Welding process abnormity early warning method based on multi-source time sequence signal dynamic analysis
By using multi-channel time-series signal analysis and a deep autoencoder model, the problem of real-time monitoring during the welding process was solved, enabling early warning of anomalies, reducing implementation costs, and improving the system's adaptability and real-time performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- EAST CHINA UNIV OF SCI & TECH
- Filing Date
- 2026-01-31
- Publication Date
- 2026-05-12
AI Technical Summary
Existing welding quality control methods cannot achieve real-time monitoring, relying on post-weld inspection or human experience, resulting in high costs for defective products to leave the factory or for rework. Furthermore, existing multi-sensor fusion methods rely on defect samples and are computationally time-consuming, making it difficult to meet real-time requirements.
By synchronously acquiring multi-channel time-series signals, performing sliding window data segmentation and preprocessing, extracting and fusing time-series features, constructing a benchmark model offline, and implementing online anomaly scoring and dynamic threshold early warning, a deep autoencoder model is used to analyze multi-source time-series signals during the welding process, enabling early anomaly detection and warning.
It enables early warning before defects appear on the surface, reduces implementation costs, improves the system's adaptability and real-time performance, reduces false alarms, and provides reliable data traceability.
Smart Images
Figure CN122020474A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial automation welding technology, and in particular to a welding process anomaly early warning method based on dynamic analysis of multi-source time-series signals. Background Technology
[0002] Welding is a critical process in equipment manufacturing, and its quality directly affects product safety. Currently, there are two main challenges in welding quality control: First, relying on post-weld non-destructive testing (such as X-rays and ultrasound) makes it impossible to detect and intervene in problems in real time during the welding process, resulting in defective products being shipped out or high rework costs; second, relying on operators' experience for online judgment makes it difficult to achieve stable and consistent automated production.
[0003] To achieve online monitoring of the welding process, existing technologies mainly fall into two categories: The first category is based on static image analysis, which involves taking a single photograph of the weld pool or weld seam and using a deep learning model to determine if defects exist in the photograph. This type of method is essentially "post-event detection," only identifying defects after they have formed and are visible to the naked eye, resulting in a significant delay in early warning. Furthermore, training such models requires a large number of labeled defect images, which are difficult to obtain in production practice. The second category is based on single signal threshold alarms, such as monitoring whether the welding current exceeds a certain fixed value. While this method is simple and real-time, it is very prone to false alarms due to normal process fluctuations and cannot identify early, minor anomalies that do not cause drastic changes in current.
[0004] In recent years, some studies have attempted to fuse information from multiple sensors, but these typically involve simply combining several signals and inputting them into a complex supervised classification model. This approach still cannot escape its dependence on defect samples, and the models are complex and computationally time-consuming, making it difficult to meet the stringent real-time requirements of industrial settings.
[0005] Therefore, there is an urgent need in industrial production for a new intelligent monitoring method for welding processes that can provide early warnings, does not require defect samples, is highly adaptable and real-time, so as to truly realize the transformation from "passive detection" to "active defense". Summary of the Invention
[0006] Based on existing technical problems, this invention proposes a welding process anomaly early warning method based on dynamic analysis of multi-source time-series signals. By analyzing multi-channel dynamic time-series signals in the welding process, early detection and early warning of abnormal process states can be achieved.
[0007] The present invention proposes a welding process anomaly early warning method based on dynamic analysis of multi-source time-series signals, including step one: synchronous acquisition of multi-channel time-series signals;
[0008] Step 2: Sliding window data segmentation and preprocessing;
[0009] Step 3: Temporal Feature Extraction and Fusion;
[0010] Step 4: Offline construction of the baseline model;
[0011] Step 5: Online anomaly scoring and dynamic threshold early warning;
[0012] Step Six: Early Warning Response and Data Traceability;
[0013] Step 7: Validation.
[0014] Preferably, in step one, the instantaneous waveform of welding current, the instantaneous waveform of arc voltage, and the sequence of changes in the molten pool size over time, calculated in real time by a vision sensor, are simultaneously acquired using sensors and a data acquisition card.
[0015] Preferably, in step two, the latest multi-channel signal data is extracted at fixed intervals to form data segments, which are then standardized and filtered.
[0016] Preferably, in step three, mathematical features describing the dynamic changes of each channel signal in the data segment are extracted, and all features are merged into a comprehensive feature vector to characterize the overall state of the welding process in the current time period.
[0017] Preferably, the specific steps of the temporal feature extraction process are as follows:
[0018] S31. Data preparation: Extract raw time series data for a fixed-duration sliding window, including current sequence, voltage sequence, and melt pool width sequence.
[0019] S32. Current and voltage signal feature extraction: The current and voltage sequences are standardized respectively. The empirical mode decomposition algorithm is used to decompose the signal into multiple intrinsic mode functions. The first four components are selected, and the normalized energy and the average value of the instantaneous frequency of each component are calculated. Each electrical signal has 8-dimensional features, for a total of 16 dimensions.
[0020] S33. Visual signal feature extraction of the molten pool: The molten pool width sequence is smoothed and filtered, its eight time-domain statistical features are calculated, the sequence is fitted using a second-order autoregressive model, two model coefficients are extracted, and a total of 10-dimensional features are obtained.
[0021] S34. Feature Fusion and Standardization: The 16-dimensional electrical signal features and the 10-dimensional visual signal features are concatenated to obtain a 26-dimensional fused feature vector, which is then standardized as a whole.
[0022] Preferably, in step four, a large amount of normal welding process data is collected and processed to obtain a normal feature vector sample library. The sample library is used to train a deep autoencoder model to reconstruct normal features, thereby establishing a benchmark for the normal welding process.
[0023] Preferably, the training steps of the baseline model (deep autoencoder) are as follows:
[0024] S41. Model Structure Construction: Construct a symmetrical deep autoencoder network. The encoder structure is: input layer → fully connected layer → fully connected layer → hidden layer.
[0025] S42. Model Training: Use all normal feature vectors extracted from the normal welding process database as the training set, and train the model using mean squared error as the loss function.
[0026] S43. Initial threshold determination: Input all samples from the training set into the trained model, calculate its reconstruction error, statistically analyze the distribution of all error values, and take the 95th percentile of all error values as the offline reference threshold.
[0027] After online monitoring is started, when the historical abnormal score queue has not reached the preset length, this offline reference threshold is used to issue an early warning. Once the queue is full, it will automatically switch to using the dynamic threshold.
[0028] Preferably, in step five, during real-time monitoring, the feature vector generated in step three is input into the trained benchmark model, and its reconstruction error is calculated as the anomaly score. At the same time, the warning threshold is dynamically calculated based on the recent historical anomaly scores. When the real-time anomaly score exceeds the dynamic threshold, an early warning is immediately triggered.
[0029] Preferably, the specific steps of the online real-time monitoring and dynamic early warning are as follows:
[0030] S51. Real-time feature calculation: Extract the latest data at fixed intervals and calculate the current standardized fusion feature vector according to the time-series feature extraction process.
[0031] S52, Anomaly Score Calculation: Input the current feature vector into the deployed deep autoencoder and calculate its reconstruction error as the anomaly score.
[0032] S53, Dynamic threshold update: The system maintains a fixed-length queue of historical abnormal scores and adds new scores to the queue.
[0033] S54. Warning Judgment and Execution: Obtain the corresponding warning threshold based on the current system status, compare the current abnormal score with the threshold, and if the abnormal score exceeds the threshold, immediately trigger the warning and save the process data before and after the warning time. If the threshold is not exceeded, continue to the next round of monitoring cycle.
[0034] Preferably, in step six, when the warning is triggered, an audible and visual alarm is immediately activated, and all original data within the key time period before and after the triggering moment is automatically saved to form a data snapshot for traceability analysis.
[0035] The beneficial effects of this invention are as follows:
[0036] 1. Early warning time: By analyzing the dynamic signals of the process, abnormalities can be detected before defects appear on the surface of the weld, changing "post-event detection" to "in-process early warning", which provides the possibility for timely adjustment of the process and avoidance of defects.
[0037] 2. No defect samples required: The model can establish a benchmark by learning from a large amount of normal data, which perfectly solves the core problem of scarce and difficult-to-collect defect samples in industrial scenarios and greatly reduces implementation costs.
[0038] 3. Adaptive anti-false alarm: It adopts dynamic early warning threshold, which can automatically adapt to the inherent normal fluctuations of different workpieces and different welding stages, significantly reducing false alarms caused by changes in working conditions, and has high reliability.
[0039] 4. High real-time performance: The entire process algorithm is highly efficient, and the process from data acquisition to judgment can be completed within hundreds of milliseconds, which fully meets the online monitoring requirements of high-speed welding.
[0040] 5. Good traceability: When an alarm is triggered, a complete snapshot of the process data is automatically saved, providing engineers with valuable first-hand information for analyzing the root cause of the problem and optimizing process parameters. Attached Figure Description
[0041] Figure 1 This is a flowchart illustrating the overall process of an early warning method for welding process anomalies based on dynamic analysis of multi-source time-series signals proposed in this invention.
[0042] Figure 2 This is a flowchart of the timing feature extraction and fusion process for a welding process anomaly early warning method based on dynamic analysis of multi-source timing signals proposed in this invention.
[0043] Figure 3 This is a schematic diagram of the baseline model structure of a deep autoencoder for a welding process anomaly early warning method based on dynamic analysis of multi-source time-series signals proposed in this invention.
[0044] Figure 4 This is a schematic diagram of the dynamic threshold calculation and early warning judgment process of a welding process anomaly early warning method based on multi-source time-series signal dynamic analysis proposed in this invention.
[0045] Figure 5 This diagram illustrates the comparison of the early warning timeliness of a welding process anomaly early warning method based on dynamic analysis of multi-source time-series signals proposed in this invention. Detailed Implementation
[0046] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0047] Reference Figures 1-5 A welding process anomaly early warning method based on dynamic analysis of multi-source time-series signals is proposed, including step one: synchronous acquisition of multi-channel time-series signals. Before the welding process, a welding signal synchronous acquisition system is set up, and synchronously triggered current sensors, voltage sensors, laser vision sensors, and high-speed synchronous acquisition cards are configured. Then, during welding, the sensors and acquisition cards are used to synchronously acquire the instantaneous waveforms of welding current and arc voltage, as well as the sequence data of the change of molten pool size (such as width) over time calculated in real time by the vision sensor, forming a "normal welding process database". This database should cover welding process data under all acceptable and defect-free process parameter fluctuation ranges to enhance the generalization ability of the model.
[0048] Step 2: Sliding window data segmentation and preprocessing: Extract the latest multi-channel signal data at fixed intervals to form data segments, and then perform standardization and filtering.
[0049] Step 3: Timing Feature Extraction and Fusion: Extract mathematical features describing the dynamic changes of each channel signal in the data segment, and merge all features into a comprehensive feature vector to characterize the overall state of the welding process in the current time period. The timing feature extraction process is executed using the "Normal Welding Process Database".
[0050] The specific steps of the time series feature extraction process are as follows:
[0051] S31. Data preparation: Extract raw time series data for a fixed-duration sliding window, including current sequence, voltage sequence, and melt pool width sequence.
[0052] S32. Current and Voltage Signal Feature Extraction: The current and voltage sequences are standardized respectively; the Empirical Mode Decomposition (EMD) algorithm is used to decompose the signal into multiple intrinsic mode functions (IMFs); the first four components are selected, and the normalized energy and the average value of the instantaneous frequency of each component are calculated. Each electrical signal yields 8-dimensional features, for a total of 16 dimensions.
[0053] S33. Melt pool visual signal feature extraction: The melt pool width sequence is smoothed and filtered; its eight time-domain statistical features are calculated; the sequence is fitted using a second-order autoregressive (AR) model, and two model coefficients are extracted. A total of 10-dimensional features are obtained.
[0054] S34. Feature Fusion and Standardization: The 16-dimensional electrical signal features and the 10-dimensional visual signal features are concatenated to obtain a 26-dimensional fused feature vector. This vector is then standardized as a whole.
[0055] like Figure 2 As shown, a standardized parallel processing flow was established to ensure the standardization, repeatability, and efficiency of feature extraction. This flow clearly defines each step, starting with the original current, voltage, and melt pool width time-series signals, and proceeding through independent channels for signal standardization, time-frequency analysis (for electrical signals) or time-domain / time-series analysis (for visual signals), ultimately fusing them into a unified feature vector through splicing and standardization. This design is key to the method's ability to adaptively extract highly discriminative process features.
[0056] Step 4: Offline Construction of the Benchmark Model: Collect a large amount of normal welding process data and process it to obtain a normal feature vector sample library. Use this sample library to train a deep autoencoder model, enabling it to accurately reconstruct normal features, thereby establishing a benchmark for the "normal welding process". Calculate the reconstruction error of all training samples and use its 95th percentile as the initial warning threshold.
[0057] Training steps for the baseline model (deep autoencoder):
[0058] S41. Model Structure Construction: Construct a symmetric deep autoencoder network. The encoder structure is: Input layer (26-dimensional) → Fully connected layer (18-dimensional) → Fully connected layer (12-dimensional) → Hidden layer (8-dimensional). The decoder part is symmetric.
[0059] S42. Model Training: Use all normal feature vectors extracted from the "Normal Welding Process Database" as the training set, and train the model using mean squared error as the loss function.
[0060] S43. Initial Threshold Determination: Input all samples from the training set into the trained model and calculate its reconstruction error. Statistically analyze the distribution of all error values and take the 95th percentile of all error values as the offline reference threshold. After online monitoring is started, when the historical abnormal score queue has not reached the preset length, this offline reference threshold is used for early warning; once the queue is full, it automatically switches to using a dynamic threshold.
[0061] like Figure 3As shown, step four selects a deep autoencoder as the core unsupervised baseline model. This model consists of a symmetric encoder and decoder. The encoder is responsible for progressively compressing the high-dimensional fused feature vector into a low-dimensional "latent space" through multiple nonlinear transformations, learning and extracting the core and essence of the dynamic patterns of the normal welding process in the process. The decoder attempts to perfectly reconstruct the original high-dimensional input features from this low-dimensional latent space representation. The model is trained on a large amount of normal data, with the goal of minimizing the reconstruction loss of normal data. Therefore, when a feature vector representing an abnormal process is input, because its pattern deviates from the learned normal distribution, the model cannot accurately reconstruct it, resulting in a significantly increased reconstruction error, which is used as the anomaly score. This mechanism allows the model to achieve anomaly detection without any defect samples. The trained deep autoencoder model, feature extraction program, and initial threshold are integrated and deployed into the control system of the welding robot.
[0062] Step 5: Online Anomaly Scoring and Dynamic Threshold Early Warning: During real-time monitoring, the feature vector generated in Step 3 is input into the trained baseline model, and its reconstruction error is calculated as the anomaly score. Simultaneously, an early warning threshold is dynamically calculated based on recent historical anomaly scores. An early warning is immediately triggered when the real-time anomaly score exceeds this dynamic threshold.
[0063] Specific steps for online real-time monitoring and dynamic early warning:
[0064] S51. Real-time feature calculation: Extract the latest data at fixed intervals and calculate the current standardized fusion feature vector according to the "time series feature extraction process".
[0065] S52, Anomaly Score Calculation: Input the current feature vector into the deployed deep autoencoder and calculate its reconstruction error as the anomaly score.
[0066] S53. Dynamic Threshold Update: The system maintains a fixed-length queue of historical abnormal scores. New scores are added to the queue. The mean and standard deviation of all scores in the current queue are calculated. Dynamic warning threshold = mean + N × standard deviation, where N is a pre-set positive coefficient (e.g., a value of 2), and the coefficient N can be adjusted according to specific requirements for false alarm rate and false negative rate.
[0067] S54. Early Warning Judgment and Execution: Based on the current system status (whether the historical score queue is full), obtain the corresponding early warning threshold (offline reference threshold or dynamic early warning threshold). Compare the current abnormal score with this threshold. If the abnormal score exceeds the threshold, an early warning is triggered immediately, and the process data before and after the early warning time is saved. If the threshold is not exceeded, continue to the next round of monitoring.
[0068] In addition, during actual monitoring, the system executes the following steps in a fixed frequency (e.g., 10Hz): synchronously collect a data segment of a fixed duration (e.g., 5 seconds) → execute the feature extraction process to obtain the fused feature vector → input the model to calculate the anomaly score → update or calculate the currently applicable warning threshold based on the status of the historical score queue → determine whether the anomaly score exceeds the current warning threshold.
[0069] like Figure 4 As shown, to effectively reduce false alarms caused by different workpieces, different welding stages, or fluctuations in normal process parameters, step five establishes an adaptive dynamic threshold early warning mechanism with complete calculation and decision-making logic. The core of this mechanism is that the early warning threshold is not a pre-set fixed value, but rather based on a real-time updated, fixed-length historical anomaly score window. Statistical indicators (such as moving average and moving standard deviation) are calculated, and this dynamically generates an early warning threshold that changes over time. This design allows the system to automatically adapt to the inherent background noise level of the welding process, increasing the threshold to reduce false alarms during stable periods and decreasing the threshold to maintain sensitivity during fluctuating periods, thereby significantly improving the system's robustness and practicality.
[0070] Step Six: Early Warning Response and Data Traceability: When an early warning is triggered, an audible and visual alarm will be activated immediately, and all original data within the key time period before and after the triggering moment will be automatically saved to form a data snapshot for traceability analysis. If the abnormal score exceeds the current early warning threshold, a real-time early warning will be triggered, saving a snapshot of the process data before and after the early warning moment, and an audible and visual alarm will be activated on the operation interface.
[0071] Step 7: Validation. For example... Figure 5 As shown in the figure, to intuitively demonstrate the advantage of the present invention in terms of early warning timeliness compared to traditional methods, a comparison can be made: The figure illustrates the process from the generation of a welding defect to its manifestation, comparing the early warning time of the present invention based on process dynamics with the alarm time of the traditional method based on image recognition. It can be seen that the present invention can issue an early warning before the defect fully manifests, gaining valuable intervention time.
[0072] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A welding process anomaly early warning method based on dynamic analysis of multi-source time-series signals, characterized in that: This includes step one: synchronous acquisition of multi-channel timing signals; Step 2: Sliding window data segmentation and preprocessing; Step 3: Temporal Feature Extraction and Fusion; Step 4: Offline construction of the baseline model; Step 5: Online anomaly scoring and dynamic threshold early warning; Step Six: Early Warning Response and Data Traceability; Step 7: Validation.
2. The welding process anomaly early warning method based on dynamic analysis of multi-source time-series signals according to claim 1, characterized in that: In step one, the instantaneous waveforms of welding current and arc voltage are simultaneously acquired using sensors and a data acquisition card, as well as the sequence of changes in the molten pool size over time calculated in real time by a visual sensor.
3. The welding process anomaly early warning method based on dynamic analysis of multi-source time-series signals according to claim 2, characterized in that: In step two, multi-channel signal data of the latest time period is extracted at a fixed period to form data segments, which are then standardized and filtered.
4. The welding process anomaly early warning method based on multi-source time-series signal dynamic analysis according to claim 3, characterized in that: In step three, mathematical features describing the dynamic changes of each channel signal in the data segment are extracted, and all features are merged into a comprehensive feature vector to characterize the overall state of the welding process in the current time period.
5. The welding process anomaly early warning method based on dynamic analysis of multi-source time-series signals according to claim 4, characterized in that: The specific steps of the temporal feature extraction process are as follows: S31. Data preparation: Extract raw time series data for a fixed-duration sliding window, including current sequence, voltage sequence, and melt pool width sequence. S32. Current and voltage signal feature extraction: The current and voltage sequences are standardized respectively. The empirical mode decomposition algorithm is used to decompose the signal into multiple intrinsic mode functions. The first four components are selected, and the normalized energy and the average value of the instantaneous frequency of each component are calculated. Each electrical signal has 8-dimensional features, for a total of 16 dimensions. S33. Visual signal feature extraction of the molten pool: The molten pool width sequence is smoothed and filtered, its eight time-domain statistical features are calculated, the sequence is fitted using a second-order autoregressive model, two model coefficients are extracted, and a total of 10-dimensional features are obtained. S34. Feature Fusion and Standardization: The 16-dimensional electrical signal features and the 10-dimensional visual signal features are concatenated to obtain a 26-dimensional fused feature vector, which is then standardized as a whole.
6. The welding process anomaly early warning method based on dynamic analysis of multi-source time-series signals according to claim 5, characterized in that: In step four, a large amount of normal welding process data is collected and processed to obtain a normal feature vector sample library. The sample library is used to train a deep autoencoder model to reconstruct normal features, thereby establishing a benchmark for the normal welding process.
7. The welding process anomaly early warning method based on dynamic analysis of multi-source time-series signals according to claim 6, characterized in that: The training steps for the baseline model (deep autoencoder) are as follows: S41. Model Structure Construction: Construct a symmetrical deep autoencoder network, wherein the encoder structure is: input layer → fully connected layer → fully connected layer → hidden layer; S42, Model Training: Use all normal feature vectors extracted from the normal welding process database as the training set, and train with mean squared error as the loss function; S43. Initial threshold determination: Input all samples from the training set into the trained model, calculate its reconstruction error, statistically analyze the distribution of all error values, and take the 95th percentile of all error values as the offline reference threshold. After online monitoring is started, when the historical abnormal score queue has not reached the preset length, this offline reference threshold is used to issue an early warning. Once the queue is full, it will automatically switch to using a dynamic threshold.
8. The welding process anomaly early warning method based on dynamic analysis of multi-source time-series signals according to claim 7, characterized in that: In step five, during real-time monitoring, the feature vector generated in step three is input into the trained benchmark model, and its reconstruction error is calculated as the anomaly score. At the same time, the warning threshold is dynamically calculated based on the recent historical anomaly scores. When the real-time anomaly score exceeds the dynamic threshold, an warning is immediately triggered.
9. The welding process anomaly early warning method based on dynamic analysis of multi-source time-series signals according to claim 8, characterized in that: The specific steps for online real-time monitoring and dynamic early warning are as follows: S51. Real-time feature calculation: Extract the latest data at fixed intervals and calculate the current standardized fusion feature vector according to the time-series feature extraction process; S52, Anomaly score calculation: Input the current feature vector into the deployed deep autoencoder and calculate its reconstruction error as the anomaly score; S53, Dynamic threshold update; The system maintains a fixed-length queue of historical abnormal scores and adds new scores to the queue; S54. Warning Judgment and Execution: Obtain the corresponding warning threshold based on the current system status, compare the current abnormal score with the threshold, and if the abnormal score exceeds the threshold, immediately trigger the warning and save the process data before and after the warning time. If the threshold is not exceeded, continue to the next round of monitoring cycle.
10. The welding process anomaly early warning method based on multi-source time-series signal dynamic analysis according to claim 9, characterized in that: In step six, when the warning is triggered, an audible and visual alarm is immediately activated, and all original data within the key time period before and after the triggering moment is automatically saved to form a data snapshot for traceability analysis.