data graveyards" due to a lack of effective structured representation and retrieval methods, failing to form a transferable process knowledge base, resulting in long development cycles and high costs.

[0008] In summary, existing technologies have made progress in addressing specific issues such as real-time process optimization, online defect classification, and service status monitoring. However, they have yet to construct a comprehensive, systematic solution covering the entire chain, from millisecond-level real-time monitoring of the tightening process to accurate prediction of long-term service performance and rapid transfer of process knowledge across vehicle models. Problems such as scenario fragmentation, data gaps, and limited predictive dimensions exist among various technical solutions, resulting in the automotive manufacturing industry still facing three core pain points in bolt tightening quality control: insufficient real-time interception, unknown long-term risks, and difficulties in knowledge reuse. Therefore, there is an urgent need for an innovative tightening quality early warning system that integrates edge-cloud architecture, combines real-time and predictive capabilities, and enables knowledge accumulation and reuse. Summary of the Invention

[0009] In view of this, the purpose of this invention is to provide an intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer. By constructing an end-edge-cloud collaborative system, it can achieve millisecond-level real-time interception of stick-slip anomalies, long-term prediction of torque attenuation based on physical constraints, and rapid cross-vehicle process transfer based on three-dimensional model features, forming a full-chain preventive quality control covering assembly process and service performance.

[0010] To achieve the above objectives, the present invention provides the following technical solution: An intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer includes: The intelligent tightening tool, acting as a data acquisition front-end, is used to collect raw timing data during the tightening process. This raw timing data includes at least the torque. ,angle and time ; An edge computing node, deployed on the production line side, communicates with the intelligent tightening tool and is used to perform real-time signal preprocessing and feature extraction on the raw time series data. When the extracted feature value exceeds a preset threshold, a local warning is triggered. The industrial brain platform, as the central hub for cloud-based data processing and analysis, communicates with the edge computing nodes to receive and store all data from the edge computing nodes. The industrial brain platform has a built-in torque decay prediction model, which incorporates a physical constraint term based on the material yield critical value into the loss function during its training process. This model is used to predict the torque decay rate within a specified time period based on the input dynamic characteristics and to generate an early warning control command when the predicted decay rate exceeds a risk threshold. The early warning feedback terminal is used to receive and execute the early warning command.

[0011] Furthermore, the features extracted in real time by the edge computing node include stick-slip oscillation feature values. Its definition is: in: For the set time window; Torque Regarding time The second derivative; when When the oscillation exceeds the preset threshold, a local alarm is triggered.

[0012] Furthermore, the features extracted by the edge computing node also include the torque fluctuation coefficient. Its definition is: in: , and These represent the maximum, minimum, and average torque values ​​within the same tightening phase; when When the fluctuation exceeds the preset threshold, a local alarm is triggered.

[0013] Furthermore, the method for real-time detection of abnormalities during the bolt tightening process performed by the edge computing node includes the following steps: Real-time reception of torque timing signals from intelligent tightening tools ; For the torque timing signal Perform temperature drift compensation and vibration filtering preprocessing; Calculate the second derivative of torque with respect to time. Calculate the stick-slip oscillation eigenvalues ​​within the sliding time window T. ; and / or, by the maximum torque within the same tightening phase. Minimum value and average Calculate the torque ripple coefficient ; The stick-slip oscillation characteristic value Compared with a preset oscillation threshold, when the stick-slip oscillation characteristic value When the oscillation exceeds a preset threshold, it is determined to be an abnormal stick-slip oscillation, triggering a local alarm and instructing the current tightening operation to be paused; and / or, Torque ripple coefficient Compared with a preset fluctuation threshold, when the torque fluctuation coefficient If the fluctuation exceeds the preset threshold, an abnormality is detected, a local alarm is triggered, and the current tightening operation is suspended.

[0014] Furthermore, the torque attenuation prediction model built into the industrial brain platform is a deep learning model that incorporates physical constraints, and its training loss function L is: Where MAE(·) represents the mean absolute error; The torque decay rate predicted by the model; This represents the actual measured torque attenuation rate. These are the constraint weighting coefficients; This is the critical decay rate set based on the yield strength of the materials being connected.

[0015] Furthermore, the critical decay rate Based on material properties; when the model predicts the torque decay rate... When the set risk threshold is reached or exceeded, the industrial brain platform generates an early warning instruction. The early warning instruction includes sending an instruction to the production line PLC control system to dynamically adjust the target torque. The adjustment range is to increase the target torque by a set percentage.

[0016] Furthermore, the method steps for the industrial brain platform to predict and warn of bolt tightening torque attenuation based on the torque attenuation prediction model are as follows: The input features are obtained, including at least the dynamic feature vector of the tightening process uploaded from the edge computing node, vehicle design parameters, and external industry collaboration parameters. The input features are fed into a torque decay prediction model trained with physical constraints, and the model outputs a predicted torque decay rate for a specified future time period. ; judge Is it greater than or equal to the set risk threshold? If so, an early warning work order will be generated and / or an instruction to dynamically adjust the tightening parameters will be sent to the production line control system.

[0017] Furthermore, the industry brain platform also includes a knowledge transfer engine for rapid adaptation of tightening process parameters across vehicle models. Its workflow includes: Eigenvectorization: used to convert a 3D model of a target connectivity structure into feature vectors; Similarity retrieval: This function matches the feature vector with the feature vector of a benchmark model in a historical process knowledge base and obtains benchmark models with a similarity higher than a preset threshold. Parameter migration verification: This is used to extract the optimal combination of tightening parameters corresponding to the benchmark model. After verifying the structural stress compliance through finite element simulation, the parameters are migrated and applied to the production line of the new model.

[0018] Furthermore, in the feature vectorization, the PointNet++ network is used to convert the three-dimensional model into a 128-dimensional feature vector; in the similarity retrieval, cosine similarity is used for matching.

[0019] Furthermore, the intelligent tightening tool integrates a multimodal sensing fusion unit for measuring the tightening angle; the fusion algorithm of the multimodal sensing fusion unit is as follows: in: for The final angle estimate after time-mapping; The angle value measured by the photoelectric encoder; This is the integral value of the gyroscope's angular velocity; The Kalman gain is dynamically adjusted to suppress measurement errors caused by mechanical backlash.

[0020] The beneficial effects of this invention are as follows: This invention is an intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer. Through the construction of a three-level collaborative system of "end-edge-cloud", it has achieved a breakthrough in the entire chain of technical effects from real-time perception and edge decision-making to cloud prediction.

[0021] First, by using intelligent tightening tools to collect raw data at high frequency and combining edge computing nodes to perform millisecond-level feature extraction and threshold judgment, the system achieves real-time identification and local interception of transient anomalies such as sticking and jamming on the physical production line, preventing defects from flowing into subsequent processes from the source and solving the problem of slow response in the traditional sampling inspection mode.

[0022] Secondly, the industry brain platform aggregates all data and runs a torque decay prediction model customized for it. The torque decay prediction model can not only integrate multi-dimensional dynamic and static characteristics for long-term performance prediction, but also ensures that the prediction results conform to the laws of material mechanics due to its embedded physical constraint mechanism. This moves quality control from "post-event remediation" to "pre-event warning", realizing true predictive maintenance.

[0023] Ultimately, a closed-loop decision-making process is formed through the early warning feedback terminal, enabling cloud-based predictive commands to be sent to the production line control unit in real time, dynamically adjusting process parameters. This "perception-decision-execution" closed-loop system organically combines real-time quality control with long-term performance management, significantly improving the reliability and stability of assembly quality and the intelligence level of the production system. Attached Figure Description

[0024] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is the overall architecture diagram of the intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer of the present invention; Figure 2 A flowchart for performing real-time detection of abnormalities in the bolt tightening process and prediction and early warning of bolt tightening torque attenuation; Figure 3 A flowchart for performing cross-vehicle process migration for the knowledge transfer engine. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0026] like Figure 1 As shown, this embodiment is a bolt tightening process intelligent monitoring system based on multi-dimensional data fusion and knowledge transfer, including intelligent tightening tools, edge computing nodes, an industry brain platform, and an early warning feedback terminal.

[0027] 1. Intelligent tightening tool The intelligent tightening tool serves as a data acquisition front-end, used to collect raw timing data during the tightening process. This raw timing data includes at least the torque. ,angle and time .

[0028] As a data acquisition front-end, the intelligent tightening tool integrates a triaxial high-precision MEMS sensor array within the electric tightening gun. The torque sensing unit employs a strain gauge Wheatstone bridge structure, featuring a wide measurement range of 5-150 Nm and a nonlinear error characteristic of ≤±0.5% FS. The angle measurement unit is equipped with a 1024 PPR photoelectric encoder and is directly coupled to the gun head drive shaft via a flexible coupling. The time synchronization module achieves microsecond-level time calibration based on the IEEE 1588 precision clock protocol. The tool's outer shell is integrally formed from forged aluminum alloy, with multiple layers of electromagnetic shielding cavities inside, ensuring stable output of a 2000 Hz sampling rate three-dimensional data stream (torque) even in the strong electromagnetic interference environment of the production line. ,angle ,time The collected data is transmitted to the edge computing nodes via armored shielded twisted-pair cables.

[0029] In this embodiment, the intelligent tightening tool integrates a multimodal sensing fusion unit for measuring the tightening angle; the fusion algorithm of the multimodal sensing fusion unit is as follows: in: for The final angle estimate after time-mapping; The angle value measured by the photoelectric encoder; This is the integral value of the gyroscope's angular velocity; The dynamically adjustable Kalman gain is used to suppress measurement errors caused by mechanical backlash, specifically to eliminate mechanical errors ≤0.3°.

[0030] 2. Edge computing nodes Edge computing nodes are deployed on the production line and communicate with the intelligent tightening tool. They are used to perform real-time signal preprocessing and feature extraction on the raw time-series data. When the extracted feature values ​​exceed a preset threshold, a local warning is triggered. Specifically, the edge computing nodes are deployed in an industrial-grade control computer within the production line control cabinet. Their core tasks include millisecond-level signal preprocessing (temperature drift compensation and vibration filtering) and real-time feature extraction.

[0031] (1) Eigenvalues ​​of stick-slip oscillations In this embodiment, the edge computing node calculates the torque differential in real time. and angle integral .

[0032] In this embodiment, the features extracted in real time by the edge computing node include stick-slip oscillation feature values. Its definition is: in: For the set time window; Torque Regarding time The second derivative of .

[0033] when When the oscillation rate exceeds a preset oscillation threshold, a local alarm is triggered. In this embodiment, the preset oscillation threshold is 15 Nm / ms. 2 .

[0034] Table 1. Performance comparison between stick-slip oscillation characteristic value calculation and traditional manual auditory judgment. (2) Torque fluctuation coefficient Features extracted from edge computing nodes also include torque ripple coefficient. Its definition is: in: , and These represent the maximum, minimum, and average torque values ​​within the same tightening phase.

[0035] when When the fluctuation exceeds a preset threshold, a local alarm is triggered. In this embodiment, the preset fluctuation threshold is 12%.

[0036] In typical abnormal scenarios, such as bolt threads being stuck with foreign objects or uneven clamping surfaces, the edge node immediately sends a level 3 alarm signal to the workshop HMI touch screen via the EtherCAT bus, triggering the corresponding workstation's red warning light to flash and freezing the assembly process for inspection.

[0037] (3) Real-time anomaly detection like Figure 2 As shown, the steps of the method for real-time detection of anomalies in the bolt tightening process performed by the edge computing node are as follows: Real-time reception of torque timing signals from intelligent tightening tools ; For the torque timing signal Perform temperature drift compensation and vibration filtering preprocessing; Calculate the second derivative of torque with respect to time. Calculate the stick-slip oscillation eigenvalues ​​within the sliding time window T. ; and / or, by the maximum torque within the same tightening phase. Minimum value and average Calculate the torque ripple coefficient ; The stick-slip oscillation characteristic value Compared with a preset oscillation threshold, when the stick-slip oscillation characteristic value When the oscillation exceeds a preset threshold, it is determined to be an abnormal stick-slip oscillation, triggering a local alarm and instructing the current tightening operation to be paused; and / or, Torque ripple coefficient Compared with a preset fluctuation threshold, when the torque fluctuation coefficient If the fluctuation exceeds the preset threshold, an abnormality is detected, a local alarm is triggered, and the current tightening operation is suspended.

[0038] 3. Industry Brain Platform The industry brain platform, as the central hub for cloud-based data processing and analysis, communicates with the edge computing nodes to receive and store all data from the edge computing nodes. The industry brain platform has a built-in torque decay prediction model, which incorporates a physical constraint term based on the material yield critical value into the loss function during its training process. This model is used to predict the torque decay rate within a specified time period based on the input dynamic characteristics and to generate an early warning control command when the predicted decay rate exceeds a risk threshold.

[0039] The Industry Brain platform, serving as the digital decision-making hub for the regional automotive industry cluster, is built on a Hadoop distributed file system. It receives real-time data streams from production lines nationwide via a Kafka message queue and simultaneously accesses four-dimensional dynamic industry data streams: supply chain data (including real-time capacity data from hinge suppliers, batch yield statistics, and logistics status), government regulatory data, after-sales market data, and resource sharing data. At the physical infrastructure level, the platform employs a high-performance private cloud cluster design, deploying at least eight computing nodes to form the core processing unit. Each node is equipped with dual Intel Xeon Gold 6330 processors, 512GB of ECC error-correcting memory, and four NVIDIA T4 GPU accelerator cards, directly connected to the provincial industrial internet identifier resolution node via a gigabit fiber optic leased line. At the software layer, it utilizes Docker containerization technology for elastic resource scheduling and integrates the Apache NiFi data stream engine to build a cross-domain data cleaning pipeline, supporting millisecond-level alignment and fusion of production equipment OEE data, supplier ERP information, government environmental databases, and market after-sales records.

[0040] Table 2 Modular Design Scheme of the Industry Brain Platform (1) Torque decay prediction model The torque decay prediction model built into the industrial brain platform is a deep learning model that incorporates physical constraints. A material yield strength constraint term is added during the training of the two-layer LSTM network, and its loss function L during training is: Where MAE(·) represents the mean absolute error; The torque decay rate predicted by the model; This represents the actual measured torque attenuation rate. These are the constraint weighting coefficients; This is the critical attenuation rate set based on the yield strength of the materials being joined. Specifically, the critical attenuation rate... Based on material properties; when the model predicts the torque decay rate... When the set risk threshold is reached or exceeded, the industrial brain platform generates an early warning instruction. The early warning instruction includes sending an instruction to the production line PLC control system to dynamically adjust the target torque. The adjustment range is to increase the target torque by a set percentage.

[0041] In this embodiment, the torque decay prediction model serves as the core intelligent analysis engine of the industry brain. Its innovative architecture integrates a multi-dimensional parameter system: the basic input layer integrates 8-dimensional dynamic feature vectors uploaded from edge nodes; the vehicle design parameter layer embeds the door mass distribution curve and the SAE 4140 steel hinge yield strength map; and the industry collaboration layer expands key external parameters—supplier batch pass rate weighting coefficients, logistics timeliness decay index, and regional carbon quota balance fluctuation curves. The main body of the model employs a two-layer LSTM neural network, and the output layer generates the predicted torque decay rate for the next 30 days using a Sigmoid activation function. The training process innovatively introduces a material mechanics constraint mechanism, adding a penalty term based on the material's yield critical value to the loss function. ,in Based on the properties of SAE 4140 steel, a 25% limiting attenuation rate was set to ensure that the predicted value strictly follows the physical laws of materials. When the predicted attenuation rate... When the risk threshold of 15% is reached or exceeded, the platform initiates dual-path closed-loop control: it pushes an early warning work order containing a bolt positioning diagram and a recommended re-tightening torque value to the enterprise office platform, and at the same time sends a dynamic adjustment instruction to the PLC control system through the OPC UA industrial protocol, such as increasing the target torque by 3% to 5%.

[0042] Table 3 Performance comparison between the torque decay prediction model and the traditional threshold method (2) Prediction and early warning of bolt tightening torque decay like Figure 2 As shown, the method steps for the industry brain platform to perform bolt tightening torque attenuation prediction and early warning based on the torque attenuation prediction model are as follows: Obtaining input features, which include at least the dynamic feature vector of the tightening process uploaded from the edge computing node, vehicle design parameters, and external industry collaboration parameters; inputting the input features into the torque attenuation prediction model trained with physical constraints, and the model outputting the predicted torque attenuation rate for a specified future time period. ;judge If the risk threshold is greater than or equal to the set risk threshold, an early warning work order is generated and / or an instruction to dynamically adjust the tightening parameters is sent to the production line control system.

[0043] (3) Knowledge transfer engine The data processing pipeline of the Industry Brain platform adopts a unified stream and batch processing architecture. Real-time data streams are processed by the Flink engine to perform minute-level statistical calculations, with typical applications including real-time tracking of the process capability index (CPK) value for each workstation. Batch data triggers a model retraining process every 24 hours, using 1 million high-quality samples accumulated over the past 30 days as the training set. An innovative A / B testing mechanism is introduced during the model deployment phase: new model versions are first tested on 5% of production lines, and only after the prediction accuracy exceeds the validation threshold of 89% for three consecutive days are they rolled out to all production lines.

[0044] like Figure 3 As shown, the industry brain platform also includes a knowledge transfer engine, which is used to quickly adapt tightening process parameters across vehicle models. Its workflow includes: feature vectorization: converting the 3D model of the target connection structure into feature vectors; similarity retrieval: matching the feature vectors with the feature vectors of the benchmark models in the historical process knowledge base, and obtaining benchmark models with similarity higher than a preset threshold; parameter transfer verification: extracting the optimal combination of tightening parameters corresponding to the benchmark model, verifying the structural stress compliance through finite element simulation, and then transferring it to the production line of the new vehicle model.

[0045] Specifically, in feature vectorization, the PointNet++ network is used to convert the 3D model into a 128-dimensional feature vector; in similarity retrieval, cosine similarity is used for matching, and models with a cosine similarity > 85% are retrieved from the historical database. The migration speed curve and target torque sequence are verified by ANSYS stress simulation.

[0046] Table 4 Knowledge Transfer Engine 4. Early warning feedback terminal The early warning feedback terminal is used to receive and execute the early warning command.

[0047] In this embodiment, the system's workflow exhibits multi-level collaborative characteristics. At the instant the bolt insertion process begins, the intelligent tightening tool captures the original signal using a 2000Hz high-frequency sampling rate. The torque sensing unit eliminates ambient temperature drift through a temperature compensation algorithm, with a temperature drift compensation coefficient of -0.02. The angle measurement unit employs a Kalman filter algorithm to fuse data from the photoelectric encoder and MEMS gyroscope, effectively suppressing the ±0.3 degree measurement error caused by mechanical backlash. After feature extraction at the edge nodes, the preprocessed data stream, including key indicators such as stick-slip oscillation values... More than 15 Nm / ms 2The preset threshold corresponds to typical abnormal working conditions such as metal chips in the thread. If these conditions are met, a local audible and visual alarm will be triggered immediately and the current assembly process will be frozen. All data is converted into Parquet columnar storage format by the Zstd high-efficiency compression algorithm and uploaded to the industry brain platform in batches for in-depth modeling and analysis.

[0048] 5. Technical Effects Compared with existing technologies, this embodiment achieves breakthrough improvements in four dimensions—quality control, cost control, production efficiency, and knowledge accumulation—by constructing a collaborative "end-edge-cloud" hinge tightening quality prediction and early warning system. Its core benefits stem from the innovative synergy of three key technological components: the intelligent tightening tool enables millisecond-level holographic process perception; the industry brain platform establishes a physical constraint attenuation prediction model; and the knowledge transfer engine opens up cross-vehicle process reuse channels. Specific benefits are presented below.

[0049] (1) Eradicating persistent quality problems and transforming to preventive control Because the intelligent tightening tool integrates a triaxial MEMS sensor array with a 2000Hz sampling frequency, its strain gauge Wheatstone bridge structure can capture microsecond-level anomalous events that traditional equipment ignores. For example, when there are metal chips in the bolt threads, the system calculates the stick-slip oscillation energy integral value in real time. The threshold characteristics trigger an audible and visual alarm and freeze the production line immediately at the assembly site. This dynamic sensing capability increases the detection rate of slippage defects from the industry average of 32% to 89%, eliminating persistent quality problems such as abnormal door noise and sagging caused by poor tightening at the source. The industry brain platform's dual-layer LSTM prediction model integrates vehicle design parameters to output the torque decay rate for the next 30 days. When the predicted value exceeds the 15% risk threshold, the system automatically sends an adjustment command to the PLC to increase the torque by 3%-5%. This predictive intervention mechanism will significantly reduce the torque attenuation failure rate, driving the transformation of quality control from passive spot checks to proactive prevention.

[0050] (2) Systematic optimization of life cycle cost Edge computing nodes utilize Intel Atom x6425E processors for millisecond-level feature extraction. Their signal preprocessing module employs second-order Butterworth filtering to eliminate equipment vibration noise, reducing the analysis time for a single tightening curve from 15 minutes manually to 50 milliseconds. This efficiency improvement, combined with a full-scale monitoring model, saves a single production line ¥1.2 million annually in manual inspection costs (based on an annual production of 500,000 vehicles, 4 bolts per vehicle, and a sampling cost of ¥30 per bolt). On the after-sales maintenance side, a physical constraint-based attenuation prediction model advances fault warnings to 30 days before occurrence, reducing repair costs from ¥1200 per unit in the after-sales stage to ¥260 per unit in preventative maintenance. This end-to-end cost optimization reduces per-vehicle quality costs by 37%, resulting in annual quality cost savings of ¥110 million based on an annual production scale of 300,000 vehicles.

[0051] (3) Production efficiency and flexible manufacturing capabilities have increased dramatically. The knowledge transfer engine uses the PointNet++ network to convert the 3D model of the new car model's hinge into a 128-dimensional feature vector. It then retrieves benchmark models with a similarity greater than 85% from the historical database, transfers their optimal tightening parameters (speed curve + target torque sequence), and verifies them through ANSYS finite element simulation. This technology will significantly reduce the process development cycle of new car models and save on prototyping costs. Simultaneously, the integrated batch processing architecture of the industry brain platform (Flink real-time calculation of CPK value + 24-hour model retraining) supports dynamic parameter tuning on the production line: when a batch of hinge hardness fluctuations is detected, the first-stage tightening speed is automatically reduced from 800 r / min to 500 r / min, reducing the stick-slip rate by 40% and increasing the average production line cycle time by 5.8%.

[0052] (4) Industrial knowledge accumulation and contributions to green manufacturing The system utilizes a distributed data lake built on Hadoop 3.0 and Parquet columnar storage to store petabytes of tightening curves. Through feature correlation mining, it has identified 12 key process guidelines, forming the industry's first hinge tightening process knowledge base. This knowledge base supports the rapid replication of best practices in new factories, reducing the process implementation cycle by 83%. In terms of green manufacturing, precise tightening control reduces bolt overtightening rate from the industry average of 7% to 0.5%, reduces ineffective torque output per bolt by 30%, saving the equivalent of 1.8 million kWh of electricity annually, while avoiding rework-related energy consumption.

[0053] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. An intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer, characterized in that: include: The intelligent tightening tool, acting as a data acquisition front-end, is used to collect raw timing data during the tightening process. This raw timing data includes at least the torque. ,angle and time ; An edge computing node, deployed on the production line side, communicates with the intelligent tightening tool and is used to perform real-time signal preprocessing and feature extraction on the raw time series data. When the extracted feature value exceeds a preset threshold, a local warning is triggered. The industrial brain platform, as the central hub for cloud-based data processing and analysis, communicates with the edge computing nodes to receive and store all data from the edge computing nodes. The industrial brain platform has a built-in torque decay prediction model, which incorporates a physical constraint term based on the material yield critical value into the loss function during its training process. This model is used to predict the torque decay rate within a specified time period based on the input dynamic characteristics and to generate an early warning control command when the predicted decay rate exceeds a risk threshold. The early warning feedback terminal is used to receive and execute the early warning command.

2. The intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer as described in claim 1, characterized in that: The features extracted in real time by the edge computing node include stick-slip oscillation feature values. Its definition is: in: For the set time window; Torque Regarding time The second derivative; when When the oscillation exceeds the preset threshold, a local alarm is triggered.

3. The intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer as described in claim 1, characterized in that: The features extracted by the edge computing nodes also include torque fluctuation coefficient. Its definition is: in: , and These represent the maximum, minimum, and average torque values ​​within the same tightening phase; when When the fluctuation exceeds the preset threshold, a local alarm is triggered.

4. The intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer according to any one of claims 1-3, characterized in that: The method for real-time detection of anomalies during the bolt tightening process performed by the edge computing node includes the following steps: Real-time reception of torque timing signals from intelligent tightening tools ; For the torque timing signal Perform temperature drift compensation and vibration filtering preprocessing; Calculate the second derivative of torque with respect to time. Calculate the stick-slip oscillation eigenvalues ​​within the sliding time window T. ; and / or, by the maximum torque within the same tightening phase. Minimum value and average Calculate the torque ripple coefficient ; The stick-slip oscillation characteristic value Compared with a preset oscillation threshold, when the stick-slip oscillation characteristic value When the oscillation exceeds the preset threshold, it is determined that an abnormal stick-slip oscillation has occurred, triggering a local alarm and instructing the current tightening operation to be paused; And / or, Torque ripple coefficient Compared with a preset fluctuation threshold, when the torque fluctuation coefficient If the fluctuation exceeds the preset threshold, an abnormality is detected, a local alarm is triggered, and the current tightening operation is suspended.

5. The intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer according to claim 1, characterized in that: The torque attenuation prediction model built into the industrial brain platform is a deep learning model that incorporates physical constraints, and its training loss function L is: Where MAE(·) represents the mean absolute error; The torque decay rate predicted by the model; This represents the actual measured torque attenuation rate. These are the constraint weighting coefficients; This is the critical decay rate set based on the yield strength of the materials being connected.

6. The intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer according to claim 5, characterized in that: The critical attenuation rate Based on material properties; when the model predicts the torque decay rate... When the set risk threshold is reached or exceeded, the industrial brain platform generates an early warning instruction. The early warning instruction includes sending an instruction to the production line PLC control system to dynamically adjust the target torque. The adjustment range is to increase the target torque by a set percentage.

7. The intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer according to claim 5, characterized in that: The steps of the industrial brain platform in performing bolt tightening torque attenuation prediction and early warning based on the torque attenuation prediction model are as follows: The input features are obtained, including at least the dynamic feature vector of the tightening process uploaded from the edge computing node, vehicle design parameters, and external industry collaboration parameters. The input features are fed into a torque decay prediction model trained with physical constraints, and the model outputs a predicted torque decay rate for a specified future time period. ; judge Is it greater than or equal to the set risk threshold? If so, an early warning work order will be generated and / or an instruction to dynamically adjust the tightening parameters will be sent to the production line control system.

8. The intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer according to claim 1, characterized in that: The industry brain platform also includes a knowledge transfer engine for rapid adaptation of tightening process parameters across vehicle models. Its workflow includes: Eigenvectorization: used to convert a 3D model of a target connectivity structure into feature vectors; Similarity retrieval: This function matches the feature vector with the feature vector of a benchmark model in a historical process knowledge base and obtains benchmark models with a similarity higher than a preset threshold. Parameter migration verification: This is used to extract the optimal combination of tightening parameters corresponding to the benchmark model. After verifying the structural stress compliance through finite element simulation, the parameters are migrated and applied to the production line of the new model.

9. The intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer according to claim 8, characterized in that: In the feature vectorization process, the PointNet++ network is used to convert the 3D model into a 128-dimensional feature vector; in the similarity retrieval process, cosine similarity is used for matching.

10. The intelligent monitoring system for bolt tightening process based on multi-dimensional data fusion and knowledge transfer according to claim 1, characterized in that: The intelligent tightening tool integrates a multimodal sensing fusion unit for measuring the tightening angle; the fusion algorithm of the multimodal sensing fusion unit is as follows: in: for The final angle estimate after time-mapping; The angle value measured by the photoelectric encoder; This is the integral value of the gyroscope's angular velocity; The Kalman gain is dynamically adjusted to suppress measurement errors caused by mechanical backlash.

Images