A welding detection method and system based on an autoencoder

Abstract

The present invention discloses a welding detection method and system based on an autoencoder ， Historical data is generated through quality determination criteria and process parameters, preprocessed data, and external influences in an offline manner; the historical data is stored to generate a database; time-coded sensor data is obtained by querying and downsampling in the database; the data of the time-coded sensor is preprocessed through cycle stage extraction to generate an input sequence #imgabs0# for use in a time series deep learning model. The input sequence #imgabs1# is subjected to dilated causal convolution in combination with FCN, causal convolution, and dilated convolution using TCN to generate a dilated causal convolution operation on the element s of the sequence x, thereby generating an output sequence u. The decoder causes two GRUs to reconstruct the input sequence #imgabs2# in forward and backward order to generate a new output sequence u; a new data set is created, new process parameters are generated, and finally a new database is generated. While predicting quality, the problem of model update due to the dynamic changes in the welding process is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of autoencoders, and particularly to a welding detection method and system based on an autoencoder. Background Art

[0002] Welding technology plays an important role in many industries such as automotive and shipbuilding. Gas Metal Arc Welding (GMAW) is one of the widely used manufacturing processes. The characteristics of the welding process are that there are complex causal relationships among material properties, process conditions, and welding quality. Traditionally, the parameters of the GMAW process, such as the welding speed, electrical cycle time, and the material composition settings of the workpiece and the welding wire, largely depend on the long-term experience and expertise of the operator. The evaluation of welding quality includes detecting the material composition and geometric characteristics of the weld seam. The invasive method is to transversely cut the weld seam and the welded workpiece to inspect the generated micro cross-section, but the cost of this inspection method is too high. The non-invasive method of material scanning with X-rays, although low in cost, requires a special working environment and has high requirements for the configuration of welding settings and material combinations. In the era of Industry 4.0 and the digitalization of welding processes, the data of electrical measurements and sensors collected will be combined with quality inspection, and an autoencoder model will be used to predict welding quality and capture the complex causal relationships among various factors during the welding process. Summary of the Invention

[0003] By providing a welding detection method and system based on an autoencoder in an embodiment of the present application, the problem in the prior art of collecting multi-sensor data, performing real-time processing and feature engineering on time series data, training and deploying an autoencoder model, and solving the model update problem due to the dynamic changes in the welding process while predicting quality is solved.

[0004] An embodiment of the present application provides a welding detection method based on an autoencoder, including,

[0005] S1: Generating historical data through quality judgment criteria, process parameters, preprocessed data, and external influences in an offline manner;

[0006] S2: Storing the historical data on TimescaleDB of PostgreSQL to generate a database;

[0007] S3: Obtaining time-encoded sensor data by querying and downsampling in the database;

[0008] S4: Preprocessing the time-encoded sensor data through cycle stage extraction, and then statistically characterizing and LSTM embedding to generate an input sequence for use in a time series deep learning model ;

[0009] S5: The input sequence The dilated causal convolution combines the FCN, causal convolution, and dilated convolution through the TCN to generate the dilated causal convolution operation on the element s of the sequence x ;

[0010] Through the dilated causal convolution operation of the matrix, an output sequence u of length n is generated;

[0011] The decoder uses two GRUs to reconstruct the input sequence in the forward and backward orders , thereby generating a new output sequence u;

[0012] Create a new data set from the new output series u to generate new process parameters, preprocessed data, and external influences;

[0013] S6: Generate online data through the new process parameters, preprocessed data, and external influences;

[0014] S7: Store the online data on TimescaleDB based on PostgreSQL to generate a new database.

[0015] Furthermore, in the cycle stage described in step S4, it includes,

[0016] S41: Pulse stage, the current source increases the current to heat up and liquefy the welding wire, and rapidly transfers sufficient energy through the rapidly increasing current to form droplets of different sizes. When reaching the plateau period, this stage ends;

[0017] S42: Droplet separation stage, when the droplet reaches a certain size, it will fall off the welding wire and be sprayed into the molten pool.

[0018] S43: Base current stage, the current further decelerates to reach 5% of the maximum power in the current cycle, ensuring that the next cycle can start in a clearly defined state and start forming new droplets.

[0019] Furthermore, the dilated causal convolution operation described in S5 also includes; inputting a one-dimensional sequence x and a filter f through the formula:

[0020] ;

[0021] where d is the dilation factor, k is the filter size, represents the number of directions for calculating from the previous node to the current node, where i represents the number of directions; u is the output sequence of length n.

[0022] Furthermore, the GRU described in S5 also includes;

[0023] The forward reconstruction and the reverse reconstruction are calculated through the attention network of time steps, and the formula is as follows:

[0024] ;

[0025] Generate the final reconstruction , where represents the weight parameter of the attention network by time step. w initializes a two-dimensional trainable weight matrix of size [T, 2]. Then, the weight matrix passes through the softmax layer to ensure that all weights at each time step are in (0, 1); where x is the sequence, is the forward sequence, is the backward sequence.

[0026] A welding detection system based on an autoencoder, comprising:

[0027] A historical data module that generates historical data through quality determination criteria and process parameters, preprocessed data, and external influences in an offline manner;

[0028] A database module: stores the historical data on TimescaleDB of PostgreSQL to generate a database;

[0029] A sensor data module: obtains time-encoded sensor data through querying and downsampling in the database;

[0030] A data preprocessing module: preprocesses the time-encoded sensor data through cycle stage extraction, and then statistically features and LSTM embeddings to generate an input sequence for use in a time series deep learning model ;

[0031] A sequence generation module: passes the input sequence through a TCN that uses dilated causal convolution combined with FCN, causal convolution, and dilated convolution to generate a dilated causal convolution operation on the elements s of the sequence x ;

[0032] Through the matrix of the dilated causal convolution operation generate an output sequence u of length n;

[0033] Use two GRUs in the decoder to reconstruct the input sequence in forward and backward order , thereby generating a new output sequence u;

[0034] A new dataset creation module: creates a new dataset through the new output series u to generate new process parameters, preprocessed data, and external influences;

[0035] Online Data Generation Module: Generate online data through new process parameters, preprocessed data, and external influences;

[0036] New Database Generation Module: Store the online data on TimescaleDB based on PostgreSQL to generate a new database.

[0037] Furthermore, in the sequence generation module, it includes,

[0038] Pulse Phase Unit: The current source increases the current to heat up and liquefy the welding wire, and rapidly transmits sufficient energy through the sharply increasing current to form droplets of different sizes. When reaching the plateau, this phase ends;

[0039] Droplet Separation Phase Unit: When the droplet reaches a certain size, it will detach from the welding wire and be ejected into the molten pool.

[0040] Base Current Phase Unit: The current further decelerates to reach 5% of the maximum power in the current cycle, ensuring that the next cycle can start in a clearly defined state and begin to form new droplets.

[0041] Furthermore, the dilated causal convolution operation in the sequence generation module also includes; Input a one-dimensional sequence x and a filter f through the formula:

[0042] ;

[0043] where d is the dilation factor and k is the filter size, represents the number of directions from the previous node to the current node, where i represents the number of directions; u is the output sequence of length n.

[0044] Furthermore, the GRU in the sequence generation module also includes;

[0045] Calculate the forward reconstruction and backward reconstruction through the attention network of time steps, and the formula is as follows:

[0046] ;

[0047] Generate the final reconstruction , where represents the weight parameter of the attention network by time step. Initialize the size of w as a two-dimensional trainable weight matrix of [T, 2], and then pass the weight matrix through the softmax layer to ensure that all weights at each time step are in (0, 1), so as to assign different weights to different time steps; where x is the sequence, is the forward sequence, is the backward sequence.

[0048] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0049] The present invention proposes a welding detection system based on an autoencoder. By collecting multi-sensor data, real-time processing and feature engineering are performed on time series data, and the autoencoder model is trained and deployed to solve the problem of model update due to the dynamic changes in the welding process while predicting quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of a welding detection system based on an autoencoder;

[0051] Figure 2 It is a schematic diagram of the device of the welding detection system based on an autoencoder;

[0052] Figure 3 It is a schematic diagram of current and voltage in three stages of a single cycle;

[0053] Figure 4 It is a schematic diagram of feature engineering and autoencoder structure, DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] A welding detection system based on an autoencoder is proposed. By collecting multi-sensor data, real-time processing and feature engineering are performed on time series data, and the autoencoder model is trained and deployed to solve the problem of model update due to the dynamic changes in the welding process while predicting quality

[0055] To better understand the above technical solutions, the above technical solutions will be described in detail below in conjunction with the accompanying drawings of the specification and specific embodiments.

[0056] A welding detection method based on an autoencoder includes:

[0057] S1: Generating historical data through quality determination criteria, process parameters, preprocessed data, and external influences in an offline manner;

[0058] See Figure 2 , specifically, the welding device includes, and the process parameters and preprocessed data include that GMAW is different from gas welding or laser welding, and it uses electrical energy to generate the heat required to melt the welding wire and the workpiece. Figure 2 It simply shows the electrical characteristics of GMAW. The most important part of the entire welding device is located between the tip of the welding wire and the workpiece. When a high voltage is applied thereto, an arc will be formed due to the ionization of the injected shielding gas. When the current density in the welding wire is high enough, it will start to heat up and liquefy, generating metal droplets. The droplets fall on the surface of the workpiece to form a molten pool. As the welding progresses, the molten pool dissipates its heat to the surrounding materials, solidifies, and forms a weld seam, fusing the previously separated metal plates together.

[0059] In terms of the quality determination criteria, weld penetration, internal fusion, cracks, porosity, and gaps are considered defects. The tolerable size and area of the defects depend on the functional load and the thickness of the parts to be welded. A binary classification of pass or fail is made for the welding quality. Generally, the product quality is determined offline by domain experts after welding is completed. In the patent process, domain experts can modify the results predicted by the model and feed them into the model for learning.

[0060] S2: Store the historical data on TimescaleDB of PostgreSQL to generate a database;

[0061] The PostgreSQL represents a relational database management system, and the TimescaleDB represents a database for querying time data of TimescaleDB;

[0062] S3: Obtain time-coded sensor data by querying and downsampling in the database;

[0063] S4: Preprocess the data of the time-coded sensor through cycle stage extraction, and then statistically extract features and perform LSTM embedding to generate an input sequence for the time series deep learning model , where the LSTM represents a long short-term memory recurrent neural network.

[0064] See Figure 3 , specifically, this patent detects the welding process behavior and infers the final welding quality by observing the sensor data during the welding process. Among all the sensor data, the current and the welding voltage have the greatest impact on the welding quality because in GMAW, electricity is responsible for the liquefaction of the welding wire. All currents and voltages are synchronously sampled at a sampling frequency of 100 kHz, with a maximum error of 0.5%. The sequence represents the corresponding multivariate time series, and X can be divided into many cycles , where , and m is the number of cycles in a welding process. Each cycle can be divided into three non-overlapping stages , and each stage has a different effect on the accumulation and separation of droplets into the molten pool. The changes in current and voltage during one cycle are as Figure 3 shown.

[0065] In the first stage, the pulse stage, the current source increases the current to heat up and liquefy the welding wire, and sufficient energy is quickly transferred through the rapidly increasing current to form droplets of different sizes. When reaching the plateau period, this stage ends.

[0066] In the second stage, the droplet separation stage, when the droplet reaches a certain size, it will detach from the welding wire and be ejected into the molten pool. Ideally, the induction of energy will be reduced, slowing down the liquefaction speed of the welding wire, thereby reducing the risk of excessive deformation of the droplet in the longitudinal direction, which may lead to unexpected short circuits.

[0067] In the third stage, the base current stage, the current further decelerates to reach 5% of the maximum power in the current cycle, ensuring that the next cycle can start in a clearly defined state and begin to form a new droplet.

[0068] In the actual scenario, the droplet may not separate from the welding wire as expected, but will continue to accumulate mass in one or more subsequent cycles. So the behavior in a single cycle is considered random and has no significant impact on the overall quality of the weld. However, the sequence across multiple cycles as a whole can provide valuable information to evaluate the welding quality.

[0069] S5: The input sequence Adopt dilated causal convolution in combination with FCN, causal convolution, and dilated convolution through TCN to generate the dilated causal convolution operation on the element s of the sequence x ; The TCN is a temporal convolutional network, and the FCN is a fully convolutional network;

[0070] Through the dilated causal convolution operation of the matrix, generate an output sequence u of length n;

[0071] Use two GRUs by the decoder to reconstruct the input sequence in the forward and backward orders to generate a new output sequence u; the GRU is a recurrent neural network;

[0072] Create a new data set through the new output series u to generate new process parameters, preprocessed data, and external influences;

[0073] Specifically, online quality prediction of welding;

[0074] See Figure 1 , the entire system structure and process are based on the GMAW process and sensor data, preprocessed through data management and feature processing, and then trained through a deep learning model to provide online prediction capabilities.

[0075] Data management and visualization require a storage solution with a large storage capacity to store time-coded sensor data obtained from a continuous process, in order to store sensor data with a high sampling rate. During the training phase, data access time is not important, but during the inference phase, when accessing online data, it must be returned immediately with a small delay. Therefore, the system uses TimescaleDB based on PostgreSQL for storage, which is a database specifically designed for time data queries. For non-serialized data such as classification quality data and digital processing data, since the sample size is small and negligible in terms of time efficiency, it can be stored in any relational database.

[0076] See Figure 4 , feature engineering, and for sensor data, preprocessing and feature extraction are also required, such as Figure 4 shown, which shows extracting important features from three stages of the droplet accumulation and separation cycle, and using the features together with a time series model to predict the welding quality. The pipeline starts from extracting different cycles and then each cycle is divided into three stages , and the method mainly focuses on the second stage because this stage is responsible for separating the droplets and has a very large impact on the final quality. To obtain information about the thermophysical effects from the previously welded metal, multiple consecutive cycles of are used as the input for the next processing block. Each is compressed using statistical measurements such as minimum, maximum, average, trend, frequency, etc., embedded and trained by an LSTM model, compressing the input sequence into the embedding, and then the statistical values and the embedding are concatenated to the feature vector of each cycle . Then these feature vectors are combined with adjacent feature vectors to form a sequence , where m represents the number of cycles. The finally obtained sequence is used as the input for the time series deep learning model.

[0077] The structure of the autoencoder is as Figure 4As shown. The encoder part is implemented using a Temporal Convolutional Network (TCN), which is a dedicated convolutional network for sequence modeling tasks. Compared with RNNs, it has the ability of parallel computing and long-dependency relationship modeling. Given an input sequence, it can generate a corresponding output sequence. The biggest feature is that the output at each time step only depends on the previously observed inputs. TCN adopts dilated causal convolutions, which combine 1D fully convolutional networks (FCNs), causal convolutions, and dilated convolutions. In an FCN, the length of each hidden layer is the same as that of the input layer, and zero-padding of the length is added to keep the lengths of subsequent layers the same as that of the previous layer. Causal convolutions are specialized convolutions where the output at time t is only related to the convolution of elements at time t and earlier in the previous layer. Dilated convolutions achieve exponentially large receptive fields by introducing a fixed stride between every two adjacent filters. Given a one-dimensional sequence input x and a filter f, the dilated causal convolution operation F on the element s of the sequence x is defined as

[0078] ;

[0079] where d is the dilation factor, k is the filter size, represents the number of directions for calculating from the previous node to the current node, where i represents the number of directions; u is the output sequence of length n.

[0080] The decoder uses two GRUs to reconstruct the input sequence in forward and backward orders respectively. By weighting these two reconstructions, a time-step attention network is adopted to generate the final reconstruction of the input. Let represent the hidden state of the forward GRU, represent the hidden state of the backward GRU. Through embedding initialization, the reconstruction is calculated according to the basic GRU formula. The difference is that the hidden state of the forward GRU is generated from to in forward order, while the backward GRU is calculated from to in the reverse order. Let the forward reconstruction and the backward reconstruction , the final reconstruction is calculated through the time-step attention network, and the formula is as follows:

[0081] ;

[0082] where x is the sequence, is the forward sequence, is the backward sequence;

[0083] where Denote the weight parameters of the attention network by time step. Initialize \(w\) as a two-dimensional trainable weight matrix of size \([T, 2]\). Then, pass the weight matrix through a softmax layer, where the softmax is an activation function layer, to ensure that all weights at each time step are in the range \((0, 1)\), thereby assigning different weights to different time steps. Finally, add an additional linear layer for classification to obtain the classification result.

[0084] The model must first be trained and validated offline on historical data. After hyperparameter optimization, the generated model can be used simultaneously with the welding process. To enable the model to be used for online prediction, the model must have real-time capabilities during the inference process. Given that the model architecture does not depend on the observation of the entire welding process but rather on a relatively short subsequence, the model can be used for online prediction. To achieve online updating of the initially offline-trained model, continual learning is utilized.

[0085] S6: Generate online data through new process parameters, preprocessed data, and external influences;

[0086] S7: Store the online data on TimescaleDB based on PostgreSQL to generate a new database.

[0087] Specifically, due to the dynamics of the welding process and the changes in boundary conditions and process parameters during production, the existing time series model must be continuously updated during the online prediction phase. In other words, once a significant process change occurs, it is necessary to observe and learn the new process and train the model.

[0088] Continual learning is one of the paradigms to solve this problem, aiming to update deep learning models with new data while retaining previously learned knowledge, and maximizing the application of knowledge learned from previous tasks to the learning of future tasks, thereby improving learning efficiency. The patent uses the regularization strategy of Elastic Weight Consolidation (EWC). By calculating the importance of the parameters in the model to the old knowledge, it restricts the update of these important weights for new tasks to reduce catastrophic forgetting. Changes in wire feeding speed, welding speed parameters, or welding equipment during the welding process will trigger model updates. Model updates include two steps. First, create a new dataset by collecting new experimental observations from the welding process. Then, retrain the model with the data using EWC. When two tasks are learned sequentially, with the old task being task A and the new task being task B, EWC adds a regularization loss function on the basis of the original network loss function. The formula is as follows:

[0089] ;

[0090] ;

[0091] Among them, represents the loss function of the current new task, represents the parameter importance on the old task, is the weight adjustment factor, and are both one-dimensional vectors of length N. D is the dataset, and L is the loss function of the old task. is the loss of the model with parameter on a data sample .

[0092] The technical solution in the embodiment of the present application has at least the following technical effects or advantages: A welding detection system based on an autoencoder, where the key technology is to propose an autoencoder model, a TCN encoder, and two RNN decoders that reconstruct the input in reverse and forward order, which can make a better trade-off between modeling short-term and long-term correlations of time series. At the same time, using continuous learning, when the device parameters or material properties change during the welding process, the model can be effectively updated to improve the model application speed.

[0093] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0094] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of processes and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for realizing the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0095] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including an instruction device that implements the functions specified in one or more of the processes Figure 1 or blocks Figure 1 specified in one or more of the processes and / or blocks.

[0096] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one or more of the processes Figure 1 or blocks Figure 1 specified in one or more of the processes and / or blocks.

[0097] Although the preferred embodiments of the present invention have been described, additional changes and modifications can be made by those skilled in the art once they learn of the basic inventive concept. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications that fall within the scope of the present invention.

[0098] It is apparent that those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A welding detection method based on an autoencoder, characterized in that including S1: Generate historical data based on quality determination criteria, process parameters, pre - processed data, and external influences in the offline mode; S2: Store the historical data on TimescaleDB of PostgreSQL to generate a database; S3: Obtain time - encoded sensor data by querying and down - sampling in the database; S4: Preprocess the data of the time-coded sensor by cycle stage extraction, then statistically characterize and LSTM-embed to generate an input sequence for use in a time series deep learning model ; Pulse stage: The current source increases the current, causing the welding wire to heat up and liquefy. Sufficient energy is rapidly transferred through the sharply increased current to form droplets of different sizes. When reaching the plateau period, this stage ends. Droplet separation stage: When the droplet reaches a certain size, it detaches from the welding wire and sprays into the molten pool; Base current stage: The current further decelerates to reach 5% of the maximum power in the current cycle, ensuring that the next cycle can start in a clearly defined state and begin to form new droplets; S5: Take the input sequence and generate the dilated causal convolution operation on the element s of the sequence x by using dilated causal convolution combined with FCN, causal convolution and dilated convolution in the TCN ; Generate an output sequence u of length n through the matrix of the dilated causal convolution operation ; The decoder uses two GRUs to reconstruct the input sequence in forward and backward orders to generate a new output sequence u; Create a new data set through the new output sequence u, generating new process parameters, pre - processed data, and external influences; S7: Generate online data based on the new process parameters, pre - processed data, and external influences; S8: Store the online data on TimescaleDB based on PostgreSQL to generate a new database. Create a new data set by collecting new experimental observations from the welding process and retrain the model using EWC. When two tasks are learned sequentially, with the old task being task A and the new task being task B, EWC adds a regularization loss function to the original network loss function; 2. A welding detection method based on an auto - encoder according to claim 1, characterized in that: The dilation causal convolution operation described in step S5 further includes: inputting a one-dimensional sequence x and a filter f through the formula: ； where d is the dilation factor and k is the filter size, represents the number of directions calculated from the previous node to the current node, where i represents the number of directions; u is the output sequence of length n.

3. The welding detection method based on an autoencoder according to claim 1, wherein: The GRU described in S5 further includes; Calculate the forward reconstruction and the reverse reconstruction through the attention network of time steps. The formula is as follows: ； Generate the final reconstruction , where represents the weight parameters of the attention network at each time step. Initialize a two-dimensional trainable weight matrix w of size [T, 2]. Then, pass the weight matrix through a softmax layer to ensure that all weights at each time step are in the range (0, 1), thereby assigning different weights to different time steps; where x is a sequence, is a forward sequence, is a backward sequence.

4. A welding detection system based on an autoencoder, characterized in that, including Historical data module: Generate historical data based on quality determination criteria, process parameters, pre - processed data, and external influences in the offline mode; Database module: Store the historical data on TimescaleDB of PostgreSQL to generate a database; Sensor data module: Obtain time - encoded sensor data by querying and down - sampling in the database; Data preprocessing module: Preprocess the data of the time-coded sensor through cycle stage extraction, then statistically characterize and LSTM embed to generate an input sequence for use in a time series deep learning model ; Pulse stage: The current source increases the current to heat and liquefy the welding wire, and rapidly transfers sufficient energy through the sharply increased current to form droplets of different sizes. When the plateau period is reached, this stage ends; Droplet separation stage: When the droplet reaches a certain size, it will fall off the welding wire and be sprayed into the molten pool; Base current stage: The current further decelerates to reach 5% of the maximum power in the current cycle, ensuring that the next cycle can start in a clearly defined state and begin to form new droplets; Sequence generation module: The input sequence uses dilated causal convolution in combination with FCN, causal convolution, and dilated convolution through TCN to generate a dilated causal convolution operation on the element s of sequence x ; Generate an output sequence u of length n through the matrix of the dilated causal convolution operation ; The decoder uses two GRUs to reconstruct the input sequence in forward and backward orders to generate a new output sequence u; New data set creation module: Create a new data set through the new output series u, generating new process parameters, pre - processed data, and external influences; Online data generation module: Generate online data based on the new process parameters, pre - processed data, and external influences; New database generation module: Store the online data on TimescaleDB based on PostgreSQL to generate a new database. Create a new data set by collecting new experimental observations from the welding process and retrain the model using EWC. When two tasks are learned sequentially, with the old task being task A and the new task being task B, EWC adds a regularization loss function to the original network loss function; 5. The welding detection system based on an autoencoder according to claim 4, characterized in that: The dilation causal convolution operation of the sequence generation module further includes: inputting a one-dimensional sequence x and a filter f through the formula: ； where d is the dilation factor and k is the filter size, represents calculating the number of directions from the previous node to the current node, where i represents the number of directions; u is the output sequence of length n.

6. The welding detection system based on an autoencoder according to claim 4, characterized in that: The GRU in the sequence generation module further includes; The forward reconstruction and the backward reconstruction are calculated by an attention network with a time step, and the formula is as follows: ； Generate the final reconstruction , where represents the weight parameter of the attention network at each time step. Initialize a two-dimensional trainable weight matrix w of size [T, 2]. Then, pass the weight matrix through a softmax layer to ensure that all weights at each time step are in the range (0, 1), thereby assigning different weights to different time steps; where x is a sequence, is a forward sequence, is a backward sequence.

Images