Welding quality monitoring method and device based on industrial internet

Abstract

The embodiment of the invention provides a welding quality monitoring method and device based on the industrial internet, a deep convolutional neural network (DCNN) and a recurrent neural network (RNN) are combined to form a hybrid model, local features and time sequence features of welding data are fully fused, the former provides high-value features, the latter realizes time sequence reasoning, and the time sequence reasoning efficiency is improved. And a complete link from original data to quality evaluation is completed together. According to the method, the accuracy of online detection of the welding quality can be improved, powerful support is provided for optimization of the welding process and control of the welding quality, the blank of efficient, intelligent and self-adaptive online detection technology of the welding quality in the field is filled up, and powerful technical support is provided for modern industrial welding production.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, in particular to a welding quality monitoring method and device based on industrial internet. BACKGROUND

[0002] In modern industrial production, welding, as one of the key manufacturing processes, is widely used in aerospace, automobile manufacturing, shipbuilding engineering and many other fields. The quality of welding directly affects product performance, structural reliability and production cost and many other key factors. Traditional welding quality detection methods rely on post-inspection, such as sampling or comprehensive inspection of welds through ultrasonic detection, radiographic detection and other methods. However, these methods often have problems such as detection lag, high cost and difficulty in real-time feedback adjustment, which cannot meet the urgent needs of today's industrial production for efficient, accurate and intelligent welding quality control.

[0003] With the acceleration of automation and intelligentization of manufacturing industry, people have higher requirements for real-time monitoring and online quality detection of the welding process. On the one hand, modern welding processes are becoming more and more complex, involving a variety of materials, different welding parameters and variable working conditions, which makes the amount of data generated during welding grow explosively. Traditional detection models are difficult to adapt to such complex and variable data characteristics, and are prone to problems such as outdated models and decreased detection accuracy. On the other hand, the acceleration of production rhythm requires the detection system to quickly respond to new situations caused by process adjustments, material changes and other changes, and to timely detect welding quality abnormalities and make appropriate treatment to ensure the continuity of production and the high-quality output of products.

[0004] Under such circumstances, there is an urgent need for a technical solution that can deeply mine welding data features, update in real time to adapt to welding dynamics, and efficiently and stably perform online quality detection. SUMMARY

[0005] To solve at least one of the above problems, the present application provides a welding quality monitoring method and device based on industrial internet to improve the accuracy and efficiency of welding quality detection and provide strong support for the optimization and quality control of the welding process.

[0006] To solve at least one of the above problems, the present application provides the following technical solution:

[0007] In a first aspect, the present application provides a welding quality monitoring method based on industrial internet, comprising:

[0008] obtaining welding data, the welding data including welding parameters and quality data; the welding parameters including welding electrical parameters and welding torch movement trajectory data; the quality data including visual images of the welding area; the visual images including molten pool visual images;

[0009] The multi-branch deep convolutional neural network is used to process the electrical parameters, visual images and movement trajectory data respectively to extract multi-modal features; the multi-modal features are fused through a cross-modal feature fusion mechanism, and the fused features are input into a recurrent neural network for time series modeling to obtain a welding quality detection model; the welding quality detection model is loaded to an edge node to monitor the welding quality;

[0010] The prediction result of the welding quality detection model is compared with the actually measured quality data, and whether the welding quality is abnormal is judged according to the comparison result; if yes, a preset rule is processed; the preset rule includes real-time compensation, process optimization and emergency intervention.

[0011] Further, the step of using a multi-branch deep convolutional neural network to process electrical parameters, visual images and movement trajectory data respectively to extract multi-modal features includes:

[0012] A one-dimensional deep convolutional neural network is used to extract features from electrical parameters; a deformable deep convolutional neural network is used to extract features from movement trajectory data; and a two-dimensional deep convolutional neural network is used to extract features from visual images.

[0013] Further, the one-dimensional deep convolutional neural network uses a layer-by-layer dimension reduction structure to extract features of different scales from movement trajectory data;

[0014] The deformable deep convolutional neural network uses deformable convolution kernels and hole convolution kernels in parallel to extract features from movement trajectory data;

[0015] The two-dimensional deep convolutional neural network uses different size convolution kernels to extract multi-scale features from visual images, and generates an attention weight map by analyzing the importance of each region of the image through a spatial attention mechanism.

[0016] Further, the method is realized by edge-cloud cooperation, and the implementation method is:

[0017] The multi-branch deep convolutional neural network is deployed on the edge; and the recurrent neural network is deployed on the cloud;

[0018] The edge is deployed in a lightweight manner: removing convolution kernels with weights lower than a set threshold in the deep convolutional neural network, and quantizing and compressing floating-point weights;

[0019] The cloud is processed in an accelerated manner: using a GPU cluster to simultaneously process feature sequences of multiple welding points, and caching the calculation results of the recurrent neural network for common working conditions for direct calling when repeated queries are made.

[0020] Further, the step of judging whether the welding quality is abnormal according to the comparison result comprises:

[0021] The comprehensive abnormality index is calculated by comprehensively considering the electric parameter deviation, the trajectory deviation and the visual defect probability, and the comprehensive abnormality index is divided into three abnormality levels according to the interval range of the comprehensive abnormality index; real-time compensation is performed for the first abnormality level, process optimization is performed for the second abnormality level, and emergency intervention is performed for the third abnormality level;

[0022] The real-time compensation comprises adjusting the parameters whose deviations exceed the threshold value through incremental PID control, and performing at the edge end;

[0023] The process optimization comprises matching similar cases from a historical library, generating multiple groups of candidate parameters based on a model of a simulation cluster, and selecting a scheme with the minimum deformation and an energy consumption increase less than a preset value for execution;

[0024] The emergency intervention comprises emergency shutdown and issuing an alarm.

[0025] Further, the welding quality monitoring method based on the industrial internet further comprises:

[0026] The bottom-layer recurrent neural network of the welding quality detection model uses a QRNN unit to process high-frequency data, and the high-layer recurrent neural network uses a GRU unit to integrate low-frequency data;

[0027] The welding quality detection model adds a thermodynamic constraint term in a loss function.

[0028] Further, the welding quality monitoring method based on the industrial internet further comprises online incremental learning; the online incremental learning comprises:

[0029] When new welding data is generated, the system automatically inputs the new welding data into the welding quality detection model for incremental training, and continuously optimizes the model parameters by learning the features and rules in the new data;

[0030] The online incremental learning further comprises setting a trigger condition for the online incremental learning; the trigger condition comprises that the confidence of new data is lower than a threshold value, process parameters are changed, and periodic triggering;

[0031] Important parameter updates are consolidated and limited by elastic weights to protect old knowledge, and a dynamic replay buffer is set to retain key samples of old data.

[0032] In a second aspect, the application provides a welding quality monitoring device based on the industrial internet, comprising:

[0033] The data acquisition module is configured to acquire welding data, wherein the welding data comprises welding parameters and quality data; the welding parameters comprise electrical parameters of welding and movement trajectory data of a welding torch; and the quality data comprises visual images of a welding area; and the visual images comprise visual images of a molten pool.

[0034] The model generation module is configured to use a multi-branch deep convolutional neural network to process the electrical parameters, the visual images and the movement trajectory data respectively, to extract multi-modal features; to fuse the multi-modal features through a cross-modal feature fusion mechanism, and to input the fused features into a recurrent neural network for time series modeling, to obtain a welding quality detection model; and to load the welding quality detection model to an edge node to monitor welding quality.

[0035] The abnormality processing module is configured to compare a prediction result of the welding quality detection model with actually measured quality data, and to determine whether welding quality is abnormal according to a comparison result; and if so, to perform processing according to a preset rule; and the preset rule comprises real-time compensation, process optimization and emergency intervention.

[0036] In a third aspect, the present application provides an electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the welding quality monitoring method based on industrial internet when executing the program.

[0037] In a fourth aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the steps of the welding quality monitoring method based on industrial internet.

[0038] In a fifth aspect, the present application provides a computer program product comprising computer programs / instructions, wherein the computer programs / instructions are executable by a processor to implement the steps of the welding quality monitoring method based on industrial internet.

[0039] According to the above technical solution, the present application provides a welding quality monitoring method and device based on industrial internet, which forms a hybrid model by combining a deep convolutional neural network (DCNN) and a recurrent neural network (RNN), fully fuses local features and time series features of welding data, provides high-value features by the former and realizes time series reasoning by the latter, and together completes a complete link from raw data to quality evaluation, accurately learns a nonlinear relationship between complex welding parameters and quality, improves the accuracy of online detection of welding quality, provides strong support for optimization of a welding process and welding quality control, fills the gap in the field of efficient, intelligent and self-adaptive online detection technology of welding quality, and provides strong technical support for modern industrial welding production. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0041] Figure 1 One of the flowcharts of the welding quality monitoring method based on industrial internet in the embodiments of the present application;

[0042] Figure 2 The second flowchart of the welding quality monitoring method based on industrial internet in the embodiments of the present application;

[0043] Figure 3 The structural diagram of the welding quality monitoring device based on industrial internet in the embodiments of the present application;

[0044] Figure 4 The structural diagram of the electronic device in the embodiments of the present application.

[0045] Reference signs:

[0046] Electronic device 9600, central processor 9100, memory 9140, communication module 9110, input unit 9120, audio processor 9130, display 9160, power supply 9170, buffer memory 9141, application / function storage unit 9142, data storage unit 9143, driver program storage unit 9144, antenna 9111, speaker 9131, microphone 9132. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0048] The acquisition, storage, use, processing and the like of data in the technical solutions of the present application all conform to the relevant provisions of national laws and regulations.

[0049] In view of the problems in the prior art, the present application provides a welding quality monitoring method and device based on an industrial internet, which combines a deep convolutional neural network (DCNN) and a recurrent neural network (RNN) to form a hybrid model, fully fuses local features and time sequence features of welding data, the former provides high-value features, and the latter realizes time sequence reasoning, together completes a complete link from original data to quality evaluation, accurately learns a nonlinear relationship between complex welding parameters and quality, improves the accuracy of online detection of welding quality, provides strong support for optimization of a welding process and welding quality control, fills a gap in the field of efficient, intelligent and adaptive online detection technology of welding quality, and provides strong technical support for modern industrial welding production.

[0050] In order to improve the accuracy and efficiency of welding quality detection and provide strong support for optimization of a welding process and quality control, an embodiment of a welding quality monitoring method based on an industrial internet is provided, as shown in Figure 1 and Figure 2 The welding quality monitoring method based on the industrial internet specifically includes the following contents:

[0051] Step S101: acquiring welding data, the welding data including welding parameters and quality data; the welding parameters including electric parameters of welding and movement trajectory data of a welding torch; the quality data including a visual image of a welding area; and the visual image including a visual image of a molten pool.

[0052] In this embodiment, the electric parameters include current and voltage data in a welding process, which can be collected by using a high-precision current sensor (such as a Hall sensor) and a voltage sensor or a HIOKI PW3390 sensor, and the sampling frequency is usually set to be higher than 10 kHz to capture high-frequency changes. The visual image can be shot by using an industrial camera (such as a CCD or CMOS camera, a FLIR Blackfly S) equipped with optical filtering technology, a light source and a light guide assembly, and the edge of the molten pool is extracted through image processing technology such as threshold segmentation and edge detection, and then the width of the molten pool is calculated, and the depth of the molten pool is calculated by using stereo vision or laser scanning technology combined with image processing algorithms. For example, 3D information of the surface of the molten pool is acquired by using a laser scanner, and then depth calculation is performed. Of course, the visual image can also include data such as weld width and thickness. The movement trajectory of the welding equipment can be recorded by installing a high-precision inertial measurement unit (such as an IMU and an Xsens MTi-670 inertial navigation module) at the end of the welding torch.

[0053] After data collection, preprocessing is performed, including: noise reduction processing of collected data at the edge computing node, for example, using wavelet transform or Kalman filter to remove noise. Normalize the current and voltage data to make the value range between 0 and 1. Grayscale, edge detection (such as Canny algorithm) and feature extraction (such as SIFT or ORB algorithm) are performed on visual images. Smooth the movement trajectory data, for example, using moving average filtering. Time alignment is performed on different modal data to ensure consistency in the time dimension.

[0054] Through the collection and preprocessing of multi-modal data, key features in the welding process can be comprehensively captured, providing a high-quality data basis for subsequent feature extraction and model training.

[0055] Step S102: Use a multi-branch deep convolutional neural network to process the electrical parameters, visual images and movement trajectory data respectively to extract multi-modal features; fuse the multi-modal features through a cross-modal feature fusion mechanism, input the fused features into a recurrent neural network for time series modeling, and obtain a welding quality detection model; load the welding quality detection model into the edge node to monitor the welding quality.

[0056] In this embodiment, a hybrid model combining deep convolutional neural network (DCNN) and recurrent neural network (RNN) is used. DCNN is used to extract local features in welding data, such as key features in current and voltage waveforms; RNN is responsible for processing time series data and capturing trends in welding process over time. A large amount of welding data under different conditions is used to train the model, so that the model can learn the nonlinear mapping relationship between complex welding parameters and quality. Through the nonlinear mapping relationship, the model can realize welding quality prediction, welding parameter optimization, real-time monitoring and feedback control, process adaptability adjustment, fault diagnosis and analysis, etc.

[0057] Optionally, in this embodiment, a one-dimensional deep convolutional neural network is used to extract features from electrical parameters; a deformable deep convolutional neural network is used to extract features from movement trajectory data; a two-dimensional deep convolutional neural network is used to extract features from visual images.

[0058] Specifically, the one-dimensional deep convolutional neural network adopts a structure of layer-by-layer dimension reduction to extract features of different scales from the mobile trajectory data, so as to help the network to gradually extract multi-scale features in the signal from macro to micro. An exemplary feature extraction process is as follows: the welding current and voltage signals are collected, 10,000 samples per second, and 2 seconds of data window (20000 data points) is processed each time. The first layer of convolution: a convolution kernel with 50 units is used, the width is 50 data points (covering a period of 5 ms), and the step is 10 data points (i.e. sliding every 0.1 ms). This layer reduces the original 20000 points to 2000 points, mainly extracting the low-frequency trend features of the current waveform (such as the slow heat accumulation effect). The second layer of convolution: a convolution kernel with a width of 20 data points is used to capture medium-time scale fluctuations (such as current fluctuations caused by several times per second of wire feed speed changes). The third layer of convolution: a 5-point narrow kernel is used to detect high-frequency transient anomalies (such as microsecond-level arc discontinuity phenomena). After global maximum pooling and fully connected layers, a 32-dimensional feature vector is generated, containing key information such as the overall energy of the waveform and the main fluctuation frequency.

[0059] In the deformable deep convolutional neural network, deformable convolution kernels and dilated convolution kernels are used in parallel to extract features from the mobile trajectory data; deformable convolution dynamically adjusts the sampling position of the convolution kernel through a learnable offset, which is particularly suitable for geometric feature extraction of non-linear trajectories such as triangles / circles; dilated convolution (dilation=2) expands the receptive field by introducing gaps in the convolution kernel, which can capture more distant feature information and correlate long-distance trajectory features, which is very useful for detecting larger defects that may occur during the welding process. An exemplary feature extraction process is as follows: record the movement coordinates (X / Y / Z) of the welding torch in three-dimensional space, 1000 samples per second, and process 2000 trajectory points in a 2-second window. A small convolutional network is used to predict the position offset of each sampling point in real time, so that the convolution kernel can dynamically adapt to the geometric mutations of the trajectory (such as the corners of a zigzag trajectory). And adjust the sampling position of the convolution kernel according to the predicted offset, for example, increase the sampling point spacing in the sharp turn area to cover a wider range. In the second layer of convolution, the dilation rate is set to 2 (i.e. the convolution kernel elements are spaced 2 points apart), so that the receptive field of a single convolution kernel is expanded to 3 times the original, effectively capturing long-distance trajectory correlation (such as the overall curvature feature of a circular trajectory). Finally, a 16-dimensional feature vector is generated, containing kinematic parameters such as trajectory smoothness, maximum curvature, and angular velocity change rate.

[0060] The two-dimensional deep convolutional neural network uses different size convolution kernels to extract multi-scale features from visual images, and analyzes the importance of each region of the image through a spatial attention mechanism to generate an attention weight map, focusing on sensitive areas such as the edge of the molten pool and pores. For example, 3x3 and 5x5 ordinary convolution kernels can be used to extract features in parallel, covering micro defects to macro morphology. An exemplary feature extraction process is as follows: a 128x128 pixel region centered on the welding torch position is intercepted, 200 frames of images per second, each frame containing RGB three-channel information. The first layer of 3x3 convolution kernel captures the local subtle changes in the welding current and voltage waveform, suitable for detecting small defects (such as spatter, pores); the second layer of 5x5 convolution kernel captures the overall morphological features of the molten pool (such as the ratio of the molten width to the molten depth), extracts more extensive features, and identifies abnormal patterns in a larger range. After global average pooling, a 24-dimensional feature vector is generated, quantifying visual indicators such as molten pool width, tail depression depth, and spatter particle number.

[0061] Then, the three features are time-synchronized and aligned, and the 32-dimensional electrical parameters, 16-dimensional trajectories, and 24-dimensional visual features are directly concatenated into a 72-dimensional vector. Dynamic weight allocation can also be performed: when the current stability decreases, the electrical parameter feature weight automatically increases (e.g., from 1.0 to 1.3); when the trajectory deviation exceeds the limit, the trajectory feature weight increases and the visual feature weight decreases; when the molten pool anomaly is significant, the visual feature weight has the highest priority. Through a two-layer fully connected network (36 nodes → 72 nodes), the importance of each dimension is learned, and the Softmax function is normalized to a weight value between 0 and 1.

[0062] Time series modeling and decision process: 1. Input sequence construction: take the last 5 seconds of 10 groups of fusion features (72 dimensions x 10 time steps) to form a time series matrix (10x72). 2. Bidirectional LSTM modeling: forward propagation: analyze the feature evolution trend from left to right (e.g., gradually increasing current indicates heat accumulation). Backward propagation: detect the initial point of the anomaly from right to left (e.g., the initial time of trajectory deviation). State preservation: each LSTM unit preserves a 32-dimensional hidden state to capture the dependence across time steps. 3. Output and decision: quality score: output a quality score between 0 and 1 through the Sigmoid function (e.g., >0.8 is qualified). Multi-level response: for example, 0.7-0.8: issue a warning and record the working conditions, do not adjust the parameters temporarily; 0.5-0.7: start the digital twin simulation to generate a parameter optimization scheme; <0.5: immediately stop and mark the defect position.

[0063] As a preferred solution, the embodiment is implemented in cooperation with edge cloud, the multi-branch deep convolutional neural network is deployed on the edge side, and the recurrent neural network is deployed on the cloud side. The edge node performs real-time preprocessing and feature extraction on the collected multi-modal data, reduces the data transmission amount, and the edge node runs a lightweight model for preliminary detection. The cloud receives the key feature data uploaded from the edge node, runs a global optimization model (such as a deep learning model or an ensemble learning model), performs more complex analysis, and the cloud regularly issues the optimized model parameters to the edge node to update the edge model. Through the edge-cloud cooperative architecture, the resources can be reasonably allocated, and the real-time performance and global optimization capability of the system can be improved.

[0064] Optionally, in the embodiment, the edge side can be deployed in a lightweight manner: removing the convolution kernels in the deep convolutional neural network whose weights are lower than a set threshold (such as the kernels whose absolute values of weights are less than 0.01), and quantizing and compressing the floating-point weights, such as converting 32-bit floating-point weights to 8-bit integers, reducing the model volume by 75%; and the cloud side is processed in an accelerated manner: using a GPU cluster to simultaneously process the feature sequences of hundreds of welding points, and caching the RNN calculation results of common working conditions, and directly calling when repeated queries.

[0065] Optionally, in the embodiment, the bottom recurrent neural network of the welding quality detection model uses QRNN units to process high-frequency data such as 1000Hz current data, current / voltage waveforms, and welding torch acceleration to extract microsecond-level arc fluctuation features; the high-level recurrent neural network uses GRU units to integrate low-frequency data such as wire feed speed and gas flow. The parallel processing of QRNN meets the real-time requirement, and GRU captures long-period process trends (such as heat accumulation effect).

[0066] At the same time, a thermodynamic constraint term (such as the thermodynamic characteristics of the molten pool need to imply the energy conservation relationship) can be added to the loss function of the welding quality detection model, that is, according to the physical principle in the welding process, the thermodynamic equation is defined as a constraint condition, and the law of conservation of energy is expressed in the form of mathematical formula, to ensure that the prediction result of the model conforms to the actual physical law. By adding the thermodynamic constraint term to the loss function, the model not only considers the fitting error of the data during the training process, but also meets the restriction of the thermodynamic equation. For example, the thermodynamic constraint term can be added as a penalty term to the loss function, and when the prediction result of the model violates the thermodynamic equation, it will be punished accordingly, thereby guiding the model to learn features that conform to the physical law. In addition, kinematic constraints can also be added to make the prediction more consistent with the actual process restrictions.

[0067] The embodiment also includes online incremental learning: when new welding data is generated, the system automatically inputs it into the model for incremental training. The model quickly learns the characteristics and rules in the new data on the basis of not affecting the original knowledge, continuously optimizes the model parameters, adapts to production changes such as welding process adjustment and material change, and always maintains good detection performance.

[0068] Specifically, the online incremental learning includes:

[0069] The trigger conditions of the online incremental learning are set, such as new data confidence lower than a threshold (such as prediction confidence < 85%), process parameter change (such as replacing welding wire material, adjusting welding speed), periodic triggering (such as automatically triggering every 100 meters of welding seam completed), and data collection: caching the welding data (current, voltage, trajectory, molten pool image) in the last 30 minutes, and labeling the new data label (automatic labeling + manual inspection review). Then, the real-time data is preprocessed in the form of denoising, standardization, feature alignment, etc.

[0070] The online incremental learning consolidates and limits important parameter updates through elastic weights to protect old knowledge. Limiting important parameter updates is to quantify the importance of parameters, and for those parameters that are critical to old tasks, constraints are imposed on them to make them change little when updated, thereby protecting old knowledge. For example, in the welding quality detection model, the key neuron connection weight used to identify the features of stable current and voltage belongs to important parameters and needs to be protected.

[0071] The online incremental learning retains key samples of old data (such as data near the decision boundary) by setting a dynamic replay buffer, and the training strategy is as follows: each batch of training data = 70% new data + 30% replayed old data; replayed data selection: based on entropy sampling (high-uncertainty samples first).

[0072] The online incremental learning of the edge lightweight model includes: parameter freezing: fixing the base feature extraction layer (such as the first 3 layers of DCNN); local fine-tuning: only updating the last 2 layers of fully connected layer; learning rate adjustment: cosine annealing scheduling (such as initial lr = 0.001, period = 50 batches).

[0073] For online incremental learning, the edge-cloud collaborative implementation method includes: 1. Edge-side rapid fine-tuning: new data preprocessing, feature extraction, etc. are sequentially performed at the edge side, after triggering incremental learning, 10 iterations of fine-tuning are performed (such as batch size = 16) using EWC constraints, and then the model parameters are synchronized to the cloud. 2. Cloud model integration: after receiving the incremental models uploaded by multiple edge nodes, when updating the global model using federated averaging (FedAvg), the weights of the models uploaded by different edge nodes are allocated according to the proportion of their data volume, which can integrate the new data features of each edge node and obtain a more comprehensive global model that adapts to the overall production. A new model version is generated every month and the last 3 versions are retained, which facilitates rolling back to the previous stable version in case of new version problems, ensuring the continuity and stability of production.

[0074] The online incremental learning also includes a degradation prevention mechanism: 5% of historical data are randomly selected as a validation set, and an alarm is triggered when the accuracy of the old task decreases by >3%, and the old data are retrained from the replay buffer; when gradient explosion or disappearance is detected, adaptive gradient clipping (AGC) is triggered, and parameter changes are constrained to limit the amplitude of a single update.

[0075] Step S103: comparing the prediction result of the welding quality detection model with the actually measured quality data, and determining whether the welding quality is abnormal according to the comparison result; if so, processing according to a preset rule; the preset rule includes real-time compensation, process optimization and emergency intervention.

[0076] In this embodiment, the step of determining whether the welding quality is abnormal according to the comparison result includes:

[0077] Comprehensive electrical parameter deviation, trajectory deviation and visual defect probability, calculate the comprehensive abnormal index, such as S = 0.5 x electrical parameter score + 0.3 x trajectory score + 0.2 x visual score, and divide it into three abnormal levels according to the interval range of the comprehensive abnormal index, such as: Level 1: 0.3≤S<0.5; Level 2: 0.5≤S<0.7; Level 3: S≥0.7; real-time compensation is performed for the first abnormal level, process optimization is performed for the second abnormal level, and emergency intervention is performed for the third abnormal level;

[0078] Among them, real-time compensation includes: adjusting the parameters whose deviation exceeds the threshold through incremental PID control to avoid excessive interference with stable working conditions and execute at the edge; Process optimization includes: matching similar cases from the historical database, and generating 3 groups of candidate parameters (current ± 5%, welding speed ± 3 mm / s, angle ± 2°) based on the simplified thermal-mechanical coupling model of the ANSYS simulation cluster deployed in the cloud, and selecting the scheme with the smallest deformation and energy consumption increase less than the preset value for execution; Emergency intervention includes directly cutting off the welding power supply through the E-Stop signal of the PLC to stop emergency and send an alarm.

[0079] Exemplarily, the embodiment also provides a specific implementation process of the hierarchical correction strategy:

[0080] 1. Level 1: Real-time dynamic compensation

[0081] The parameters can be fine-tuned through incremental PID control to avoid excessive interference with stable working conditions and execute at the edge. The specific operation can include proportional term fine-tuning, integral term fine-tuning and differential term fine-tuning, wherein: the proportional term fine-tuning is to linearly adjust the welding current according to the current deviation amplitude (such as increasing the current by 1.5% for every 0.1 increase in deviation); the integral term fine-tuning is to gradually correct the wire feeding speed by accumulating the continuous deviation in the past 10 seconds; and the differential term fine-tuning is to adjust the welding gun movement acceleration in advance by monitoring the deviation change rate.

[0082] Example of actuator operation: send angle correction instructions to the servo motor through the CAN bus, with a step accuracy of 0.01°; adjust the wire feeder speed using the Modbus TCP protocol, with a control accuracy of ±0.5 mm / s.

[0083] 2. Level 2: Digital twin optimization

[0084] Similar case matching: use a feature matching algorithm (such as K-NN) to find similar cases in the historical database under the current welding conditions.

[0085] Simplified thermal-mechanical coupling model: heat conduction calculation: solve the transient temperature field based on the actual material properties (such as the thermal conductivity of 304L stainless steel 16.2 W / m·K); stress analysis: predict the residual stress distribution in the weld area by combining the thermal expansion coefficient and the constraint condition; deformation evaluation: calculate the maximum deformation under each parameter combination and select the candidate scheme with deformation <0.15 mm.

[0086] Candidate parameter generation: generate 3 different parameter combinations based on the current welding parameters, each parameter combination including changes in current, welding speed and angle.

[0087] Optimal parameter decision: Perform penetration prediction (e.g., require ≥ 2 mm) and heat-affected zone width evaluation (e.g., require ≤ 1.2 mm) on candidate solutions; score based on deformation, energy consumption increase, penetration, and heat-affected zone width, and select the parameter set with the highest score.

[0088] Parameter delivery: Send the selected optimal parameters to the controller of the welding equipment through OPC UA protocol to achieve automatic parameter update.

[0089] 3. Level 3: Emergency intervention

[0090] Safety protection sequence: Emergency stop trigger: send an emergency stop instruction to the PLC to cut off the welding power supply and terminate the arc within 50 ms; gas protection maintenance: open the standby gas valve to continuously supply argon for 15 seconds to prevent the molten pool from oxidizing; mechanical arm retreat: execute the preset safety path planning to retreat to the Home position at a speed of 0.5 m / s; sound and light alarm: start double-frequency flashing light and pulsed buzzing.

[0091] Optionally, the effective correction parameter combination is stored in the process database, and the applicable material (such as carbon steel / aluminum alloy) and thickness range (such as 3-12 mm) are labeled. And the model incremental learning can be started when the set conditions are met, such as when the same type of deviation is successfully corrected for 3 times in a row, 500 groups of feature data are automatically extracted to fine-tune the detection model for 30 minutes.

[0092] The hierarchical correction strategy ensures that the welding process can be effectively controlled and optimized under different deviation levels through three levels of correction measures. Level 1 quickly responds to small deviations through real-time dynamic compensation, Level 2 optimizes the selection of optimal parameter combinations through digital twinning, and Level 3 ensures the safety of the welding process through emergency intervention. This hierarchical strategy can significantly improve the stability of welding quality and production efficiency.

[0093] As can be seen from the above description, the welding quality monitoring method based on industrial internet provided by the embodiments of the present application forms a hybrid model by combining deep convolutional neural network (DCNN) and recurrent neural network (RNN), fully fuses the local features and time sequence features of the welding data, the former provides high-value features, and the latter realizes time sequence reasoning, together completing the complete link from raw data to quality evaluation, accurately learning the nonlinear relationship between complex welding parameters and quality, to improve the accuracy of online detection of welding quality, providing strong support for optimization and welding quality control of the welding process, filling the gap in this field for efficient, intelligent, and self-adaptive online welding quality detection technology, and providing strong technical support for modern industrial welding production.

[0094] In order to improve the accuracy and efficiency of the welding quality detection, and provide strong support for the optimization and quality control of the welding process, the application provides an embodiment of an industrial internet-based welding quality monitoring device for implementing all or part of the contents of the industrial internet-based welding quality monitoring method, see Figure 3 , which specifically includes the following contents:

[0095] A data acquisition module 10 is configured to acquire welding data, wherein the welding data includes welding parameters and quality data; the welding parameters include welding electrical parameters and welding torch movement trajectory data; and the quality data includes visual images of a welding area; and the visual images include molten pool visual images.

[0096] A model generation module 20 is configured to use a multi-branch deep convolutional neural network to process electrical parameters, visual images and movement trajectory data respectively, extract multi-modal features, fuse the multi-modal features through a cross-modal feature fusion mechanism, input the fused features into a recurrent neural network for time series modeling, and obtain a welding quality detection model; and the welding quality detection model is loaded into an edge node to monitor welding quality.

[0097] An abnormality processing module 30 is configured to compare a prediction result of the welding quality detection model with actually measured quality data, and determine whether welding quality is abnormal according to a comparison result; if so, processing is performed according to a preset rule; and the preset rule includes real-time compensation, process optimization and emergency intervention.

[0098] As can be seen from the above description, the industrial internet-based welding quality monitoring device provided by the embodiments of the application can form a hybrid model by combining a deep convolutional neural network (DCNN) and a recurrent neural network (RNN), fully fuse local features and time series features of welding data, the former provides high-value features, and the latter realizes time series reasoning, together complete the complete link from raw data to quality evaluation, accurately learn the nonlinear relationship between complex welding parameters and quality, improve the accuracy of online welding quality detection, provide strong support for the optimization and quality control of the welding process, fill the gap in the field of efficient, intelligent and adaptive welding quality online detection technology, and provide strong technical support for modern industrial welding production.

[0099] From the hardware level, in order to improve the accuracy and efficiency of the welding quality detection, and provide strong support for the optimization and quality control of the welding process, the application provides an embodiment of an electronic device for implementing all or part of the contents of the industrial internet-based welding quality monitoring method, which specifically includes the following contents:

[0100] The processor, the memory, the communications interface and the bus; wherein the processor, the memory, the communications interface complete the communication among each other through the bus; the communications interface is used for realizing the information transmission between the welding quality monitoring device based on the industrial internet and the core business system, the user terminal and the related database and other related equipment; the logic controller can be a desktop computer, a tablet computer and a mobile terminal and the like, and the embodiment is not limited thereto. In the embodiment, the logic controller can be implemented with reference to the embodiment of the welding quality monitoring method based on the industrial internet in the embodiment and the embodiment of the welding quality monitoring device based on the industrial internet, the contents of which are incorporated herein, and the repeated parts will not be described herein.

[0101] It can be understood that the user terminal can include a smart phone, a tablet electronic device, a network set-top box, a portable computer, a desktop computer, a personal digital assistant (PDA), a vehicle-mounted device, a smart wearable device and the like. The smart wearable device can include smart glasses, a smart watch, a smart bracelet and the like.

[0102] In actual application, part of the welding quality monitoring method based on the industrial internet can be executed on the electronic device as described above, or all operations can be completed in the client device. Specifically, the selection can be made according to the processing capacity of the client device and the limitation of the user's use scene, and the like. The present application is not limited thereto. If all operations are completed in the client device, the client device can further include a processor.

[0103] The client device described above can have a communication module (i.e. a communication unit) and can be communicatively connected with a remote server to realize the data transmission with the server. The server can include a server of the task scheduling center side, and the server of the intermediate platform can also be included in other implementation scenarios, such as the server of the third-party server platform communicatively connected with the server of the task scheduling center. The server can include a single computer device, a server cluster composed of multiple servers, or a distributed server structure.

[0104] Figure 4 A schematic block diagram of the system configuration of the electronic device 9600 of the embodiment of the present application is shown in FIG. 9. As shown in FIG. 9, the electronic device 9600 can include a central processor 9100 and a memory 9140; the memory 9140 is coupled to the central processor 9100. It is worth noting that the structure shown in FIG. 9 is exemplary; other types of structures can also be used to supplement or replace the structure to realize the telecommunication function or other functions. Figure 4 Figure 4 The structure shown in FIG. 9 is exemplary; other types of structures can also be used to supplement or replace the structure to realize the telecommunication function or other functions. ​

[0105] In one embodiment, the welding quality monitoring method based on the Industrial Internet can be integrated into the central processing unit 9100. The central processing unit 9100 can be configured to perform the following control:

[0106] Step S101: Acquire welding data, wherein the welding data includes welding parameters and quality data; the welding parameters include welding electrical parameters and welding gun movement trajectory data; the quality data includes a visual image of the welding area; and the visual image includes a molten pool visual image;

[0107] Step S102: Using a multi-branch deep convolutional neural network to process electrical parameters, visual images, and motion trajectory data, respectively, to extract multimodal features; fusing the multimodal features through a cross-modal feature fusion mechanism, inputting the fused features into a recurrent neural network for time series modeling, and obtaining a welding quality detection model; loading the welding quality detection model into an edge node to monitor welding quality;

[0108] Step S103: Compare the prediction results of the welding quality detection model with the quality data actually measured, and determine whether the welding quality is abnormal based on the comparison result; if so, process it according to preset rules; the preset rules include real-time compensation, process optimization and emergency intervention.

[0109] From the above description, it can be seen that the electronic device provided in the embodiment of the present application forms a hybrid model by combining a deep convolutional neural network (DCNN) and a recurrent neural network (RNN), fully integrating the local features and time series features of the welding data. The former provides high-value features, and the latter realizes temporal reasoning, and together completes the complete link from raw data to quality assessment, accurately learns the nonlinear relationship between complex welding parameters and quality, so as to improve the accuracy of online detection of welding quality, provide strong support for the optimization of the welding process and welding quality control, fill the gap in this field for efficient, intelligent, and adaptive online detection technology of welding quality, and provide strong technical support for modern industrial welding production.

[0110] In another embodiment, the welding quality monitoring device based on the industrial Internet can be configured separately from the central processing unit 9100. For example, the welding quality monitoring device based on the industrial Internet can be configured as a chip connected to the central processing unit 9100, and the function of the welding quality monitoring method based on the industrial Internet can be realized through the control of the central processing unit.

[0111] like Figure 4 As shown, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily have to includeFigure 4 All the components shown in FIG. 9A; in addition, the electronic device 9600 can further include Figure 4 components not shown in FIG. 9A, reference can be made to the prior art.

[0112] As shown in FIG. 9B, the central processing unit 9100, which is also sometimes referred to as a controller or operation control, can include a microprocessor or other processor device and / or logic device, which receives input and controls the operation of the various components of the electronic device 9600. Figure 4 The memory 9140, for example, can be one or more of a buffer, a flash memory, a hard drive, a removable media, a volatile memory, a non-volatile memory, or other suitable device. Information relating to failures can be stored, and in addition, programs for executing the information can be stored. The central processing unit 9100 can execute the programs stored in the memory 9140 to achieve information storage or processing, etc.

[0113] The input unit 9120 provides input to the central processing unit 9100. The input unit 9120 is, for example, a key or a touch input device. The power supply 9170 is used to supply power to the electronic device 9600. The display 9160 is used to display display objects such as images and text. The display can be, for example, an LCD display, but is not limited thereto.

[0114] The memory 9140 can be a solid state memory, such as a read only memory (ROM), a random access memory (RAM), a SIM card, etc. It can also be a memory that retains information even when power is off, can be selectively erased, and is provided with more data, an example of which is sometimes referred to as an EPROM, etc. The memory 9140 can also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 can include an application / function storage section 9142 for storing application programs and function programs or for storing a flow for executing the operation of the electronic device 9600 by the central processing unit 9100.

[0115] The memory 9140 can further include a data storage section 9143 for storing data such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. A driver program storage section 9144 of the memory 9140 can include various driver programs of the electronic device for communication functions and / or for executing other functions of the electronic device (such as a messaging application, an address book application, etc.).

[0116]

[0117] ​The communication module 9110 is a transmitter / receiver that transmits and receives signals via the antenna 9111. The communication module 9110 (transmitter / receiver) is coupled to the central processor 9100 to provide input signals and receive output signals, as in the case of a conventional mobile communication terminal.

[0118] Based on different communication technologies, a plurality of communication modules 9110, such as a cellular network module, a Bluetooth module, and / or a wireless local area network module, etc., can be provided in the same electronic device. The communication module 9110 (transmitter / receiver) is also coupled to the speaker 9131 and the microphone 9132 via the audio processor 9130 to provide audio output via the speaker 9131 and receive audio input from the microphone 9132, thereby implementing the usual telecommunication functions. The audio processor 9130 can include any suitable buffers, decoders, amplifiers, etc. In addition, the audio processor 9130 is also coupled to the central processor 9100, so that it is possible to record on the local machine via the microphone 9132 and play the sound stored on the local machine via the speaker 9131.

[0119] The embodiments of the present application also provide a computer readable storage medium capable of implementing all steps of the above-mentioned welding quality monitoring method based on industrial internet, wherein the execution subject is a server or a client. The computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement all steps of the welding quality monitoring method based on industrial internet, wherein the execution subject is a server or a client. For example, the processor executes the computer program to implement the following steps:

[0120] Step S101: acquiring welding data, wherein the welding data includes welding parameters and quality data; the welding parameters include electrical parameters of welding and movement trajectory data of a welding torch; the quality data includes visual images of a welding area; and the visual images include visual images of a molten pool;

[0121] Step S102: using a multi-branch deep convolutional neural network to process the electrical parameters, the visual images, and the movement trajectory data respectively to extract multi-modal features; fusing the multi-modal features through a cross-modal feature fusion mechanism, inputting the fused features into a recurrent neural network for time series modeling to obtain a welding quality detection model; and loading the welding quality detection model to an edge node to monitor welding quality;

[0122] Step S103: comparing a prediction result of the welding quality detection model with actually measured quality data, and judging whether welding quality is abnormal according to a comparison result; if yes, processing according to a preset rule; and the preset rule includes real-time compensation, process optimization, and emergency intervention.

[0123] From the above description, the computer readable storage medium provided by the embodiments of the application can form a hybrid model by combining a deep convolutional neural network (DCNN) and a recurrent neural network (RNN), fully fuse the local features and time sequence features of the welding data, the former provides high-value features, and the latter realizes time sequence reasoning, together complete the complete link from the original data to the quality evaluation, accurately learn the nonlinear relationship between the complex welding parameters and the quality, and improve the accuracy of the online detection of the welding quality, provide strong support for the optimization of the welding process and the welding quality control, fill the gap in the field of efficient, intelligent and adaptive welding quality online detection technology, and provide strong technical support for modern industrial welding production.

[0124] The embodiments of the application also provide a computer program product capable of realizing all steps in the welding quality monitoring method based on industrial internet in which the execution subject in the above embodiments is a server or a client. The computer program / instruction is executed by a processor to realize the steps of the welding quality monitoring method based on industrial internet. For example, the computer program / instruction realizes the following steps:

[0125] Step S101: acquiring welding data, the welding data including welding parameters and quality data; the welding parameters including welding electrical parameters and welding torch movement trajectory data; the quality data including visual images of a welding area; the visual images including molten pool visual images;

[0126] Step S102: using a multi-branch deep convolutional neural network to process the electrical parameters, the visual images and the movement trajectory data respectively to extract multi-modal features; fusing the multi-modal features through a cross-modal feature fusion mechanism, inputting the fused features into a recurrent neural network for time sequence modeling to obtain a welding quality detection model; and loading the welding quality detection model to an edge node to monitor the welding quality;

[0127] Step S103: comparing the prediction result of the welding quality detection model with the actually measured quality data, and judging whether the welding quality is abnormal according to the comparison result; if yes, processing according to a preset rule; the preset rule including real-time compensation, process optimization and emergency intervention.

[0128] From the above description, the computer program product provided by the embodiment of the present application forms a hybrid model by combining a deep convolutional neural network (DCNN) and a recurrent neural network (RNN), fully fuses the local features and time sequence features of the welding data, the former provides high-value features, and the latter realizes time sequence reasoning, and together completes the complete link from the original data to the quality evaluation, accurately learns the nonlinear relationship between the complex welding parameters and the quality, improves the accuracy of the welding quality online detection, provides strong support for the optimization of the welding process and the welding quality control, fills the gap of efficient, intelligent and self-adaptive welding quality online detection technology in the field, and provides strong technical support for modern industrial welding production.

[0129] Those skilled in the art will understand that the embodiments of the present application can be provided as methods, apparatuses, or computer program products. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0130] The present application is described with reference to flowcharts and / or block diagrams of the methods, devices (apparatuses), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus generate a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The apparatus that performs the functions specified in one or more flows and / or blocks.

[0131] These computer program instructions can also be stored in a computer-readable memory that can guide the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The apparatus that performs the functions specified in one or more flows and / or blocks.

[0132] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide a means for implementing the functions specified in the flowcharts and / or block diagrams.Figure 1 one or more processes and / or functions specified in one or more blocks Figure 1 one or more blocks or steps of the functions specified in the one or more blocks.

[0133] The principles and implementation manners of the present application are described in the specific embodiments. The above embodiment descriptions are only used to help understand the method and core idea of the present application. Meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manners and application ranges will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A welding quality monitoring method based on industrial Internet, characterized in that: The method comprises: Acquire welding data, the welding data including welding parameters and quality data; the welding parameters including welding electrical parameters and welding gun movement trajectory data; the quality data including a visual image of the welding area; the visual image including a molten pool visual image; A multi-branch deep convolutional neural network is used to process electrical parameters, visual images, and motion trajectory data respectively to extract multimodal features. The multimodal features are fused through a cross-modal feature fusion mechanism, and the fused features are input into a recurrent neural network for time series modeling to obtain a welding quality detection model. The welding quality detection model is loaded into an edge node to monitor welding quality. The prediction results of the welding quality detection model are compared with the actual measured quality data, and whether the welding quality is abnormal is determined based on the comparison result; if so, it is processed according to preset rules; the preset rules include real-time compensation, process optimization and emergency intervention.

2. The welding quality monitoring method based on the Industrial Internet according to claim 1 is characterized in that: The step of using a multi-branch deep convolutional neural network to process electrical parameters, visual images, and movement trajectory data respectively to extract multimodal features includes: A one-dimensional deep convolutional neural network is used to extract features from electrical parameters; a deformable deep convolutional neural network is used to extract features from movement trajectory data; and a two-dimensional deep convolutional neural network is used to extract features from visual images.

3. The welding quality monitoring method based on the Industrial Internet according to claim 2 is characterized in that: The one-dimensional deep convolutional neural network uses a layer-by-layer dimension reduction structure to extract features of different scales from the movement trajectory data; The deformable deep convolutional neural network uses a deformable convolution kernel and a dilated convolution kernel in parallel to extract features from the movement trajectory data; The two-dimensional deep convolutional neural network uses convolution kernels of different sizes to extract multi-scale features from visual images, and analyzes the importance of each area of ​​the image through a spatial attention mechanism to generate an attention weight map.

4. The welding quality monitoring method based on the industrial Internet according to claim 1 is characterized in that: The method is implemented by edge-cloud collaboration, and the implementation method is: The multi-branch deep convolutional neural network is deployed on the edge; the recurrent neural network is deployed on the cloud; Lightweight deployment is performed on the edge: convolution kernels with weights lower than a set threshold in the deep convolutional neural network are removed, and floating-point weights are quantized and compressed; Accelerate the processing on the cloud: use a GPU cluster to process the feature sequences of multiple welding points at the same time, and cache the recurrent neural network calculation results of common working conditions, which can be directly called when repeated queries are made.

5. The welding quality monitoring method based on the industrial Internet according to claim 1 is characterized in that: The step of judging whether the welding quality is abnormal according to the comparison result includes: The system calculates a comprehensive anomaly index based on electrical parameter deviation, trajectory deviation, and visual defect probability. The index is then divided into three levels based on the range of the index. Real-time compensation is performed for the first level, process optimization is performed for the second level, and emergency intervention is performed for the third level. The real-time compensation includes: adjusting parameters whose deviation exceeds a threshold value through incremental PID control and executing it at the edge; The process optimization includes: matching similar cases from a historical database, generating multiple sets of candidate parameters based on the simulation cluster model, and selecting a solution with the smallest deformation and an energy consumption increase less than a preset value for execution; The emergency intervention includes: emergency shutdown and alarm.

6. The welding quality monitoring method based on the industrial Internet according to claim 1 is characterized in that: Also includes: The underlying recurrent neural network of the welding quality detection model uses QRNN units to process high-frequency data; High-level recurrent neural networks use GRU units to integrate low-frequency data; The welding quality detection model adds a thermodynamic constraint term into the loss function.

7. The welding quality monitoring method based on the industrial Internet according to claim 1 is characterized in that: Also includes: Online incremental learning; The online incremental learning includes: When new welding data is generated, the system automatically inputs it into the welding quality detection model for incremental training, and continuously optimizes the model parameters by learning the characteristics and patterns in the new data; The online incremental learning further includes: setting trigger conditions for the online incremental learning; the trigger conditions include new data confidence being lower than a threshold, process parameter changes, and periodic triggering; Elastic weight consolidation limits important parameter updates and protects old knowledge; a dynamic replay buffer is set to retain key samples of old data.

8. A welding quality monitoring device based on the industrial Internet, characterized in that: The device comprises: A data acquisition module is used to acquire welding data, wherein the welding data includes welding parameters and quality data; the welding parameters include welding electrical parameters and welding gun movement trajectory data; the quality data includes a visual image of the welding area; and the visual image includes a visual image of the molten pool; A model generation module is configured to use a multi-branch deep convolutional neural network to process electrical parameter, visual image, and motion trajectory data, respectively, to extract multimodal features; fuse the multimodal features through a cross-modal feature fusion mechanism, input the fused features into a recurrent neural network for time series modeling, and obtain a welding quality detection model; and load the welding quality detection model into an edge node to monitor welding quality. The abnormality handling module is used to compare the prediction results of the welding quality detection model with the actual measured quality data, and determine whether the welding quality is abnormal based on the comparison results; if so, it is processed according to preset rules; the preset rules include real-time compensation, process optimization and emergency intervention.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the welding quality monitoring method based on the industrial Internet as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the welding quality monitoring method based on the industrial Internet as described in any one of claims 1 to 7 are implemented.

Images