A method and system for intelligent perception, prediction and decision of vehicle body welding quality based on multi-source industrial data

Abstract

The present application relates to a kind of based on multi-source industrial data's car body welding quality intelligent perception, prediction and decision-making method and system, belong to the technical field of intelligent manufacturing and industrial data analysis.The present application obtains multi-source high-frequency synchronous original signal by sensor array and robot bus;Perform space-time alignment and multidimensional feature mining;Quality perception model is constructed based on multi-path parallel network and cross-modal attention;Quality evolution trend and electrode cap life are evaluated using time series prediction model;Integrate expert rules and reinforcement learning to make process compensation decision.The present application realizes the accurate perception of welding quality, forward warning and closed-loop real-time compensation, significantly improves the perception accuracy, production robustness and prolongs the service life of electrode.

Description

Technical Field

[0001] This invention relates to a method and system for intelligent sensing, prediction and decision-making of vehicle body welding quality based on multi-source industrial data, belonging to the technical field of intelligent manufacturing and industrial data analysis. Background Technology

[0002] In the precision manufacturing chain of the modern automotive industry, body welding, as a core process determining the structural strength, collision resistance, and long-term driving safety of the entire vehicle, has become a key indicator of an OEM's core competitiveness, particularly in terms of its automation level and quality control precision. With the widespread application of lightweight materials and the increasing complexity of vehicle body structures, thousands of resistance spot welds not only bear the physical rigidity requirements of the vehicle body but also serve as the first line of defense for ensuring the safety of passengers under extreme conditions. Therefore, achieving real-time monitoring and accurate evaluation of welding quality on high-speed automated production lines has always been a research focus in the fields of industrial sensing and process control.

[0003] For assessing the quality of spot welding on automotive bodies, traditional production models primarily rely on offline sampling inspections after welding. Methods such as measuring the weld nugget diameter using destructive tensile testing or employing ultrasonic flaw detectors for non-destructive testing of weld points offer direct physical evidence for quality assessment. However, in modern, high-paced, large-scale production environments, their inherent lag prevents real-time responses to production fluctuations. Any process deviation can easily lead to batch-wide quality incidents. To address this, subsequent technologies have gradually shifted towards online monitoring. This involves integrating sensors into the welding circuit to collect key process parameters such as welding current, voltage, and electrode pressure. Statistical process control or simple threshold comparison logic is then used to preliminarily determine the controlled state of the welding process. With the rise of artificial intelligence, some existing solutions are incorporating traditional machine learning algorithms or deep neural networks to attempt to establish a mapping relationship between welding electrical signal characteristics and weld strength, striving to achieve online quality prediction without damaging the workpiece.

[0004] However, a deeper analysis of the physical mechanisms of the welding process and the complex working conditions in industrial settings reveals that existing technical solutions still suffer from deep-seated contradictions at the fundamental level. Firstly, spot welding is essentially a very short, dynamic evolutionary process involving strong coupling of multiple physical fields such as electricity, heat, force, and materials. Its quality is comprehensively affected by a variety of factors, including the welding torch actuator, the motion accuracy of the industrial robot, the stability of the welding power supply waveform, and the matching gap of the sheet metal. Most existing monitoring solutions are limited to shallow correlations of single-source signals or a few explicit process parameters, lacking the ability to deeply integrate welding torch servo characteristics, robot pose feedback, and heterogeneous data from multiple sensors. This results in models that, when faced with complex electromagnetic interference, power grid fluctuations, or workpiece assembly deviations on the production line, often fail to meet the comprehensiveness of their perception and the robustness of their identification to meet industrial-grade application requirements.

[0005] More critically, current technological paradigms remain highly focused on "post-construction assessment" of predetermined weld points, neglecting the "temporal evolution trend" of weld quality affected by equipment condition drift. This monitoring logic, lacking proactive early warning capabilities, prevents production systems from intervening when potential quality issues first emerge. Particularly in the management of critical consumable materials—electrode caps—existing production practices generally employ a static replacement system based on a fixed number of welds. This linear, experience-based maintenance strategy completely ignores the non-linear physical wear process of electrode caps under different materials and current densities. This fundamental misalignment between static management logic and dynamic physical wear on-site either leads to excessive electrode wear causing incomplete welds and spatter, or results in premature scrapping of electrodes before their service life, severely restricting production efficiency and cost control.

[0006] Furthermore, existing diagnostic systems generally exhibit "decision silos." Even when a system identifies welding anomalies, it often only outputs simple logical judgments, lacking the ability to combine historical process big data with domain expert knowledge for causal inference. Because it cannot provide clear process compensation suggestions or equipment adjustment plans for the causes of anomalies, the three core links of perception, prediction, and decision-making are essentially disconnected in physical links and logical loops, failing to form a truly closed-loop intelligent control system. Therefore, how to construct a comprehensive technical solution that can deeply integrate multi-source heterogeneous industrial data, possess full lifecycle evolution modeling capabilities, and achieve closed-loop control from accurate perception to intelligent decision-making has become a key technical bottleneck in improving the level of intelligence and quality consistency in vehicle body manufacturing. Summary of the Invention

[0007] To address the aforementioned problems, this invention discloses a method and system for intelligent perception, prediction, and decision-making regarding vehicle body welding quality based on multi-source industrial data. The specific technical solution is as follows: A method for intelligent sensing, prediction, and decision-making regarding vehicle body welding quality based on multi-source industrial data includes the following steps: Step 1: Establish a high-frequency synchronous acquisition system for multi-source industrial data: Arrange a sensor array, a welding power communication interface and an industrial robot controller (11) bus at the welding execution end to acquire the original time domain signal of the entire welding cycle in real time. Step 2: Perform spatiotemporal alignment and feature mining of multi-source heterogeneous data: Input the acquired time-domain signals into the preprocessing module to filter out high-frequency electromagnetic noise, and then, taking the welding current triggering time as zero point, perform spatiotemporal alignment and feature mining of the electrode pressure signal. Welding current signal Welding voltage signal High-frequency acoustic emission signals Furthermore, time-axis resampling and alignment of robot state data are performed to form a multi-dimensional spatiotemporal feature vector sequence; time-frequency domain feature extraction is then performed on welding current and voltage signals, and the energy entropy of each frequency band is calculated. For electrode pressure signals Peak extraction pressure Pressure rise slope and time constant For acoustic emission signals Extract spectral features, including centroid frequency. and peak count rate ; Step 3: Construct a welding quality intelligent perception model based on a multimodal fusion deep network topology. The welding quality intelligent perception model includes multiple parallel feature extraction channels and a cross-modal attention fusion layer. The output vectors of each feature extraction channel are weighted and aggregated in the cross-modal attention fusion layer, and the perception weights are adaptively allocated to finally output the perception quality evaluation index of the current weld point. The perceived quality index Including predicted melting core diameter Predicted shear strength and defect probability assessment matrix ; Step 4: Perform online prediction of welding quality evolution trend and assessment of electrode cap remaining life: Establish a time series prediction model based on nonlinear degradation process, use historical quality data series as input, introduce a hidden Markov chain to characterize the physical degradation state of the electrode cap, and calculate the wear amount on the electrode cap end face. With welding cycle number The mapping relationship to predict the future Quality trend curve of individual solder joints ; Step 5, Implement intelligent adjustment decision-making based on industrial knowledge reasoning and reinforcement learning: Integrate and fuse expert rule bases with a deep reinforcement learning decision engine, when perceiving quality indicators... When an anomaly or prediction warning is triggered, the decision engine calls the expert rule base for preliminary diagnosis and uses a deep reinforcement learning model to search for the optimal compensation scheme in the action space, and sends it to the welding controller (12) to realize closed-loop real-time compensation of process parameters. The action space includes: welding current increment control Welding time correction Electrode pressure adjustment And electrode repair / replacement instructions.

[0008] Furthermore, in step 1, the sensor array includes a piezoelectric pressure sensor (7), a Rogowski coil current transformer (8), a Hall voltage sensor (9), and an acoustic emission sensor (10); the robot controller (11) reads the robot's six-axis pose data of the current welding point in real time. Servo motor current value and welding torch opening and closing stroke ; The piezoelectric pressure sensor (7) is installed at the force fulcrum of the welding torch electrode arm to collect electrode pressure signals during the welding process. ; The Rogowski coil current transformer (8) and the Hall voltage sensor (9) are respectively coupled to the secondary side welding circuit; The acoustic emission sensor (10) is adsorbed onto the side wall of the welding torch electrode holder; During the data acquisition process, all devices within the network are synchronized, and the moment when the pressure signal rises to 10% of the set pressure is defined as... The moment when the current rises to 10% of the set current is defined as... Calculate the time difference between the two. ,Will As a characteristic parameter characterizing the mechanical response properties of the welding torch actuator.

[0009] Furthermore, in the spatiotemporal alignment process of step 2, an algorithm based on dynamic time warping is used to eliminate the signal phase difference caused by the small offset of the robot's motion beat. By finding the minimum cost path between multidimensional signal sequences, it is ensured that the pressure rise segment and the current rise segment physically coincide in the feature space. During feature mining, the welding voltage signal is targeted. By calculating the correlation dimension and Lyapunov exponent after phase space reconstruction, the stability of the welding arc and the degree of oxidation on the electrode surface are identified; for acoustic emission signals... The empirical mode decomposition method is used to decompose the signal into several intrinsic mode functions, and the instantaneous frequency spectrum is obtained by Hilbert transform to capture the splashing impact sound caused by the pressure imbalance inside the liquid molten core.

[0010] Furthermore, in the multi-parallel feature extraction channels of step 3, the first channel uses a residual convolutional neural network to process the dynamic waveform data of welding current and voltage, and the second channel uses a long short-term memory network to process the temporal evolution data of pressure and displacement. The residual convolutional neural network channels use alternating stacks of 3x3 and 1x1 convolutional kernels to extract electrophysical characteristics. The long short-term memory network channel uses a bidirectional gated recurrent unit to extract thermal coupling characteristics. The cross-modal attention fusion layer employs a multi-head attention mechanism, where the query matrix... Derived from pressure characteristics, bond matrix Sum matrix Derived from the characteristics of electrical signals, nonlinear coupling of data from different physical dimensions in the feature space is achieved by calculating attention scores: ,in The feature dimension of the key matrix K represents the hidden layer representation dimension when the electrical signal features are encoded as key vectors.

[0011] Furthermore, in step 3, the intelligent welding quality perception model introduces a loss function based on physical constraints. The loss function consists of the prediction error term. and physical law constraint terms Together they constitute: in, To input the total heat for welding, For material density, For specific heat capacity, For the temperature rise, This is the heat loss correction factor; These are the data fitting weight coefficients, used to adjust the optimization intensity of the model's fitting accuracy to historical data; These are the physical constraint weighting coefficients, used to adjust the constraint strength that the model prediction results follow the law of energy conservation; By introducing physical constraints from the first law of thermodynamics, the prediction output of the intelligent welding quality sensing model in regions with scarce data samples conforms to the law of energy conservation.

[0012] Furthermore, in step 4, the remaining effective lifespan of the electrode cap is evaluated online using the following formula. : in, This is the electrode wear rate coefficient. and These are the effective values ​​of the current. and average electrode pressure Sensitivity factor to the effect of wear This refers to the number of critical failures defined in the process specifications. To evaluate the cumulative number of welds completed by the electrode cap at the start time; The system monitors the resistance between electrodes in real time. The changing trend was observed by calculating the contact resistance distribution through applying a low-amplitude, high-frequency probe current signal during the extrusion phase at the beginning of the welding cycle. When the detected contact resistance continuously increased and was accompanied by… When the peak value occurs early, the electrode wear rate coefficient is automatically increased. The weight of the value.

[0013] Furthermore, in step 4, the time series prediction model adopts a combined architecture based on variational autoencoder and bidirectional gated recurrent unit. The variational autoencoder is used to learn the low-dimensional latent space representation of welding process data and remove observation noise, while the bidirectional gated recurrent unit is used to learn the evolution law of latent space features in long-cycle production process.

[0014] Furthermore, in step 5, the reward function of the deep reinforcement learning model... Set as: in, The target melt core size as required by the process specifications; This is an indicator function for no splashing. This represents the economic cost resulting from increased electrode wear. , , These are the weight coefficients of the reward function, where It is the melt core size accuracy weight, used to reward the actual melt core size for being close to the target value; It is a spatter-free reward weight, used to reward welding processes that produce no spatter. It is a cost penalty weight, used to penalize the increased economic cost caused by electrode wear; When performing parameter compensation, the decision engine calculates the current splash critical current. And set the compensation cap as All control commands are sent through the PLC interface, and the actuator is required to complete a handshake response within 5ms.

[0015] Furthermore, in step 5, during the intelligent adjustment decision-making process, the system executes a dynamic wear-repair decision-making mechanism for the maintenance of the electrode cap: based on the remaining effective lifespan of the electrode cap... Prediction results and current weld quality slope A comprehensive evaluation is conducted when perceived quality indicators are considered. In the critical qualified zone and When the absolute value exceeds a preset threshold, the decision engine outputs an "instant grinding" command and calculates the optimal grinding depth based on the current wear level, a preset process knowledge base, and an optimization algorithm. and grinding pressure The control electrode grinding machine (13) actuator completes maintenance.

[0016] A system for implementing the above-mentioned intelligent sensing, prediction, and decision-making method for vehicle body welding quality based on multi-source industrial data includes: The data acquisition module (1) is connected to the piezoelectric pressure sensor (7), Rogowski coil current transformer (8), Hall voltage sensor (9), acoustic emission sensor (10) and robot controller (11) bus at its input end, and is used to acquire multi-source raw signals of the welding process in real time. The data acquisition module (1) includes an edge computing unit (5), which adopts a field programmable gate array architecture to realize real-time FFT transformation of 1MHz acoustic emission signal and send a high-priority interrupt signal within 10ms after the spatter occurs. The core processing server (2) integrates a high-performance graphics processor and a dedicated integrated circuit to run the welding quality intelligent perception model, time series prediction model and decision engine; the core processing server (2) is equipped with a distributed database (6), which adopts a time-series data storage structure to store the original waveform, feature vector, perception result, prediction trajectory and decision action executed for each weld point in a related manner. Industrial gateway (3) is used for bidirectional data interaction between the core processing server (2) and the workshop field layer equipment. The workshop field layer equipment includes robot controller (11), welding controller (12) and electrode grinding machine (13). The human-computer interaction terminal (4) is used to visualize the welding quality map, electrode cap status trend chart and decision suggestions.

[0017] The working principle of this invention is: This invention first deploys high-frequency data acquisition nodes within the robotic welding workstation. The physical wiring of each sensor strictly adheres to electromagnetic compatibility (EMC) standards, and double-shielded twisted-pair cables are used for signal transmission to suppress strong magnetic interference generated by the intermediate-frequency welding inverter power supply. The acquisition nodes communicate with the central processing server via an EtherCAT industrial bus, ensuring that the synchronization delay between the multi-axis robot pose data and welding physical signals is less than 1ms.

[0018] In the data processing flow, the system first initiates a clock alignment (i.e., spacetime alignment mechanism) thread. Time synchronization is performed on all devices within the network using a GPS clock source or the IEEE 1588 protocol. Upon receiving the "pre-pressure start" signal from the welding controller, each channel begins buffering data. Because electrode pressure builds up earlier than current is applied, the system defines the moment when the pressure signal rises to 10% of the set pressure as... The moment when the current rises to 10% of the set current is defined as... Calculate the time difference between the two. This parameter is then used as a characteristic parameter to characterize the mechanical response of the welding torch actuator and input into the subsequent model.

[0019] For the feature mining stage, this invention not only extracts conventional statistical features (mean, standard deviation, peak-to-peak value), but also delves into the nonlinear dynamic features of the signal. For example, for welding voltage signals... By calculating the correlation dimension and Lyapunov exponent after phase space reconstruction, the stability of the welding arc and the degree of oxidation on the electrode surface are identified. For acoustic emission signals, the empirical mode decomposition (EMD) method is used to decompose the signal into several intrinsic mode functions (IMFs), and the instantaneous frequency spectrum is obtained through Hilbert transform to accurately capture the subtle splashing impact sound caused by the internal pressure imbalance of the liquid molten nugget.

[0020] In the construction of deep learning perception models, residual convolutional layers use alternating stacks of 3x3 and 1x1 convolutional kernels to capture transient change features in voltage and current waveforms. The cross-modal attention fusion layer employs a multi-head attention mechanism, where the query matrix... Derived from pressure features, bond matrix Sum matrix Based on electrical signal characteristics. Attention scores are calculated as follows: Nonlinear coupling of data from different physical dimensions in the feature space is achieved, ensuring that the model can still accurately infer the formation state of the melt nucleus through compensatory changes in the current waveform even under conditions of large electrode pressure fluctuations.

[0021] In the predictive model, this invention employs a combined architecture based on a variational autoencoder (VAE) and a bidirectional gated recurrent unit (Bi-GRU). The VAE is used to learn the low-dimensional latent space representation of the welding process data, eliminating observation noise in the stochastic process; the Bi-GRU is responsible for learning the evolution of latent space features in the long-cycle production process. The model output includes not only the mean of the quality indicators but also the confidence interval, providing a risk probability reference for the decision-making layer.

[0022] The intelligent decision engine adheres to strict physical consistency constraints when performing parameter compensation. For example, when the system detects incomplete weld nugget development due to excessive gaps in the plate material matching, the decision engine calculates the appropriate current compensation. The heat requirements must comply with Joule's law. Simultaneously, to prevent excessive compensation current from causing splashing, the decision engine will calculate the current critical splashing current. And set the compensation cap as All control commands are sent via the PLC interface in the form of structured messages, and the actuator is required to complete a handshake response within 5ms to ensure that the command takes effect before the next welding cycle.

[0023] In electrode cap life assessment, if the system monitors 10 consecutive electrode pressure peaks at solder joints... It exhibits a monotonically decreasing trend, and the ratio of mid-to-high frequency energy in the acoustic emission signal is abnormally increased, even affecting the current sensing quality of the solder joint. Even if the electrode cap is still within acceptable limits, the predictive model will still determine that it has entered a period of severe deformation. At this point, the decision engine will automatically insert a short grinding cycle during the next workpiece loading and unloading interval, based on the current production schedule, instead of waiting for quality defects before stopping the machine. This dynamic predictive maintenance mode maximizes the balance between production efficiency and product quality.

[0024] In summary, this invention constructs a highly intelligent automotive body welding quality control ecosystem through deep fusion perception of multi-source data, forward-looking prediction based on evolution trends, and closed-loop decision-making based on knowledge reasoning. This invention can adapt to different combinations of sheet metal thicknesses, different material properties (such as galvanized steel, high-strength steel, and aluminum alloys), and complex on-site working conditions, providing crucial technical support for the digital transformation of automotive manufacturing enterprises.

[0025] The beneficial effects of this invention are: This invention constructs a comprehensive welding dynamic characteristic sensing system by integrating multi-source heterogeneous data such as current, voltage, pressure, displacement, and acoustic emission. Compared with traditional single-parameter monitoring, it can capture minute physical field fluctuations, and improves the accuracy of sensing weld nugget diameter and shear strength by more than 15% in complex production environments, significantly enhancing the robustness of the system.

[0026] This invention introduces a quality prediction model based on time series evolution, achieving a leap from "post-event detection" to "pre-event early warning." By online modeling of the electrode cap degradation trend, it can accurately predict potential quality deviation risks up to 100 solder joints in advance, allowing sufficient intervention time for production line maintenance and effectively avoiding the generation of batch welding defects.

[0027] This invention establishes a complete closed loop from perception and prediction to decision-making through a reinforcement learning decision engine. The system can not only automatically identify anomalies but also provide precise parameter compensation suggestions and electrode maintenance plans based on real-time operating conditions, achieving dynamic optimization of the electrode cap replacement cycle. While ensuring welding quality, the average service life of the electrode caps is increased by 20%, significantly reducing spare parts costs and downtime for maintenance.

[0028] This invention adopts an architecture that combines edge computing and central cloud collaboration, ensuring the real-time performance of ultra-high frequency signal processing and the depth of analysis of massive historical data. It can adapt to the high-speed and flexible production requirements of modern automotive welding lines and provides standardized technical support for quality control in smart factories. Attached Figure Description

[0029] Figure 1 This is a block diagram of the overall structure of the system of the present invention; Figure 2 This is a schematic diagram of the overall process of the method of the present invention; Figure 3 This is a schematic diagram of the network topology of the intelligent sensing model for welding quality in this invention; Figure 4 This is a logical framework diagram of the time series prediction model and decision engine in this invention; Figure 5 This is a flowchart illustrating the decision-making process for electrode cap life assessment and dynamic maintenance in an embodiment of the present invention.

[0030] The attached diagram is labeled as follows: 1. Data acquisition module, 2. Core processing server, 3. Industrial gateway, 4. Human-machine interaction terminal, 5. Edge computing unit, 6. Distributed database, 7. Piezoelectric pressure sensor, 8. Rogowski coil current transformer, 9. Hall voltage sensor, 10. Acoustic emission sensor, 11. Industrial robot controller, 12. Welding controller, 13. Electrode grinding machine. Detailed Implementation

[0031] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0032] This intelligent sensing, prediction, and decision-making method and system for vehicle body welding quality based on multi-source industrial data achieves full-stack closed-loop control from bottom-level perception to high-level decision-making in specific engineering practices through in-depth deconstruction and digital reconstruction of the physical field signals of the welding production line. To enable those skilled in the art to more clearly understand the technical solution and implementation details of this invention, a detailed description is provided below, combining specific hardware configurations, algorithm logic, and experimental data.

[0033] In the production environment of a car body welding workshop, the hardware architecture of this invention primarily relies on the physical deployment of a high-frequency synchronous acquisition system. The data acquisition module 1 constructs a multi-dimensional physical quantity capture network through precise placement at the robot's end effector. Specifically, a piezoelectric pressure sensor 7 is installed at the force fulcrum of the welding torch electrode arm. This position minimizes the inertial impact during the torch's activation process while accurately capturing minute pressure fluctuations during the welding extrusion, thermal expansion, and sustaining stages. The selected piezoelectric sensor has a high dynamic response frequency, with a sampling frequency set to 1kHz, sufficient to distinguish the electrode bounce phenomenon caused by the weld nugget growth process. Simultaneously, a Rogowski coil current transformer 8 and a Hall voltage sensor 9 are coupled to the secondary output circuit of the welding transformer. The Rogowski coil is non-contactly mounted on the secondary flexible copper busbar, utilizing the Faraday electromagnetic induction principle to acquire the welding current signal. The Hall voltage sensor is connected in parallel between the positive and negative electrode handles to acquire the welding voltage signal in real time. The sampling frequency of these two electrical signals is uniformly set to 10kHz to ensure complete capture of each half-wave characteristic of the intermediate frequency inverter welding power supply (typically 1kHz). For the stress wave signal generated during welding, the acoustic emission sensor 10 is fixed to the side wall of the welding torch electrode holder by magnetic adsorption. Its internal piezoelectric ceramic element can sense high-frequency vibrations in the 100kHz to 1MHz frequency band, specifically used to identify the splashing of molten metal and the initiation of internal cracks.

[0034] At the data transmission level, this invention utilizes the bus interface of the industrial robot controller 11 to synchronously read the robot's real-time status via Profinet or EtherCAT protocols. This data includes the absolute encoder values ​​for the robot's six axes. This is used to calculate the welding torch's attitude vector in space; simultaneously, it reads the motor current value of the servo electrode driver, which is strongly correlated with the electrode pressure and can be used as a redundancy verification of the piezoelectric pressure signal; the welding torch opening and closing stroke... The data is acquired using a high-precision displacement sensor with a resolution of 0.01 mm, used to characterize the thermoplastic deformation of the weld joint. All raw signals are initially processed by the edge computing unit 5, which uses a field-programmable gate array (FPGA) as its core operator and is capable of performing real-time pipelined parallel computing at a sampling rate of 1 MHz.

[0035] For the spatiotemporal alignment and feature mining of multi-source heterogeneous data, the system performs refined preprocessing in the core processing server 2. Due to differences in the physical response time and signal transmission link delay of different sensors, the system first uses a zero-phase digital filter to denoise the original signal, ensuring that the filtering process does not produce phase shift. Subsequently, the moment when the welding current jumps from zero to 10% of its rated value is used as the time reference zero point. To eliminate the phase difference caused by small shifts in the robot's motion rhythm, an alignment algorithm based on Dynamic Time Warping (DTW) is further introduced. This algorithm constructs a cost matrix to find the shortest warped path between two time-series signals, ensuring that the inflection point of the slope of the pressure rise segment and the starting point of the current heating segment are physically strictly coincident. In the feature mining stage, the... and Wavelet packet decomposition was performed, selecting the db4 wavelet as the basis function to decompose the signal into three levels, generating coefficients for eight sub-bands. The energy entropy of each band was then calculated. This quantifies the disorder of the welding energy distribution in the frequency domain. Specifically, it addresses the pressure signal. The system uses a feature point detection algorithm to pinpoint the pressure peak caused by welding thermal expansion. And calculate the slope of pressure rise from the start of energization to the peak pressure. This indicator directly reflects the temperature rise rate in the weld nugget area. Acoustic emission signal The processing focuses on transient characteristics, obtaining spectral features through short-time Fourier transform (STFT), with a particular emphasis on monitoring the centroid frequency. The drift situation, when Significant high-frequency shifts were observed, accompanied by peak count rates. When there is a sudden increase, it is determined to be a metal splashing incident.

[0036] In constructing the intelligent sensing model for welding quality, this invention employs a multimodal fusion deep network topology. The model is deployed on the GPU cluster of the core processing server 2. The first channel of the sensing model is a Residual Convolutional Neural Network (ResNet), designed with 18 convolutional layers, specifically for processing dual-channel one-dimensional time series consisting of current and voltage. Through 3x3 one-dimensional convolutional kernels, the model can adaptively extract the changing characteristics of welding dynamic impedance. The second channel is a Bidirectional Long Short-Term Memory Network (Bi-LSTM), containing two hidden layers, each with 128 neurons, used to handle stress with significant temporal dependencies. and displacement Signal. The forgetting gate mechanism of LSTM can effectively preserve the initial contact state information of the squeeze segment formation and pass it to the subsequent nucleus development stage. The feature vectors of the two channels are aggregated in the cross-modal attention fusion layer. This layer calculates the query matrix. (Derived from pressure characteristics) and bond matrix The attention weight matrix is ​​generated by the dot product of electrical features. This mechanism ensures that when electrode wear causes abnormal contact resistance, the model can assign higher perception weight to pressure signals, thereby correcting misjudgments caused by purely electrical signals. Finally, the model outputs a perception quality evaluation index. This includes the predicted value of the molten core diameter. Shear strength And a defect probability matrix covering cold solder joints, overheating, and spatter. To further improve the reliability of the model under extreme conditions, this invention introduces a physical constraint term based on the first law of thermodynamics into the loss function. This constraint requires that the predicted melt core diameter must be consistent with its input heat. (Obtained through integration of current, voltage, and time) and maintain logical consistency with the thermal properties of the material. If the prediction results violate the law of conservation of energy, the physical constraint terms will incur a high penalty, forcing the model weights to converge to a physically reasonable interval.

[0037] Building upon real-time sensing, this invention further performs quality evolution trend prediction. The core of this process lies in modeling the electrode cap degradation trajectory. With the number of welding cycles... With increasing temperature and pressure, the electrode cap end face undergoes alloying wear and plastic deformation. This system characterizes the physical state transition of the electrode cap using a hidden Markov chain (HMM), thus mitigating the unobservable wear amount. This is mapped to an observable sequence of quality indicators. Utilizing the aforementioned... The evaluation formula comprehensively considers the effective value of the current. The Joule heating effect and mean pressure generated The resulting mechanical wear effect. Sensitivity factor. and The system is trained offline using historical production big data. Furthermore, by applying a low-amplitude probe current of 100A and 1kHz during the pre-pressure phase of each welding cycle, the system uses a Hall voltage sensor to capture the minute voltage drop at this point, and then calculates the contact resistance. The evolution trend of this resistance value is an important basis for judging the degree of oxidation of the electrode cap. When the detected value... When the slope of the curve exceeds a preset threshold, the system will automatically increase the base wear rate coefficient. This makes remaining life predictions more conservative, ensuring production safety.

[0038] The intelligent adjustment decision layer embodies the closed-loop control capability of this invention. The decision engine is integrated into the core processing server 2 and employs a deep reinforcement learning (DRL) algorithm for policy optimization. The reinforcement learning agent takes the perceived quality state, the predicted degradation trend, and the current process parameters as its input state space. Its action space encompasses current compensation. Time compensation Pressure adjustment And maintenance instructions. At the end of each welding cycle, the system... The deviation from the set target is used to calculate a reward value. If the current weld shows a trend of shrinking weld nugget diameter, the AI ​​will attempt to provide a small current increment compensation. To prevent spatter caused by blind compensation, the decision engine will search the built-in expert rule base. The rule base contains pre-stored spatter critical curves for different material combinations (such as DC51D + duplex steel DP780). Only when the compensation action suggested by reinforcement learning is within the safe physical boundaries will the instruction be issued to the welding controller 12 for execution via the industrial gateway 3. For electrode maintenance, the system abandons the traditional fixed-frequency grinding and instead implements a dynamic grinding mechanism. Based on the prediction results, when the system anticipates that the quality of the next 50 weld points will fall below the critical threshold, it will automatically trigger the electrode grinding machine 13 during idle periods on the production line. The decision engine will calculate the optimal grinding parameters, such as grinding pressure. The pressure is 0.15 MPa, and the grinding depth is... The diameter is 0.05mm, thus restoring the electrode geometry with minimal end face cutting and maximizing the electrode cap life.

[0039] To verify the technical effects of the present invention, a complete embodiment is presented below. In this embodiment, the welding object is an automotive side panel assembly, and the material combination is 1.2mm galvanized steel sheet + 1.5mm high-strength steel sheet. The experiment was conducted on an automated production line with a cycle time of 60 JPH, comparing the system of the present invention with a traditional monitoring scheme based on constant current control and fixed number of grinding cycles.

[0040] In this embodiment, the system first established a data baseline for 20,000 weld points. Data acquisition module 1 continuously monitored acoustic emission signals at a frequency of 1 MHz. During a specific welding cycle, the sensing model detected a slight dip in the current waveform during the later stages of heating, while the pressure signal... At the corresponding moment, a high-frequency pulse appeared. The cross-modal attention fusion layer recognized that this combination of features matched the physical pattern of "latent splashing," i.e., small-scale splashing occurred inside the molten core but did not eject the electrode pressure ring. At this time, the perception model output... The predicted value is 4.2mm, slightly lower than the lower limit of the process requirement of 4.5mm. Meanwhile, the prediction model, through analysis of historical sequences of the past 500 solder joints, found that the peak electrode pressure... It has shown a continuous, slow downward trend. The evaluation results showed that the electrode cap could still maintain approximately 120 qualified solder joints.

[0041] In response to this situation, the decision engine immediately initiated an intervention process. First, it determined that the current quality degradation was caused by current density dilution due to the increased diameter of the electrode cap end face. Based on a reinforcement learning strategy, the system issued a... The current compensation command was issued and applied to the next weld joint. Subsequently, the sensing model fed back feedback that the adjusted weld nugget diameter had recovered to 4.7mm, and the quality level returned to "Excellent". When welding reached the 12100th weld, the system monitored the contact resistance. An abnormal fluctuation occurred, and the predictive model warned that the remaining lifespan was only 10 cycles. The decision engine then instructed the electrode grinding machine 13 to perform a precision grinding to a depth of 0.08 mm during the 20-second interval when the robot was changing workpieces. The pressure curve of the first weld point after grinding completely returned to its original state, and the system successfully avoided potential batch defects.

[0042] To quantitatively illustrate the advancement of this invention, the table below (Table 1) lists the key performance indicators of this embodiment compared with those of a conventional monitoring scheme (control group). The control group uses a conventional current threshold alarm mechanism, and the electrode caps are ground every 200 weld points.

[0043] Table 1: Performance Comparison Table of Embodiments of the Invention and Traditional Solutions

[0044] The data in Table 1 clearly demonstrates the overwhelming advantage of this invention in sensing accuracy. Traditional monitoring solutions, unable to process pressure and acoustic emission signals, have virtually no ability to identify complex defects such as "cold solder joints" and "internal spatter." This invention, however, significantly improves defect identification accuracy to 97.8% through multimodal feature fusion. Regarding electrode cap lifespan management, this invention... The formula and reinforcement learning decision-making not only extended the physical life of the electrode cap itself, but also reduced unnecessary production downtime through a dynamic compensation mechanism.

[0045] In terms of system software architecture, a high-concurrency processing engine runs inside the core processing server 2. To support real-time feature extraction at a 1MHz sampling rate, the FPGA logic inside the edge computing unit 5 is divided into a preprocessing pipeline and a feature mapping matrix. The raw acoustic emission signal undergoes a 512-point Fast Fourier Transform (FFT) within the FPGA, and the calculated spectral energy features are then transmitted at high speed to the core processing server 2 via DMA (Direct Memory Access). The distributed database 6 employs a combined architecture of the high-performance time-series database InfluxDB and the relational database PostgreSQL. The former stores millisecond-level waveform data, while the latter stores structured solder joint quality files. Each file carries a unique workpiece identification code (VIN) and solder joint number, achieving full lifecycle traceability of quality data.

[0046] Furthermore, as a deep optimization method for the perception model of this invention, the ResNet channel incorporates a transfer learning strategy during training. For different vehicle models, the system can first be pre-trained on a general welding dataset to obtain basic convolutional kernel weights, and then fine-tuned with small samples under specific sheet metal combinations. Experiments show that, using this method, the system can achieve a perception accuracy of over 90% when facing new materials (such as aluminum alloys or ultra-high-strength steel) with only about 200 samples collected, significantly reducing the deployment and debugging cycle on-site.

[0047] In the specific engineering implementation details, to ensure rigorous synchronization of multi-source data, the industrial gateway 3 employs a hardware-level clock alignment mechanism. Whenever the robot control bus sends a "welding torch in position" signal, the system generates a globally unique trigger pulse, which simultaneously latches the counter values ​​of the piezoelectric sensor, current sensor, and robot encoder. In this way, even with minor jitter in network transmission, the server can rearrange all signals onto the same timeline based on the hardware timestamps. This nanosecond-level synchronization precision is a prerequisite for achieving physical alignment in the DTW algorithm.

[0048] In summary, this invention not only provides a technical means but also constructs a complete data-driven industrial control paradigm. Through in-depth analysis of physical signals, it shields the uncertainties caused by complex electromagnetic interference and mechanical wear on the production line. From centimeter-level prediction accuracy of the weld nugget diameter to millisecond-level decision feedback response speed, this invention demonstrates outstanding engineering application value. The real-time welding quality map presented by the human-machine interface terminal 4 not only allows on-site operators to intuitively grasp the "health status" of each weld point but also provides a scientific basis for the factory's spare parts planning through electrode cap status trend charts. This transformation from data perception to logical prediction and then to autonomous decision-making is precisely the core technological support for achieving zero-defect production in modern smart car body factories.

[0049] Further in-depth technical analysis of the system configuration of this invention is provided. In the data acquisition module 1, the piezoelectric pressure sensor 7 is preferably a high-sensitivity model with integrated circuit piezoelectric (IEPE) output, whose linearity is better than 0.5% FS, and can withstand the instantaneous high overload generated by the rapid closing of the welding torch. At the welding controller 12, the system acquires the welding voltage through a dedicated high-speed analog output port, and performs potential offset processing through an isolation operational amplifier to ensure that it enters the signal range of the acquisition card while achieving electrical isolation from the high-voltage circuit. The six-axis pose data provided by the industrial robot controller 11 is used in this invention to calculate the perpendicularity angle between the welding torch electrode and the surface of the plate. When the perception model detects that the perpendicularity deviation exceeds 3 degrees, the decision engine will automatically fine-tune the electrode pressure. The set value is used to compensate for the uneven distribution of contact stress caused by the deflection angle.

[0050] In the design of the loss function for the perceptual model, the formula is... Possesses a strong engineering physics background. Weighting coefficients. and It will be dynamically adjusted based on the convergence progress during the training phase. Among these, the physical law constraint term... The coefficient takes into account the heat loss carried away by the water-cooled electrodes. In practical applications, the system dynamically corrects this by monitoring the temperature difference between the inlet and outlet cooling water of the electrodes. The value of is determined by this hybrid architecture of physical modeling and deep learning, which effectively solves the problem of outrageous predictions that may occur when deep learning models encounter unseen combinations of materials ("cold start problem"), thus ensuring the robustness of industrial control.

[0051] Regarding the electrode cap lifetime assessment described in step four, the system maintains a dynamically evolving degradation baseline library. Slight differences in the material of electrode caps from different batches will be detected through the probe current signal. It is captured in real time. When the first 10 solder joints after replacing the electrode cap exhibit abnormally high initial resistance, the system will automatically determine that the batch of electrode caps contains a lot of impurities and automatically raise the lifespan warning threshold. The reduction is 10%. This intelligent adaptive capability enables the system to be compatible with spare parts from different suppliers, maintaining a stable level of quality control.

[0052] When implementing the intelligent adjustment decision in step five, the reward mechanism of the deep reinforcement learning model is designed as follows: ,in This is an indicator function for no splashing. This represents the economic cost of increased electrode wear. The function guides the agent to minimize unnecessary current increases while ensuring welding quality, thereby protecting the electrode cap.

[0053] As another key technological improvement of this invention, the transient spatter detection logic integrated within the edge computing unit 5 employs a detection algorithm based on multi-scale energy operators. When the energy mutation rate of a 1MHz acoustic emission signal exceeds five times the average value of conventional welding within a specific frequency band (e.g., 300kHz-500kHz), the FPGA immediately generates a hard real-time interrupt. This interrupt signal bypasses the ordinary network protocol stack and reaches the core processing server 2 directly through a dedicated fast channel. Upon receiving the interrupt, the server immediately locks the current weld point into a "suspected spatter" state and retrieves the concurrent robot servo current waveform for secondary confirmation. If spatter is confirmed, the system not only records the defect but also determines whether the problem stems from excessive plate gap or current overload based on the timing of the spatter occurrence (e.g., early or late heating), and provides targeted diagnostic suggestions to the decision engine.

[0054] The core processing server 2 of this invention adopts a containerized deployment mode, and the various model components (perception, prediction, and decision-making) communicate through lightweight message queues. This architecture supports horizontal scaling; as the number of production line robots increases, computing power can be linearly increased simply by adding computing nodes. The distributed database 6 employs a hot and cold data separation storage strategy. The full raw waveforms of the past 7 days are stored in a real-time database composed of high-speed solid-state drives, while feature data and perception results older than 7 days are migrated to a massive storage cluster, providing a data mining foundation for subsequent process optimization and long-cycle quality traceability.

[0055] On the human-machine interface terminal 4, this system provides a three-dimensional visualized digital twin interface. This interface synchronously maps the actual pose of the production line robot and directly marks the perceived quality of each weld point on the vehicle body CAD model using a color heatmap (green represents excellent, yellow represents acceptable, and red represents defective). When the operator clicks on a red dot, the system automatically pops up a multi-source feature comparison curve for that weld point and provides rectification suggestions recommended by the decision engine (e.g., suggesting checking the alignment of welding clamp #3 or suggesting immediate replacement of the electrode cap). This highly integrated information display method significantly reduces the technical threshold for on-site maintenance personnel and shortens the troubleshooting cycle.

[0056] In summary, the intelligent sensing, prediction, and decision-making method and system for automotive body welding quality based on multi-source industrial data described in this invention completely solves the long-standing problem of welding quality fluctuations in the automotive manufacturing industry through the extreme application of underlying sensing technologies, the deep integration of mid-level multimodal algorithms, and the scientific construction of top-level closed-loop decision-making. It can not only detect problems in real time but also accurately predict risks and automatically eliminate hidden dangers. This complete mapping from sensory perception to the brain and then to limb movements marks a qualitative leap in industrial welding process control from traditional "experience-driven" to "intelligent-driven," possessing extremely high industrial application value and economic and social benefits.

[0057] In long-term operation on actual production lines, the system of this invention has also demonstrated extremely strong environmental adaptability. Due to the presence of numerous frequency converters, servo motors, and strong electromagnetic pulses generated by high-frequency spot welding in the body shop, all analog front-ends of the system's acquisition link adopted balanced circuit inputs, coupled with differential amplifiers with high common-mode rejection ratios, ensuring that the signal-to-noise ratio remained consistently above 40dB. The perception model running inside the core processing server 2 exhibited a perception accuracy drift of less than 1% after experiencing four quarters of temperature changes (corresponding to fluctuations in workshop humidity and temperature), fully demonstrating the superior robustness of the physically constrained deep learning model in engineering practice.

[0058] Furthermore, this invention supports remote OTA (Over-the-Air) updates. When a new vehicle platform is introduced or new materials (such as the heterogeneous bonding of carbon fiber reinforced composites and aluminum alloys) are launched, experts at the R&D center can distribute updated model weights and expert rules via the cloud. Edge computing units and core servers at the factory can complete algorithm upgrades without interrupting production, enabling continuous evolution of the control system. Through this digital closed loop, automotive manufacturers can manage their core production processes with unprecedented flexibility and precision, truly entering the deeper waters of Industry 4.0.

[0059] The experimental data, formulas, and specific hardware models cited in this specification are all optimal choices for specific industrial scenarios, intended to provide a practical reference benchmark for those skilled in the art. In practical applications, depending on the specific working conditions such as the welding gun type (e.g., C-type or X-type welding gun), the type of sheet coating, and the production cycle speed, the parameter configuration of each component and the hyperparameters within the algorithm (e.g., convolution kernel stride, number of LSTM neurons, reinforcement learning learning rate, etc.) can be specifically optimized through engineering. The scope of protection of this invention should cover all variations based on the core ideas of multi-source fusion, trend prediction, and intelligent closed-loop decision-making.

[0060] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0061] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for intelligent sensing, prediction, and decision-making regarding vehicle body welding quality based on multi-source industrial data, characterized in that, Includes the following steps: Step 1: Establish a high-frequency synchronous acquisition system for multi-source industrial data: Arrange a sensor array, a welding power communication interface and an industrial robot controller (11) bus at the welding execution end to acquire the original time domain signal of the entire welding cycle in real time. Step 2: Perform spatiotemporal alignment and feature mining of multi-source heterogeneous data: Input the acquired time-domain signals into the preprocessing module to filter out high-frequency electromagnetic noise, and then, taking the welding current triggering time as zero point, perform spatiotemporal alignment and feature mining of the electrode pressure signal. Welding current signal Welding voltage signal High-frequency acoustic emission signals Furthermore, time-axis resampling and alignment of robot state data are performed to form a multi-dimensional spatiotemporal feature vector sequence; time-frequency domain feature extraction is then performed on welding current and voltage signals, and the energy entropy of each frequency band is calculated. For electrode pressure signals Peak extraction pressure Pressure rise slope and time constant For acoustic emission signals Extract spectral features, including centroid frequency. and peak count rate ; Step 3: Construct a welding quality intelligent perception model based on a multimodal fusion deep network topology. The welding quality intelligent perception model includes multiple parallel feature extraction channels and a cross-modal attention fusion layer. The output vectors of each feature extraction channel are weighted and aggregated in the cross-modal attention fusion layer, and the perception weights are adaptively allocated to finally output the perception quality evaluation index of the current weld point. The perceived quality index Including predicted melting core diameter Predicted shear strength and defect probability assessment matrix ; Step 4: Perform online prediction of welding quality evolution trend and assessment of electrode cap remaining life: Establish a time series prediction model based on nonlinear degradation process, use historical quality data series as input, introduce a hidden Markov chain to characterize the physical degradation state of the electrode cap, and calculate the wear amount on the electrode cap end face. With welding cycle number The mapping relationship to predict the future Quality trend curve of individual solder joints ; Step 5, Implement intelligent adjustment decision-making based on industrial knowledge reasoning and reinforcement learning: Integrate and fuse expert rule bases with a deep reinforcement learning decision engine, when perceiving quality indicators... When an anomaly or prediction warning is triggered, the decision engine calls the expert rule base for preliminary diagnosis and uses a deep reinforcement learning model to search for the optimal compensation scheme in the action space, and sends it to the welding controller (12) to realize closed-loop real-time compensation of process parameters. The action space includes: welding current increment control, welding time correction, and electrode pressure adjustment. And electrode repair / replacement instructions.

2. The intelligent sensing, prediction, and decision-making method for vehicle body welding quality based on multi-source industrial data according to claim 1, characterized in that, In step 1, the sensor array includes a piezoelectric pressure sensor (7), a Rogowski coil current transformer (8), a Hall voltage sensor (9), and an acoustic emission sensor (10); the robot controller (11) reads the robot's six-axis pose data of the current welding point in real time. Servo motor current value and welding torch opening and closing stroke ; The piezoelectric pressure sensor (7) is installed at the force fulcrum of the welding torch electrode arm to collect electrode pressure signals during the welding process. ; The Rogowski coil current transformer (8) and the Hall voltage sensor (9) are respectively coupled to the secondary side welding circuit; The acoustic emission sensor (10) is adsorbed onto the side wall of the welding torch electrode holder; During the data acquisition process, all devices within the network are synchronized, and the moment when the pressure signal rises to 10% of the set pressure is defined as... The moment when the current rises to 10% of the set current is defined as... Calculate the time difference between the two. ,Will As a characteristic parameter characterizing the mechanical response properties of the welding torch actuator.

3. The intelligent sensing, prediction, and decision-making method for vehicle body welding quality based on multi-source industrial data according to claim 1, characterized in that, In the spatiotemporal alignment process of step 2, an algorithm based on dynamic time warping is used to eliminate the signal phase difference caused by the small offset of the robot's motion beat. By finding the minimum cost path between multidimensional signal sequences, it is ensured that the pressure rise segment and the current rise segment physically coincide in the feature space. During feature mining, the welding voltage signal is targeted. By calculating the correlation dimension and Lyapunov exponent after phase space reconstruction, the stability of the welding arc and the degree of oxidation on the electrode surface are identified; for acoustic emission signals... The empirical mode decomposition method is used to decompose the signal into several intrinsic mode functions, and the instantaneous frequency spectrum is obtained by Hilbert transform to capture the splashing impact sound caused by the pressure imbalance inside the liquid molten core.

4. The intelligent sensing, prediction, and decision-making method for vehicle body welding quality based on multi-source industrial data according to claim 1, characterized in that, In the multi-parallel feature extraction channel of step 3, the first channel uses a residual convolutional neural network to process the dynamic waveform data of welding current and voltage, and the second channel uses a long short-term memory network to process the time-series evolution data of pressure and displacement. The residual convolutional neural network channels use alternating stacks of 3x3 and 1x1 convolutional kernels to extract electrophysical characteristics. The long short-term memory network channel uses a bidirectional gated recurrent unit to extract thermal coupling characteristics. The cross-modal attention fusion layer employs a multi-head attention mechanism, where the query matrix... Derived from pressure characteristics, bond matrix Sum matrix Derived from the characteristics of electrical signals, nonlinear coupling of data from different physical dimensions in the feature space is achieved by calculating attention scores: ,in The feature dimension of the key matrix K represents the hidden layer representation dimension when the electrical signal features are encoded as key vectors.

5. The intelligent sensing, prediction, and decision-making method for vehicle body welding quality based on multi-source industrial data according to claim 1, characterized in that, In step 3, the intelligent sensing model for welding quality introduces a loss function based on physical constraints. The loss function consists of the prediction error term. and physical law constraint terms Together they constitute: in, To input the total heat for welding, For material density, For specific heat capacity, For the temperature rise, This is the heat loss correction factor; These are the data fitting weight coefficients, used to adjust the optimization intensity of the model's fitting accuracy to historical data; These are the physical constraint weighting coefficients, used to adjust the constraint strength that the model prediction results follow the law of energy conservation; By introducing physical constraints from the first law of thermodynamics, the prediction output of the intelligent welding quality sensing model in regions with scarce data samples conforms to the law of energy conservation.

6. The intelligent sensing, prediction, and decision-making method for vehicle body welding quality based on multi-source industrial data according to claim 1, characterized in that, In step 4, the remaining effective lifespan of the electrode cap is evaluated online using the following formula. : in, This is the electrode wear rate coefficient. and These are the effective values ​​of the current. and average electrode pressure Sensitivity factor to the effect of wear This refers to the number of critical failures defined in the process specifications. To evaluate the cumulative number of welds completed by the electrode cap at the start time; The system monitors the resistance between electrodes in real time. The changing trend was observed by calculating the contact resistance distribution through applying a low-amplitude, high-frequency probe current signal during the extrusion phase at the beginning of the welding cycle. When the detected contact resistance continuously increased and was accompanied by… When the peak value occurs early, the electrode wear rate coefficient is automatically increased. The weight of the value.

7. The intelligent sensing, prediction, and decision-making method for vehicle body welding quality based on multi-source industrial data according to claim 1, characterized in that, In step 4, the time series prediction model adopts a combined architecture based on variational autoencoder and bidirectional gated recurrent unit. The variational autoencoder is used to learn the low-dimensional latent space representation of welding process data and remove observation noise, while the bidirectional gated recurrent unit is used to learn the evolution law of latent space features in long-cycle production process.

8. The intelligent sensing, prediction, and decision-making method for vehicle body welding quality based on multi-source industrial data according to claim 1, characterized in that, In step 5, the reward function of the deep reinforcement learning model Set as: in, The target melt core size as required by the process specifications; This is an indicator function for no splashing. This represents the economic cost resulting from increased electrode wear. , , These are the weight coefficients of the reward function, where It is the melt core size accuracy weight, used to reward the actual melt core size for being close to the target value; It is a spatter-free reward weight, used to reward welding processes that produce no spatter. It is a cost penalty weight, used to penalize the increased economic cost caused by electrode wear; When performing parameter compensation, the decision engine calculates the current splash critical current. And set the compensation cap as All control commands are sent through the PLC interface, and the actuator is required to complete a handshake response within 5ms.

9. The intelligent sensing, prediction, and decision-making method for vehicle body welding quality based on multi-source industrial data according to claim 1, characterized in that, In step 5, during the intelligent adjustment decision-making process, the system executes a dynamic wear-repair decision-making mechanism for the maintenance of the electrode cap: based on the remaining effective lifespan of the electrode cap... Prediction results and current weld quality slope A comprehensive evaluation is conducted when perceived quality indicators are considered. In the critical qualified zone and When the absolute value exceeds a preset threshold, the decision engine outputs an "instant grinding" command and calculates the optimal grinding depth based on the current wear level, a preset process knowledge base, and an optimization algorithm. and grinding pressure The control electrode grinding machine (13) actuator completes maintenance.

10. A system for intelligent sensing, prediction, and decision-making of vehicle body welding quality based on multi-source industrial data, implementing the method described in any one of claims 1-9, characterized in that, include: The data acquisition module (1) is connected to the piezoelectric pressure sensor (7), Rogowski coil current transformer (8), Hall voltage sensor (9), acoustic emission sensor (10) and robot controller (11) bus at its input end, and is used to acquire multi-source raw signals of the welding process in real time. The data acquisition module (1) includes an edge computing unit (5), which adopts a field programmable gate array architecture to realize real-time FFT transformation of 1MHz acoustic emission signal and send a high-priority interrupt signal within 10ms after the spatter occurs. The core processing server (2) integrates a high-performance graphics processor and a dedicated integrated circuit to run the welding quality intelligent perception model, time series prediction model and decision engine; the core processing server (2) is equipped with a distributed database (6), which adopts a time-series data storage structure to store the original waveform, feature vector, perception result, prediction trajectory and decision action executed for each weld point in a related manner. Industrial gateway (3) is used for bidirectional data interaction between the core processing server (2) and the workshop field layer equipment. The workshop field layer equipment includes robot controller (11), welding controller (12) and electrode grinding machine (13). The human-computer interaction terminal (4) is used to visualize the welding quality map, electrode cap status trend chart and decision suggestions.

Images