Abnormity identification method and system in ultrasonic welding process
By employing time-series reconstruction and multi-dimensional error feature extraction, this method addresses the issues of insufficient dynamic feature capture, limited feature extraction, and noise interference in the welding process in existing technologies, achieving high accuracy and low over-alarm welding quality detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-31
- Publication Date
- 2026-03-27
AI Technical Summary
Existing technologies cannot effectively capture the dynamic characteristics of welding power time series, resulting in insufficient identification of hidden welding defects. The limitations of feature extraction lead to low detection accuracy. High-frequency noise masks abnormal signals, and the simplistic anomaly judgment logic leads to excessive alarms.
A time-series reconstruction model is used to reconstruct power time-series data, generating reconstructed time-series data. Multi-dimensional error feature vectors are extracted, and combined with wavelet denoising and adaptive threshold generation mechanisms, welding quality anomalies are determined through multi-dimensional error feature vectors.
It achieves accurate identification of hidden welding defects, with a detection accuracy rate of over 90.2%, a 30% increase in sensitivity for identifying minor anomalies, a 40% reduction in the over-alarm rate, and stable detection performance under complex interference.
Smart Images
Figure CN121732969A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of lithium ion battery cell manufacturing, and particularly relates to an ultrasonic welding process anomaly identification method and system. BACKGROUND
[0002] In the lithium ion battery cell manufacturing process, the ultrasonic welding quality of the tab directly determines the electrical performance, safety performance and service life of the battery cell. The current industry monitoring of the ultrasonic welding process mainly relies on two types of technical solutions: Traditional physical parameter monitoring scheme: through sensors, discrete physical parameters such as amplitude, pressure, welding time, etc. in the welding process are collected, and fixed threshold values are set to judge whether the welding is normal. This type of scheme only focuses on a single or a small number of discrete parameters, and cannot capture the dynamic characteristics of the welding power changing over time, and is prone to misjudgment of "qualified parameters but welding failure".
[0003] Traditional machine learning detection scheme: some technologies use machine learning algorithms such as support vector machine (SVM) and random forest to extract features (such as mean, variance, peak, etc. statistical features) from the welding power time series, and then use a classification model to judge the anomaly. However, this type of scheme relies on manual feature design and cannot mine hidden nonlinear dynamic correlations (such as time series correlation of power fluctuations, mutation patterns) in the time series. When the welding process is disturbed by complex factors such as incoming material differences (such as tab thickness deviation) and equipment aging, the detection accuracy decreases significantly.
[0004] Deep learning preliminary application scheme: in recent years, the industry has begun to try to use deep learning for welding quality detection, such as using convolutional neural network (CNN) for one-dimensional convolution operation on power time series data. However, this type of scheme does not reconstruct the time series, and directly inputs the original time series data into the model, which is easily affected by high-frequency noise (such as sensor electromagnetic interference), resulting in insufficient sensitivity of the model to slight power anomalies (such as slow drift of power caused by minor short circuits).
[0005] Problems in the prior art include: Insufficient dynamic feature capture: the traditional physical parameter monitoring scheme only focuses on discrete parameters and ignores the dynamic changes of the welding power time series (such as the power ramp-up rate in the welding start-up stage and the fluctuation frequency in the stable stage), and cannot identify hidden welding defects (such as virtual welding and local overwelding) that have "parameters within the threshold but abnormal time series patterns".
[0006] Feature extraction limitations: the traditional machine learning scheme relies on manual design of statistical features and cannot automatically mine deep nonlinear correlations (such as the hidden mapping relationship between power mutation and tab oxidation) in the time series data. In the presence of multiple interference factors (such as incoming material impurities and equipment temperature drift), the feature discrimination is low and the false detection rate is as high as 15%-20%.
[0007] Weak noise interference resistance: existing deep learning solutions directly use raw power time series data without effective time series reconstruction and noise reduction processing. High-frequency noise (such as 50Hz power grid interference) introduced by sensors can mask the real power anomaly signal, making it difficult to identify slight anomalies (such as 5%-8% deviation in welding power).
[0008] Single abnormality determination logic: existing solutions mostly use "model direct classification" to determine anomalies, lack quantitative analysis of "reconstructed time series and original time series difference", and cannot distinguish between "equipment accidental fluctuations" and "persistent welding defects", which can easily lead to excessive alarms (such as single electromagnetic interference being determined as a welding anomaly), increasing production line downtime costs.
[0009] To solve the above technical problems, the present application provides an ultrasonic welding process anomaly identification method and system. SUMMARY
[0010] To solve the above problems, the present application provides an ultrasonic welding process anomaly identification method, which comprises: Collecting power time series data during the welding process of the to-be-detected battery cell; Pretreating the power time series data to obtain pretreated time series data; Reconstructing the pretreated time series data using a time series reconstruction model to generate reconstructed time series data; Based on the reconstructed time series data and the pretreated time series data, a multi-dimensional error feature vector is extracted; According to the multi-dimensional error feature vector, the welding quality anomaly is determined.
[0011] According to some embodiments of the present disclosure, the power time series data is pretreated by sequentially performing the following operations: Wavelet threshold denoising is used to remove high-frequency electromagnetic interference; Aligning the time series data based on the welding start time; Normalizing the aligned time series data.
[0012] According to some embodiments of the present disclosure, the time series reconstruction model is composed of a local convolution layer and an improved Transformer encoder; The local convolution layer includes a plurality of convolution blocks; The improved Transformer encoder includes multiple layers of encoders, each layer of encoder being provided with a time series attention mask and a feedback neural network.
[0013] According to some embodiments of the present disclosure, the time series reconstruction model is obtained by loss function optimization training; The expression for the loss function L is: ; in, This is the original power timing data; To reconstruct power timing data; N is the number of sampling points.
[0014] According to some embodiments of this disclosure, the extraction of multi-dimensional error feature vectors includes: Acquire the power data to be detected; Based on the power data to be detected and the reconstructed power time series data, determine the mean square error and the mean absolute error; A power error sequence is generated based on the overall error characteristics, and the peak value is extracted by first-order difference of the power error sequence to determine the peak value of the absolute difference. Linear fitting was performed on the reconstructed time series data and the preprocessed time series data respectively to determine the slope difference; By standardizing the mean square error, mean absolute error, peak value of absolute difference, and slope difference, a multi-dimensional error feature vector is obtained.
[0015] According to some embodiments of this disclosure, the expression for determining the peak value of the absolute difference is: ; Wherein, PEAK represents the peak value of the absolute difference; This represents the power error value at the Nth sampling point.
[0016] According to some embodiments of this disclosure, the expression for determining the slope difference is: ; Where SLOPE_DIFF represents the slope difference; The slope of the fitted original power time series is represented. This represents the slope of the reconstructed power time series.
[0017] According to some embodiments of this disclosure, the determination of abnormal welding quality includes: Based on the error feature vector of historical normal power data, kernel density estimation is used to generate an initial threshold; The initial threshold is adjusted in real time based on the multi-dimensional error feature vector of the battery cell to be tested. If multiple dimensions of the multi-dimensional error feature vector of the battery cell to be tested are greater than the adjusted threshold, it is determined to be a persistent anomaly. If only one dimension is greater than the adjusted threshold but less than a set multiple of the adjusted threshold, it is judged as an accidental fluctuation.
[0018] According to some embodiments of this disclosure, after determining the welding quality abnormality based on multi-dimensional error characteristics, the method further includes: Real-time output of welding quality anomaly judgment results; When the judgment result is a continuous abnormality, an alarm is triggered and the welding station is suspended.
[0019] This application also provides an ultrasonic welding process anomaly identification system, the system comprising: The acquisition module is used to acquire power timing data during the welding process of the battery cell under test; The preprocessing module is used to preprocess the power timing data to obtain preprocessed timing data; The reconstruction module is used to reconstruct preprocessed time series data using a time series reconstruction model, generating reconstructed time series data. The extraction module is used to extract multi-dimensional error feature vectors based on the reconstructed time series data and the preprocessed time series data; The judgment module is used to determine welding quality abnormalities based on multi-dimensional error feature vectors.
[0020] Compared with the prior art, this application has the following advantages: (1) This application achieves accurate identification of hidden welding defects (such as cold welds and local over-welding) through time series reconstruction and multi-dimensional error feature extraction, improving the detection accuracy to over 90.2%, which is 15%-20% higher than traditional machine learning solutions; (2) The combination of wavelet denoising and time-series reconstruction in this application effectively filters out high-frequency noise and improves the sensitivity of identifying slight power anomalies (deviation of 5%-8%) by more than 30%, avoiding the risk of "missed detection". (3) The optimized processing of the single sensor power timing and the adaptive threshold adjustment in this application can cope with complex interferences such as material differences (e.g., tab thickness deviation ±0.02mm) and equipment aging (e.g., amplitude attenuation 10%), and maintain stable detection performance under different production conditions. (4) Using this application, the excessive alarm rate is reduced by more than 40%, reducing unnecessary production line downtime; at the same time, it enables the tracing of the root cause of abnormalities (locating the type of abnormality through the error feature dimension), helping engineers to quickly troubleshoot problems (such as "PEAK over the limit" corresponding to electrode impurities), and improving maintenance efficiency by 50%; (5) This application can be directly transferred to single-sensor power detection in other ultrasonic welding scenarios (such as capacitor tab welding and semiconductor wire bonding). Only the training data of the corresponding scenario needs to be replaced. There is no need to make significant modifications to the model structure. It has a wide range of applications.
[0021] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1 A diagram of an ultrasonic welding process anomaly identification method according to an embodiment of this application is shown; Figure 2 A detailed flowchart of the ultrasonic welding process anomaly identification method according to an embodiment of this application is shown; Figure 3 A diagram of an ultrasonic welding process anomaly identification system according to an embodiment of this application is shown. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0025] The technical problem that this application aims to solve is as follows: Existing technologies cannot effectively capture the dynamic characteristics of welding power time series, thus hindering the accurate identification of hidden welding defects (such as incomplete welds and local over-welding). To address the issue of low detection accuracy in complex interference scenarios due to the limitations of manually designed features, this paper proposes to automatically extract deep nonlinear temporal features. High-frequency noise masks abnormal signals, thus improving the sensitivity to detect slight power anomalies; The problem of excessive alarms caused by a simplistic anomaly judgment logic is addressed by achieving accurate differentiation between "random fluctuations" and "persistent defects".
[0026] like Figure 1 As shown, this application provides a method for identifying anomalies in an ultrasonic welding process, the method comprising: Collect power timing data during the welding process of the battery cell under test; The power time series data is preprocessed to obtain preprocessed time series data; The time series reconstruction model is used to reconstruct the preprocessed time series data to generate reconstructed time series data. Based on the reconstructed time series data and the preprocessed time series data, multi-dimensional error feature vectors are extracted; Welding quality abnormalities are determined based on multi-dimensional error feature vectors.
[0027] Specifically, the preprocessing of the power timing data includes performing the following operations in sequence: Wavelet threshold denoising is used to remove high-frequency electromagnetic interference; The timing data is aligned based on the welding vibration initiation time. The aligned time-series data is then normalized.
[0028] Specifically, the temporal reconstruction model consists of local convolutional layers and an improved Transformer encoder; The local convolutional layer includes multiple convolutional blocks; The improved Transformer encoder includes a multi-layer encoder, each layer having a temporal attention mask and a feedback neural network.
[0029] Specifically, the temporal reconstruction model is obtained through training optimization using a loss function; The expression for the loss function L is: ; in, This is the original power timing data; To reconstruct power timing data; N is the number of sampling points.
[0030] Specifically, the extraction of multi-dimensional error feature vectors includes: Acquire the power data to be detected; Based on the power data to be detected and the reconstructed power time series data, determine the mean square error and the mean absolute error; A power error sequence is generated based on the overall error characteristics, and the peak value is extracted by first-order difference of the power error sequence to determine the peak value of the absolute difference. Linear fitting was performed on the reconstructed time series data and the preprocessed time series data respectively to determine the slope difference; By standardizing the mean square error, mean absolute error, peak value of absolute difference, and slope difference, a multi-dimensional error feature vector is obtained.
[0031] Specifically, the expression for determining the peak value of the absolute difference is: ; Wherein, PEAK represents the peak value of the absolute difference; This represents the power error value at the Nth sampling point.
[0032] Specifically, the expression for determining the slope difference is: ; Where SLOPE_DIFF represents the slope difference; The slope of the fitted original power time series is represented. This represents the slope of the reconstructed power time series.
[0033] Specifically, the determination of abnormal welding quality includes: Based on the error feature vector of historical normal power data, kernel density estimation is used to generate an initial threshold; The initial threshold is adjusted in real time based on the multi-dimensional error feature vector of the battery cell to be tested. If multiple dimensions of the multi-dimensional error feature vector of the battery cell to be tested are greater than the adjusted threshold, it is determined to be a persistent anomaly. If only one dimension is greater than the adjusted threshold but less than a set multiple of the adjusted threshold, it is judged as an accidental fluctuation.
[0034] Specifically, after determining the welding quality abnormality based on multi-dimensional error characteristics, the process further includes: Real-time output of welding quality anomaly judgment results; When the judgment result is a continuous abnormality, an alarm is triggered and the welding station is suspended.
[0035] like Figure 2 As shown, this application also provides an ultrasonic welding process anomaly identification system, the system comprising: The acquisition module is used to acquire power timing data during the welding process of the battery cell under test; The preprocessing module is used to preprocess the power timing data to obtain preprocessed timing data; The reconstruction module is used to reconstruct preprocessed time series data using a time series reconstruction model, generating reconstructed time series data. The extraction module is used to extract multi-dimensional error feature vectors based on the reconstructed time series data and the preprocessed time series data; The judgment module is used to determine welding quality abnormalities based on multi-dimensional error feature vectors.
[0036] like Figure 3As shown, this system includes a data acquisition module, a data preprocessing module, a time series reconstruction module, an error feature extraction module, an anomaly detection module, and a result output module. The overall technology includes a data acquisition module, a data preprocessing module, a time series reconstruction module, an error feature extraction module, an anomaly detection module, and a result output module.
[0037] 1. Data Acquisition Module Sensor selection: Only high-frequency current sensors (sampling frequency 1kHz) are used to collect ultrasonic welding power signals, focusing on the accurate acquisition of single sensor data; Data storage: The collected power time series data is indexed and stored according to "welding batch-cell number-welding time". The length of a single welding time series data is 500 sampling points (covering the three stages of welding start-up, stabilization and finish).
[0038] 2. Data Preprocessing Module Noise Removal: Wavelet threshold denoising (using db4 wavelet basis, decomposition level 5) is employed to remove high-frequency electromagnetic interference and enhance the noise suppression effect of single sensor data. Timing alignment: Based on the welding start-up time (the moment when the power first exceeds 5W), the power timing data of different cells are aligned to ensure consistent timing length; Normalization: Min-Max normalization is used to map the power data to the [0,1] interval, as shown in the formula:
[0039] in, This is the raw power data. , These are the minimum and maximum values of historical normal power data, respectively. This is the result after normalization.
[0040] 3. Timing Reconstruction Module Model Structure: Employing a hybrid architecture of "local convolution + improved Transformer," this model optimizes feature extraction capabilities for single-sensor power time-series analysis, including: a. Local convolutional layer: 3 convolutional blocks (kernel sizes of 3, 5, and 7, stride of 1) to enhance the accuracy of capturing short-term power fluctuation features; b. Improved Transformer encoder: 6-layer encoder, each layer containing a "temporal attention mask" (allowing attention only to the temporal sequence before the current time step) and a "feedforward neural network" to accurately model the long-term correlation of power time series; Model training: c. Training dataset: Collect 10,000 power time-series data points of normal welding (covering cells with different tab materials and thicknesses) to ensure the generalization of single-sensor data training; d. Loss function L: The mean squared error (MSE) loss is used, and the formula is:
[0041] in, This is the original power timing data. To reconstruct the power timing data, N is the number of sampling points; e. Training optimization: The Adam optimizer is used with an initial learning rate of 0.001, which decays by 10% every 100 epochs until the loss converges (fluctuation is less than 0.001).
[0042] 4. Error Feature Extraction Module Overall error characteristics: Calculate the mean square error (MSE) and mean absolute error (MAE) of the reconstructed power time series and the power time series to be detected. Mean Squared Error (MSE) formula: ,in The power data to be detected. To reconstruct the power data, N is the number of sampling points; The formula for Mean Absolute Error (MAE) is: |, among which The power data to be detected. To reconstruct the power data, N is the number of sampling points; Local mutation error characteristics: Perform first-order difference on the power error sequence and take the peak value (PEAK) of the absolute value of the difference as the local mutation index; First-order difference formula for error sequence: ,in For the first Power error value at each sampling point; Peak value of absolute difference (PEAK) formula: ; Trend error characteristics: Linear fitting was used to calculate the slope of the original power time series and the reconstructed power time series respectively, and the slope difference (SLOPE_DIFF) was taken as the trend deviation index; Linear fitting slope calculation: For power time series data Based on sampling time Using the least squares method to fit a straight line as the independent variable. , where the slope The calculation formula is:
[0043] Slope Difference (SLOPE_DIFF) Formula: ,in The slope of the fitted original power time series is denoted as . The slope for fitting the reconstructed power time series; Feature vector construction: The above four indicators (MSE, MAE, PEAK, SLOPE_DIFF) are standardized to form a 4-dimensional error feature vector. ; Standardized formula: ,in For the first One original index value, , These are the first in the historical normal data. The minimum and maximum values of each indicator.
[0044] 5. Anomaly Detection Module Adaptive threshold generation: Based on the error feature vector of 5000 historical normal power data, kernel density estimation (KDE) is used to model the distribution of each feature dimension, and the upper limit of the 99.5% confidence interval is taken as the initial threshold for that dimension. ; Dynamic threshold adjustment: Real-time statistics are performed on the average power error characteristics of the first 100 cells in each batch. If the average fluctuation exceeds 5%, the threshold is adjusted according to the following formula to adapt to differences in incoming materials or changes in equipment status: ; in, The adjusted threshold The threshold before adjustment, This represents the difference between the current batch average and the historical average. For symbolic functions ( Take 1 at time. (Take -1 at time). Anomaly detection rules: If any two or more dimensions in the error feature vector of the cell under test exceed the corresponding threshold, it is judged as "persistent anomaly"; if only one dimension exceeds the threshold and the value of that dimension does not exceed 1.2 times the threshold, it is judged as "random fluctuation" (no alarm is triggered).
[0045] 6. Result Output Module Real-time output: Display the welding status (normal / occasional fluctuation / continuous abnormality) of each cell on the production line display screen in real time, and mark the abnormal dimensions (such as "MSE, PEAK exceeds the limit"). Historical traceability: The original power timing sequence, reconstructed power timing sequence, and error feature vector of abnormal cells are stored in the database, supporting historical queries and traceability by batch and time. Alarm Trigger: When a "persistent abnormality" is detected, an audible and visual alarm is triggered, the corresponding welding station is automatically suspended, and the abnormality information is pushed to the production line management system.
[0046] This application proposes a two-stage detection framework of "time sequence reconstruction - error quantification": breaking through the traditional detection approach of "direct classification", it first reconstructs the normal welding power time sequence through a deep learning model, and then realizes the anomaly determination by quantifying the error between the "reconstructed time sequence and the time sequence to be detected", which greatly improves the interpretability and accuracy of anomaly identification. This application designs a timing reconstruction model based on an improved Transformer: targeting the long-term correlation characteristics of welding power timing, a "timing attention mask" (focusing only on timing information before the current moment) is introduced into the Transformer encoder, and combined with a "local convolution module" (capturing short-term power fluctuations) to achieve accurate modeling of the timing dynamics, reducing the reconstruction error by more than 30%. This application constructs a multi-dimensional error feature vector: breaking through the limitations of traditional single error indicators (such as mean square error), constructing an error feature vector from three dimensions: "overall error" (mean square error, mean absolute error), "local mutation error" (differential error peak), and "trend error" (time series slope deviation), to achieve comprehensive coverage of different types of anomalies (such as sudden short circuit, slow over-soldering). This application introduces an adaptive abnormal threshold generation mechanism: based on the error characteristic distribution of historical normal welding data, kernel density estimation (KDE) is used to dynamically generate an adaptive threshold to replace the fixed threshold. The judgment criteria can be automatically adjusted according to the difference in incoming materials and equipment status, reducing the excessive alarm rate by more than 40%. This application optimizes the single-sensor power timing processing flow: focusing on single-sensor ultrasonic power data, by enhancing the noise suppression capability and error feature discrimination design of the timing reconstruction model, it solves the problem of misjudgment caused by interference of a single sensor, and improves the detection accuracy to over 90.2%.
[0047] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A method for identifying anomalies in an ultrasonic welding process, characterized in that, The method includes: Collect power timing data during the welding process of the battery cell under test; The power time series data is preprocessed to obtain preprocessed time series data; The time series reconstruction model is used to reconstruct the preprocessed time series data to generate reconstructed time series data. Based on the reconstructed time series data and the preprocessed time series data, multi-dimensional error feature vectors are extracted; Welding quality abnormalities are determined based on multi-dimensional error feature vectors.
2. The ultrasonic welding process anomaly identification method according to claim 1, characterized in that, The preprocessing of the power time-series data includes performing the following operations in sequence: Wavelet threshold denoising is used to remove high-frequency electromagnetic interference; The timing data is aligned based on the welding vibration initiation time. The aligned time-series data is then normalized.
3. The method for identifying abnormalities in the ultrasonic welding process according to claim 1, characterized in that, The temporal reconstruction model consists of local convolutional layers and an improved Transformer encoder. The local convolutional layer includes multiple convolutional blocks; The improved Transformer encoder includes a multi-layer encoder, each layer having a temporal attention mask and a feedback neural network.
4. The ultrasonic welding process anomaly identification method according to claim 3, characterized in that, The temporal reconstruction model is obtained through training optimization using a loss function; The expression for the loss function L is: ； in, This is the original power timing data; To reconstruct power timing data; N is the number of sampling points.
5. The method for identifying abnormalities in the ultrasonic welding process according to claim 1, characterized in that, The extraction of multi-dimensional error feature vectors includes: Acquire the power data to be detected; Based on the power data to be detected and the reconstructed power time series data, determine the mean square error and the mean absolute error; A power error sequence is generated based on the overall error characteristics, and the peak value is extracted by first-order difference of the power error sequence to determine the peak value of the absolute difference. Linear fitting was performed on the reconstructed time series data and the preprocessed time series data respectively to determine the slope difference; By standardizing the mean square error, mean absolute error, peak value of absolute difference, and slope difference, a multi-dimensional error feature vector is obtained.
6. The ultrasonic welding process anomaly identification method according to claim 5, characterized in that, The expression for determining the peak value of the absolute difference is: ； Wherein, PEAK represents the peak value of the absolute difference; This represents the power error value at the Nth sampling point.
7. The ultrasonic welding process anomaly identification method according to claim 5, characterized in that, The expression for determining the slope difference is: ； Where SLOPE_DIFF represents the slope difference; The slope of the fitted original power time series is represented. This represents the slope of the reconstructed power time series.
8. The method for identifying abnormalities in the ultrasonic welding process according to claim 1, characterized in that, The determination of abnormal welding quality includes: Based on the error feature vector of historical normal power data, kernel density estimation is used to generate an initial threshold; The initial threshold is adjusted in real time based on the multi-dimensional error feature vector of the battery cell to be tested. If multiple dimensions of the multi-dimensional error feature vector of the battery cell to be tested are greater than the adjusted threshold, it is determined to be a persistent anomaly. If only one dimension is greater than the adjusted threshold but less than a set multiple of the adjusted threshold, it is judged as an accidental fluctuation.
9. The method for identifying abnormalities in the ultrasonic welding process according to claim 1, characterized in that, After determining welding quality abnormalities based on multi-dimensional error characteristics, the process also includes: Real-time output of welding quality anomaly judgment results; When the judgment result is a continuous abnormality, an alarm is triggered and the welding station is suspended.
10. An ultrasonic welding process anomaly identification system, characterized in that, The system includes: The acquisition module is used to acquire power timing data during the welding process of the battery cell under test; The preprocessing module is used to preprocess the power timing data to obtain preprocessed timing data; The reconstruction module is used to reconstruct preprocessed time series data using a time series reconstruction model, generating reconstructed time series data. The extraction module is used to extract multi-dimensional error feature vectors based on the reconstructed time series data and the preprocessed time series data; The judgment module is used to determine welding quality abnormalities based on multi-dimensional error feature vectors.