CN-122020474-A - Welding process abnormity early warning method based on multisource time sequence signal dynamic analysis

CN122020474ACN 122020474 ACN122020474 ACN 122020474ACN-122020474-A

Abstract

The invention belongs to the technical field of industrial automatic welding, in particular to a welding process abnormality early warning method based on multi-source time sequence signal dynamic analysis, which is used for synchronously collecting continuous time sequence signals of welding current, voltage and molten pool forms, intercepting and preprocessing data in real time through a sliding window, extracting time sequence characteristics of each signal in parallel and fusing the time sequence characteristics into a comprehensive characteristic vector, inputting the characteristic vector into an unsupervised reference model trained only by normal data, calculating a reconstruction error as an abnormality score, dynamically calculating an early warning threshold based on a statistic value of recent abnormality score, and triggering real-time early warning and saving process data when the abnormality score exceeds the dynamic threshold. The welding process abnormity early warning method based on the multisource time sequence signal dynamic analysis realizes the transition from 'post detection' to 'early warning in the event', and has the advantages of early warning time, no need of a defect sample, self-adaption error report resistance, strong real-time performance, good traceability and the like.

Inventors

WANG XIAOQI
ZHU MINGLIANG
XUAN FUZHEN
LU WENQING
WANG YITING
LI ZIXIN

Assignees

华东理工大学

Dates

Publication Date: 20260512
Application Date: 20260131

Claims (10)

1. A welding process abnormity early warning method based on multi-source time sequence signal dynamic analysis is characterized by comprising the following steps of firstly, synchronously collecting multi-channel time sequence signals; Step two, data segmentation and pretreatment of a sliding window; step three, extracting and fusing time sequence features; step four, constructing a reference model offline; step five, online abnormal scoring and dynamic threshold early warning; Step six, early warning response and data tracing; and step seven, verifying the validity.
2. The method for dynamically analyzing abnormal welding process according to claim 1, wherein in the first step, the instantaneous waveform of the welding current, the instantaneous waveform of the arc voltage and the sequence of the size change of the molten pool with time calculated in real time by the vision sensor are synchronously acquired by the sensor and the acquisition card.
3. The method for warning abnormal welding process based on multi-source time sequence signal dynamic analysis according to claim 2, wherein in the second step, multichannel signal data of the latest period of time is intercepted in a fixed period to form data fragments, and standardization and filtering are performed.
4. The method for warning abnormal welding process based on dynamic analysis of multi-source time sequence signals according to claim 3, wherein in the third step, mathematical features describing dynamic changes of each channel signal in the data segment are extracted respectively, and all the features are combined into a comprehensive feature vector to represent the overall state of the welding process in the current time period.
5. The welding process anomaly pre-warning method based on the multi-source time sequence signal dynamic analysis of claim 4, wherein the time sequence feature extraction flow comprises the following specific steps: s31, preparing data, namely intercepting original time sequence data of a sliding window with fixed duration, wherein the original time sequence data comprises a current sequence, a voltage sequence and a molten pool width sequence; S32, extracting characteristics of current and voltage signals, respectively standardizing the current and voltage sequences, decomposing the signals into a plurality of eigen mode functions by using an empirical mode decomposition algorithm, selecting the first 4 components, calculating the average value of the normalized energy and instantaneous frequency of each component, and obtaining 8-dimensional characteristics of each electric signal, wherein the total is 16 dimensions; S33, extracting visual signal characteristics of a molten pool, carrying out smooth filtering on the molten pool width sequence, calculating 8 time domain statistical characteristics of the molten pool width sequence, using a second-order autoregressive model fitting sequence, extracting 2 model coefficients, and obtaining 10-dimensional characteristics altogether; And S34, feature fusion and standardization, namely splicing the 16-dimensional electrical signal features and the 10-dimensional visual signal features to obtain 26-dimensional fusion feature vectors, and carrying out integral standardization on the vectors.
6. The method for warning abnormal welding process based on multi-source time sequence signal dynamic analysis according to claim 5, wherein in the fourth step, a large amount of normal welding process data is collected and processed to obtain a normal feature vector sample library, a deep self-encoder model is trained by using the sample library, normal features are reconstructed, and accordingly a standard of a normal welding process is established.
7. The welding process anomaly pre-warning method based on the multi-source time sequence signal dynamic analysis of claim 6, wherein the training step of the reference model (depth self-encoder) is as follows: S41, constructing a model structure, namely constructing a symmetrical depth self-encoder network, wherein the encoder structure comprises an input layer, a full connection layer and a hidden layer; s42, model training, namely training by using all normal feature vectors extracted from a normal welding process database as a training set and using a mean square error as a loss function; S43, determining an initial threshold value, namely inputting all samples of a training set into a trained model, calculating a reconstruction error of the model, counting the distribution of all error values, and taking the 95 th percentile of all error values as an offline reference threshold value; after the on-line monitoring is started, when the historical anomaly score queue does not reach the preset length, the off-line reference threshold is used for early warning, and after the queue is filled, the operation is automatically switched to the use of the dynamic threshold.
8. The welding process anomaly early warning method based on the multi-source time sequence signal dynamic analysis of claim 7, wherein in the fifth step, during real-time monitoring, the feature vector generated in the third step is input into a trained reference model, the reconstruction error is calculated as an anomaly score, meanwhile, an early warning threshold is dynamically calculated according to the recent historical anomaly score, and when the real-time anomaly score exceeds the dynamic threshold, early warning is immediately triggered.
9. The welding process anomaly pre-warning method based on the multi-source time sequence signal dynamic analysis of claim 8, wherein the specific steps of online real-time monitoring and dynamic pre-warning are as follows: s51, calculating real-time characteristics, namely intercepting the latest data according to a fixed period, and calculating according to a time sequence characteristic extraction flow to obtain a current standardized fusion characteristic vector; S52, calculating an anomaly score by inputting the current feature vector into a deployed depth self-encoder and calculating a reconstruction error of the current feature vector as the anomaly score; S53, updating a dynamic threshold value, wherein the system maintains a historical abnormal score queue with a fixed length, and adds a new score into the queue; S54, early warning judgment and execution, wherein a corresponding early warning threshold value is obtained according to the current state of the system, the current abnormal score is compared with the threshold value, if the abnormal score exceeds the threshold value, early warning is immediately triggered, meanwhile, process data before and after the early warning moment are stored, and if the abnormal score does not exceed the threshold value, the next monitoring cycle is continued.
10. The method for welding process anomaly early warning based on multi-source time sequence signal dynamic analysis of claim 9, wherein in the step six, when early warning is triggered, audible and visual warning is immediately carried out, and all original data in key time periods before and after the triggering time are automatically saved to form a data snapshot for retrospective analysis.

Description

Welding process abnormity early warning method based on multisource time sequence signal dynamic analysis Technical Field The invention relates to the technical field of industrial automatic welding, in particular to a welding process abnormity early warning method based on multisource time sequence signal dynamic analysis. Background Welding is a key process in the manufacture of equipment, the quality of which directly affects product safety. At present, the welding quality control mainly has two major problems, namely, depending on post-welding nondestructive detection (such as X-ray and ultrasonic), the problem cannot be found in real time and is interfered in the welding process, so that the outflow or repair cost of a defective product is high, and depending on the experience of an operator, the online judgment is carried out, so that stable and consistent automatic production is difficult to realize. In order to realize online monitoring of the welding process, the prior art is mainly divided into two types, namely a method based on static image analysis, namely, shooting a single photo of a welding pool or a welding line, and judging whether defects exist in the photo by using a deep learning model. Such methods are essentially "post-detection", which can only be identified after a defect has formed and is visible to the naked eye, with a serious lag in early warning. At the same time, training such models requires a large number of annotated defect pictures, which are difficult to obtain in production practice. The second type is based on single signal threshold alerting methods, such as monitoring whether the welding current exceeds a certain fixed value. This method, while simple and real-time, is very prone to false positives due to normal process fluctuations and is unable to identify early, minor anomalies that do not cause severe current changes. In recent years, there have been studies attempting to integrate a plurality of sensor information, but it is common to simply combine several signals and input them into a complex supervised classification model. The method still cannot get rid of dependence on a defect sample, and has the defects of complex model and long calculation time consumption, and is difficult to meet the severe requirements of industrial sites on real-time performance. Therefore, a new method for intelligent monitoring of a welding process, which can early warn in advance, does not need a defect sample, has strong adaptability and high instantaneity, is needed in industrial production, so as to truly realize the transition from 'passive detection' to 'active defense'. Disclosure of Invention Based on the prior art, the invention provides a welding process abnormality early warning method based on multi-source time sequence signal dynamic analysis, and early detection and early warning of process abnormal states are realized by analyzing multi-channel dynamic time sequence signals in the welding process. The invention provides a welding process abnormity early warning method based on multi-source time sequence signal dynamic analysis, which comprises the following steps of firstly, synchronously collecting multi-channel time sequence signals; Step two, data segmentation and pretreatment of a sliding window; step three, extracting and fusing time sequence features; step four, constructing a reference model offline; step five, online abnormal scoring and dynamic threshold early warning; Step six, early warning response and data tracing; and step seven, verifying the validity. Preferably, in the first step, the sensor and the acquisition card are used to synchronously acquire the welding current instantaneous waveform, the arc voltage instantaneous waveform, and the sequence of the molten pool size changing with time calculated in real time by the vision sensor. Preferably, in the second step, the multichannel signal data of the latest period of time is intercepted at a fixed period to form a data segment, and the normalization and the filtering are performed. Preferably, in the third step, mathematical features describing dynamic changes of each channel signal in the data segment are extracted respectively, and all the features are combined into a comprehensive feature vector to represent the overall state of the welding process in the current time period. Preferably, the specific steps of the time sequence feature extraction flow are as follows: S31, data preparation, namely intercepting original time sequence data of a fixed-time-length sliding window, wherein the original time sequence data comprises a current sequence, a voltage sequence and a molten pool width sequence. And S32, extracting characteristics of current and voltage signals, respectively standardizing the current and voltage sequences, decomposing the signals into a plurality of eigen mode functions by using an empirical mode decomposition algorithm, selecting the first 4 components, calculating the average v