Welding detection method and system based on auto-encoder

Abstract

The invention discloses a welding detection method and system based on an auto-encoder, and the method comprises the steps: generating historical data through a quality judgment standard and flow parameters, preprocessing data and external influences in an offline mode; storing the historical data to generate a database; obtaining time-coded sensor data by querying and down-sampling in a database; the method comprises the following steps: preprocessing data of a time coding sensor through periodic phase extraction, generating an input sequence # imgabs0 # serving as a time sequence deep learning model, generating an expansion causal convolution operation on an element s of a sequence x by combining the input sequence # imgabs1 # with FCN, causal convolution and expansion convolution through TCN by adopting expansion causal convolution, so as to generate an output sequence u, enabling the two GRUs to reconstruct an input sequence # imgabs2 # according to a forward sequence and a backward sequence through a decoder, and generating a new output sequence u; and creating a new data set, generating new process parameters, and finally generating a new database. And the problem of model updating due to dynamic change in the welding process is solved while the quality is predicted.

Description

Technical Field

[0001] The present invention relates to the technical field of autoencoders, and in particular 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 automobile and shipbuilding. Gas shielded arc welding (GMAW) is one of the widely used manufacturing processes. The welding process is characterized by a complex causal relationship between material properties, process conditions and welding quality. Traditionally, the setting of GMAW process parameters such as welding speed, electrical cycle time, material composition of workpiece and welding wire depends largely on the operator's long-term experience and expertise. The evaluation of welding quality includes detecting the material composition and geometric characteristics of the weld. The invasive method is to cross-cut the weld and inspect the small cross-sections produced by the welded workpiece, but this inspection method is too expensive. The non-invasive method of material scanning with x-rays is low-cost but 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 collected electrical measurement and sensor data will be combined with quality detection, and the welding quality will be predicted through the autoencoder model to capture the complex causal relationship of various factors in the welding process. Summary of the invention

[0003] The embodiments of the present application provide a welding detection method and system based on an autoencoder, which solves the problem in the prior art of collecting multi-sensor data, performing real-time processing and feature engineering of time series data, training and deploying autoencoder models, and solving the problem of model updating due to dynamic changes in the welding process while predicting quality.

[0004] The embodiment of the present application provides a welding detection method based on an autoencoder, comprising: S1: Generate historical data through quality judgment criteria and process parameters, preprocessing data, and external influences in an offline manner; S2: Store historical data on TimescaleDB of PostgreSQL to generate a database; S3: Obtain time-encoded sensor data by querying and downsampling in the database; S4: Preprocessing the time-coded sensor data by period phase extraction, followed by statistical features and LSTM embedding to generate the input sequence used as the time series deep learning model ; S5: Input sequence The dilated causal convolution operation on the element s of the sequence x is generated by combining FCN, causal convolution and dilated convolution with TCN. ; Through dilated causal convolution operation , generating an output sequence u of length n; Reconstruct the input sequence in forward and backward order using two GRUs through the decoder , thereby generating a new output sequence u; Create a new data set through the new output series u, generate new process parameters, preprocessing data, external influences; S6: Generate online data through new process parameters, preprocessing data, and external influences; S7: Store online data in TimescaleDB based on PostgreSQL to generate a new database.

[0005] Furthermore, in step S4, the cycle phase includes: S41: Pulse stage, the current source increases the current to heat up and liquefy the welding wire, and quickly transfers sufficient energy through the sharply increased current to form droplets of different sizes. When the plateau is reached, this stage ends; S42: Droplet separation stage, when the droplets reach a certain size, they fall off the welding wire and spray into the molten pool.

[0006] S43: Base flow phase, the current is further decelerated to 5% of the maximum power in the current cycle, ensuring that the next cycle can start in a well-defined state and start to form new droplets.

[0007] Furthermore, the dilated causal convolution operation described in S5 Also includes; a one-dimensional sequence input x and filter f through the formula: ; Where d is the dilation factor, k is the filter size, It means 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.

[0008] Furthermore, the GRU described in S5 also includes: The forward reconstruction and the reverse reconstruction are calculated through the attention network of the time step, and the formula is as follows: ;

[0009] Generate final reconstruction ,in represents the weight parameters of the attention network by time step, w initializes a two-dimensional trainable weight matrix of size [T, 2], and then passes the weight matrix through the softmax layer to ensure that all weights of each time step are in (0, 1); where x is a sequence, is the forward sequence, is a backward sequence.

[0010] A welding detection system based on an autoencoder, comprising: The historical data module generates historical data through quality judgment criteria and process parameters, preprocessing data, and external influences in an offline manner; Database module: Store historical data on PostgreSQL's TimescaleDB to generate a database; Sensor data module: obtains time-encoded sensor data by querying and downsampling in the database; Data preprocessing module: preprocesses the data of time-coded sensors through period phase extraction, statistical features and LSTM embedding to generate input sequences used as time series deep learning models ; Sequence generation module: input sequence The dilated causal convolution is used by TCN to combine FCN, causal convolution and dilated convolution to generate the dilated causal convolution operation on the element s of the sequence x. ; Through dilated causal convolution operation , generating an output sequence u of length n; Reconstruct the input sequence in forward and backward order using two GRUs through the decoder , thereby generating a new output sequence u; New dataset creation module: creates a new dataset through a new output series u, generates new process parameters, preprocessing data, and external influences; Online data generation module: Generate online data through new process parameters, preprocessed data, and external influences; New database generation module: Store online data on TimescaleDB based on PostgreSQL to generate a new database.

[0011] Furthermore, in the sequence generation module, it includes: In the pulse stage unit, the current source increases the current to heat up and liquefy the welding wire. By rapidly increasing the current, sufficient energy is quickly delivered to form droplets of different sizes. When the plateau is reached, this stage ends. In the droplet separation stage unit, when the droplets reach a certain size, they will fall off the welding wire and spray into the molten pool.

[0012] In the base flow stage unit, the current is further decelerated to 5% of the maximum power in the current cycle, ensuring that the next cycle can start in a well-defined state and begin to form new droplets.

[0013] Furthermore, the sequence generation module dilates the causal convolution operation Also includes; a one-dimensional sequence input x and filter f through the formula: ; Where d is the dilation factor, k is the filter size, It means 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.

[0014] Furthermore, the GRU in the sequence generation module also includes: The forward reconstruction and the reverse reconstruction are calculated through the attention network of the time step, and the formula is as follows: ; Generate final reconstruction ,in Represents the weight parameters of the attention network by time step, w initializes a two-dimensional trainable weight matrix of size [T, 2], and then passes the weight matrix through the softmax layer to ensure that all weights of each time step are in (0, 1), thereby assigning different weights to different time steps; where x is a sequence, is the forward sequence, is a backward sequence.

[0015] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: The present invention proposes a welding detection system based on autoencoder, which collects multi-sensor data, performs real-time processing and feature engineering on time series data, trains and deploys autoencoder models, and solves the model update problem caused by the dynamic changes of the welding process while predicting the quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic diagram of a welding detection system based on an autoencoder; Figure 2 It is a schematic diagram of a welding detection system based on an autoencoder; Figure 3 Schematic diagram of current and voltage in three stages in a single cycle; Figure 4 Schematic diagram of feature engineering and autoencoder structure, DETAILED DESCRIPTION

[0017] A welding inspection system based on autoencoder is proposed. By collecting multi-sensor data, real-time processing and feature engineering of time series data, training and deploying autoencoder models, the system can solve the model update problem caused by the dynamic changes of the welding process while predicting the quality. In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0018] A welding detection method based on an autoencoder, comprising: S1: Generate historical data through quality judgment criteria and process parameters, preprocessing data, and external influences in an offline manner; See also Figure 2 Specifically, the welding device includes, the process parameters and preprocessing data include, GMAW is different from gas welding or laser welding, which uses electrical energy to generate the heat required to melt the welding wire and the workpiece. Figure 2 The electrical characteristics of GMAW are simply shown. The most important part of the entire welding device is located between the tip of the wire and the workpiece. When high voltage is applied to it, an arc is formed due to the ionization of the injected shielding gas. When the current density in the wire is high enough, it will begin to heat up and liquefy, producing metal droplets, which fall on the surface of the workpiece to form a molten pool. As the welding progresses, the molten pool dissipates its heat into the surrounding material, solidifies and forms a weld, fusing the previously separated metal sheets together.

[0019] In terms of quality judgment criteria, weld penetration, internal fusion, cracks, pores and gaps are considered defects. The tolerable size and area of ​​defects depend on the functional load and the thickness of the parts to be welded, and the welding quality is classified into a binary pass or fail. Generally speaking, 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 send them into the model for learning.

[0020] S2: Store historical data on TimescaleDB of PostgreSQL to generate a database; The PostgreSQL represents a relational database management system, and the TimescaleDB represents a database for TimescaleDB time data query; S3: Obtain time-encoded sensor data by querying and downsampling in the database; S4: Preprocessing the time-coded sensor data by period phase extraction, followed by statistical features and LSTM embedding to generate the input sequence used as the time series deep learning model , the LSTM represents a long short-term memory recurrent neural network.

[0021] See also 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 welding voltage The greatest impact on welding quality is that GMAW is powered by electricity to liquefy the wire. All currents and voltages are sampled synchronously at a sampling frequency of 100 kHz with a maximum error of 0.5%. Represents the corresponding multivariate time series, X can be divided into many periods ,in , m is the number of cycles in a welding process. Each cycle can be divided into three non-overlapping stages Each stage has different effects on the accumulation and separation of droplets into the molten bed. The changes of current and voltage in one cycle are as follows: Figure 3 shown.

[0022] In the first stage, the pulse stage, the current source increases the current to heat up and liquefy the welding wire, and quickly delivers enough energy through the sharply increased current to form droplets of different sizes. This stage ends when the plateau is reached.

[0023] In the second stage, the droplet separation stage, when the droplets reach a certain size, they will fall off the wire and spray into the molten pool. Ideally, the energy induction will be reduced, slowing down the wire liquefaction speed, thereby reducing the risk of excessive deformation of the droplets in the longitudinal direction, which may cause unintended short circuits.

[0024] In the third phase, the base flow phase, the current is further decelerated to 5% of the maximum power in the current cycle, ensuring that the next cycle can start in a well-defined state and begin to form new droplets.

[0025] In real-world scenarios, droplets may not separate from the wire as expected, but continue to accumulate mass in one or more subsequent cycles. Therefore, 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 weld quality.

[0026] S5: Input sequence The dilated causal convolution operation on the element s of the sequence x is generated by combining FCN, causal convolution and dilated convolution with TCN. ; The TCN is a time domain convolutional network, and the FCN is a full convolutional network; Through dilated causal convolution operation , generating an output sequence u of length n; Reconstruct the input sequence in forward and backward order using two GRUs through the decoder , thereby generating a new output sequence u; the GRU is a recurrent neural network; Create a new data set through the new output series u, generate new process parameters, preprocessing data, external influences; Specifically, online quality prediction of welding; See also Figure 1 ,The entire system structure and process are built on the GMAW process and ,sensor data, which are pre-processed through data management and feature ,processing, and then trained through deep learning model to ,provide online prediction capabilities.

[0027] Data management and visualization Because it is necessary to store time-encoded sensor data obtained from continuous processes, a storage solution with large storage capacity is selected to store sensor data with high sampling rates. In the training phase, data access time is not important, but in the inference phase, access to online data must be returned immediately with a small delay, so the system uses TimescaleDB based on PostgreSQL for storage, which is a database specifically used for time data queries. For non-serialized data such as classification quality data and digital processing data, the time efficiency is negligible due to the small sample size and can be stored in any relational database.

[0028] See also Figure 4 ,Feature engineering, sensor data also needs to be preprocessed to extract features, such as Figure 4 The figure shows the extraction of important features from the three stages of the droplet accumulation and separation cycle and the use of the features together with the time series model to predict the welding quality. Each cycle is divided into three phases: , the method mainly focuses on the second stage , as this phase is responsible for the separation of the droplets and has a very large influence on the final quality. In order to obtain information about the thermophysical effects from the previously welded metal, it is necessary to use multiple consecutive cycles of Used as input to the next processing block. All use statistical measurements such as minimum, maximum, average, trend, frequency, etc. for compression, and the embedding is trained by the LSTM model to convert the input sequence Compress into embeddings, then concatenate the statistics and embeddings to the feature vector for each cycle Then these eigenvectors Combined with adjacent feature vectors into a sequence , m represents the number of cycles. The final sequence Used as input to time series deep learning models.

[0029] The structure of the autoencoder is as follows Figure 4 As shown. The encoder part is implemented using the temporal convolutional network TCN, which is a dedicated convolutional network for sequence modeling tasks. Compared with RNN, it has the ability of parallel computing and long dependency modeling. Given an input sequence, it can generate a corresponding output sequence. The biggest feature is that the output of each time step depends only on the previously observed input. TCN uses dilated causal convolution, which combines 1D fully convolutional network (FCN), causal convolution and dilated convolution. In FCN, the length of each hidden layer is the same as the input layer, and zero padding of the length is added to keep the subsequent layer the same length as the previous layer. Causal convolution is a special convolution in which the output at time t is only related to the convolution of elements at time t and earlier in the previous layer. Dilated convolution achieves an exponentially large receptive field by introducing a fixed step size 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 ; Where d is the dilation factor, k is the filter size, It means 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.

[0030] The decoder uses two GRUs to reconstruct the input sequence in forward and backward order respectively, and uses a time-step attention network to produce the final reconstruction of the input by weighting these two reconstructions. represents the hidden state of the forward GRU, Represents 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 calculated in forward order from arrive Generate, and the backward GRU is in reverse order from arrive Calculate. Assume forward reconstruction and reverse reconstruction , and finally reconstruct Calculated by the attention network of the time step, the formula is as follows: ; Where x is a sequence, is the forward sequence, is a backward sequence; in Represents the weight parameters of the attention network according to the time step, w initializes a two-dimensional trainable weight matrix of size [T, 2], and then passes the weight matrix through the softmax layer, which is an activation function layer to ensure that all weights of each time step are in (0, 1), thereby assigning different weights to different time steps. Finally, an additional linear layer is added for classification to obtain the classification result.

[0031] The model must initially be trained and validated on historical data in an offline manner, and after hyperparameter optimization, the resulting model can be used in parallel with the welding process. In order to be able to use the model for online predictions, the model must have real-time capabilities during inference. Given that the model architecture does not rely on observations of the entire welding process, but rather on subsequences of fairly short length, the model can be used for online predictions. In order to achieve online updates of the initially offline trained model, continuous learning is utilized.

[0032] S6: Generate online data through new process parameters, preprocessing data, and external influences; S7: Store online data in TimescaleDB based on PostgreSQL to generate a new database.

[0033] Specifically, due to the dynamic nature 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 major process change occurs, observation learning and model training of the new process are required.

[0034] Continuous learning is one of the paradigms to solve this problem. The purpose is to allow deep learning models to be updated through new data while retaining previously learned knowledge, and to apply the knowledge learned from previous tasks to the learning of future tasks to the greatest extent possible, thereby improving learning efficiency. The patent uses a regularization strategy called Plastic Weight Consolidation (EWC) to limit the update of these important weights in new tasks by calculating the importance of the parameters in the model to the old knowledge to reduce catastrophic forgetting. Changes in wire feed speed, welding speed parameters, or welding equipment during welding will trigger model updates. Model updates include two steps. First, a new data set is created by collecting new experimental observations from the welding process, and then the model is retrained on the data using EWC. When two tasks are learned sequentially, the old task is Task A and the new task is Task B. EWC adds a regularization loss function based on the original network loss function. The formula is as follows: ; ; in, Represents the loss function of the current new task, Representative parameters Importance in old tasks, is the weight adjustment factor, and are all one-dimensional vectors of length N, D is the dataset, L is the loss function of the old task, The parameter is The model in a data sample on losses.

[0035] 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, wherein 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 the short-term and long-term correlation of the time series. At the same time, using continuous learning, when the equipment parameters or material properties change during the welding process, the model can be effectively updated to improve the speed of model application.

[0036] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may 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-ROM, optical storage, etc.) containing computer-usable program code.

[0037] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, 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 device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0038] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1A function specified in one or more boxes.

[0039] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0040] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0041] Obviously, 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 equivalents, 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: include: S1: Generate historical data through quality judgment criteria and process parameters, preprocessing data, and external influences in an offline manner; S2: Store historical data on TimescaleDB of PostgreSQL to generate a database; S3: Obtain time-encoded sensor data by querying and downsampling in the database; S4: Preprocessing the time-coded sensor data by period phase extraction, followed by statistical features and LSTM embedding to generate the input sequence used as the time series deep learning model ; S5: Input sequence The dilated causal convolution is used by TCN to combine FCN, causal convolution and dilated convolution to generate the dilated causal convolution operation on the element s of the sequence x. ; Through dilated causal convolution operation , generating an output sequence u of length n; Reconstruct the input sequence in forward and backward order using two GRUs through the decoder , thereby generating a new output sequence u; Create a new data set through the new output series u, generate new process parameters, preprocessing data, external influences; S6: Generate online data through new process parameters, preprocessing data, and external influences; S7: Store online data in TimescaleDB based on PostgreSQL to generate a new database.

2. A welding detection method based on an autoencoder according to claim 1, characterized in that: The cycle phase in step S4 includes: S41: Pulse stage, the current source increases the current to heat up and liquefy the welding wire, and quickly transfers sufficient energy through the sharply increased current to form droplets of different sizes. When the plateau is reached, this stage ends; S42: Droplet separation stage, when the droplets reach a certain size, they fall off the welding wire and spray into the molten pool; S43: Base flow phase, the current is further decelerated to 5% of the maximum power in the current cycle, ensuring that the next cycle can start in a well-defined state and start to form new droplets.

3. The welding detection method based on the autoencoder according to claim 1, characterized in that: Step S5: dilated causal convolution operation Also includes; a one-dimensional sequence input x and filter f through the formula: ； Where d is the dilation factor, k is the filter size, It means 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.

4. The welding detection method based on the autoencoder according to claim 1, characterized in that: The GRU described in S5 also includes; The forward reconstruction and the reverse reconstruction are calculated through the attention network of the time step, and the formula is as follows: ； Generate final reconstruction ,in Represents the weight parameters of the attention network by time step, w initializes a two-dimensional trainable weight matrix of size [T, 2], and then passes the weight matrix through the softmax layer to ensure that all weights of each time step are in (0, 1), thereby assigning different weights to different time steps; Where x is a sequence, is the forward sequence, is a backward sequence.

5. A welding detection system based on an autoencoder, characterized in that: include, The historical data module generates historical data through quality judgment criteria and process parameters, preprocessing data, and external influences in an offline manner; Database module: Store historical data on PostgreSQL's TimescaleDB to generate a database; Sensor data module: obtains time-encoded sensor data by querying and downsampling in the database; Data preprocessing module: preprocesses the data of time-coded sensors through period phase extraction, statistical features and LSTM embedding to generate input sequences used as time series deep learning models ; Sequence generation module: input sequence The dilated causal convolution is used by TCN to combine FCN, causal convolution and dilated convolution to generate the dilated causal convolution operation on the element s of the sequence x. ; Through dilated causal convolution operation , generating an output sequence u of length n; Reconstruct the input sequence in forward and backward order using two GRUs through the decoder , thereby generating a new output sequence u; New dataset creation module: creates a new dataset through a new output series u, generates new process parameters, preprocessing data, and external influences; Online data generation module: Generate online data through new process parameters, preprocessed data, and external influences; New database generation module: Store online data on TimescaleDB based on PostgreSQL to generate a new database.

6. A welding detection system based on an autoencoder according to claim 5, characterized in that: In the sequence generation module, including: In the pulse stage unit, the current source increases the current to heat up and liquefy the welding wire. By rapidly increasing the current, sufficient energy is quickly delivered to form droplets of different sizes. When the plateau is reached, this stage ends. Droplet separation stage unit, when the droplets reach a certain size, they will fall off the welding wire and spray into the molten pool; In the base flow stage unit, the current is further decelerated to 5% of the maximum power in the current cycle, ensuring that the next cycle can start in a well-defined state and begin to form new droplets.

7. The welding detection system based on the autoencoder according to claim 5, characterized in that: Dilated causal convolution operation described in the sequence generation module Also includes; a one-dimensional sequence input x and filter f through the formula: ； Where d is the dilation factor, k is the filter size, It means 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.

8. The welding detection system based on the autoencoder according to claim 5, characterized in that: The GRU described in the sequence generation module also includes: The forward reconstruction and the reverse reconstruction are calculated through the attention network of the time step, and the formula is as follows: ； Generate final reconstruction ,in Represents the weight parameters of the attention network by time step, w initializes a two-dimensional trainable weight matrix of size [T, 2], and then passes the weight matrix through the softmax layer to ensure that all weights of each time step are in (0, 1), thereby assigning different weights to different time steps; Where x is a sequence, is the forward sequence, is a backward sequence.

Images