Resistance spot welding quality monitoring method and system based on edge calculation

Abstract

The invention discloses a resistance spot welding quality monitoring method and system based on edge calculation, and the method comprises the steps: obtaining the welding voltage data of a resistance spot welding machine and the welding current data of the resistance spot welding machine, and constructing a welding dynamic resistance curve of the resistance spot welding machine; based on the welding dynamic resistance curve of the resistance spot welding machine and a Tensor Flow framework, an embedded neural network model is constructed; and on the basis of the embedded neural network model, welding spot quality prediction is conducted on the welding dynamic resistance curve of the resistance spot welding machine, and a welding spot quality prediction result is obtained. Dynamic resistance data and spot welding quality characteristics in the spot welding process can be automatically extracted, and the welding quality detection precision of the resistance spot welding machine is improved. The resistance spot welding quality monitoring method and system based on edge calculation can be widely applied to the technical field of intelligent detection of resistance welding manufacturing.

Description

Technical Field

[0001] The present invention relates to the field of intelligent detection technology for resistance welding manufacturing, and in particular to a resistance spot welding quality monitoring method and system based on edge computing. Background Art

[0002] In the manufacturing industries such as home appliances, automobiles, electronics, and aerospace, resistance spot welding technology is a key technology for connecting structural parts. For example, in the automobile manufacturing process, the welded parts used in the car body are directly related to the structural rigidity and collision safety performance of the vehicle, and the quality of the welds has become a key factor in ensuring that the overall quality of the car meets the standards. At present, the research on resistance welding quality detection has conducted in-depth analysis of various types of data in the welding process, including electrical signals, electrode displacement, ultrasonic signals, and image data. A large number of studies have shown that defective welds are reflected in the corresponding welding process data, which makes it possible to apply data analysis methods on actual production lines. In addition, some studies on spot welding quality assessment methods based on shallow machine learning models require manual feature extraction, feature analysis selection, and feature dimension reduction of welding process data. This method can only extract limited features, and the internal and external connections of data changes cannot be analyzed through empirical feature extraction. In addition, after the machine learning model is deployed, the model needs to be updated, and this process also requires manual operation. Summary of the invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide a resistance spot welding quality monitoring method and system based on edge computing, which can automatically extract the dynamic resistance data and spot welding quality characteristics during the spot welding process, thereby improving the welding quality detection accuracy of the resistance spot welding machine.

[0004] The first technical solution adopted by the present invention is: a resistance spot welding quality monitoring method based on edge computing, comprising the following steps:

[0005] Acquire welding voltage data and welding current data of the resistance spot welding machine, and construct a welding dynamic resistance curve of the resistance spot welding machine;

[0006] Based on the welding dynamic resistance curve of the resistance spot welder and the TensorFlow framework, an embedded neural network model is constructed;

[0007] Based on the embedded neural network model, the welding dynamic resistance curve of the resistance spot welder is used to predict the weld quality and the weld quality prediction result is obtained.

[0008] Further, the step of obtaining the welding voltage data and the welding current data of the resistance spot welder and constructing the welding dynamic resistance curve of the resistance spot welder specifically includes:

[0009] Obtain welding voltage data of the resistance spot welding machine through the voltage acquisition module;

[0010] Acquire welding current data of the resistance spot welding machine through the current acquisition module;

[0011] The dynamic resistance curve is calculated according to the welding voltage data and the welding current data of the resistance spot welding machine to obtain the welding dynamic resistance curve of the resistance spot welding machine.

[0012] Furthermore, the calculation expression of the dynamic resistance curve is specifically as follows:

[0013]

[0014] In the above formula, R(t) represents the change of dynamic resistance over time, V(t) and I(t) represent the change of welding voltage and welding current over time respectively.

[0015] Furthermore, the step of constructing an embedded neural network model based on the welding dynamic resistance curve of the resistance spot welder and the Tensor Flow framework specifically includes:

[0016] Performing neural network model training based on the welding dynamic resistance curve of the resistance spot welder to obtain an original neural network model, wherein the original neural network model represents a one-dimensional convolutional neural network model built under the Tensor Flow framework;

[0017] The original neural network model is subjected to lightweight operations such as quantization and pruning to obtain an embedded neural network model.

[0018] Furthermore, the embedded neural network model specifically includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer and an output layer, wherein the input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the third pooling layer, the first fully connected layer, the second fully connected layer, the third fully connected layer and the output layer are connected in sequence.

[0019] Further, the step of predicting the quality of the weld spot based on the dynamic resistance curve of the resistance spot welder based on the embedded neural network model to obtain the weld spot quality prediction result specifically includes:

[0020] Input the welding dynamic resistance curve of the resistance spot welder into the embedded neural network model;

[0021] Based on the input layer of the embedded neural network model, the welding dynamic resistance curve of the resistance spot welder is obtained;

[0022] Based on the first convolution layer of the embedded neural network model, feature extraction processing is performed on the welding dynamic resistance curve of the resistance spot welder to obtain the welding dynamic resistance characteristic curve of the first resistance spot welder;

[0023] Based on the first pooling layer of the embedded neural network model, a pooling process is performed on the welding dynamic resistance characteristic curve of the first resistance spot welding machine to obtain the welding dynamic resistance characteristic curve of the resistance spot welding machine after the first pooling process;

[0024] Based on the second convolution layer of the embedded neural network model, feature extraction processing is performed on the welding dynamic resistance characteristic curve of the resistance spot welder after the first pooling to obtain the welding dynamic resistance characteristic curve of the second resistance spot welder;

[0025] Based on the second pooling layer of the embedded neural network model, a pooling process is performed on the welding dynamic resistance characteristic curve of the second resistance spot welding machine to obtain the welding dynamic resistance characteristic curve of the resistance spot welding machine after the second pooling process;

[0026] Based on the third convolution layer of the embedded neural network model, feature extraction processing is performed on the welding dynamic resistance characteristic curve of the resistance spot welder after the second pooling to obtain the welding dynamic resistance characteristic curve of the third resistance spot welder;

[0027] Based on the third pooling layer of the embedded neural network model, the welding dynamic resistance characteristic curve of the third resistance spot welding machine is pooled to obtain the welding dynamic resistance characteristic curve of the resistance spot welding machine after the third pooling;

[0028] Based on the first fully connected layer, the second fully connected layer, and the third fully connected layer of the embedded neural network model, feature mapping processing is performed on the welding dynamic resistance characteristic curve of the resistance spot welder after the third pooling to obtain a weld quality prediction result;

[0029] Based on the output layer of the embedded neural network model, the solder joint quality prediction results are output.

[0030] Furthermore, it also includes combining the welding voltage data of the resistance spot welding machine, the welding current data of the resistance spot welding machine and the welding spot quality prediction result, and performing a visual display process.

[0031] Furthermore, it also includes:

[0032] Select and process the abnormal solder joint data in the solder joint quality prediction results to construct an abnormal solder joint data set;

[0033] Perform data enhancement processing on the abnormal solder joint data set to obtain an enhanced abnormal solder joint data set;

[0034] Based on the enhanced abnormal solder joint data set, the embedded neural network model is trained and updated to construct an updated embedded neural network model;

[0035] The updated embedded neural network model and the embedded neural network model are tested for accuracy, and the embedded neural network model corresponding to the higher accuracy is saved.

[0036] The second technical solution adopted by the present invention is: a resistance spot welding quality monitoring system based on edge computing, comprising:

[0037] The first module is used to obtain the welding voltage data and the welding current data of the resistance spot welding machine, and construct the welding dynamic resistance curve of the resistance spot welding machine;

[0038] The second module is used to build an embedded neural network model based on the welding dynamic resistance curve of the resistance spot welder and the Tensor Flow framework;

[0039] The third module is used to predict the quality of the weld spot based on the welding dynamic resistance curve of the resistance spot welder based on the embedded neural network model to obtain the weld spot quality prediction result.

[0040] The beneficial effects of the method and system of the present invention are as follows: the present invention constructs a welding dynamic resistance curve of the resistance spot welder by acquiring the welding voltage data and the welding current data of the resistance spot welder, further constructs an embedded neural network model based on the welding dynamic resistance curve of the resistance spot welder and the Tensor Flow framework, analyzes and processes the dynamic resistance data in the spot welding process through an end-to-end deep learning model algorithm, takes the original resistance curve data as input and the spot welding quality type as output, realizes automatic extraction of the characteristics of the dynamic resistance data and the spot welding quality in the spot welding process, avoids complex feature extraction and analysis processes, and improves the welding quality detection accuracy of the resistance spot welder. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flow chart of the steps of a resistance spot welding quality monitoring method based on edge computing of the present invention;

[0042] Figure 2 It is a structural block diagram of a resistance spot welding quality monitoring system based on edge computing of the present invention;

[0043] Figure 3 It is a schematic diagram of the hardware system structure provided by a specific embodiment of the present invention;

[0044] Figure 4 is a schematic diagram of a system architecture provided by a specific embodiment of the present invention;

[0045] Figure 5It is a schematic diagram of a WEB management interface of a cloud platform provided by a specific embodiment of the present invention;

[0046] Figure 6 is a schematic diagram of a system network topology provided by a specific embodiment of the present invention;

[0047] Figure 7 It is a schematic diagram of the workflow of the embedded platform provided by a specific embodiment of the present invention;

[0048] Figure 8 It is a schematic diagram of a process of self-updating a neural network model on a cloud platform provided by a specific embodiment of the present invention;

[0049] Fig. 9 It is a schematic diagram of the structure of an embedded neural network model provided by a specific embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only provided for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0051] First, it should be noted that the embodiment of the present invention includes an embedded platform and a cloud platform. The embedded platform pre-processes the sensor data to obtain the dynamic resistance curve data of the resistance spot welding machine and inputs it into the neural network for prediction and outputs the quality type of the weld. The embedded platform transmits the dynamic resistance curve data and the weld prediction results to the cloud platform through the local area network. The cloud platform displays the dynamic resistance curves and prediction results of multiple resistance spot welding devices in real time, and saves them to the database for model training and updating.

[0052] The system can transmit data by connecting multiple embedded platforms and cloud platforms to the same LAN network.

[0053] The embedded platform includes a microprocessor, an edge computing module, a voltage acquisition module, a current acquisition module, an Ethernet chip and an Ethernet interface, and a display module. The microprocessor is responsible for coordinating the operation of the embedded platform, the voltage acquisition module is used to collect voltage data during welding, the current acquisition module is used to obtain current data during welding, the edge computing module is used for calculating the neural network model, the Ethernet chip and the Ethernet interface are used for communication between the embedded platform and the cloud platform, and the display module is used to display the operating status of the embedded platform. The microprocessor receives the voltage and current signals collected by the sensor through the data bus; the Ethernet chip communicates with the microprocessor through the RMII interface, and transmits data with the Ethernet interface socket through two pairs of sending and receiving differential lines; the display module transmits data with the microprocessor through the FMSC bus, and displays the voltage, current data, dynamic resistance curve, and solder joint quality prediction results on the display screen. The solder joint quality prediction results are calculated by the edge computing module through the neural network model. The neural network model is a one-dimensional convolutional neural network model built under the Tensor Flow framework. After training on the cloud platform, a lightweight embedded neural network model is obtained by quantization and cutting.

[0054] The cloud platform includes a neural network model training module, a neural network model clipping module, a neural network model push module, a database module, and a WEB management interface; the neural network model training module automatically generates a data set by reading data in the database and performs neural network model training to obtain an original neural network model; the original neural network model is a one-dimensional convolutional neural network model built under the Tensor Flow framework; the neural network model clipping module performs lightweight operations such as quantization and clipping on the trained original neural network model to obtain an embedded neural network model; the neural network model push module sends the embedded neural network model to each embedded platform through the network; the database module receives data from the embedded platform in real time and stores the data; the WEB management interface can view the sensor data and solder joint quality prediction results of all embedded platforms in real time, view the historical records of welding, enter or modify the actual detection quality of solder joints, and view and switch each historical version of the neural network model.

[0055] Reference Figure 1 The present invention provides a resistance spot welding quality monitoring method based on edge computing, the method comprising the following steps:

[0056] S100, obtaining welding voltage data and welding current data of the resistance spot welding machine, and constructing a welding dynamic resistance curve of the resistance spot welding machine;

[0057] Specifically, the welding voltage data of the resistance spot welder is obtained through the voltage acquisition module; the welding current data of the resistance spot welder is obtained through the current acquisition module; the dynamic resistance curve is calculated according to the welding voltage data and the welding current data of the resistance spot welder to obtain the welding dynamic resistance curve of the resistance spot welder.

[0058] The calculation expression of the dynamic resistance curve is as follows:

[0059]

[0060] In the above formula, R(t) represents the change of dynamic resistance over time, V(t) and I(t) represent the change of welding voltage and welding current over time respectively.

[0061] S200, based on the welding dynamic resistance curve of the resistance spot welder and the Tensor Flow framework, build an embedded neural network model;

[0062] Specifically, a neural network model is trained based on the welding dynamic resistance curve of a resistance spot welder to obtain an original neural network model, wherein the original neural network model represents a one-dimensional convolutional neural network model built under the Tensor Flow framework; lightweight operations such as quantization and trimming are performed on the original neural network model to obtain an embedded neural network model.

[0063] In this embodiment, if Fig. 9 As shown, the embedded neural network model specifically includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer and an output layer, wherein the input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the third pooling layer, the first fully connected layer, the second fully connected layer, the third fully connected layer and the output layer are connected in sequence.

[0064] S300, based on the embedded neural network model, the welding dynamic resistance curve of the resistance spot welder is used to predict the weld quality and obtain the weld quality prediction result.

[0065] Specifically, the welding dynamic resistance curve of the resistance spot welder is input into the embedded neural network model; based on the input layer of the embedded neural network model, the welding dynamic resistance curve of the resistance spot welder is obtained; based on the first convolution layer of the embedded neural network model, the welding dynamic resistance curve of the resistance spot welder is subjected to feature extraction processing to obtain the welding dynamic resistance characteristic curve of the first resistance spot welder; based on the first pooling layer of the embedded neural network model, the welding dynamic resistance characteristic curve of the first resistance spot welder is subjected to pooling processing to obtain the welding dynamic resistance characteristic curve of the resistance spot welder after the first pooling; based on the second convolution layer of the embedded neural network model, the welding dynamic resistance characteristic curve of the resistance spot welder after the first pooling is subjected to feature extraction processing to obtain the welding dynamic resistance characteristic curve of the second resistance spot welder; based on the second pooling layer of the embedded neural network model, the welding dynamic resistance characteristic curve of the second resistance spot welder is subjected to feature extraction processing to obtain the welding dynamic resistance characteristic curve of the second resistance spot welder The welding dynamic resistance characteristic curve of the welding machine is pooled to obtain the welding dynamic resistance characteristic curve of the resistance spot welding machine after the second pooling; based on the third convolution layer of the embedded neural network model, the characteristic curve of the welding dynamic resistance characteristic curve of the resistance spot welding machine after the second pooling is subjected to feature extraction processing to obtain the welding dynamic resistance characteristic curve of the third resistance spot welding machine; based on the third pooling layer of the embedded neural network model, the welding dynamic resistance characteristic curve of the third resistance spot welding machine is pooled to obtain the welding dynamic resistance characteristic curve of the resistance spot welding machine after the third pooling; based on the first fully connected layer, the second fully connected layer, and the third fully connected layer of the embedded neural network model, the welding dynamic resistance characteristic curve of the resistance spot welding machine after the third pooling is subjected to feature mapping processing to obtain the weld quality prediction result; based on the output layer of the embedded neural network model, the weld quality prediction result is output.

[0066] In this embodiment, the embedded platform transmits the dynamic resistance curve of the welding to the edge computing module for prediction through the neural network model and outputs the prediction result of the solder joint quality, and transmits the dynamic resistance curve and the quality prediction result to the cloud platform; if the prediction result is normal, the solder joint is judged to be qualified and proceeds to the next welding; if the prediction result is abnormal (defects such as cold solder joint, burn-through, small diameter, etc.), the solder joint is judged to be unqualified and an alarm is issued; perform multiple welding operations, record the defective solder joints and enter them into the cloud platform.

[0067] Another embodiment of the present invention also includes selecting and processing abnormal solder joint data in the solder joint quality prediction results to construct an abnormal solder joint data set; performing data enhancement processing on the abnormal solder joint data set to obtain an enhanced abnormal solder joint data set; training and updating the embedded neural network model based on the enhanced abnormal solder joint data set to construct an updated embedded neural network model; performing accuracy testing on the updated embedded neural network model and the embedded neural network model, and saving the corresponding embedded neural network model with higher accuracy.

[0068] Specifically, after collecting data from multiple welding spots, the cloud platform generates a data set from all the welding data stored in the database and divides it into a training set and a test set in a ratio of 7:3; based on the low probability of defects when welding with a resistance spot welder, the data enhancement method is used to expand the data of defective welds and optimize the model's recognition of defective welds; based on the simple structure of spot welding dynamic resistance data, an original one-dimensional convolutional neural network model is built under the TensorFlow framework and trained using the training set. After training, the model is quantized and trimmed to obtain an embedded neural network model and tested using the test set.

[0069] The embedded neural network model is pushed to the embedded platform, and the embedded platform receives the neural network model and updates the model; then the front-end interface of the cloud platform is built to realize the dynamic resistance curve display of resistance spot welding of multiple embedded platforms, the display and statistical recording of spot welding quality prediction results.

[0070] like Figure 3 As shown in the figure, the embedded platform collects the data of the resistance welding robot and transmits the data to the cloud server of the cloud platform through Ethernet. The cloud platform includes WEB management, database system and model update modules. One embedded platform is responsible for the data collection, analysis and transmission of one resistance welding robot, and the cloud platform manages multiple embedded platforms.

[0071] like Figure 4 The system architecture diagram of the present invention is shown, including an embedded platform and a cloud platform. The embedded platform pre-processes the sensor data to obtain the dynamic resistance curve data of the resistance spot welding machine and inputs it into the neural network for prediction and outputs the quality type of the weld. The embedded platform transmits the dynamic resistance curve data and the weld prediction results to the cloud platform through the local area network. The cloud platform displays the dynamic resistance curves and prediction results of multiple resistance spot welding devices in real time, and saves them to the database for model training and updating.

[0072] like Figure 5As shown, the WEB management interface of the cloud platform includes a real-time data viewing page, a historical data viewing page, an actual detection defect entry interface, a historical version model viewing and switching interface, and an embedded platform detection parameter modification page; the real-time data viewing page can view the real-time welding data and solder joint quality detection results of multiple embedded platforms; the historical data viewing page can view the historical welding data of all embedded platforms and modify their data; the actual detection defect entry interface can enter defective solder joints found during actual manual inspection; the historical version model viewing and switching interface can view the accuracy curves and model parameter setting values ​​of all historical versions of neural network models, and choose to switch the detection model used by the current embedded platform to a certain historical version; the embedded platform detection parameter modification page can adjust the detection parameters of the embedded platform.

[0073] Further, if Figure 6 As shown, the system can be composed of one or more embedded platforms and a cloud platform, and the platforms are connected through a switch to form a local area network for data transmission. An embedded platform is used to collect and process the welding information of a resistance spot welder. The cloud server stores the data information of all resistance spot welders.

[0074] In some embodiments, the cloud server is connected to the local area network where the embedded platform is located via Ethernet, and the embedded platform and the cloud server use the TCP protocol to exchange data. The cloud server stores the received data in the database of the cloud platform and displays the data visually on the cloud platform.

[0075] like Figure 7 As shown in the figure, after the resistance spot welder starts working, the embedded platform collects the welding current and voltage data in real time, pre-processes the voltage and current data to generate a dynamic resistance curve, and inputs the dynamic resistance curve data into the embedded neural network model to obtain the weld quality prediction result and display it on the display module of the embedded platform. If the prediction result is abnormal, an alarm is issued. At the same time, the welding dynamic resistance data and the welding prediction result are transmitted to the cloud server, and the welding situation is displayed on the cloud platform WEB page. When all the welding work is completed, the cloud platform will optimize the neural network model. Before the model optimization, the actual detected defective welds need to be entered into the cloud platform.

[0076] The actual detection of defective welds and entering them into the cloud platform includes manual random inspection of welds after welding is completed. If the welds are found to be defective during manual inspection, they will be recorded and entered into the cloud platform.

[0077] The optimization of neural network models includes neural network model training, model quantization and pruning, and model push. Figure 8As shown, first extract all the solder joint data stored in the cloud server, perform data enhancement on the solder joint data actually detected as defects to obtain more defect data, convert all the data into the input format of Tensor Flow, generate training sets and test sets in a ratio of 7:3, and train the model. In some embodiments of the present invention, the cloud server stores the solder joint data in a MySQL server, each embedded platform has its corresponding number, each embedded platform has multiple solder joints with different welding parameters, each solder joint has its corresponding model, and when training the model, the corresponding model is trained according to the device number and solder joint number.

[0078] Since the probability of welding defects in resistance spot welders is low in actual scenarios, resulting in a small amount of defective data, data enhancement is required. In some embodiments of the present invention, periodic noise, Gaussian white noise or random noise is added to simulate the situation where data collection is interfered with by electromagnetic interference during welding. In some embodiments, a generative adversarial network is used to generate new welding data. In theory, high-quality synthetic data can be generated, but due to the complexity of the training process of the generative adversarial network, the problem of mode collapse is prone to occur. The signal-to-noise ratio is a measure used to describe the relationship between signal and noise, and is defined as the ratio of signal power to noise power. The higher the signal-to-noise ratio, the greater the signal power compared to the noise power, indicating that the data quality is better. The calculation formula is as follows:

[0079]

[0080] In the above formula, P s Represents the power of the signal, P z Represents the power of the noise.

[0081] In some embodiments, Gaussian white noise is randomly added to the data in the training set to generate new data with different S values, and the model evaluation accuracy of different S values ​​is statistically analyzed to obtain the S value parameter with the highest test accuracy. After completing the selection of the training set, it is necessary to select a suitable model for training. In some embodiments of the present invention, the dynamic resistance data in the resistance spot welding process studied does not have obvious periodicity, and the length of the original data sample is one-dimensional time series data, and its structure is simple. Therefore, a one-dimensional convolutional neural network is designed to construct an online detection model for resistance welding quality. In some embodiments, the specific implementation method of model training is as follows:

[0082] The solder joint quality evaluation model is constructed using three convolutional feature extractors and three fully connected layers, such as Fig. 9As shown in the figure, the dynamic resistance curve data is input, and the last pooling layer is obtained after 3 convolutions and maximum pooling. The multi-channel feature mapping matrix obtained by multiple convolutions is concatenated using the flattening layer to obtain the output vector as the input of the fully connected layer. After the dimension is adjusted by 3 layers of fully connected layers, a 5-dimensional feature vector is obtained as the input of the SoftMax layer. In order to enhance the nonlinear mapping ability of the model, the ReLU activation function is added after each convolution layer and fully connected layer of the network. In addition, technologies such as BatchNorm layer and Dropout layer are introduced to prevent overfitting and improve the generalization ability of the model.

[0083] In some instances, the open source deep learning framework Tensor Flow is selected to write an algorithm model, and the relevant algorithm modules are selected to be trained and tested on the device. After the model training is completed, it needs to be quantized and cropped to be converted into a neural network model that can be used by the embedded platform and the model is tested using a test set. In some embodiments, the microprocessor model used by the embedded platform is RK3588, and the model trained using the Tensor flow framework is saved in .pb format. The TensorFlow model is converted to a TFLite model using TensorFlow Lite Converter, and the default optimization and INT16 data types are selected to perform preliminary quantization of the model. The TFLite model is converted to an RKNN model using RKNN-Toolkit2, and the model is further quantized and optimized. The RKNN model is a neural network model format supported by the Rockchip series of microprocessors. After the embedded model is generated, the simulator provided by RKNN-Toolkit2 is used to simulate the Rockchip NPU to run the RKNN model and the model is tested using a test set.

[0084] After the test is completed, the accuracy of the model will be generated. If the accuracy of the model is higher than that of the old version of the model, the model will be pushed to the embedded platform and the old version of the model will be backed up. Since the resistance spot welding quality inspection model belongs to a classification task, there are many evaluation indicators for the model accuracy, such as accuracy, precision, recall, etc. In some embodiments, the accuracy rate (Ac) is used to evaluate the model accuracy, and its calculation formula is as follows:

[0085]

[0086] In the above formula, TP represents a true positive example, TN represents a true negative example, FP represents a false positive example, and FN represents a false negative example.

[0087] The push of the embedded neural network model includes push by the cloud platform and reception by the embedded platform. In some embodiments, the cloud platform publishes a TCP message so that the embedded platform receives a command to update the model, downloads the model file located on the cloud platform to the memory of the embedded platform, stops the model prediction program, deletes the original model file, and restarts the model prediction program to complete the loading of the new version of the model.

[0088] In summary, the embodiment of the present invention analyzes and processes the dynamic resistance data in the spot welding process through an end-to-end deep learning model algorithm, takes the original resistance curve data as input, and the spot welding quality type as output, thereby realizing automatic extraction of the characteristics of the dynamic resistance data and the spot welding quality in the spot welding process, avoiding the complex feature extraction and analysis process, achieving higher detection accuracy, and solving the problems of high detection costs and limitations of manual feature extraction in traditional detection technology. This method deploys the neural network model on the edge device for calculation, saving computing resources for cloud model training. The cloud automatically trains and updates the model in an idle state, and automatically lightweights and converts the updated model and pushes it to the edge device, eliminating the need for manual upgrades of the edge device. This further improves the automation and intelligence of spot welding spot quality detection.

[0089] Therefore, compared with the prior art, the embodiments of the present invention have the following advantages:

[0090] 1) It can be applied to robot and manual resistance spot welding. After the welding of the weld spot, the current welding condition of the workpiece is displayed immediately through the embedded platform, and the location of the weld spot defect is displayed on the cloud platform, so that the inspection personnel can have a more comprehensive understanding of the welding condition of the current workpiece and carry out targeted inspection and repair of the weld spot. The cloud platform can self-learn the defects caused by the welding of the workpiece and regularly update the spot welding detection model. Compared with the traditional defect detection that requires manual intervention to retrain and deploy the model, it saves human resources and improves inspection efficiency.

[0091] 2) The collection of solder joint data and the prediction of solder joint quality are realized on the embedded platform, making the quality inspection of resistance spot welding more dependent on big data information rather than the experience of inspection engineers, reducing labor costs, improving the level of automated inspection of resistance spot welding, and thus improving the production efficiency of large-scale resistance spot welding. The edge computing implemented on the embedded platform is highly efficient and low-cost. Compared with cloud computing, it requires the deployment of a large number of servers, which reduces the cost of large-scale spot welding inspection.

[0092] 3) The welding information during the welding process and the spot welding quality prediction results formed by model calculation are obtained and saved through the cloud platform, which is convenient for inspection personnel to analyze the quality of the welds and also enables the data of the welding process to be quantified and traced back.

[0093] Reference Figure 2 , a resistance spot welding quality monitoring system based on edge computing, comprising:

[0094] The first module 201 is used to obtain welding voltage data and welding current data of the resistance spot welder and construct a welding dynamic resistance curve of the resistance spot welder;

[0095] The second module 202 is used to build an embedded neural network model based on the welding dynamic resistance curve of the resistance spot welder and the Tensor Flow framework;

[0096] The third module 203 is used to predict the quality of the weld spot based on the welding dynamic resistance curve of the resistance spot welder based on the embedded neural network model to obtain the weld spot quality prediction result.

[0097] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0098] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A resistance spot welding quality monitoring method based on edge computing, characterized in that: The following steps are involved: Acquire welding voltage data and welding current data of the resistance spot welding machine, and construct a welding dynamic resistance curve of the resistance spot welding machine; Based on the welding dynamic resistance curve of the resistance spot welder and the Tensor Flow framework, an embedded neural network model is constructed; Based on the embedded neural network model, the welding dynamic resistance curve of the resistance spot welder is used to predict the weld quality and the weld quality prediction result is obtained.

2. According to the edge computing-based resistance spot welding quality monitoring method of claim 1, it is characterized in that: The step of obtaining the welding voltage data and the welding current data of the resistance spot welder and constructing the welding dynamic resistance curve of the resistance spot welder specifically includes: Obtain welding voltage data of the resistance spot welding machine through the voltage acquisition module; Acquire welding current data of the resistance spot welding machine through the current acquisition module; The dynamic resistance curve is calculated according to the welding voltage data and the welding current data of the resistance spot welding machine to obtain the welding dynamic resistance curve of the resistance spot welding machine.

3. According to the method for monitoring resistance spot welding quality based on edge computing as claimed in claim 2, it is characterized in that: The calculation expression of the dynamic resistance curve is specifically as follows: In the above formula, R(t) represents the change of dynamic resistance over time, V(t) and I(t) represent the change of welding voltage and welding current over time respectively.

4. According to the method for monitoring resistance spot welding quality based on edge computing as claimed in claim 3, it is characterized in that: The step of constructing an embedded neural network model based on the welding dynamic resistance curve of the resistance spot welder and the Tensor Flow framework specifically includes: Performing neural network model training based on the welding dynamic resistance curve of the resistance spot welder to obtain an original neural network model, wherein the original neural network model represents a one-dimensional convolutional neural network model built under the Tensor Flow framework; The original neural network model is subjected to lightweight operations such as quantization and pruning to obtain an embedded neural network model.

5. A resistance spot welding quality monitoring method based on edge computing according to claim 4, characterized in that: The embedded neural network model specifically includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer and an output layer, wherein the input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the third pooling layer, the first fully connected layer, the second fully connected layer, the third fully connected layer and the output layer are connected in sequence.

6. A resistance spot welding quality monitoring method based on edge computing according to claim 5, characterized in that: The step of predicting the quality of the weld spot based on the dynamic resistance curve of the resistance spot welder based on the embedded neural network model to obtain the prediction result of the weld spot quality specifically includes: Input the welding dynamic resistance curve of the resistance spot welder into the embedded neural network model; Based on the input layer of the embedded neural network model, the welding dynamic resistance curve of the resistance spot welder is obtained; Based on the first convolution layer of the embedded neural network model, feature extraction processing is performed on the welding dynamic resistance curve of the resistance spot welder to obtain the welding dynamic resistance characteristic curve of the first resistance spot welder; Based on the first pooling layer of the embedded neural network model, a pooling process is performed on the welding dynamic resistance characteristic curve of the first resistance spot welding machine to obtain the welding dynamic resistance characteristic curve of the resistance spot welding machine after the first pooling process; Based on the second convolution layer of the embedded neural network model, feature extraction processing is performed on the welding dynamic resistance characteristic curve of the resistance spot welder after the first pooling to obtain the welding dynamic resistance characteristic curve of the second resistance spot welder; Based on the second pooling layer of the embedded neural network model, a pooling process is performed on the welding dynamic resistance characteristic curve of the second resistance spot welding machine to obtain the welding dynamic resistance characteristic curve of the resistance spot welding machine after the second pooling process; Based on the third convolution layer of the embedded neural network model, feature extraction processing is performed on the welding dynamic resistance characteristic curve of the resistance spot welder after the second pooling to obtain the welding dynamic resistance characteristic curve of the third resistance spot welder; Based on the third pooling layer of the embedded neural network model, the welding dynamic resistance characteristic curve of the third resistance spot welding machine is pooled to obtain the welding dynamic resistance characteristic curve of the resistance spot welding machine after the third pooling; Based on the first fully connected layer, the second fully connected layer, and the third fully connected layer of the embedded neural network model, feature mapping processing is performed on the welding dynamic resistance characteristic curve of the resistance spot welder after the third pooling to obtain a weld quality prediction result; Based on the output layer of the embedded neural network model, the solder joint quality prediction results are output.

7. A resistance spot welding quality monitoring method based on edge computing according to claim 6, characterized in that: The method also includes combining the welding voltage data and the welding current data of the resistance spot welding machine with the welding spot quality prediction result, and performing a visual display process.

8. A resistance spot welding quality monitoring method based on edge computing according to claim 7, characterized in that: Also includes: Select and process the abnormal solder joint data in the solder joint quality prediction results to construct an abnormal solder joint data set; Perform data enhancement processing on the abnormal solder joint data set to obtain an enhanced abnormal solder joint data set; Based on the enhanced abnormal solder joint data set, the embedded neural network model is trained and updated to construct an updated embedded neural network model; The updated embedded neural network model and the embedded neural network model are tested for accuracy, and the embedded neural network model corresponding to the higher accuracy is saved.

9. A resistance spot welding quality monitoring system based on edge computing, characterized in that: Includes the following modules: The first module is used to obtain the welding voltage data and the welding current data of the resistance spot welding machine, and construct the welding dynamic resistance curve of the resistance spot welding machine; The second module is used to build an embedded neural network model based on the welding dynamic resistance curve of the resistance spot welder and the Tensor Flow framework; The third module is used to predict the quality of the weld spot based on the welding dynamic resistance curve of the resistance spot welder based on the embedded neural network model to obtain the weld spot quality prediction result.

Images