Steel bar welding quality abnormality unsupervised identification method and system based on one-class SVM

Abstract

The present application relates to a steel bar welding quality anomaly unsupervised identification method and system based on One-Class SVM, comprising: collecting multi-source time series data; preprocessing and normalizing the multi-source time series data; based on the preset feature importance weight, fusing the extracted time domain features and frequency domain features to form a high-dimensional feature vector representing the welding process state; training an OCSVM model to build a parameter feature domain corresponding to qualified welding process; inputting the real-time high-dimensional feature vector into the trained OCSVM model, and judging whether the welding process to be measured deviates from the parameter feature domain based on the decision function, if it deviates, it is determined that the welding quality is abnormal; generating a rectification guidance scheme. The beneficial effects of the present application are: the present application is suitable for the scene of defect sample scarcity in actual construction, and significantly improves the practicability of detection.

Description

Technical Field

[0001] This invention relates to the field of steel bar welding inspection technology, and more specifically, to an unsupervised identification method and system for steel bar welding quality anomalies based on One-Class SVM. Background Technology

[0002] Steel bar welding is a core process in civil engineering (buildings, bridges, rail transit, etc.). Welding quality directly determines the structural bearing capacity and safety. Welding defects (such as slag inclusions, incomplete penetration, cracks, and porosity) can lead to insufficient steel bar connection strength, which can easily cause structural collapse, fracture, and other safety accidents, seriously threatening construction safety and the service life of the project.

[0003] Currently, the quality inspection of steel bar welding mainly relies on two methods: one is manual visual inspection, which depends entirely on the experience of the inspectors, resulting in a high rate of missed detections and false detections, and it cannot identify hidden defects inside the steel bar weld, leading to poor inspection reliability; the other is non-destructive testing methods such as ultrasonic testing and radiographic testing, which are post-construction inspections that need to be carried out after welding. These methods are inefficient, costly, and require professional personnel to operate, making them unsuitable for real-time inspection needs on construction sites and difficult to achieve full-process control of welding quality.

[0004] With the development of artificial intelligence technology, some studies have attempted to use AI models for steel bar welding quality inspection. However, existing AI inspection technologies mostly adopt supervised learning models, requiring a large number of labeled "welding defect samples" for model training. In actual construction, however, while qualified welding samples are readily available, welding defect samples (such as slag inclusions and incomplete penetration) are extremely scarce. On the one hand, the incidence of welding defects is low, making it difficult to collect a sufficient number of defect samples; on the other hand, labeling defect samples requires professional personnel to combine non-destructive testing results, which is time-consuming and labor-intensive, further limiting the application of supervised AI models.

[0005] Furthermore, existing technologies can only determine whether a defect exists, but cannot accurately pinpoint the type of defect or the cause of the abnormality (such as incomplete penetration due to abnormal current or cracks due to excessive temperature). This makes it difficult to guide on-site operators to make timely corrections and prevents the formation of a closed-loop management system of "detection-early warning-rectification," resulting in the inability to eliminate potential welding quality hazards in a timely manner.

[0006] In view of the shortcomings of the existing technologies, there is an urgent need for a steel bar welding quality inspection method and system that does not require defect samples, can identify welding quality anomalies in real time, and can realize defect tracing and rectification guidance, so as to solve the core pain points of the existing technologies. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and to provide an unsupervised identification method and system for steel bar welding quality anomalies based on One-Class SVM.

[0008] Firstly, an unsupervised method for identifying weld quality anomalies in reinforcing bars based on One-Class SVM is provided, including:

[0009] S1. By deploying a multi-source sensor acquisition unit on the steel bar welding equipment, collect multi-source time-series data of the standard welding process that has been confirmed as qualified by non-destructive testing, and build a qualified sample library; the multi-source time-series data includes at least welding current, welding voltage and welding area temperature.

[0010] S2. Preprocess and normalize the multi-source time-series data, and extract multi-dimensional time-domain and frequency-domain features for current, voltage and temperature time-series data respectively; based on the preset feature importance weights, fuse the extracted time-domain features and frequency-domain features to form a high-dimensional feature vector characterizing the welding process state.

[0011] S3. Input the high-dimensional feature vector as the only training set into the OCSVM model, and train the OCSVM model by optimizing the parameters of the OCSVM model to construct the parameter feature domain corresponding to the qualified welding process.

[0012] S4. Collect multi-source time-series data of the welding process to be tested in real time, and repeat step S2 to obtain real-time high-dimensional feature vectors; input the real-time high-dimensional feature vectors into the trained OCSVM model, and determine whether the welding process to be tested deviates from the parameter feature domain based on the decision function. If it deviates, it is determined to be an abnormal welding quality.

[0013] S5. When the welding quality is determined to be abnormal, the defect identification rule base based on statistical thresholds is called. The real-time high-dimensional feature vector is matched with the multi-level quantitative criteria in the defect identification rule base, and the unique corresponding defect type and parameter deviation value are output, and the corresponding rectification guidance plan is generated.

[0014] Preferably, S2 includes:

[0015] S201. Using the global minimum and global maximum values ​​of the same type of parameters in the qualified sample library, the time series data of the qualified samples and real-time samples are mapped to the [0,1] interval;

[0016] S202. For the normalized current, voltage and temperature time series data, extract time domain features including mean, variance, peak value, valley value, fluctuation coefficient and trend slope respectively, and extract frequency domain features including characteristic frequency, frequency peak value and frequency domain energy ratio based on wavelet transform.

[0017] S203. Using a combination of feature concatenation and weight allocation, the current feature vector, voltage feature vector, and temperature feature vector are fused into the high-dimensional feature vector, with the current feature having the largest weight.

[0018] Preferably, in S3, the parameters to be optimized in the OCSVM model include the kernel function and slack variables. With kernel width parameter .

[0019] Preferably, in S4, the decision function calculates the positional relationship between the real-time feature vector and the parameter feature domain based on the support vector coefficients and kernel function; when the output of the decision function is positive, it is determined to be qualified; when the output is negative, it is determined to be abnormal, triggering an audible and visual warning signal and outputting a pause command to the welding execution mechanism.

[0020] Simultaneously, the deviation of the real-time high-dimensional feature vector from the parameter feature domain is calculated, and the warning level is divided according to the numerical range of the deviation. The warning level includes at least a minor anomaly warning and a severe anomaly warning.

[0021] Preferably, in S5, the defect identification rule base includes quantitative criteria for five defect types: incomplete penetration, slag inclusion, crack, porosity, and lack of fusion. Each quantitative criterion is calculated based on the mean and standard deviation of the corresponding feature in the qualified sample library. The matching process is carried out according to a preset priority order. When the real-time high-dimensional feature vector satisfies all the quantitative criteria for the current defect type, the defect type and the deviation value of the dominant parameter that caused the judgment result are output.

[0022] Secondly, an unsupervised identification system for steel bar welding quality anomalies based on One-Class SVM is provided for performing any of the methods described in the first aspect, including:

[0023] The first acquisition module is used to acquire multi-source time-series data of the standard welding process that has been confirmed as qualified by non-destructive testing through a multi-source sensor acquisition unit deployed on the steel bar welding equipment, and to build a qualified sample library; the multi-source time-series data includes at least welding current, welding voltage and welding area temperature.

[0024] The preprocessing module is used to preprocess and normalize multi-source time-series data, and extract multi-dimensional time-domain and frequency-domain features for current, voltage and temperature time-series data respectively; based on the preset feature importance weights, the extracted time-domain features and frequency-domain features are fused to form a high-dimensional feature vector characterizing the welding process state.

[0025] The input module is used to input the high-dimensional feature vector as a unique training set into the OCSVM model, and to train the OCSVM model by optimizing the parameters of the OCSVM model to construct the parameter feature domain corresponding to the qualified welding process.

[0026] The second acquisition module is used to acquire multi-source time-series data of the welding process under test in real time, and repeatedly execute the preprocessing module to obtain real-time high-dimensional feature vectors; input the real-time high-dimensional feature vectors into the trained OCSVM model, and determine whether the welding process under test deviates from the parameter feature domain based on the decision function. If it deviates, it is determined to be an abnormal welding quality.

[0027] The calling module is used to call the defect identification rule base based on statistical thresholds when the welding quality is determined to be abnormal. It matches the real-time high-dimensional feature vector with the multi-level quantitative criteria in the defect identification rule base, outputs the unique corresponding defect type and parameter deviation value, and generates the corresponding rectification guidance plan.

[0028] Thirdly, a computer storage medium is provided, wherein a computer program is stored therein; when the computer program is run on a computer, the computer causes the computer to perform any of the methods described in the first aspect.

[0029] Fourthly, electronic devices are provided, including:

[0030] Memory, used to store computer programs;

[0031] A processor for executing the computer program to implement the method as described in any of the first aspects.

[0032] The beneficial effects of this invention are:

[0033] 1. This invention features a unique "welding timing parameters + OCSVM" unsupervised identification mode, which does not require welding defect samples. It constructs a normal feature domain only through qualified welding data, solving the pain point of existing technologies that "rely on defect samples and cannot detect in real time". It is suitable for scenarios where defect samples are scarce in actual construction and significantly improves the practicality of detection.

[0034] 2. This invention adopts multi-dimensional time-series feature fusion optimization, which integrates the time-domain and frequency-domain features of three core parameters: welding current, voltage, and temperature. It introduces a weight allocation mechanism to increase the feature ratio of key parameters, avoid misjudgment and omission caused by a single parameter, and improve the identification accuracy by more than 30% compared with traditional methods.

[0035] 3. After the OCSVM anomaly identification, this invention independently sets up a defect identification module based on statistical thresholds and expert rules (without the need for defect sample training). Through parameter deviation pattern matching, it can accurately distinguish five common defect types: incomplete penetration, slag inclusion, cracks, porosity, and lack of fusion. This avoids the limitation of the traditional OCSVM method, which can only output "anomaly" but cannot trace the specific type. It simultaneously outputs the anomaly parameter deviation value and rectification guidance plan, guiding operators to adjust welding parameters in real time, thus solving the limitation of the existing technology that "can only identify defects but cannot trace the source and rectify them".

[0036] 4. This invention is adaptable to various steel bar welding scenarios. Through adaptive optimization of model parameters, it can adapt to different specifications of steel bars (HRB400, HRB500, etc.) and different welding methods (arc welding, flash butt welding, electroslag pressure welding). There is no need to retrain the model for different scenarios, resulting in low deployment costs, simple operation, and adaptability to complex construction site environments.

[0037] 5. This invention is highly real-time and efficient, enabling real-time monitoring and early warning of anomalies during the welding process. It eliminates the need for post-process testing, significantly improving the efficiency of steel bar welding quality control, reducing construction safety hazards, and possessing extremely high engineering application value. Attached Figure Description

[0038] Figure 1 This is a flowchart of the unsupervised identification method for steel bar welding quality anomalies based on One-Class SVM according to the present invention;

[0039] Figure 2 This is a schematic diagram of the multi-dimensional feature fusion process of the present invention;

[0040] Figure 3 This is a schematic diagram illustrating the principle of OCSVM model training and anomaly detection in this invention;

[0041] Figure 4 This is a structural block diagram of the unsupervised identification system for steel bar welding quality anomalies based on One-Class SVM, as described in this invention. Detailed Implementation

[0042] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0043] Example 1:

[0044] To address the technical problems of existing steel bar welding quality inspection methods, which rely on defect samples, are inefficient, cannot identify defects in real time, and are difficult to trace and rectify, Embodiment 1 of this application provides an unsupervised identification method for steel bar welding quality anomalies based on One-Class SVM. This method enables unsupervised, real-time, and accurate detection of welding quality, while simultaneously providing defect tracing and rectification guidance, thereby improving the level of steel bar welding quality control.

[0045] Specifically, such as Figure 1 As shown, the method provided in this application includes:

[0046] S1. Acquisition of qualified welding data: By using a multi-source sensor acquisition unit deployed on the steel bar welding equipment, multi-source time-series data of the standard welding process that has been confirmed as qualified by non-destructive testing are collected to build a qualified sample library; the multi-source time-series data includes at least welding current, welding voltage and welding area temperature.

[0047] The core objective of this step is to obtain normal samples required for unsupervised training, eliminating the need to collect defective samples and adapting to scenarios where defective samples are scarce in actual construction. Current, voltage, and temperature are chosen as the three core parameters because they directly determine welding quality—excessive welding current easily leads to burn-through and cracks, while insufficient current easily leads to incomplete penetration; unstable voltage easily produces slag inclusions and porosity; improper temperature control affects the strength of the welded joint and easily causes cracks. Collecting qualified samples of different specifications of steel bars and different welding methods can improve the model's generalization ability and adapt to the diverse needs of construction sites.

[0048] Specifically, in S1, current sensors, voltage sensors, and infrared temperature sensors are deployed on the steel bar welding equipment to collect multi-source time-series data of qualified welding processes that have been confirmed to be defect-free by ultrasonic testing under different specifications of steel bars and different welding methods. The multi-source time-series data includes welding current, welding voltage, and temperature of the weld joint area, with sampling frequencies of 10Hz, 10Hz, and 5Hz, respectively. Complete time-series data of at least 100 sets of qualified welds are collected to construct a qualified weld sample library.

[0049] Furthermore, the steel reinforcement specifications in S1 include HRB400 and HRB500, and the welding methods include arc welding, flash butt welding, and electroslag pressure welding. The infrared temperature sensor has a monitoring range of 800-1500℃, which is suitable for the steel reinforcement welding temperature range.

[0050] S2. Preprocess and normalize the multi-source time-series data, and extract multi-dimensional time-domain and frequency-domain features for current, voltage and temperature time-series data respectively; based on the preset feature importance weights, fuse the extracted time-domain features and frequency-domain features to form a high-dimensional feature vector characterizing the welding process state.

[0051] S2 includes:

[0052] S201. Preprocess the multi-source time series data, and map the time series data of the qualified samples and real-time samples to the [0,1] interval by using the global minimum and global maximum values ​​of the same type of parameters in the qualified sample library.

[0053] Specifically, data preprocessing includes: removing glitch data and abnormal jump data (caused by sensor errors or instantaneous fluctuations in welding equipment) from multi-source time-series data, and supplementing missing data with equal-time interpolation to ensure data continuity.

[0054] The normalization formula is: in, These are the original timing parameters. These are the normalized parameters. , These are the global minimum and maximum values ​​of this type of parameter in the qualified sample library, respectively.

[0055] S202. Feature Extraction: For the normalized current, voltage, and temperature time-series data, time-domain features including mean, variance, peak value, valley value, fluctuation coefficient, and trend slope are extracted respectively. Frequency-domain features including characteristic frequency, frequency peak value, and frequency domain energy ratio are extracted based on wavelet transform. The time-domain features reflect the overall level, stability, and trend of the parameters; the frequency-domain features reflect the frequency characteristics of parameter fluctuations (e.g., high-frequency current fluctuations correspond to welding instability).

[0056] Specifically, the calculation formulas for each time-domain feature are as follows:

[0057] Mean: ,in The length of the time series. The normalized time-series data reflects the overall level of the parameters;

[0058] variance: This reflects the degree of fluctuation in the parameters;

[0059] Peak value: This reflects the maximum value of the parameter, preventing welding defects caused by excessively high parameters;

[0060] Valley value: This reflects the minimum value of the parameter, avoiding welding defects caused by parameters being too low;

[0061] Volatility coefficient: The coefficient of variation reflects the relative degree of fluctuation of the parameter; the larger the coefficient of variation, the more unstable the parameter.

[0062] Trend slope: The least squares method is used to calculate and reflect the changing trend of parameters (such as a continuous rise or fall in temperature).

[0063] The frequency domain feature extraction described in S202 employs wavelet transform, specifically performing a three-level wavelet decomposition on the normalized time-series data to extract the feature frequencies, frequency peaks, and frequency domain energy proportions of the detail coefficients at each level.

[0064] S203. Using a combination of feature concatenation and weight allocation, the current feature vector, voltage feature vector, and temperature feature vector are fused into the high-dimensional feature vector. ,in The current feature vector has a weight of 40%. This is the voltage feature vector, with a weighting of 25%. This is a temperature feature vector with a weight of 35%.

[0065] The weight allocation is based on the priority setting of factors affecting welding quality—current has the greatest impact on weld penetration and width, followed by temperature, while voltage has a relatively smaller impact. By allocating weights, the characteristic influence of key parameters can be enhanced, feature redundancy can be avoided, and the model's identification accuracy can be improved.

[0066] S3. Input the high-dimensional feature vector as the only training set into the OCSVM model, and train the OCSVM model by optimizing the parameters of the OCSVM model to construct the parameter feature domain corresponding to the qualified welding process.

[0067] In S3, the parameters to be optimized in the OCSVM model include the kernel function and slack variables. With kernel width parameter .

[0068] Specifically, a grid search + 5-fold cross-validation approach is used for parameter optimization, and the RBF kernel is selected as the kernel function, with the following formula: ,in , , For different high-dimensional feature vectors, slack variables The optimization range is [0.01, 0.1]. The RBF kernel has strong nonlinear mapping capability, can adapt to the nonlinear distribution of high-dimensional fusion features, and can accurately construct the feature boundaries of qualified welds. Relaxation variables. The optimization range is [0.01, 0.1], which is used to control the coverage of the normal feature domain and ensure the model's sensitivity to minor anomalies.

[0069] The model training process includes: using the high-dimensional feature vectors obtained from S203 as the unique training set, inputting them into the optimized OCSVM model, and training the model to construct the qualified welding feature domain; the optimization objective function of the OCSVM model is...

[0070]

[0071] The constraints are:

[0072] in, Let be the normal vector of the hyperplane. This is the offset threshold. As slack variables, The number of training samples. It is a high-dimensional mapping function that achieves high-dimensional space mapping through the RBF kernel function without explicit computation.

[0073] S4. Collect multi-source time-series data of the welding process to be tested in real time, and repeat step S2 to obtain real-time high-dimensional feature vectors; input the real-time high-dimensional feature vectors into the trained OCSVM model, and determine whether the welding process to be tested deviates from the parameter feature domain based on the decision function. If it deviates, it is determined to be an abnormal welding quality.

[0074] In S4, the decision function calculates the positional relationship of the real-time feature vector relative to the parameter feature domain based on the support vector coefficients and kernel function; when the output of the decision function is positive, it is determined to be qualified, and when the output is negative, it is determined to be abnormal, triggering an audible and visual warning signal and outputting a pause command to the welding execution mechanism;

[0075] Simultaneously, the deviation of the real-time high-dimensional feature vector from the parameter feature domain is calculated, and the warning level is divided according to the numerical range of the deviation. The warning level includes at least a minor anomaly warning and a severe anomaly warning.

[0076] S5. When the welding quality is determined to be abnormal, the defect identification rule base based on statistical thresholds is called. The real-time high-dimensional feature vector is matched with the multi-level quantitative criteria in the defect identification rule base, and the unique corresponding defect type and parameter deviation value are output, and the corresponding rectification guidance plan is generated.

[0077] In S5, the defect identification rule base contains quantitative criteria for five defect types: incomplete penetration, slag inclusion, crack, porosity, and lack of fusion. Each quantitative criterion is calculated based on the mean and standard deviation of the corresponding feature in the qualified sample library. The matching process is carried out in a preset priority order. When the real-time high-dimensional feature vector satisfies all the quantitative criteria of the current defect type, the defect type and the deviation value of the dominant parameter that caused the judgment result are output.

[0078] Specifically, the defect type discrimination rules include at least the following: Incomplete penetration: the current fluctuation coefficient exceeds the mean of qualified samples plus 2 standard deviations, and the current trend slope is negative and the absolute value is greater than 0.1; Slag inclusion: the current fluctuation coefficient exceeds the mean of qualified samples plus 3 standard deviations, the voltage fluctuation coefficient exceeds the mean of qualified samples plus 2 standard deviations, and the temperature mean is lower than 90% of the mean of qualified samples; Cracks: the temperature rise rate exceeds 200℃ / s or the fall rate is lower than -150℃ / s, and the proportion of the component greater than 50Hz in the frequency domain energy exceeds 0.3; Porosity: the temperature fluctuation coefficient exceeds the mean of qualified samples plus 2.5 standard deviations, and the peak frequency of the temperature frequency domain deviates from the mean of qualified samples by more than 3 standard deviations; Lack of fusion: the voltage fluctuation coefficient exceeds the mean of qualified samples plus 3 standard deviations, and the voltage zero-crossing frequency exceeds 2Hz.

[0079] Furthermore, the specific quantitative criteria for the defect type discrimination rule table include: Incomplete penetration is categorized as a current fluctuation coefficient greater than the mean of qualified samples plus 2 standard deviations and a current trend slope less than -0.1; Slag inclusions are categorized as a current fluctuation coefficient greater than the mean of qualified samples plus 3 standard deviations, a voltage fluctuation coefficient greater than the mean of qualified samples plus 2 standard deviations, and a temperature mean lower than 90% of the mean of qualified samples; Cracks are categorized as a temperature rise rate greater than 200℃ / s or a fall rate less than -150℃ / s, and a frequency domain energy greater than 50Hz accounting for more than 0.3%; Porosity is categorized as a temperature fluctuation coefficient greater than the mean of qualified samples plus 2.5 standard deviations and a temperature frequency domain peak frequency deviating from the mean of qualified samples by more than 3 standard deviations; Lack of fusion is categorized as a voltage fluctuation coefficient greater than the mean of qualified samples plus 3 standard deviations and a voltage zero-crossing frequency greater than 2Hz.

[0080] Example 2:

[0081] Based on Example 1, Example 2 of this application provides a more specific unsupervised identification method for steel bar welding quality anomalies based on One-Class SVM, including:

[0082] S1. By using a multi-source sensor acquisition unit deployed on the steel bar welding equipment, multi-source time-series data of the standard welding process that has been confirmed as qualified by non-destructive testing are collected to build a qualified sample library; the multi-source time-series data includes at least welding current, welding voltage and welding area temperature.

[0083] For example, a current sensor, a voltage sensor, and an infrared temperature sensor are deployed on an arc welding machine. HRB400 steel bars are selected, and arc welding is used to perform qualified welding operations (confirmed to be defect-free by ultrasonic testing). Time-series data of 50 sets of qualified welds are collected. Then, HRB500 steel bars are selected, and time-series data of 30 sets of qualified welds are collected for each of flash butt welding and electroslag pressure welding, for a total of 110 sets of qualified weld samples, to construct a qualified weld sample library. The sampling frequency of the current sensor is 10Hz, the sampling frequency of the voltage sensor is 10Hz, and the sampling frequency of the infrared temperature sensor is 5Hz, with a monitoring range of 800-1500℃.

[0084] S2. Preprocess and normalize the multi-source time-series data, and extract multi-dimensional time-domain and frequency-domain features for current, voltage and temperature time-series data respectively; based on the preset feature importance weights, fuse the extracted time-domain features and frequency-domain features to form a high-dimensional feature vector characterizing the welding process state.

[0085] For example, firstly, based on all qualified welding samples collected by S1, calculate the global minimum value for each type of parameter (current, voltage, temperature). and global maximum value Then, the qualified samples and subsequent real-time samples are all normalized using the same set of global extrema, mapped to... The interval eliminates dimensional differences while retaining the absolute deviation information of the parameter relative to the qualified benchmark.

[0086] Subsequently, six time-domain features—mean, variance, peak value, trough value, fluctuation coefficient, and trend slope—were extracted from the normalized time-series data. Three-level wavelet decomposition was then used to extract three frequency-domain features—feature frequency, frequency peak value, and frequency domain energy percentage—resulting in a total of nine features for each parameter class.

[0087] Then, the feature vectors of current, voltage, and temperature are concatenated and weighted (current 40%, temperature 35%, voltage 25%) to construct a high-dimensional feature vector with dimension 27.

[0088] S3. Input the high-dimensional feature vector as the only training set into the OCSVM model, and train the OCSVM model by optimizing the parameters of the OCSVM model to construct the parameter feature domain corresponding to the qualified welding process.

[0089] For example, grid search + 5-fold cross-validation is used to optimize the OCSVM parameters: RBF kernel. slack variables High-dimensional feature vectors are used as the training set to train the model and construct the qualified welding feature domain.

[0090] S4. Collect multi-source time-series data of the welding process to be tested in real time, and repeat step S2 to obtain a real-time high-dimensional feature vector; input the real-time high-dimensional feature vector into the trained OCSVM model, judge whether the welding process to be tested deviates from the parameter feature domain based on the decision function, and if it deviates, determine that the welding quality is abnormal.

[0091] Specifically, S4 includes:

[0092] S401. Real-time data collection: The sensors deployed in step S1 are used to collect time-series data of current, voltage and temperature in the actual welding process in real time, and steps S21 to S23 are repeated to obtain the high-dimensional feature vector of real-time welding;

[0093] S402. Abnormality judgment: Input the real-time high-dimensional feature vector into the trained OCSVM model, and judge whether the welding quality is abnormal through the decision function; the decision function is:

[0094]

[0095] wherein, is the support vector coefficient; when , the welding is determined to be qualified; when , the welding is determined to be abnormal, an acousto-optic warning is triggered and welding operation is suspended, so as to avoid further expansion of defects;

[0096] S403. Hierarchical warning: The deviation degree between the real-time feature vector and the qualified welding feature domain is calculated through a deviation degree formula, so as to realize hierarchical warning, which is convenient for operators to take different disposal measures according to the warning level; the deviation degree formula is:

[0097]

[0098] The hierarchical judgment criteria are as follows: is normal and qualified; is a slight abnormality, corresponding to a first-level warning, which prompts the operator to check the welding parameters; is a serious abnormality, corresponding to a second-level warning, which requires immediate shutdown for troubleshooting.

[0099] For example, adopting HRB400 specification steel bars to carry out real-time welding in an arc welding mode, collecting real-time current, voltage and temperature data, inputting the data into the trained OCSVM model after preprocessing and feature fusion; when the welding current fluctuates too much (the fluctuation coefficient exceeds the normal range), the model outputs , the deviation degree , the welding is determined to be seriously abnormal, the acousto-optic warning is triggered, and the welding operation is suspended.

[0100] S5. When the welding quality is determined to be abnormal, the defect identification rule base based on statistical thresholds is called. The real-time high-dimensional feature vector is matched with the multi-level quantitative criteria in the defect identification rule base, and the unique corresponding defect type and parameter deviation value are output, and the corresponding rectification guidance plan is generated.

[0101] This step operates independently and sequentially from the OCSVM anomaly detection step. When S402 determines that the welding quality is abnormal ( After that, defect type identification is automatically triggered. Based on the statistical threshold discrimination rule table (see Table 1 below), multi-rule matching is performed on the real-time extracted feature vectors to output a unique corresponding defect type and abnormal parameter deviation value. Specifically:

[0102] First, calculate the statistical characteristics (mean) of each characteristic parameter of the real-time welding process (current fluctuation coefficient, temperature rise rate, voltage zero-crossing frequency, etc.) relative to the qualified sample library. Standard deviation The deviation multiple;

[0103] Then, match the five defect discrimination rules in Table 1 in sequence (in the order of incomplete penetration → slag inclusion → crack → porosity → lack of fusion), and stop when a match is found;

[0104] Finally, the defect type and the deviation value of the dominant parameter that caused the judgment (such as "current fluctuation coefficient exceeds the upper limit of 0.25") are output, and rectification suggestions are given (such as "reduce welding current and stabilize wire feed speed").

[0105] It should be noted that the OCSVM model is trained using only qualified samples and does not participate in defect type identification. Defect type identification relies entirely on a post-hoc rule model based on the statistical characteristics of qualified samples (which can be regarded as a deterministic expert system). Therefore, no defect samples are required for training, but it can still achieve accurate differentiation of multiple types of defects. The discrimination rule table is constructed based on the principles of welding metallurgy and the experience of field process experts. The specific values ​​of each threshold are shown in Table 1.

[0106] Table 1 Defect Type Identification Rules

[0107]

[0108]

[0109] For example, if the deviation mainly comes from the current characteristics and the current fluctuation coefficient exceeds the normal range, it is determined to be incomplete penetration or slag inclusion, and the operator is instructed to adjust the welding current to stabilize the current output.

[0110] For example, after the model determines a welding anomaly, it enters the defect type matching process. The real-time current fluctuation coefficient is calculated. In the qualified sample library , ,but Established; Current trend slope Successful; voltage and temperature characteristics are normal. According to the discrimination rule table, the "incomplete penetration" defect type is matched. The output defect type is incomplete penetration, and the output current deviates from the normal range by 20A (currently 180A; since the normalization uses global extrema, the normalized value of the real-time sample can be directly calculated back to the absolute physical quantity, thus outputting information with practical guidance significance such as "current deviates from the normal range by 20A"). This instructs the operator to adjust the welding current to 160A to stabilize the current output.

[0111] It should be noted that the parts in this embodiment that are the same as or similar to those in Embodiment 1 can be referred to each other, and will not be repeated in this application.

[0112] Example 3:

[0113] Based on Example 2, Example 3 of this application provides an unsupervised identification system for steel bar welding quality anomalies based on One-Class SVM, including:

[0114] The first acquisition module is used to acquire multi-source time-series data of the standard welding process that has been confirmed as qualified by non-destructive testing through a multi-source sensor acquisition unit deployed on the steel bar welding equipment, and to build a qualified sample library; the multi-source time-series data includes at least welding current, welding voltage and welding area temperature.

[0115] The preprocessing module is used to preprocess and normalize multi-source time-series data, and extract multi-dimensional time-domain and frequency-domain features for current, voltage and temperature time-series data respectively; based on the preset feature importance weights, the extracted time-domain features and frequency-domain features are fused to form a high-dimensional feature vector characterizing the welding process state.

[0116] The input module is used to input the high-dimensional feature vector as a unique training set into the OCSVM model, and to train the OCSVM model by optimizing the parameters of the OCSVM model to construct the parameter feature domain corresponding to the qualified welding process.

[0117] The second acquisition module is used to acquire multi-source time-series data of the welding process under test in real time, and repeatedly execute the preprocessing module to obtain real-time high-dimensional feature vectors; input the real-time high-dimensional feature vectors into the trained OCSVM model, and determine whether the welding process under test deviates from the parameter feature domain based on the decision function. If it deviates, it is determined to be an abnormal welding quality.

[0118] The calling module is used to call the defect identification rule base based on statistical thresholds when the welding quality is determined to be abnormal. It matches the real-time high-dimensional feature vector with the multi-level quantitative criteria in the defect identification rule base, outputs the unique corresponding defect type and parameter deviation value, and generates the corresponding rectification guidance plan.

[0119] Furthermore, it also includes an audible and visual early warning module, which triggers an audible and visual early warning and sends a signal to the welding equipment to suspend welding operations when a welding abnormality is detected.

[0120] For example, the first acquisition module includes a current sensor (model: ACS712), a voltage sensor (model: LV25-P), and an infrared temperature sensor (model: MLX90614); the input module and the second acquisition module are implemented using a microcontroller (model: STM32F103), which can be embedded in the welding equipment to realize real-time data processing and early warning; the audible and visual early warning module uses LED warning lights and a buzzer. When an abnormal warning is triggered, the LED light flashes, the buzzer sounds an alarm, and a signal is sent to the welding equipment to suspend the welding operation.

[0121] Tests have shown that the accuracy rate of this system in identifying welding quality anomalies reaches over 95%, with an early warning lead time of 1-2 seconds. It can effectively identify common welding defects such as incomplete penetration, slag inclusion, and cracks, without requiring defect samples. It is adaptable to different specifications of steel bars and different welding methods, and is simple to operate and has low deployment costs. It can be widely used in steel bar welding quality control scenarios in civil engineering projects such as buildings and bridges.

[0122] To verify the accuracy of the method of this invention in identifying five common welding defects, a total of 175 sets of field welding test samples were collected, including:

[0123] Qualified samples: 50 sets (confirmed by ultrasonic testing and tensile testing);

[0124] Incomplete penetration samples: 25 groups (insufficient penetration confirmed by metallographic microscopy);

[0125] Slag inclusion samples: 25 groups (internal slag inclusions confirmed by X-ray inspection);

[0126] Crack samples: 25 groups (confirmed by dye penetrant testing and metallography);

[0127] Stomatal samples: 25 groups (confirmed by ultrasonic testing);

[0128] Unfused samples: 25 groups (confirmed by cross-sectional inspection).

[0129] All defective samples were excluded from OCSVM model training (OCSVM uses only 50 lattice samples as the training set). The verification process is as follows:

[0130] The time-series data of 175 samples were processed according to steps S21-S23 to obtain high-dimensional feature vectors;

[0131] First, anomaly / normal classification is performed using the OCSVM model;

[0132] For samples that are determined to be abnormal, the specific defect type is output by the defect type identification submodule (statistical threshold rule table, i.e., Table 1).

[0133] To demonstrate the superiority of the method of this invention, the following two comparative methods are provided:

[0134] Traditional OCSVM: Only uses time-domain statistical features (mean, variance, peak, valley), without frequency-domain feature fusion and weight allocation, and the kernel function and parameter range are the same;

[0135] Supervised SVM: It adopts the same 27-dimensional features as the present invention, but uses all defective samples (125 groups) and qualified samples (50 groups) for supervised training (binary classification: normal / abnormal) as a reference for the theoretical upper limit of performance.

[0136] All experiments were conducted on an embedded platform (STM32F103, 72MHz). Real-time latency was defined as the time from the acquisition of the last data point to the output of the discrimination result (excluding sensor sampling latency). All metrics were calculated using 5-fold cross-validation.

[0137] The verification results are presented in the form of a confusion matrix as shown in Table 2.

[0138] Table 2. Confusion Matrix for Defect Type Identification in the Method of the Invention

[0139]

[0140] The performance indicators are compared as shown in Tables 3 and 4.

[0141] Table 3 Performance Index Comparison Table

[0142]

[0143] Table 4 Performance Comparison of Anomaly Detection Phase (OCSVM Binary Classification)

[0144]

[0145] The method of this invention has an accuracy of (48+119) / 175 = 95.4%, a recall of 119 / 125 = 95.2%, and a specificity of 48 / 50 = 96.0% in the anomaly detection stage.

[0146] In addition, this application also provides the performance of the defect identification stage (rule module multi-classification), as shown in Table 5.

[0147] Table 5 Performance of the Defect Identification Stage (Multi-classification of Rule Module)

[0148]

[0149] As can be seen, the accuracy of this invention in the anomaly detection stage (95.4%) is 13.1 percentage points higher than that of the traditional OCSVM (82.3%), representing a relative improvement of approximately 15.9%. Simultaneously, its performance is close to that of supervised SVM (96.7%), without requiring any defect samples for training. In the defect identification stage, the rule module achieves a macro-average recall of 88.0% for the five types of defects, with cracks and lack of fusion showing the best identification performance (recall rate 92%). Regarding real-time latency, the average latency of this invention is 55ms, meeting the real-time monitoring requirements of the welding process (the timescale of welding process feature changes is approximately on the order of seconds). The above verification results demonstrate that this invention, without requiring defect samples to train the OCSVM model, can still achieve high-precision defect type identification through a post-implemented statistical threshold rule module, meeting the quality control requirements of construction sites.

[0150] It should be noted that the system provided in this embodiment is the same as the system corresponding to the method provided in embodiment 2. Therefore, the parts that are the same or similar to those in embodiment 2 in this embodiment can be referred to each other, and will not be described again in this application.

Claims

1. An unsupervised identification method for steel bar welding quality anomalies based on One-Class SVM, characterized in that, include: S1. By deploying a multi-source sensor acquisition unit on the steel bar welding equipment, collect multi-source time-series data of the standard welding process that has been confirmed as qualified by non-destructive testing, and build a qualified sample library; the multi-source time-series data includes at least welding current, welding voltage and welding area temperature. S2. Preprocess and normalize the multi-source time series data, and extract multi-dimensional time domain features and frequency domain features for current, voltage and temperature time series data respectively. Based on the preset feature importance weights, the extracted time-domain features and frequency-domain features are fused to form a high-dimensional feature vector characterizing the welding process state. S3. Input the high-dimensional feature vector as the only training set into the OCSVM model, and train the OCSVM model by optimizing the parameters of the OCSVM model to construct the parameter feature domain corresponding to the qualified welding process. S4. Collect multi-source time-series data of the welding process to be tested in real time, and repeat step S2 to obtain real-time high-dimensional feature vectors; input the real-time high-dimensional feature vectors into the trained OCSVM model, and determine whether the welding process to be tested deviates from the parameter feature domain based on the decision function. If it deviates, it is determined to be an abnormal welding quality. S5. When the welding quality is determined to be abnormal, the defect identification rule base based on statistical thresholds is called. The real-time high-dimensional feature vector is matched with the multi-level quantitative criteria in the defect identification rule base, and the unique corresponding defect type and parameter deviation value are output, and the corresponding rectification guidance plan is generated.

2. The unsupervised identification method for steel bar welding quality anomalies based on One-Class SVM according to claim 1, characterized in that, S2 include: S201. Using the global minimum and global maximum values ​​of the same type of parameters in the qualified sample library, the time series data of the qualified samples and real-time samples are mapped to the [0,1] interval; S202. For the normalized current, voltage and temperature time series data, extract time domain features including mean, variance, peak value, valley value, fluctuation coefficient and trend slope respectively, and extract frequency domain features including characteristic frequency, frequency peak value and frequency domain energy ratio based on wavelet transform. S203. Using a combination of feature concatenation and weight allocation, the current feature vector, voltage feature vector, and temperature feature vector are fused into the high-dimensional feature vector, with the current feature having the largest weight.

3. The unsupervised identification method for steel bar welding quality anomalies based on One-Class SVM according to claim 2, characterized in that, In S3, the parameters to be optimized in the OCSVM model include the kernel function and slack variables. With kernel width parameter .

4. The unsupervised identification method for steel bar welding quality anomalies based on One-Class SVM according to claim 3, characterized in that, In S4, the decision function calculates the positional relationship between the real-time feature vector and the parameter feature domain based on the support vector coefficients and kernel function; when the output of the decision function is positive, it is determined to be qualified; when the output is negative, it is determined to be abnormal, triggering an audible and visual warning signal and outputting a pause command to the welding execution mechanism. Simultaneously, the deviation of the real-time high-dimensional feature vector from the parameter feature domain is calculated, and the warning level is divided according to the numerical range of the deviation. The warning level includes at least a minor anomaly warning and a severe anomaly warning.

5. The unsupervised identification method for steel bar welding quality anomalies based on One-Class SVM according to claim 4, characterized in that, In S5, the defect identification rule base contains quantitative criteria for five defect types: incomplete penetration, slag inclusion, crack, porosity, and lack of fusion. Each quantitative criterion is calculated based on the mean and standard deviation of the corresponding feature in the qualified sample library. The matching process is carried out in a preset priority order. When the real-time high-dimensional feature vector satisfies all the quantitative criteria of the current defect type, the defect type and the deviation value of the dominant parameter that caused the judgment result are output.

6. An unsupervised identification system for abnormal steel bar welding quality based on One-Class SVM, characterized in that, For performing the method as described in any one of claims 1 to 5, comprising: The first acquisition module is used to acquire multi-source time-series data of the standard welding process that has been confirmed as qualified by non-destructive testing through a multi-source sensor acquisition unit deployed on the steel bar welding equipment, and to build a qualified sample library; the multi-source time-series data includes at least welding current, welding voltage and welding area temperature. The preprocessing module is used to preprocess and normalize multi-source time-series data, and extract multi-dimensional time-domain and frequency-domain features for current, voltage and temperature time-series data respectively; based on the preset feature importance weights, the extracted time-domain features and frequency-domain features are fused to form a high-dimensional feature vector characterizing the welding process state. The input module is used to input the high-dimensional feature vector as a unique training set into the OCSVM model, and to train the OCSVM model by optimizing the parameters of the OCSVM model to construct the parameter feature domain corresponding to the qualified welding process. The second acquisition module is used to acquire multi-source time-series data of the welding process under test in real time, and repeatedly execute the preprocessing module to obtain real-time high-dimensional feature vectors; input the real-time high-dimensional feature vectors into the trained OCSVM model, and determine whether the welding process under test deviates from the parameter feature domain based on the decision function. If it deviates, it is determined to be an abnormal welding quality. The calling module is used to call the defect identification rule base based on statistical thresholds when the welding quality is determined to be abnormal. It matches the real-time high-dimensional feature vector with the multi-level quantitative criteria in the defect identification rule base, outputs the unique corresponding defect type and parameter deviation value, and generates the corresponding rectification guidance plan.

7. A computer storage medium, characterized in that, The computer storage medium stores a computer program; when the computer program is run on the computer, it causes the computer to perform the method described in any one of claims 1 to 5.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1 to 5.

Images