Intelligent online flash butt welding quality evaluation method

Abstract

The invention discloses an intelligent online flash butt welding quality evaluation method which comprises the following steps: dividing a welding time sequence according to the total number of stages of flash butt welding processing to obtain a time cluster, extracting in-channel features and inter-channel features from the time cluster, and constructing an initial statistical model and an unsupervised model; based on a statistical model, an unsupervised model and a deep learning model, an anomaly detection model is constructed, and weights of the three models are determined according to the positive sample number and the negative sample number of the welding time sequence; performing anomaly detection on data acquired by the sensor after flash butt welding; introducing a weight correction coefficient to obtain a final abnormal score; if the final exception score is greater than a set threshold value, adding an exception label; otherwise, adding a normal label; post-welding quality detection is conducted, and the correction coefficient is corrected according to a post-welding quality detection result; and judging that the added label is consistent with the detection quality result, if so, regularly training the anomaly detection model, and if not, re-training the deviation model.

Description

Technical Field

[0001] This invention relates to methods for detecting the quality of flash butt welding, and more particularly to an intelligent online evaluation method for the quality of flash butt welding. Background Technology

[0002] With the mature application of machine learning and deep learning technologies in industrial inspection, their nonlinear modeling and complex pattern recognition capabilities offer new solutions to the aforementioned problems. However, in the flash butt welding scenario of mooring chains, qualified welds account for over 99%, while defect samples such as incomplete penetration and microcracks are scarce. Traditional machine learning models are prone to overfitting, resulting in insufficient generalization ability for defect recognition. Therefore, addressing the common inspection needs of flash butt welding processes and considering the special service requirements of high-end components such as mooring chains, developing an intelligent online evaluation method that can adapt to complex working conditions and overcome the limitations of small sample sizes has become an urgent need to ensure the quality of high-end equipment manufacturing and promote the upgrading of welding inspection technology.

[0003] Chinese invention patent CN202010309872.1 discloses an "Online Detection Method for Mooring Chain Flash Welding Quality Based on Incremental Learning." This method addresses the modeling difficulties caused by sample imbalance through piecewise linear interpolation, normalization, and Euclidean distance. It employs incremental learning for continuous model training, effectively mitigating the model update problem caused by the increasing number of mooring chain flash welding samples as production cycles progress. It also avoids catastrophic forgetting of the model to some extent, demonstrating practical value. However, this technical solution still has the following shortcomings: when facing complex and variable multi-channel time-series signals, piecewise linear interpolation can smooth or weaken key anomalous changes, making it difficult for the model to effectively identify point anomalies, short-term mutations, or cross-channel coupling anomalies. Especially when multi-sensor time-series data exhibits nonlinear correlation characteristics, this solution does not consider intra-channel variations and inter-channel coupling relationships in multi-channel time-series signals, significantly limiting its anomaly identification capability.

[0004] Chinese patent CN117900687A discloses a "Method and System for Detecting Anomalies in Flash Welding of Mooring Chains." This method involves constructing a semi-supervised learning network model (GAN) for anomaly detection, determining the optimal hyperparameter model and the comprehensive anomaly score S3 and threshold score T, identifying fault samples, and using CAM visualization technology for anomaly localization. However, a drawback is that in the actual process of monitoring anomalies in flash welding of mooring chains, there is a typical problem of extreme sample imbalance, where qualified samples far outnumber abnormal samples. Directly applying semi-supervised learning can lead to the model further reinforcing incorrect decision boundaries in subsequent learning, causing model degradation. Furthermore, when facing the production of new specifications of mooring chains, sufficient sample data is required to train the model, making rapid deployment of the model for anomaly detection impossible. Summary of the Invention

[0005] Purpose of the invention: This invention aims to provide an online detection method for the quality of flash welding of mooring chains based on multi-model fusion, which involves retraining models through multiple methods and fusing multiple feature extractions. This method addresses issues such as insufficient sample data in new systems, severe imbalance in flash welding data of mooring chains, and model overfitting.

[0006] Technical solution: The intelligent online evaluation method for flash butt welding quality described in this invention includes the following steps:

[0007] (1) Create a new database of specification D mooring chains, accumulate F positive sample data, and calculate the missing values ​​of the F positive sample welding data of specification D mooring chains in turn by linear interpolation and fill them into their corresponding positions to obtain the welding time series of specification D mooring chains.

[0008] (2) According to the total number of flash butt welding stages n, the welding time sequence of the specification D mooring chain is divided into n time clusters;

[0009] (3) Extract intra-channel and inter-channel features from time clusters, and construct an initial statistical model and a One Class SVM unsupervised model based on the feature mean, variance, normal range and threshold.

[0010] (4) Based on the statistical model, the One-Class SVM unsupervised model and the deep learning model, an anomaly detection model is constructed, and the number of positive samples of the welding time series of the D-specification mooring chain is obtained according to step (1). and negative sample count The weights of the statistical model, the One Class SVM unsupervised model, and the deep learning model are determined; the welding time series of the D-specification mooring chain obtained in step (1) is divided into training set and validation set for model training, and anomaly detection is performed on the data collected by the sensor after flash welding.

[0011] (5) Input the data collected by the sensor in real time into the statistical model for anomaly detection, and obtain an anomaly score of 1. The data collected in real time by the sensor is input into the unsupervised learning model One Class SVM for anomaly detection, and an anomaly score of 2 is obtained. The data collected in real time by the sensor is input into the deep learning model for anomaly detection, and an anomaly score is obtained. Weight correction coefficients are introduced for statistical models, unsupervised models, and deep learning models. , , When the model's detection results deviate from the actual situation, the model weights are adjusted using a weight correction coefficient to obtain the final anomaly score. If the final anomaly score exceeds a set threshold, the corresponding input data is archived and labeled as 1. If the condition is not met, it is marked as an exception; otherwise, the corresponding input data is archived and labeled with tag 1. This is recorded as normal.

[0012] (6) After the mooring chain completes flash butt welding, post-weld quality inspection is carried out, and the correction coefficient is corrected according to the post-weld quality inspection results;

[0013] (7) Determine whether the labels added in step (5) are consistent with the detection quality results in step (6). If they are consistent, accumulate data and train the anomaly detection model regularly. If they are inconsistent, retrain the deviation model.

[0014] Furthermore, in step (2), the flash butt welding process includes the following stages: A preheating-up phase Each warm-up stage A preheating and subsequent retreat phase, The process consists of one equilibrium stage, one continuous flashing stage, one electrically charged upsetting stage, one electrically de-charged upsetting stage, and one reset stage.

[0015] Furthermore, in step (3), the intra-channel features and inter-channel features of the time clusters are extracted, including intra-cluster features of time clusters, inter-cluster features of different time clusters in a single dimension, inter-cluster features of different dimensions of the same time cluster, and inter-cluster Pearson correlation coefficients of different dimensions of the same time cluster.

[0016] Furthermore, features within time clusters include A preheating-up phase The mean, standard deviation, mean of first derivative, variance of first derivative, peak factor, and peak value of the time clusters corresponding to the preheating retreat stage, one electrically charged upsetting stage, one electrically de-charged upsetting stage, and one reset stage. Each warm-up stage The mean, standard deviation, minimum, maximum, and linear trend slope of the time clusters corresponding to each equilibrium stage and each continuous flash stage; the inter-cluster characteristics of different time clusters in a single dimension include the difference in mean, relative rate of change, variance, and difference in the first derivative mean of adjacent time clusters in each single dimension of welding current, welding voltage, welding electrode position, and upsetting force sequence; the inter-cluster characteristics of different dimensions within the same time cluster include the... time clusters Dimensions With dimension The mean difference, relative rate of change, variance difference, and first derivative mean difference; the inter-cluster Pearson correlation coefficients of different dimensions of the same time cluster, including the first derivative. time clusters Dimensions With dimension Pearson correlation coefficient .

[0017] Furthermore, in step (4), the weights of the statistical model... for

[0018]

[0019] Weights of a One-Class SVM Unsupervised Model for

[0020]

[0021] Weights of deep learning models for

[0022]

[0023] in, , Normal samples at the start of custom phase two Quantity, For the abnormal samples at the start of custom phase three Quantity, Normal samples at the start of custom phase three Quantity, Normal samples at the start of custom phase four Quantity, For the abnormal samples at the start of custom phase four Quantity, The weights of the statistical model eventually converge to these values. The weights of the unsupervised One-Class SVM model eventually converge to these values. This represents the final approximate value of the weights in the deep learning model.

[0024] Furthermore, in step (5), the final abnormal score for

[0025]

[0026]

[0027]

[0028]

[0029] in, , , These represent the corrected weights for the statistical model, the unsupervised model, and the deep learning model, respectively. , , After performing proportional normalization, the following was obtained , , , , , These represent the weight correction coefficients for the statistical model, the unsupervised model, and the deep learning model, respectively.

[0030] Furthermore, step (6) is as follows:

[0031] like Proceed to step (7), where, This is the accumulated value of the system test data. The threshold for adjusting the weight adjustment coefficient;

[0032] like Furthermore, if the static tensile strength of the welded mooring chain links is substandard, then... , ;like ,but ,in The weighting adjustment coefficient is 1 and ;like ,but ;like ,but ;like ,but ,in The weighting adjustment factor is 2 and ;like ,but ;like ,but ;

[0033] like And the static tensile strength of the welded mooring chain links meets the requirements, if ,but ,in The weighting adjustment factor is 2 and ;like ,but ;like ,but ;like ,but ,in The weighting adjustment coefficient is 1 and ;like ,but ;like ,but ;

[0034] Archive the data that fails the post-weld quality inspection and add tag 2. Record it as abnormal; archive the data that passes the post-weld quality inspection and add label 2. This is recorded as normal.

[0035] Furthermore, in step (7), if the labels added in step (5) are consistent with the detection quality results in step (6), then data is accumulated, and the anomaly detection model is trained periodically. The specific training of the model is as follows:

[0036] Let the model determine the correct data accumulation value Add one, if If the correct accumulated data value is zeroed out, and based on the welding data of the specification D mooring chain, features of all positive samples are extracted, and the mean and variance of each feature are recalculated to update the normal range and threshold of the statistical model. Threshold for retraining statistical models;

[0037] like If the model determines the data to be valid, let the accumulated value of the positive sample data in the data where the model correctly determines the data be valid. Add one; if the model determines the data to be unqualified, add the accumulated value of the negative sample data in the data that the model correctly determined. Add one;

[0038] like If the accumulated value of positive sample data is zeroed out, the features of all normal samples are extracted based on the welding data of the D-specification mooring chain. The mean and variance of each feature are recalculated to update the normal range and threshold of the statistical model. At the same time, the features of all positive samples are extracted to retrain the One-Class SVM unsupervised model, and the hyperparameters are manually adjusted periodically to update its decision boundary and save the new model parameters. This is the threshold for retraining the statistical model and the One-Class SVM unsupervised model; otherwise, the step ends.

[0039] like Then the accumulated value of negative sample data is cleared to zero. Based on the welding data of the D-specification mooring chain, the amount of negative sample data at this time is... One, sampled There are positive sample data, among which... , To set the threshold for retraining the deep learning model, For each set of positive and negative samples, 80% is used as the training set for model training and 20% is used as the validation set for training the deep learning model, with hyperparameters adjusted manually periodically; otherwise, the process ends.

[0040] Furthermore, in step (7), if the labels added in step (5) are inconsistent with the detection quality results in step (6), the bias model is retrained, as follows:

[0041] when If the model is deemed abnormal, but passes the tensile test, the model with the closest abnormal score to 0 during retraining is selected; when If the model is judged to be normal, but fails the tensile test, the model with the closest abnormal score to 1 will be retrained.

[0042] Let the model determine the accumulated value of erroneous data. Add one, if If the process continues, it proceeds to retraining the statistical model, the OneClass SVM unsupervised model, and the deep learning model; otherwise, the process ends. Thresholds for retraining statistical models, OneClass SVM unsupervised models, and deep learning models.

[0043] Beneficial effects: Compared with the prior art, the significant advantages of this invention are: 1. Lightweight start-up and gradual performance enhancement. In the early stages of production of new specification mooring chains, data is scarce. This invention, through the use of... The system uses a statistical model built from a number of normal samples to quickly construct a preliminary anomaly detection capability, ensuring lightweight deployment and avoiding model instability due to insufficient data. Furthermore, as data accumulates and the number of anomaly samples gradually increases, the anomaly detection capability of the One-Class SVM model is improved. Finally, a deep learning model based on a Siamese dual-attention mechanism for feature fusion is introduced to further enhance diagnostic accuracy. 2. Based on a data-driven dynamic model fusion mechanism, this invention can adjust the model fusion based on the number of positive samples. With the number of negative samples To address the changes, a model adaptive model fusion strategy and weight formula are proposed to dynamically adjust the weights of the statistical model, unsupervised model, and supervised model. , , At the same time, a weight correction coefficient is proposed. , , The accuracy of the three models is periodically adjusted to obtain the final outlier score. This strategy can adapt to the number of positive samples in real time. With the number of negative samples Changes in data structure enable a smooth transition from simple to complex models, improving system diagnostic capabilities and stability; 3. Incremental training and maintenance: This invention supports phased triggering of incremental model training, targeting data... , , growth, abnormal samples The invention selectively updates parts or the entire model to avoid retraining on all data, reducing the risk of overfitting and model degradation. Under new specifications or operating conditions, only the relevant models are updated, significantly reducing system maintenance costs. 4. Deep feature mining and model innovation: This invention divides time clusters according to the characteristics of flash welding on mooring chains. By analyzing the intra-channel variations and inter-channel coupling relationships of multi-channel time-series signals, a multi-dimensional representation of the system state is achieved, enhancing the ability to identify complex anomalies and provide early warnings of faults. Furthermore, an innovative feature fusion model based on twin structures and dual attention mechanisms is proposed to enhance the representational differences between normal and abnormal data, thereby improving discrimination capabilities and detection accuracy. Attached Figure Description

[0044] Figure 1 This is a flowchart of the method of the present invention;

[0045] Figure 2 This is a flowchart of the flash welding process;

[0046] Figure 3 Dynamic anomaly scores for statistical models, unsupervised One-Class SVM models, and deep learning models. , , Schematic diagram of the changes;

[0047] Figure 4 This is a structural diagram of a twin-structure dual-attention feature fusion model. Detailed Implementation

[0048] like Figure 1 As shown, the intelligent online evaluation method for flash butt welding quality of the present invention includes the following steps:

[0049] (1) Build a model to achieve lightweight startup.

[0050] (11) Set the current specifications on the system side. Before starting the mooring chain welding, set the relevant parameters for flash butt welding of the mooring chain on the host computer system side, and set the welding mooring chain specification to the diameter. Transformer connection Clamping position Burning length Upsetting length Electric upsetting time Cooling time Threshold current Number of preheating cycles Then proceed to step (12).

[0051] (12) Real-time data acquisition by sensors. The welding process is monitored by sensors to obtain the welding parameter sequence, which includes the welding time sequence. The number of sampling points is 300, and the welding current sequence is... Welding voltage sequence welding electrode position sequence and upsetting force sequence Simultaneously, the control system collects welding status sequences. ,in, This is the preheating stage. This is the preheating stage. This is the preheating and subsequent retreat phase. This is the equilibrium phase. This is the continuous flashing phase. This is the electric upsetting stage. This is the stage of upsetting without electricity. For the reset phase, such as Figure 2 As shown. Then for At any given moment, the set of sensor observation signals ,in , , , , These represent the welding current dimension, welding voltage dimension, welding electrode position dimension, upsetting force dimension, and stage dimension, respectively, before proceeding to step (13).

[0052] (13) Query the database to see if there is a mooring chain with a diameter of The sub-database. It contains mooring chain specifications with a diameter of... The sub-database is then processed, proceeding to step (15).

[0053] (14) New specifications Sub-database Then proceed to step (15), specifications. The sub-database is used to store specifications. Positive and negative sample data.

[0054] (15) Determine specifications Is the sub-database data larger than If greater than Proceed to step (16), if it is not greater than Proceed to step (12) to continue collecting welding data until the sensor collects the data. The purpose of this process is to manually determine the positive sample data. The data points are positive samples and are used to build a statistical model for anomaly detection in the initial stage of the system.

[0055] (16) Time cluster partitioning. The data in the database is processed, and all missing values ​​are calculated sequentially by linear interpolation. All missing values ​​are then filled into their corresponding positions.

[0056] Furthermore, regarding the time cluster division, as can be seen from steps (11) and (12), the welding specification is the diameter. The number of preheating cycles for the mooring chain is The welding sequence consists of 3 preheating advance stages, 3 preheating stages, 3 preheating retreat stages, 2 equalization stages, 1 continuous flashing stage, 1 electrically charged upsetting stage, 1 electrically neutral upsetting stage, and 1 reset stage, for a total of 15 stages.

[0057] Furthermore, the time series It is divided into 15 time clusters according to different stages, denoted as , No. Each time cluster is recorded as ,in For the first Time cluster sensor Observation signal set, ,in , , , , These represent the welding current dimension, welding voltage dimension, welding electrode position dimension, upsetting force dimension, and stage dimension, respectively, before proceeding to step (17).

[0058] (17) Feature extraction within and between channels. Based on prior knowledge, features are extracted from the flash welding data of the mooring chain collected by the sensor from within and between channels, and then proceed to step (18).

[0059] Furthermore, the number of preheating cycles was extracted. Preheating time Preheating displacement change Continuous flash welding time Continuous flash displacement change Upsetting speed Upsetting maintains overall welding displacement variation Burning length Preheating intensity .

[0060] Furthermore, features within time clusters are extracted. A preheating-up phase The mean, standard deviation, mean of first derivative, variance of first derivative, and peak factor of the time clusters corresponding to one preheating retreat stage, one electrically charged upsetting stage, one electrically de-charged upsetting stage, and one reset stage; extract Each warm-up stage The mean, standard deviation, minimum, maximum, and linear trend slope of the time clusters corresponding to each equilibrium phase and each continuous flash phase.

[0061] Furthermore, the inter-cluster features of different time clusters in a single dimension are extracted, and the mean difference, relative rate of change, variance difference, and mean difference of first derivative of adjacent time clusters in each single dimension of welding current, welding voltage, welding electrode position, and upsetting force sequence are extracted respectively.

[0062] Furthermore, inter-cluster features of different dimensions within the same time cluster are extracted, and the first-order features are extracted respectively. time clusters Dimensions With dimension The mean difference, relative rate of change, variance difference, and first derivative mean difference, among which and .

[0063] Furthermore, the inter-cluster Pearson correlation coefficients of different dimensions of the same time cluster are extracted, and the first-order correlation coefficients are extracted respectively. time clusters Dimensions With dimension Pearson correlation coefficient ,in and .

[0064] (18) Determine whether the number of manually calibrated normal data for this specification is 20. The purpose of this process is to manually determine the first 20 positive sample data, and based on this, establish a statistical model for anomaly detection in the initial stage of the system. If the number of normal data is 20, then the 20 manually calibrated normal data will proceed to step (19) to establish a statistical model; if the number of normal data is not 20, then the data is the data to be detected, proceed to step (2), and then calculate the weights of the three models.

[0065] (19) Establish a statistical model. The specifications are established based on the aforementioned steps. Sub-database We obtained 20 manually labeled normal data points and built a statistical model based on them. We calculated the mean and variance of the features of the 20 manually labeled normal data points, and set the normal range and threshold to build the statistical model.

[0066] (2) Calculate the weights of the three models to achieve dynamic fusion of the models.

[0067] Retrieve the positive and negative sample sizes from the database for this specification, and calculate the weights of the three models. Retrieve Specification The number of positive samples in the sub-database is The number of negative samples is The dynamic weights of the computational statistical model, the unsupervised model OneClass SVM, and the deep learning model are respectively , , Then, proceed to step (31) to perform anomaly detection on the data to be detected obtained in step (18).

[0068] Furthermore, as the system continues to operate, specifications are continuously accumulated. Based on the amount of normal and abnormal data, we divided the mooring chain welding data into four phases, making... This marks the dividing point between Phase One and Phase Two. , This marks the dividing point between Phase Two and Phase Three. , This marks the dividing point between Phase Three and Phase Four:

[0069] make ,when and This is Phase 1, the initial stage of system operation, where the number of positive and negative samples is close to 0. In this phase, anomaly detection is mainly based on statistical models. As the number of positive sample data accumulates, the weight of the unsupervised model One Class SVM is gradually increased, while the weight of the statistical model is reduced. Since the number of negative sample data is close to 0 in this phase, the weight of the deep learning model is 0.

[0070] when and This is Phase Two, the early stage of system operation. Positive sample data has accumulated to a certain extent, while negative samples are close to zero. In this phase, anomaly detection is mainly based on the unsupervised One Class SVM model. As the number of positive sample data accumulates, the weight of the unsupervised One Class SVM model is gradually increased, while the weight of the statistical model is reduced. Due to the severe imbalance in the mooring chain flash welding data, the negative sample data is still close to zero in this phase, so the weight of the deep learning model is set to 0.

[0071] when For phase three, when At this time At this time Then at this time and During the mid-stage of system operation, there is sufficient positive sample data and a certain amount of negative sample data has been accumulated. At this stage, anomaly detection is mainly based on the unsupervised model One Class SVM. As negative sample data accumulates, the weight of the deep learning model gradually increases, while the weight of the statistical model and the unsupervised model One Class SVM decreases.

[0072] when For stage four, when At this time At this time Then at this time and In the later stages of system operation, both positive and negative sample data are sufficient. At this stage, anomaly detection is primarily based on a deep learning model. As negative sample data accumulates, the weights of the deep learning model gradually increase, while the weights of the statistical model and the unsupervised One-Class SVM model decrease. When the weights are large enough, the weights of a deep learning model tend to be large enough. The weights of the unsupervised model One Class SVM approach 100%. The weights based on the statistical model tend to be ,in and ;

[0073] Furthermore, For weights based on statistical models, The weights of the unsupervised One-Class SVM model. The weights of the deep learning model are known. ,make

[0074]

[0075]

[0076]

[0077] Among them when hour, The value reaches its maximum at this time. , , ;when hour, , ;when , hour, The value reaches its maximum at this time. , , ;when , hour, , , ;when , When it approaches infinity, , , ,like Figure 3 As shown, the weight function is designed to change rapidly in the early stages and then level off in the later stages. In the initial stages of each phase, the relevant model is primarily used, while in the later stages, a smooth transition to the next model is achieved. This ensures that the model switching process is both adaptive and avoids abrupt changes that could lead to performance instability. Furthermore, the function transitions between phases exhibit good continuity and differentiability, guaranteeing a smooth transition in weight changes and preventing inconsistencies in model fusion or performance jumps.

[0078] The number of positive samples at this time is The number of negative samples is ,but , , .

[0079] (3) Real-time anomaly detection is performed by extracting features from multiple angles.

[0080] (31) The anomaly detection part consists of three parts: a statistical model, an unsupervised model based on One-Class SVM, and a deep learning model based on Siamese dual attention mechanism feature fusion, which yields anomaly score 1, anomaly score 2, and anomaly score 3, respectively. , , Then proceed to step (32).

[0081] (311) Based on the features extracted in step (17), anomaly detection is performed using a statistical model to obtain an anomaly score of 1. ;

[0082] (312) Based on the features extracted in step (17), anomaly detection is performed using the unsupervised learning model One Class SVM. The features extracted from one class of data (positive samples) are used to train the model to obtain the decision boundary, thereby performing anomaly detection on the time series data and obtaining anomaly score 2. ;

[0083] (313) Based on the time series data obtained in step (17), anomaly detection is performed using a deep learning model that fuses features based on a Siamese dual attention mechanism, such as... Figure 4 As shown, multivariate time series data are input into a deep learning network model;

[0084] Furthermore, the input data is first normalized by the data preprocessing module, and then the network performs channel splitting and time clustering on the normalized data to prepare it for the next step of inputting into the self-attention mechanism network for feature extraction.

[0085] Furthermore, the processed data is input into two self-attention mechanism networks to obtain data features from two perspectives: inter-cluster and intra-cluster.

[0086] Furthermore, the similarity between the features from the two angles is calculated.

[0087] Furthermore, the abnormal score was obtained as 3. .

[0088] (32) Correct the model weights and perform normalization to obtain the anomaly scores. From step (31), the anomaly scores for the statistical model, the unsupervised model based on One-Class SVM, and the deep learning model based on Siamese dual-attention mechanism feature fusion are: , , Step 2 yields the dynamic weights of the statistical model, the unsupervised One-Class SVM model, and the deep learning model, respectively. , , ;set up , , These are the weight correction coefficients for the statistical model, the unsupervised model, and the deep learning model, respectively, and are the default values ​​at the beginning of the system. (When the weight correction coefficient is 1, the weights do not need to be corrected.) When the model's detection results deviate from the actual situation, the model weights are corrected using the weight correction coefficient.

[0089]

[0090]

[0091]

[0092] in , , Different weights are applied to statistical models, unsupervised models, and deep learning models. , , After performing proportional normalization, the following was obtained , , The final anomaly score of the sample is obtained by weighting the samples using the following formula. ,

[0093]

[0094] Then proceed to step (33).

[0095] (33) Determine whether an anomaly is based on a threshold. The threshold is manually set as follows: From step (32) we obtain If the sample data is normal, proceed to step (35).

[0096] (34) Archive the data and add tag 1. If it is recorded as abnormal, proceed to step (41) to perform post-weld quality inspection and judgment model and correct model weights.

[0097] (35) Archive the data and add tag 1. If it is recorded as normal, proceed to step (41) to perform post-weld quality inspection and judgment model and correct model weights.

[0098] (4) Post-weld quality inspection and correction of weight correction coefficient.

[0099] (41) After the mooring chain welding is completed, the static tensile strength of the welded mooring chain links is verified by methods such as tensile testing, magnetic particle testing, and ultrasonic testing. If the strength is not qualified, proceed to step (42); if the strength is qualified, proceed to step (43). At the same time, sample data Accumulate sample data for systematic testing. The accumulated value of the system test data, the initial value of the new system. =0;

[0100] (411) If ,in To adjust the threshold of the weight adjustment coefficient, proceed to the next step;

[0101] (412) If Furthermore, a tensile test confirmed that the static tensile strength of the welded mooring chain links did not meet the qualification requirements, therefore... , As can be seen from step (32), the weight correction coefficients of the statistical model, the unsupervised model, and the deep learning model are respectively , , ;

[0102] Furthermore, if Then let ,in The weighting adjustment coefficient is 1 and ;like Then let ;like Then let ;

[0103] Furthermore, if Then let ,in The weighting adjustment factor is 2 and ;like Then let ;like Then let ;

[0104] (413) If Furthermore, the static tensile strength of the welded mooring chain links was verified to meet the qualification requirements through tensile testing, thus making , As can be seen from step (32), the weight correction coefficients of the statistical model, the unsupervised model, and the deep learning model are respectively , , ;

[0105] Furthermore, if Then let ,in The weighting adjustment factor is 2 and ;like Then let ;like Then let ;

[0106] Furthermore, if Then let ,in The weighting adjustment coefficient is 1 and ;like Then let ;like Then let .

[0107] (42) Archive the data and add tag 2. If it is recorded as an anomaly, proceed to step (5) to determine whether the model detection is accurate;

[0108] (43) Archive the data and add tag 2. If it is recorded as normal, proceed to step (5) to determine whether the model detection is accurate.

[0109] (5) Determine if the model is correct. This is obtained from steps (34) and (35). And obtained from steps (42) and (43) ,if If the model's judgment is correct, proceed to step (61) to accumulate data and train the model periodically; if If the model judgment is incorrect, proceed to step (71) to retrain the biased model.

[0110] (6) Accumulate data and retrain relevant models regularly.

[0111] (61) Correct data Accumulate data that correctly identifies the model. Accumulated data values ​​for model judgment, initial values ​​for the new system. If the value is 0, proceed to step (62).

[0112] (62) Correct data Determine if the data is correct. Is it greater than ,like Then proceed to step (63), otherwise proceed to step (64); where The threshold for retraining the statistical model is set manually. The larger the value, the longer the model retraining cycle; among which The training cycle coefficient of the retraining statistical model, and the specification The value is related to the amount of data in the sub-database.

[0113] ,

[0114] in .

[0115] (63) Retrain the statistical model. The statistical model is retrained periodically as it is used initially and built upon all positive sample data. At that time, the correct data is cleared to zero, that is... and based on specifications The sub-database data is used to extract features from all normal samples. The mean and variance of each feature are recalculated to update the normal range and threshold of the statistical model, and the process ends.

[0116] (64) Check if it is normal. If the model judges the data to be correct, if the data is a positive sample, proceed to step (133); if the data is a negative sample, proceed to step (66).

[0117] (65) Positive sample data Accumulate positive sample data from the data in which the model correctly identifies the data. The accumulated values ​​of positive sample data in the data from which the model correctly judges the data are the initial values ​​of the new system. If the value is 0, proceed to step (651).

[0118] (651) Positive sample data Determine positive sample data. Is it greater than ,like Then proceed to step (652); otherwise, end the process. The thresholds for retraining statistical models and unsupervised models are set manually. The larger the value, the longer the model retraining cycle; among which The training cycle coefficients for retraining statistical models and unsupervised models, and the specifications. The value is related to the amount of data in the sub-database.

[0119] ,

[0120] in .

[0121] (652) Retrain the statistical model and the unsupervised model. Retrain the statistical model and the unsupervised model. Since the statistical model and the supervised learning model are built based on all positive sample data, the model should be retrained periodically when a certain amount of positive sample data has accumulated. At that time, the accumulation of positive sample data is reset to zero, that is... and based on specifications The sub-database data is used to extract features from all normal samples. The mean and variance of each feature are recalculated to update the normal range and threshold of the statistical model. At the same time, the features of all positive samples are extracted to retrain the One Class SVM model. The hyperparameters are manually adjusted periodically, the decision boundary is updated, and the new model parameters are saved to end the process.

[0122] (66) Negative sample data Accumulate negative sample data from the data in which the model correctly identifies the data. The accumulated value of negative sample data in the data from which the model correctly judges the data; the initial value of the new system. If the value is 0, proceed to step (661).

[0123] (661) Negative sample data Determine negative sample data. Is it greater than ,like Then proceed to step (662); otherwise, end the process. The threshold for retraining a deep learning model is set manually. The larger the value, the longer the model retraining cycle; among which The training cycle coefficients for retraining deep learning models, and the specifications The value is related to the amount of data in the sub-database.

[0124] ,

[0125] in .

[0126] (662) Retrain the deep learning model. The deep learning model is retrained because it is built upon a combination of positive and negative sample data. Therefore, when a certain amount of negative sample data accumulates, the model should be retrained periodically. Negative sample data accumulation is cleared to zero, that is Based on specifications The sub-database data, at this time the amount of negative sample data is One, sampled There are positive sample data, among which ,Will The positive and negative sample data is divided into 80% as the training set for model training and 20% as the validation set for retraining the model. The hyperparameters are manually adjusted periodically before the process ends.

[0127] (7) Retrain the bias model.

[0128] (71) Check if it is normal. Determine if the chain link is qualified. If the model indicates an anomaly, but passes the tensile test, proceed to step (711). If the model is deemed normal, but fails the tensile test, proceed to step (712).

[0129] (711) Retrain the model whose anomaly score is closest to 0. From step (31), we obtain the statistical model, the unsupervised model based on OneClass SVM, and the deep learning model based on Siamese dual attention mechanism feature fusion, respectively, and obtain anomaly score 1, anomaly score 2, and anomaly score 3, respectively, with corresponding symbols representing: , , ,make ;

[0130] Furthermore, if Based on specifications The sub-database data is used to extract the features of all normal samples. The mean and variance of each feature are recalculated to update the normal range and threshold of the statistical model to retrain the model, and then proceed to step (72).

[0131] Furthermore, if Based on specifications The sub-database data is used to extract features from all positive samples, retrain the One Class SVM model, and manually adjust the hyperparameters periodically, update its decision boundary, and save the new model parameters before proceeding to step (72).

[0132] Furthermore, if Based on specifications The sub-database data, at this time the amount of negative sample data is One, sampled There are positive sample data, among which ,Will 80% of the positive and negative sample data is divided into a training set for model training and 20% is divided into a validation set for retraining the model. The hyperparameters are manually adjusted periodically. Proceed to step (72).

[0133] (712) Retrain the model whose anomaly score is closest to 1. It is known that anomaly scores 1, 2, and 3 are obtained from a statistical model, an unsupervised model based on One-Class SVM, and a deep learning model based on Siamese dual-attention feature fusion, respectively. The corresponding symbols are: , , ,make ;

[0134] Furthermore, if Based on specifications The sub-database data is used to extract the features of all normal samples. The mean and variance of each feature are recalculated to update the normal range and threshold of the statistical model to retrain the model, and then proceed to step (72).

[0135] Furthermore, if Based on specifications The sub-database data is used to extract features from all positive samples, retrain the One Class SVM model, and manually adjust the hyperparameters periodically, update its decision boundary, and save the new model parameters before proceeding to step (72).

[0136] Furthermore, if Based on specifications The sub-database data, at this time the amount of negative sample data is One, sampled There are positive sample data, among which ,Will 80% of the positive and negative sample data is divided into a training set for model training and 20% is divided into a validation set for retraining the model. The hyperparameters are manually adjusted periodically. Proceed to step (72).

[0137] (72) Accumulated error data Accumulate data on errors in model judgment. The accumulated value of erroneous data in model judgment, the initial value of the new system. If the value is 0, proceed to step (73).

[0138] (73) Error data Determine positive sample data. Is it greater than ,like Then proceed to step (74); otherwise, end the process. The threshold for retraining the anomaly detection model is set manually. The larger the value, the longer the model retraining cycle; among which The training cycle coefficients for retraining the anomaly detection model are related to the specifications. The value is related to the amount of data in the sub-database.

[0139] .

[0140] (74) Retrain the anomaly detection model. Retrain the statistical model, unsupervised model, and deep learning model. When a certain amount of erroneous data accumulates, the detection accuracy may decrease due to catastrophic forgetting, degradation, or overfitting of the model. Therefore, at this time, all models should be retrained, and the relevant parameters of the models should be manually adjusted. Error data is cleared to zero, that is Based on specifications Using the sub-database data, we retrained the statistical model, unsupervised model, and deep learning model. For the statistical model, we extracted features from all normal samples, recalculated the mean and variance of each feature, updated the normal range and threshold of the statistical model, and retrained the model. For the unsupervised model, we extracted features from all positive samples, retrained the One-Class SVM model, and periodically adjusted the hyperparameters manually, updated its decision boundary, and saved the new model parameters. For the deep learning models, the amount of negative sample data was [data missing]. One, sampled There are positive sample data, among which ,Will The positive and negative sample data is divided into 80% as the training set for model training and 20% as the validation set for retraining the model. The hyperparameters are manually adjusted periodically before the process ends.

[0141] This invention addresses the issues of insufficient sample data, severe data imbalance in mooring chain flash welding, and model overfitting in online detection of mooring chain flash welding quality based on multi-model fusion, multiple model retraining methods, and multi-feature extraction fusion. In the initial stages of new specification production operation, when there is insufficient sample data (especially abnormal samples), statistical models can be quickly deployed and provide preliminary anomaly detection capabilities, avoiding performance instability caused by forcibly using complex models when data is scarce. After the system has run for a period and accumulated some data but still lacks abnormal samples, One-Class SVM (unsupervised model) is used for training. Unsupervised models do not require labeled data and can learn normal patterns using normal data to detect anomalies deviating from those patterns. Although the performance may not be as good as supervised models, it provides a feasible solution when labels are missing. When enough abnormal samples are accumulated, a deep learning model based on Siamese dual-attention mechanism feature fusion (supervised learning model) is trained. This model can learn complex features and dependencies within and between time series channels from labeled data, thereby enabling more accurate anomaly diagnosis. Meanwhile, by continuously updating and calibrating the weights of each model, retraining the model through multiple methods, and using different models at different stages, the accuracy and recall of the system detection can be significantly improved.

Claims

1. A smart online evaluation method for flash butt welding quality, characterized in that, Includes the following steps: (1) Create a new database of specification D mooring chains, accumulate F positive sample data, and calculate the missing values ​​of the F positive sample welding data of specification D mooring chains in turn by linear interpolation and fill them into their corresponding positions to obtain the welding time series of specification D mooring chains. (2) According to the total number of flash butt welding stages n, the welding time sequence of the specification D mooring chain is divided into n time clusters; (3) Extract intra-channel and inter-channel features from time clusters, and construct an initial statistical model and a One Class SVM unsupervised model based on the feature mean, variance, normal range and threshold. (4) Based on the statistical model, the One-Class SVM unsupervised model and the deep learning model, an anomaly detection model is constructed, and the number of positive samples of the welding time series of the D-specification mooring chain is obtained according to step (1). and negative sample count The weights of the statistical model, the One Class SVM unsupervised model, and the deep learning model are determined; the welding time series of the D-specification mooring chain obtained in step (1) is divided into training set and validation set for model training, and anomaly detection is performed on the data collected by the sensor after flash welding. (5) Input the data collected by the sensor in real time into the statistical model for anomaly detection, and obtain an anomaly score of 1. The data collected in real time by the sensor is input into the unsupervised learning model One-Class SVM for anomaly detection, and an anomaly score of 2 is obtained. The data collected in real time by the sensor is input into the deep learning model for anomaly detection, and an anomaly score is obtained. Weight correction coefficients are introduced for statistical models, unsupervised models, and deep learning models. , , When the model detection results deviate from the actual situation, the model weights are corrected by the weight correction coefficient to obtain the final anomaly score. If the final abnormal score exceeds the set threshold, the corresponding input data will be archived and labeled with tag 1. This is recorded as an anomaly; Conversely, the corresponding input data is archived and labeled with tag 1. This is recorded as normal. (6) After the mooring chain completes flash butt welding, post-weld quality inspection is carried out, and the correction coefficient is corrected according to the post-weld quality inspection results; (7) Determine whether the labels added in step (5) are consistent with the detection quality results in step (6). If they are consistent, accumulate data and train the anomaly detection model regularly. If they are inconsistent, retrain the deviation model.

2. The intelligent online evaluation method for flash butt welding quality according to claim 1, characterized in that, In step (2), the flash butt welding process includes the following stages: A preheating-up phase Each warm-up stage A preheating and subsequent retreat phase, The process consists of one equilibrium stage, one continuous flashing stage, one electrically charged upsetting stage, one electrically de-charged upsetting stage, and one reset stage.

3. The intelligent online evaluation method for flash butt welding quality according to claim 2, characterized in that, In step (3), the features within and between channels of the time cluster are extracted, including features within the time cluster, inter-cluster features of different time clusters in a single dimension, inter-cluster features of different dimensions of the same time cluster, and inter-cluster Pearson correlation coefficients of different dimensions of the same time cluster.

4. The intelligent online evaluation method for flash butt welding quality according to claim 3, characterized in that, Features within a time cluster include A preheating-up phase The mean, standard deviation, mean of first derivative, variance of first derivative, and peak factor of the time clusters corresponding to the preheating retreat stage, one electrically charged upsetting stage, one electrically de-charged upsetting stage, and one reset stage. Each warm-up stage The mean, standard deviation, minimum, maximum, and linear trend slope of the time clusters corresponding to each equilibrium stage and each continuous flash stage; the inter-cluster characteristics of different time clusters in a single dimension include the difference in mean, relative rate of change, variance, and difference in the first derivative mean of adjacent time clusters in each single dimension of welding current, welding voltage, welding electrode position, and upsetting force sequence; the inter-cluster characteristics of different dimensions within the same time cluster include the... time clusters Dimensions With dimension The mean difference, relative rate of change, variance difference, and first derivative mean difference; the inter-cluster Pearson correlation coefficients of different dimensions of the same time cluster, including the first derivative. time clusters Dimensions With dimension Pearson correlation coefficient .

5. The intelligent online evaluation method for flash butt welding quality according to claim 1, characterized in that, In step (4), the weights of the statistical model for Weights of a One-Class SVM Unsupervised Model for Weights of deep learning models for in, , Normal samples at the start of custom phase two Quantity, For the abnormal samples at the start of custom phase three Quantity, Normal samples at the start of custom phase three Quantity, Normal samples at the start of custom phase four Quantity, For the abnormal samples at the start of custom phase four Quantity, The weights of the statistical model eventually converge to these values. The weights of the unsupervised One-Class SVM model eventually converge to these values. This represents the final approximate value of the weights in the deep learning model.

6. The intelligent online evaluation method for flash butt welding quality according to claim 1, characterized in that, In step (5), the final abnormal score for in, , , These represent the corrected weights for the statistical model, the unsupervised model, and the deep learning model, respectively. , , After performing proportional normalization, the following was obtained , , , , , These represent the weight correction coefficients for the statistical model, the unsupervised model, and the deep learning model, respectively.

7. The intelligent online evaluation method for flash butt welding quality according to claim 6, characterized in that, Step (6) is as follows: like Proceed to step (7), where, This is the accumulated value of the system test data. The threshold for adjusting the weight adjustment coefficient; like Furthermore, if the static tensile strength of the welded mooring chain links is substandard, then... , ;like ,but ,in The weighting adjustment coefficient is 1 and ;like ,but ;like ,but ;like ,but ,in The weighting adjustment factor is 2 and ;like ,but ;like ,but ; like And the static tensile strength of the welded mooring chain links meets the requirements, if ,but ,in The weighting adjustment factor is 2 and ;like ,but ;like ,but ;like ,but ,in The weighting adjustment coefficient is 1 and ;like ,but ;like ,but ; Archive the data that fails the post-weld quality inspection and add tag 2. Record it as abnormal; archive the data that passes the post-weld quality inspection and add a normal label 2. This is recorded as normal.

8. The intelligent online evaluation method for flash butt welding quality according to claim 1, characterized in that, In step (7), if the labels added in step (5) are consistent with the detection quality results in step (6), then data is accumulated, and the anomaly detection model is trained periodically. The specific training of the model is as follows: Let the model determine the correct data accumulation value Add one, if If the correct accumulated data value is zeroed out, and based on the welding data of the specification D mooring chain, features of all positive samples are extracted, and the mean and variance of each feature are recalculated to update the normal range and threshold of the statistical model. Threshold for retraining statistical models; like If the model determines the data to be valid, let the accumulated value of the positive sample data in the data where the model correctly determines the data be valid. Add one; if the model determines the data to be unqualified, add the accumulated value of the negative sample data in the data that the model correctly determined. Add one; like If the accumulated value of positive sample data is zeroed out, the features of all normal samples are extracted based on the welding data of the D-specification mooring chain. The mean and variance of each feature are recalculated to update the normal range and threshold of the statistical model. At the same time, the features of all positive samples are extracted to retrain the One-Class SVM unsupervised model, and the hyperparameters are manually adjusted periodically to update its decision boundary and save the new model parameters. This is the threshold for retraining the statistical model and the One-Class SVM unsupervised model; otherwise, the step ends. like Then the accumulated value of negative sample data is cleared to zero. Based on the welding data of the D-specification mooring chain, the amount of negative sample data at this time is... One, sampled There are positive sample data, among which... , To set the threshold for retraining the deep learning model, For each set of positive and negative samples, 80% is used as the training set for model training and 20% is used as the validation set for training the deep learning model, with hyperparameters adjusted manually periodically; otherwise, the process ends.

9. The intelligent online evaluation method for flash butt welding quality according to claim 7, characterized in that, In step (7), if the labels added in step (5) are inconsistent with the detection quality results in step (6), the bias model is retrained, as follows: when If the model is deemed abnormal, but passes the tensile test, the model with the closest abnormal score to 0 during retraining is selected; when If the model is judged to be normal, but fails the tensile test, the model with the closest abnormal score to 1 will be retrained. Let the model determine the accumulated value of erroneous data. Add one, if If the process continues, it proceeds to retraining the statistical model, the One-Class SVM unsupervised model, and the deep learning model; otherwise, the process ends. Thresholds for retraining statistical models, One Class SVM unsupervised models, and deep learning models.

Images