A method and system for identifying welding defects in sealing components of engineering vehicles

Abstract

This invention discloses a method and system for identifying welding defects in sealing components of engineering vehicles, relating to the field of welding defect detection. The method includes: collecting a global monitoring dataset of the tamping device housing during railway tamping vehicle operation and generating local monitoring feature vectors through dimensionality reduction; using a welding defect anomaly discriminator to determine the presence of defects; if an anomaly warning mechanism is triggered, collecting the global monitoring dataset within a dynamic monitoring window, inputting it into a welding defect anomaly identification model, predicting the degree of welding defect anomalies in several key parts of the tamping device housing, and outputting a predicted defect probability distribution; identifying high-risk defect locations based on the predicted defect probability distribution, and setting an adaptive defect detection mechanism according to the predicted defect probability to identify welding defects in high-risk defect locations. This solves the problems of low detection efficiency, low detection accuracy, and poor adaptability of detection schemes to different detection locations in existing welding defect detection technologies.

Description

Technical Field

[0001] This invention relates to the field of welding defect detection, and specifically to a method and system for identifying welding defects in sealing components of engineering vehicles. Background Technology

[0002] As a core piece of equipment for railway track maintenance, the tamping machine's tamping device housing, a key sealing component, houses high-speed rotating parts such as vibrators and eccentric shafts. During operation, it is subjected to high-frequency alternating loads and complex dynamic stresses. After long-term service, the welded areas of the housing are highly susceptible to welding defects such as fatigue cracks. Failure of these defects can lead to lubricant leakage, damage to the housing's sealing performance, and even serious safety accidents.

[0003] Currently, the detection of welding defects in tamping device housings mainly includes dynamic sensor detection and static welding defect detection. Dynamic monitoring has low detection accuracy; while static detection uses simple threshold alarms, which easily generate a large number of false alarms and missed alarms, resulting in low detection efficiency and the inability to achieve early warning. Summary of the Invention

[0004] This application provides a method and system for identifying welding defects in sealing components of engineering vehicles, addressing the problems of low efficiency, low accuracy, and poor adaptability of detection schemes to different detection locations in existing technologies.

[0005] In view of the above problems, this application provides a method and system for identifying welding defects in sealing components of engineering vehicles.

[0006] In a first aspect, this application provides a method for identifying welding defects in sealing components of engineering vehicles, the method comprising: Collect a global monitoring dataset of the tamping device housing during the operation of the railway tamping machine, and perform dimensionality reduction on the global monitoring dataset to generate local monitoring feature vectors; The local monitoring feature vector is input into a pre-constructed welding defect anomaly discriminator to determine whether there is a defect anomaly. If the abnormal warning mechanism is triggered, the global monitoring dataset in the dynamic monitoring window is collected, the pre-built welding defect abnormality identification model is input, the degree of welding defect abnormality of several key parts of the tamping device box is predicted, and the predicted defect probability distribution is output. The predicted defect probability distribution includes several predicted defect probabilities of several key parts. Based on the predicted defect probability distribution, high-risk defect locations are identified, and an adaptive defect detection mechanism is set according to the predicted defect probability corresponding to the high-risk defect locations to identify welding defects in the high-risk defect locations.

[0007] Secondly, the present invention provides a welding defect identification system for sealing components of engineering vehicles, the system comprising: The feature vector acquisition module is used to collect the global monitoring dataset of the tamping device box when the railway tamping machine is working, and to reduce the dimensionality of the global monitoring dataset to generate local monitoring feature vectors. The defect anomaly discrimination module is used to input the local monitoring feature vector into a pre-constructed welding defect anomaly discriminator to determine whether there is a defect anomaly. The defect probability prediction module is used to collect the global monitoring dataset in the dynamic monitoring window if the abnormal warning mechanism is triggered, input the pre-built welding defect abnormality identification model, predict the degree of welding defect abnormality of several key parts of the tamping device box, and output the predicted defect probability distribution, wherein the predicted defect probability distribution includes several predicted defect probabilities of several key parts. The defect detection mechanism configuration module is used to identify and determine high-risk defect locations based on the predicted defect probability distribution, and to set an adaptive defect detection mechanism according to the predicted defect probability corresponding to the high-risk defect locations to identify welding defects in the high-risk defect locations.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application first performs dimensionality reduction processing on the global monitoring dataset of the tamping device box to generate local monitoring feature vectors, achieving information condensation and feature extraction from massive high-dimensional monitoring data, laying a data foundation for subsequent analysis. Second, the local monitoring feature vectors are input into the welding defect anomaly discriminator for rapid judgment and first-level rapid screening, only determining whether anomalies exist, avoiding continuous operation that would waste computing power. Third, when the anomaly warning mechanism is triggered, the global monitoring dataset within the dynamic monitoring window is collected and input into the welding defect anomaly identification model, outputting the predicted defect probability distribution of multiple key parts, providing a scientific basis for decision-making. Finally, high-risk defect parts are identified by predicting the defect probability distribution, and corresponding defect detection mechanisms are configured to ensure that high-risk parts receive in-depth detection commensurate with their risk level, improving the detection efficiency and accuracy of welding defect identification, realizing the organic integration of dynamic working condition monitoring and static welding defect detection, and configuring the optimal welding defect identification scheme. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a method for identifying welding defects in sealing components of engineering vehicles according to this application. Figure 2 This is a schematic diagram of the structure of a welding defect identification system for sealing components of engineering vehicles according to this application.

[0010] In the attached diagram, the components represented by each number are as follows: Feature vector acquisition module 11, defect and anomaly discrimination module 12, defect probability prediction module 13, and defect detection mechanism configuration module 14. Detailed Implementation

[0011] This application provides a method and system for identifying welding defects in sealing components of engineering vehicles, addressing the problems of low efficiency, low accuracy, and poor adaptability of detection schemes to different detection locations in existing technologies.

[0012] The present invention will now be described in detail with reference to the accompanying drawings.

[0013] Example 1, as Figure 1 As shown, this application provides a method for identifying welding defects in sealing components of engineering vehicles, the method comprising: S10: Collect the global monitoring dataset of the tamping device box when the railway tamping machine is working, and reduce the dimensionality of the global monitoring dataset to generate local monitoring feature vectors; In this embodiment, the tamping device housing is the core sealing component of the railway tamping car. Since the housing itself bears a huge high-frequency alternating load, the weld is a weak link in the structural strength. The railway tamping car is a large mechanical equipment used for railway line maintenance.

[0014] Specifically, all relevant monitoring data from the tamping device housing of the railway tamping machine are collected to construct a global monitoring dataset. However, due to the high dimensionality and large volume of this dataset, direct analysis would impose a significant computational burden. Therefore, dimensionality reduction is performed on the global monitoring dataset through information compression and feature extraction, eliminating redundant information and noise to obtain low-dimensional local monitoring feature vectors.

[0015] In step S10 of the method provided in this embodiment of the invention, the global monitoring dataset is collected based on a preset global monitoring index set. The preset global monitoring index set includes vibration monitoring index, temperature monitoring index, strain monitoring index, and working condition monitoring index. The vibration monitoring index includes the triaxial vibration acceleration of key measuring points, the effective value of vibration velocity of key measuring points, and the peak value of vibration acceleration of key measuring points. The temperature monitoring index includes the temperature of multiple measuring points on the surface of the housing and the lubricating oil temperature. The strain monitoring index includes the strain value of measuring points near key welds. The working condition monitoring index includes tamping depth, clamping force, operating speed, and operating mode.

[0016] In this embodiment, the vibration monitoring indicators are parameters that reflect the mechanical vibration state of the tamping device housing and its components, collected by devices such as accelerometers. The presence and propagation of welding defects can significantly alter the local stiffness of the structure, thereby affecting its vibration response characteristics. Therefore, vibration signals are the most sensitive and commonly used indicators for judging the health status of a structure.

[0017] Vibration monitoring indicators include triaxial vibration acceleration at key measuring points, effective value of vibration velocity at key measuring points, and peak value of vibration acceleration at key measuring points. Triaxial vibration acceleration at key measuring points involves installing triaxial accelerometers at the locations most sensitive to vibration response, measuring the instantaneous vibration acceleration in the three orthogonal directions (X, Y, and Z), reflecting the vibration intensity and frequency components in each direction. Effective value of vibration velocity at key measuring points is the value obtained by calculating the root mean square of the vibration velocity signal at the key measuring point; it is typically used to assess the energy level or intensity of vibration. Peak value of vibration acceleration at key measuring points is the maximum value of vibration acceleration at the key measuring point within a specific monitoring period.

[0018] For example, there is a key measuring point A. The vibration state of the tamping device box is monitored, and the three-dimensional vibration acceleration of the key measuring point A is obtained, such as: X-axis 12.5m / s², Y-axis 8.3m / s², Z-axis 15.2m / s². The mean value of the vibration velocity signal is calculated to obtain the effective value of the vibration velocity of the key measuring point A as 5.6mm / s, and the peak value of the vibration acceleration of the key measuring point can be 28.3m / s².

[0019] Temperature monitoring indicators are parameters that reflect the thermal state of the enclosure and its internal medium, collected by temperature sensors or infrared thermal imagers. Abnormal temperature rises may be caused by increased friction, poor cooling, or energy dissipation during crack propagation.

[0020] Temperature monitoring indicators include the temperature at multiple measuring points on the housing surface and the lubricating oil temperature. The multiple measuring points on the housing surface are achieved by placing temperature sensors at different locations on the housing surface to monitor the local temperature distribution. Changes in temperature difference may reveal abnormal localized heating. The lubricating oil temperature monitors the temperature of the lubricating oil inside the tamping device housing. It reflects the frictional heat generation of the entire transmission system; an abnormally high temperature may indicate lubrication failure or damage to internal components. For example, a key measuring point A has a measured temperature of 45℃ and a lubricating oil temperature of 50℃.

[0021] Strain monitoring parameters are parameters that directly measure the minute deformations occurring in key parts of the enclosure under load, using devices such as resistance strain gauges or fiber optic grating sensors. They directly reflect local stress levels and have extremely high sensitivity for monitoring crack initiation and propagation in stress concentration areas. Strain values ​​at measuring points near critical welds involve deploying strain sensors in the area near welds most prone to welding defects to monitor strain changes in stress concentration areas during tamping operations. Abnormal increases in strain values ​​or changes in fluctuation patterns are often precursors to defects such as cracks. For example, the strain value measured at measuring points near the critical welds of the tamping device enclosure of a railway tamping car was 235µε.

[0022] The working condition monitoring indicators are parameters that reflect the current operating status and load conditions of the vehicle, obtained directly from the tamping machine control system or measured by independent sensors. These indicators include tamping depth, clamping force, operating speed, and operating mode. Tamping depth is the depth to which the tamping device is inserted into the track bed, directly affecting the load on the box girder. Clamping force refers to the force with which the tamping device clamps the rail or sleeper, and is the direct external load that causes strain in the box girder. Operating speed refers to the forward speed of the tamping machine, affecting the frequency of dynamic load application. Operating modes, such as track lifting, track shifting, and tamping, correspond to different load characteristics and stress patterns. For example, a key measuring point A measures the working condition indicators as follows: tamping depth 325mm, clamping force 11.2kN, operating speed 0.48km / h, and operating mode is track lifting mode.

[0023] In step S10 of the method provided in this embodiment of the invention, the construction step of the welding defect anomaly identification model includes: Several global monitoring datasets were collected, and the proportion of welding defect abnormal events in each key part of the tamping device box under the same working conditions was calculated for each global monitoring dataset. The proportion of abnormal events was recorded as the sample defect probability corresponding to the key part, and several sample defect probabilities were obtained. The key parts of the tamping device box include the weld connecting the vibrator mounting base and the box side plate, the weld at the intersection of the box stiffener plate and the bottom plate, the radial weld around the bearing seat, the fillet weld connecting the front and rear end plates and the side plate of the box, and the weld near the lubricating oil passage. The sample training dataset is constructed based on the aforementioned global monitoring dataset and the corresponding sample defect probabilities. The deep neural network is trained under supervision using the sample training dataset until the preset convergence condition is met, thereby generating a welding defect anomaly recognition model.

[0024] In this embodiment, a welding defect anomaly is an event confirmed during the service of the tamping device housing through post-service non-destructive testing methods such as ultrasonic testing and magnetic particle testing, revealing a welding defect in a critical area. The anomaly event percentage is the frequency of welding defect anomalies occurring in a critical area within a large amount of historical operational data under identical working conditions; that is, the probability of defect occurrence in that area under specific working conditions. The sample defect probability is the anomaly event percentage used as the label value corresponding to that sample. Since each sample has its corresponding working condition, its defect probability is a condition-dependent conditional probability.

[0025] Specifically, a global monitoring dataset of historical operational data is collected, where each sample records the vibration, temperature, strain response, and corresponding operating parameters of the tamping device housing at a specific moment. Subsequently, based on the post-incident monitoring records, the frequency of welding defect anomalies occurring in key components of each sample under the same operating conditions is statistically analyzed, and the percentage of anomalies for each component under the same conditions is calculated. Then, the percentage of anomalies is used as the sample defect probability for the corresponding component, assigning a defect risk label to each original monitoring data sample.

[0026] The key components include the structural mechanics analysis, finite element simulation, and on-site fault statistics of the tamping device housing, which identified the weakest areas most prone to fatigue cracks or welding defects. These weakest areas include: the weld connecting the vibrator mounting base and the housing side plate, the weld at the intersection of the housing stiffener plate and the bottom plate, the radial weld around the bearing housing, the fillet weld connecting the front and rear end plates and the side plates, and the weld near the lubrication oil passages.

[0027] The connection between the vibrator mounting base and the side plate of the box can withstand the high-frequency alternating excitation force generated by the vibrator, which is one of the most concentrated areas of fatigue load and is prone to fatigue cracks; the junction of the box stiffener plate and the bottom plate is the point where the structural stiffness changes abruptly, the stress concentration is severe, and it is subjected to complex bending and torsional stresses during tamping operations.

[0028] The bearing housing is where the drive shaft bearing is installed and bears radial loads. The stress state of the welds around it is complex, and failure will lead to a decrease in the bearing support stiffness and trigger a chain of failures. The connection angle between the front and rear end plates and the side plates of the housing forms the basic framework of the housing and bears the overall bending and torsional loads. Its integrity is crucial to the overall stiffness of the housing. If there are microcracks in the welds near the lubrication oil passages, it may affect the structural strength and may even lead to lubrication oil leakage, damage the sealing performance of the housing, and cause serious consequences such as lubrication failure.

[0029] For example, suppose there is a global monitoring dataset of 1000 samples collected under the working conditions of tamping depth 300-350mm and clamping force 10-12kN. Through post-operational inspection records, it was found that within the time period corresponding to the 1000 operations, there were 20 events of cracks occurring in the vibrator mounting seat weld. For the 1000 samples, the sample defect probability corresponding to the vibrator mounting seat weld is 20 / 1000 = 2%. The weld at the junction of the stiffening plate and the base plate had 30 defects, labeled as 3%. Finally, the sample labels for each sample are obtained as follows: [Vibrator mounting seat weld 2%, weld at the junction of the stiffening plate and the base plate 3%, weld around the bearing seat 0.5%, fillet weld connecting the front and rear end plates and the side plates of the housing 3%, weld near the lubrication oil passage 0.6%].

[0030] Secondly, several global monitoring datasets and corresponding sample defect probabilities are paired and organized to form a dataset for training machine learning models, namely the sample training dataset, which provides standard input-output pairs for subsequent deep neural network training.

[0031] Finally, the sample training dataset is input into the deep neural network for supervised training. During training, the network takes the global monitoring dataset as input and calculates the predicted values ​​of defect probabilities for each key part through forward propagation. Then, the error between the model's predicted values ​​and the actual sample defect probabilities is calculated using a loss function. Next, the error is propagated forward layer by layer using the backpropagation algorithm, and the network's connection weights are updated. This process is repeated iteratively, while the model's generalization ability is monitored using a validation set. Training terminates when the model's performance on the validation set no longer improves or reaches other preset convergence conditions. The deep neural network with optimal parameters is the trained welding defect anomaly recognition model.

[0032] For example, a fully connected neural network can be used to construct a welding defect anomaly recognition model. Its structure includes an input layer, a hidden layer, and an output layer. The input layer is used to receive input data, such as an input layer with 50 nodes. The hidden layer is used to perform complex nonlinear transformations to extract deep features of the data, such as three hidden layers with 128, 64, and 32 nodes respectively. The output layer is used to generate the final prediction result, such as an output layer with 5 nodes, corresponding to the defect probability of 5 key parts.

[0033] A welding defect anomaly identification model is trained using a fully connected neural network. The global monitoring dataset and sample defect probabilities are used as inputs, with the sample defect probabilities serving as the supervision target. The training and validation sets are divided in a 7:3 ratio. An initial learning rate of 0.001 is set, and the Adam optimizer is used for training. Forward propagation involves weighted summation of the input data with weights and biases, followed by a nonlinear transformation using an activation function. Then, backpropagation calculates the error gradient, and gradient descent updates the weights and biases in the network to reduce the model's prediction error. After training, the model's performance on the reserved validation set is evaluated. If the mean absolute error between the prediction results and the supervision target decreases to within 0.05, the requirement is met, and the welding defect anomaly identification model is obtained. For example, after 500 iterations, if the model's loss function on the validation set no longer decreases, the preset convergence condition is met, and the welding defect anomaly identification model is obtained.

[0034] In step S10 of the method provided in this embodiment of the invention, the dimensionality reduction of the global monitoring dataset is performed to generate local monitoring feature vectors, including: Using the global monitoring dataset of the sample training dataset as the training sample, the preset data compression ratio as the constraint condition of the encoder output dimension, and minimizing the difference between the global monitoring dataset of the sample and the reconstructed dataset after encoder-decoder reconstruction as the optimization objective, the encoder and decoder are trained. During training, the encoder compresses the high-dimensional global monitoring dataset of samples into a low-dimensional feature representation that conforms to the preset data compression ratio, and the decoder reconstructs the low-dimensional feature representation into a reconstructed dataset with the same dimension as the original input. The trained encoder is used as a data dimensionality reduction plugin to reduce the dimensionality of the real-time collected global monitoring dataset and generate local monitoring feature vectors.

[0035] In this embodiment, the preset data compression ratio refers to the ratio between the original data dimension and the dimension of the reduced feature vector. For example, if the original data is 50-dimensional and the compression ratio is 5:1, the reduced feature vector will be 10-dimensional. The preset data compression ratio is a constraint on the encoder's output dimension, determining the strength of the data compression. The encoder maps high-dimensional input data to a low-dimensional feature space. Through multiple nonlinear transformations, the encoder gradually extracts the core features from the input data, discarding redundant information and noise.

[0036] The decoder is a neural network structure symmetrical to the encoder. It reconstructs the low-dimensional feature representation output by the encoder back into the original high-dimensional space, recovering as much of the original data as possible from the compressed features. The reconstructed dataset is a dataset with the same dimension as the original input. By comparing the difference between the reconstructed dataset and the original sample global monitoring dataset, the degree to which the encoder-decoder retains the core information of the data can be evaluated.

[0037] Specifically, the training dataset is first used as input, and the encoder and decoder are trained with the goal of minimizing the reconstruction loss, constrained by a preset data compression ratio. During training, a corresponding preset data compression ratio needs to be set, which determines the feature dimension after dimensionality reduction. The encoder receives the high-dimensional original input and compresses it into a low-dimensional feature representation that meets the required compression ratio through layer-by-layer abstraction and feature extraction.

[0038] Subsequently, the decoder receives low-dimensional features and reconstructs them back to the original dimensions. The training objective function is to minimize the difference between the original input and the reconstructed output, making the reconstructed data as close as possible to the original data. Through backpropagation, the weights within the encoder and decoder are adjusted, enabling the encoder to learn to extract core information from the original data, and the decoder to reconstruct the data from this core information. After multiple iterations of training, the loss function converges to a small value, resulting in the trained encoder and decoder.

[0039] For example, suppose the training dataset contains 10,000 sets of global monitoring data, each set containing 50 dimensions of monitoring indicators. The preset data compression ratio is set to 5:1, meaning the reduced feature dimension is 10. An autoencoder is constructed with an input layer of 50 nodes, two hidden layers of 128 and 64 nodes respectively, and an output layer of 10 nodes. The decoder is symmetrical to the encoder, with a final output layer of 50 nodes. 10,000 sets of 50-dimensional data are input into the network for training, using the mean squared error between the original and reconstructed data as the loss function. After 500 iterations of training, the loss function converges to a small value, the encoder learns to compress the 50-dimensional original data into 10-dimensional core features, and the decoder can recover the original data from the 10-dimensional features.

[0040] Secondly, in each iteration, the high-dimensional global monitoring dataset is input into the encoder. After nonlinear transformation and dimensionality reduction by a multi-layer neural network, it is compressed into a low-dimensional feature representation that meets the preset data compression ratio, retaining only the core features that best characterize the box state. Subsequently, the low-dimensional features are fed into the decoder. The decoder, through a symmetrical network structure, gradually restores the low-dimensional features, ultimately generating a reconstructed dataset with the exact same dimensions as the original input. By comparing the differences between the reconstructed data and the original data, the degree to which the low-dimensional features retain the original information is evaluated, and the parameters of the encoder and decoder are adjusted accordingly.

[0041] Finally, the trained encoder is used as a standardized data processing module, namely the data dimensionality reduction plugin, which can receive high-dimensional monitoring data collected in real time as input and output the corresponding low-dimensional feature representation. The real-time global monitoring dataset is the latest monitoring data collected in real time from the preset global monitoring index set during the actual operation of the tamping machine. The local monitoring feature vector is a low-dimensional feature vector obtained by dimensionality reduction of the real-time data through the data dimensionality reduction plugin.

[0042] After completing the offline training of the autoencoder, the trained encoder part is isolated, encapsulated into a data dimensionality reduction plugin, and deployed into the real-time monitoring system. This plugin receives real-time high-dimensional monitoring data as input and outputs the corresponding low-dimensional feature representation. When the tamping machine is operating, it collects a real-time global monitoring dataset. Each time a new set of high-dimensional monitoring data is collected, it is input into the data dimensionality reduction plugin. Through nonlinear mapping parameters, the high-dimensional input data is converted into a low-dimensional local monitoring feature vector, which is then output to the subsequent processing module.

[0043] For example, the trained encoder can be packaged into a data dimensionality reduction plugin and deployed into the on-board monitoring system of the tamping machine. Real-time monitoring data was collected for key monitoring point A, such as: [Vibration acceleration X-axis 12.5m / s², Y-axis 8.3m / s², Z-axis 15.2m / s², effective velocity 5.6mm / s, peak velocity 28.3m / s², temperature at measuring point 1 45.3°C, temperature at measuring point 2 47.1°C, lubricating oil temperature 52.8°C, strain value 235με, tamping depth 325mm, clamping force 11.2kN, operating speed 0.48km / h, track-starting operation mode]. The 50 monitoring data points from the five monitoring points were input into the data dimensionality reduction plugin to calculate the corresponding local monitoring feature vector: [0.23, -0.56, 0.78, 0.12, -0.33, 0.45, -0.67, 0.89, -0.21, 0.34].

[0044] In step S10 of the method provided in this embodiment of the invention, the preset data compression ratio is determined through the following steps: Principal component analysis is performed on each monitoring indicator in the preset global monitoring indicator set to calculate the eigenvalues ​​and variance contribution rate of each principal component, and the variance contribution rates are sorted from high to low. The number of principal components required to accumulate the variance contribution rate to reach the preset contribution threshold is determined as the basic compression dimension, and the range of values ​​for the candidate compression ratio is determined based on the basic compression dimension. Within the range of values, for each candidate compression ratio, the data compression performance at that compression ratio is evaluated. The data compression performance is measured by the reconstruction error of the original data by the dimensionality-reduced data. The smaller the reconstruction error, the higher the data compression performance. Among all candidate compression ratios that satisfy the condition that the reconstruction error is less than the preset reconstruction error threshold, the candidate compression ratio with the largest value is selected as the most suitable data compression ratio, so as to minimize the number of monitoring indicators while ensuring that the data compression performance meets the standard. The optimal data compression ratio is used as the preset data compression ratio.

[0045] In this embodiment, principal component analysis (PCA) is a classic linear dimensionality reduction and feature extraction method. Through orthogonal transformation, it converts multiple potentially correlated indicators into a few linearly uncorrelated composite indicators, called principal components. Each principal component is a linear combination of the original indicators and can reflect some information in the original data. Eigenvalues ​​are unique features in PCA; each principal component corresponds to one eigenvalue, and its magnitude represents the variance of the original data that the principal component can explain. Larger eigenvalues ​​indicate more information contained in the principal component. The variance contribution rate is the proportion of the eigenvalue of a single principal component to the sum of all principal component eigenvalues.

[0046] First, a large amount of historical monitoring data was collected, and principal component analysis was performed on each monitoring indicator. Through mathematical transformation, the original interrelated indicators were converted into linearly independent principal components, and the eigenvalues ​​and variance contribution rates of each principal component were calculated. The magnitude of the eigenvalues ​​reflects the amount of information carried by the principal component, and the variance contribution rate of the principal component reflects the proportion of the total variance of the original data. Subsequently, all principal components were sorted in descending order of variance contribution rate.

[0047] Secondly, the preset contribution threshold is a pre-defined target value for the cumulative variance contribution rate, representing the proportion of the original information retained after dimensionality reduction to the total information. It is typically set to 80%, 85%, or 90%, etc. Based on the preset contribution threshold, the variance contribution rates are accumulated in descending order, starting with the principal component ranked first, and then sequentially accumulating their variance contribution rates until the preset contribution threshold is reached. The number of principal components accumulated at this point is the basic compression dimension. For example, if the cumulative variance contribution rate of the first 5 principal components reaches 87%, then the basic compression dimension is 5. Subsequently, the range of candidate compression ratios is determined based on the basic compression dimension. That is, taking the basic compression dimension as the center, considering the compression space that nonlinear dimensionality reduction may bring, the range of candidate compression ratios is determined. For example, if the basic compression dimension is 5 and the original dimension is 50, then the basic compression ratio is 10:1. Considering the advantages of nonlinear methods, the range of candidate compression ratios can be set between 8:1 and 15:1.

[0048] Secondly, reconstruction error is the degree of difference between the dimensionality-reduced data and the original data after the data is restored to the original dimension by the decoder. Commonly used metrics include mean squared error and mean absolute error. The smaller the reconstruction error, the less information is lost during the dimensionality reduction process and the higher the data compression efficiency. Data compression efficiency is the ability of the dimensionality reduction operation to retain the core information of the original data at a specific compression ratio. It is an indicator for evaluating the quality of dimensionality reduction.

[0049] Specifically, each candidate compression ratio within a defined range is quantitatively evaluated. For each candidate compression ratio, a corresponding autoencoder is trained according to the training steps, where the output dimension of the encoder is determined by the corresponding candidate compression ratio. After training, the difference between the original input data and the decoder output data is calculated using a validation dataset to obtain the reconstruction error of each autoencoder. The magnitude of the reconstruction error reflects the degree to which the dimensionality-reduced data retains the original information at the corresponding candidate compression ratio.

[0050] A smaller reconstruction error indicates less loss of original information during dimensionality reduction and higher data compression efficiency; conversely, a larger reconstruction error indicates more severe information loss and lower compression efficiency. By calculating the reconstruction error, the data compression efficiency corresponding to each candidate compression ratio is obtained, and the dimensionality reduction quality of the candidate compression ratio is evaluated.

[0051] Simultaneously, a preset reconstruction error threshold is set as a benchmark value for acceptable dimensionality reduction quality. Then, all candidate compression ratios that meet the condition of reconstruction error less than this threshold are selected. The selected candidate compression ratios all ensure that the dimensionality-reduced data can be reconstructed back to the original data with acceptable accuracy, effectively preserving core information. Finally, from the qualified candidate ratios, the one with the largest value is selected as the most suitable data compression ratio, representing the highest compression degree achievable while ensuring information preservation quality. The preset reconstruction error threshold is a pre-set, acceptable maximum reconstruction error value, serving as a benchmark value for measuring whether the dimensionality reduction quality meets the standards.

[0052] Finally, the optimal data compression ratio is selected as the preset data compression ratio in the autoencoder for subsequent formal training and online applications. For example, a preset reconstruction error threshold of 0.05 is set. Assuming three candidate compression ratios are obtained: 5:1, 6:1, and 7:1, 7:1 is chosen as the optimal data compression ratio. The original 50-dimensional monitoring data is compressed to 7 dimensions, and the autoencoder is subsequently trained using 7:1 as the preset data compression ratio.

[0053] In step S10 of the method provided in this embodiment of the invention, the construction step of the welding defect anomaly discriminator includes: The data dimensionality reduction plugin is used to reduce the dimensionality of several global monitoring datasets of samples in the sample training dataset to obtain several local monitoring feature vectors of samples. A first defect probability threshold is set, and the defect probability distribution of the plurality of samples is classified into several samples by binary classification using the first defect probability threshold to obtain several defect discrimination results, wherein the defect discrimination results are welding defects or non-welding defects. Using the local monitoring feature vectors of several samples and the defect discrimination results of several samples as training data, a regression model is trained to generate a welding defect anomaly discriminator.

[0054] In this embodiment of the application, the dimensionality of several global monitoring datasets of samples in the sample training dataset is first reduced using a data dimensionality reduction plugin to obtain several local monitoring datasets of samples. The local monitoring feature vectors of the samples are feature vectors with significantly reduced dimensionality obtained after processing by the data dimensionality reduction plugin.

[0055] Specifically, all samples from the global monitoring dataset in the training dataset are input one by one into the trained data dimensionality reduction plugin. An encoder learns the nonlinear mapping parameters, and forward computation is performed on each high-dimensional sample data point, compressing and converting it into a low-dimensional sample local monitoring feature vector. This transformation of the original high-dimensional monitoring data into a low-dimensional feature representation forms the corresponding set of sample local monitoring feature vectors.

[0056] Secondly, a first defect probability threshold is set, and the defect probability distribution of several samples is binary classified using the first defect probability threshold to obtain several defect discrimination results. The first defect probability threshold is a pre-set probability threshold that transforms the refined probability prediction of multiple parts into a whole binary discrimination result. It can be adjusted according to different tolerance levels for false negatives and false positives, for example, set to 5% or 10%.

[0057] The sample defect probability distribution is a vector labeled for each sample, containing the probability of defects in multiple key areas. Binary classification is the process of transforming the original multi-valued probability distribution into two category labels. The result of binary classification is either welding defect or non-welding defect, where welding defect indicates that there is an abnormal state that needs attention at the corresponding time of the sample; non-welding defect indicates that the equipment is in normal condition at the corresponding time of the sample.

[0058] Specifically, a first defect probability threshold is first set, for example, 5%. Then, the sample defect probability distribution of each sample is traversed, and the labels containing probabilities of multiple parts are converted into binary labels through binary classification, reflecting whether there are any noteworthy abnormal risks in the overall equipment at the sample time.

[0059] Finally, a regression model is trained using several sample local monitoring feature vectors and several defect discrimination results as training data to generate a welding defect anomaly discriminator. The generated sample local monitoring feature vectors are used as input features, and the defect discrimination results generated by binary classification are used as supervision labels, together forming a complete training dataset. Logistic regression, support vector machines, or shallow neural networks are used for supervised training of the training dataset.

[0060] During training, the internal parameters are continuously adjusted to make the predicted output of the input features as close as possible to the true label. After training, a mapping relationship from the low-dimensional feature space to the normal / abnormal decision boundary is obtained, forming a welding defect anomaly discriminator. Subsequently, the welding defect anomaly discriminator can be deployed in a real-time monitoring system for rapid screening.

[0061] In step S20 of the method provided in this embodiment of the invention, the defect probability distributions of the plurality of samples are classified into two categories using the first defect probability threshold, including: For any sample defect probability distribution, if the sample defect probability of any key part in the sample defect probability distribution is greater than the first defect probability threshold, then the discrimination result corresponding to the sample defect probability distribution is marked as a welding defect. If the sample defect probability of all key parts in the sample defect probability distribution is less than or equal to the first defect probability threshold, then the discrimination result corresponding to the sample defect probability distribution is marked as a non-welding defect.

[0062] In this embodiment, a first defect probability threshold is first set. When traversing the sample defect probability distribution labeled for each sample, for any current sample, the defect probability values ​​of each key part are checked. If the sample defect probability distribution of the current sample contains any key part whose sample defect probability value is greater than the preset first defect probability threshold, then the judgment result corresponding to the current sample is marked as a welding defect.

[0063] If the defect probability of all critical parts in the current sample is less than or equal to the first defect probability threshold, it is marked as a non-welding defect. For example, the first defect probability threshold is set to 5%. The sample defect probability distribution of the existing sample is as follows: [vibrator mounting seat weld 2%, weld at the junction of stiffener plate and base plate 3%, weld around bearing seat 1%, fillet weld connecting front and rear end plates and side plates of the housing 0.5%, weld near lubricating oil passage 0.3%]. Traversing all probability values ​​in the sample defect probability distribution, all values ​​are less than 5%. Therefore, the current sample is marked as a non-welding defect.

[0064] In this embodiment, a multi-data fusion monitoring dataset is established by pre-setting a global monitoring index set, laying a high-quality data foundation for subsequent analysis. Then, a deep neural network is used to construct and supervise the training of a welding defect anomaly identification model, enabling the labeling of defect probabilities in key areas. An encoder-decoder is trained with a pre-set data compression ratio as a constraint and minimizing reconstruction error as the optimization objective. The trained encoder is used as a data dimensionality reduction plugin to achieve nonlinear feature extraction and information condensation from high-dimensional monitoring data, effectively removing redundancy and noise while retaining core state features, providing standardized low-dimensional input for subsequent rapid discrimination. Finally, the sample defect probability distribution is binary-classified to obtain several defect discrimination results. A training dataset is constructed based on the defect discrimination results and the sample local monitoring feature vectors, and a welding defect anomaly discriminator is built and trained, achieving maximum data compression while ensuring information retention quality.

[0065] S20: Input the local monitoring feature vector into the pre-constructed welding defect anomaly discriminator to determine whether there is a defect anomaly; In this embodiment, the local monitoring feature vector is used as input data and input into a pre-constructed welding defect anomaly discriminator. The welding defect anomaly discriminator then determines the welding defect anomaly and outputs a binary classification result to determine whether there is a defect anomaly at the current monitoring time.

[0066] In step S20 of the method provided in this embodiment of the invention, if the welding defect anomaly discriminator outputs a judgment result indicating the presence of defect anomalies at K consecutive monitoring nodes, an anomaly warning mechanism is triggered, where K is a positive integer.

[0067] In this embodiment, a monitoring node is a point in time or a time period for data acquisition and analysis. In a real-time monitoring system, data is acquired and processed in discrete time points or time windows, with each time unit constituting a monitoring node. For example, if data is acquired and analyzed once per second, then each second corresponds to one monitoring node.

[0068] Specifically, K is a preset positive integer parameter that indicates how many consecutive anomalies are required before triggering subsequent actions. The specific value can be set according to the requirements for false alarm tolerance in the actual application scenario, such as K=3, K=5, or K=10. A smaller K value results in a faster system response to anomalies and greater sensitivity to false alarms; a larger K value makes the system more robust and reduces the false alarm rate, but may slightly delay the warning.

[0069] During real-time monitoring, the welding defect anomaly discriminator quickly identifies the local monitoring feature vectors collected and dimensionality-reduced at each monitoring node, outputting a judgment result indicating whether a defect or anomaly exists. The system does not trigger an alert based on a single anomaly judgment result. An anomaly alert mechanism is triggered only when the discriminator outputs a defect or anomaly in every single one of the K consecutive monitoring nodes. If any one of the K consecutive nodes outputs a normal result, no alert is triggered; normal monitoring continues, the counter is reset, and counting restarts until the next anomaly alert trigger condition is met or monitoring resumes.

[0070] In this embodiment, an abnormal early warning triggering mechanism is used to trigger an early warning when the welding defect anomaly discriminator outputs the presence of defect anomalies at K consecutive monitoring nodes. This effectively filters out single misjudgments caused by accidental factors. Only continuous and stable abnormal patterns can trigger the subsequent in-depth diagnostic process, reducing the false alarm rate while achieving a balance between early warning sensitivity and stability, thus making the early warning results more reliable.

[0071] S30: If the abnormal warning mechanism is triggered, the global monitoring dataset in the dynamic monitoring window is collected, the pre-built welding defect abnormal identification model is input, the degree of welding defect abnormality of several key parts of the tamping device box is predicted, and the predicted defect probability distribution is output. The predicted defect probability distribution includes several predicted defect probabilities of several key parts. In this embodiment of the application, the dynamic monitoring window continuously collects global monitoring datasets over a period of time after an early warning is triggered, with the aim of obtaining sufficiently rich time-series data that is relevant to the current anomaly.

[0072] Specifically, the dimensionality-reduced data is first input into a pre-trained welding defect anomaly recognition model for deep analysis, outputting a predicted defect probability distribution. When the judgment result of defect anomaly is output at K consecutive monitoring nodes, an anomaly warning mechanism is triggered, and a secondary judgment is performed using the global monitoring dataset. After determining the dynamic monitoring window, a more comprehensive global monitoring dataset is collected within the dynamic monitoring window. Subsequently, the global monitoring dataset is input into the pre-trained welding defect anomaly recognition model for deep analysis, outputting a predicted defect probability distribution to obtain the predicted defect probability of key parts.

[0073] In this embodiment, global monitoring data is collected through a dynamic monitoring window, enabling the model to make judgments based on more complete dynamic process characteristics. Through a pre-built welding defect anomaly identification model, the defect risks of multiple parts are predicted in parallel, and the predicted defect probability of key parts is output, providing a scientific decision-making basis for the subsequent identification of high-risk parts and the setting of appropriate detection schemes, thereby improving the operability of defect identification.

[0074] S40: Based on the predicted defect probability distribution, identify high-risk defect locations, and set up an adaptive defect detection mechanism according to the predicted defect probability corresponding to the high-risk defect locations to identify welding defects in the high-risk defect locations.

[0075] In step S40 of the method provided in this embodiment of the invention, high-risk defect locations are identified and determined based on the predicted defect probability distribution, and an adaptive defect detection mechanism is set according to the predicted defect probability corresponding to the high-risk defect locations, including: A second defect probability threshold is set, wherein the second defect probability threshold is greater than the first defect probability threshold; Traverse the predicted defect probability distribution and mark key parts with predicted defect probabilities greater than the second defect probability threshold as high-risk defect parts; A mapping relationship is constructed between the predicted defect probability and the defect detection scheme, wherein the detection depth of the defect detection scheme and the predicted defect probability are positively correlated in the mapping relationship; Based on the predicted defect probability corresponding to the high-risk defect location, the mapping relationship is queried to determine the matching defect detection scheme, and welding defect identification is performed on the high-risk defect location.

[0076] In this embodiment, a second defect probability threshold is first set. The first defect probability threshold, used for rapid screening, is relatively low to ensure sensitivity and avoid missing any possible anomalies; its value is small. The second defect probability threshold, used to identify the most critical areas with the highest risk when an anomaly is confirmed, has a larger value to ensure that the selected areas indeed have a high defect probability. Therefore, the second defect probability threshold is greater than the first defect probability threshold used for initial screening. For example, if the first defect probability threshold is set to 5%, the second defect probability threshold can be set to 20% or 30%.

[0077] Secondly, the predicted defect probability distribution output by the welding defect anomaly identification model is obtained. This process is then iterated through each critical area, comparing the predicted defect probability with a set second defect probability threshold. If the predicted defect probability is greater than the second threshold, the area is marked as a high-risk defect; conversely, if the predicted probability is less than or equal to the second threshold, it is either not marked or marked as a normal-risk area. By comparing these factors, the area with the most prominent risk is identified from among multiple critical areas and used as the target for subsequent detection.

[0078] Finally, a mapping relationship is established between the predicted defect probability and the defect detection scheme, where the detection depth is positively correlated with the predicted defect probability. For areas with a low predicted defect probability, a scheme with shallow detection depth, low cost, and high speed should be matched; for areas with a medium predicted defect probability, a detection scheme with medium depth should be matched; and for areas with an extremely high predicted defect probability, a detection scheme with the deepest detection depth and highest accuracy should be matched. This mapping relationship can be determined comprehensively based on factors such as detection cost analysis and safety requirements, and can be dynamically adjusted and optimized according to actual application results.

[0079] For example, the classification mapping table can be as follows: If the risk level is Level 1, the predicted defect probability range is (P2, 0.85), the solution is a rapid scanning solution, and the detection depth is achieved by using an automated device equipped with a conventional ultrasonic probe to quickly scan the marked high-risk defect area to confirm the presence of obvious defects. If the risk level is Level 2, the predicted defect probability range is (0.85, 0.95), and the defect detection solution can be a fine flaw detection solution, where the detection depth is achieved by professional inspectors using a phased array ultrasonic detector or digital X-ray inspection equipment to perform high-precision imaging flaw detection on the marked area, quantifying the size, shape, and orientation of the defects. If the risk level is Level 3, the predicted defect probability range is (0.95, 1.0), and the defect detection solution is a shutdown depth detection solution, where the detection depth triggers an alarm signal, recommending immediate shutdown. While the system is shut down, the high-risk defect area and its surrounding area are disassembled, and surface flaw detection methods such as magnetic particle testing or penetrant testing are used for final confirmation and evaluation.

[0080] In this embodiment, a second defect probability threshold higher than the first defect probability threshold is set, and key parts with predicted probabilities exceeding this threshold are marked as high-risk defect parts, thereby achieving accurate screening of high-risk parts. By constructing a mapping relationship between predicted defect probabilities and defect detection schemes, an association between risk levels and detection strategies is established, thereby optimizing the allocation of detection resources and improving cost-effectiveness. Finally, based on the predicted probability of high-risk parts, a matching detection scheme is determined to ensure that high-risk parts receive in-depth detection appropriate to their risk level, thereby improving the accuracy and efficiency of welding defect identification.

[0081] The embodiments of this application, through the above specific implementation methods, achieve the following technical effects: In this embodiment, firstly, a multi-data fusion monitoring data set is established by pre-setting a global monitoring index set, laying a high-quality data foundation for subsequent analysis. Then, a deep neural network is used to construct and supervise the training of a welding defect anomaly identification model, enabling the labeling of defect probabilities in key areas. An encoder-decoder is trained with a pre-set data compression ratio as a constraint and minimizing reconstruction error as the optimization objective. The trained encoder is used as a data dimensionality reduction plugin to achieve nonlinear feature extraction and information condensation of high-dimensional monitoring data, effectively removing redundancy and noise while retaining core state features, providing standardized low-dimensional input for subsequent rapid discrimination. Finally, the sample defect probability distribution is binary-classified to obtain several defect discrimination results. A training dataset is constructed based on the defect discrimination results and the sample local monitoring feature vectors, and a welding defect anomaly discriminator is built and trained, achieving maximum data compression while ensuring information retention quality.

[0082] Secondly, through the abnormal early warning triggering mechanism, when the welding defect anomaly discriminator outputs the presence of defect anomalies under K consecutive monitoring nodes, an early warning is triggered, effectively filtering out single misjudgments caused by accidental factors. Only continuous and stable abnormal patterns can trigger the subsequent in-depth diagnostic process, reducing the false alarm rate while achieving a balance between early warning sensitivity and stability, making the early warning results more reliable.

[0083] Furthermore, by collecting global monitoring data through a dynamic monitoring window, the model can make judgments based on more complete dynamic process characteristics. Through a pre-built welding defect anomaly identification model, the risk of defects in multiple locations is predicted in parallel. The predicted defect probability of key locations provides a scientific basis for subsequent identification of high-risk locations and the setting of appropriate detection schemes, thereby improving the operability of defect identification.

[0084] Finally, by setting a probability threshold higher than the second defect probability threshold, high-risk defect locations exceeding the second defect probability threshold are identified, enabling accurate screening of high-risk locations. By constructing a mapping relationship between predicted defect probabilities and defect detection schemes, a correlation between risk levels and detection strategies is established, achieving optimized allocation of detection resources and improved cost-effectiveness. Ultimately, based on the predicted probability of high-risk locations, a matching detection scheme is determined to ensure that high-risk locations receive in-depth detection commensurate with their risk level, thereby improving the accuracy and efficiency of welding defect identification.

[0085] Example 2, as Figure 2 As shown, based on the same inventive concept as the welding defect identification method for sealing components of engineering vehicles provided in Embodiment 1, this embodiment of the invention also provides a welding defect identification system for sealing components of engineering vehicles, the system comprising: Feature vector acquisition module 11 is used to collect the global monitoring dataset of the tamping device box when the railway tamping car is working, and to reduce the dimensionality of the global monitoring dataset to generate local monitoring feature vectors. The defect anomaly discrimination module 12 is used to input the local monitoring feature vector into a pre-constructed welding defect anomaly discriminator to determine whether there is a defect anomaly. The defect probability prediction module 13 is used to collect the global monitoring dataset in the dynamic monitoring window if the abnormal warning mechanism is triggered, input the pre-built welding defect abnormality identification model, predict the degree of welding defect abnormality of several key parts of the tamping device box, and output the predicted defect probability distribution, wherein the predicted defect probability distribution includes several predicted defect probabilities of several key parts. The defect detection mechanism configuration module 14 is used to identify and determine high-risk defect locations based on the predicted defect probability distribution, and to set an adaptive defect detection mechanism according to the predicted defect probability corresponding to the high-risk defect locations to identify welding defects in the high-risk defect locations.

[0086] In one embodiment, the feature vector acquisition module 11 is used for: The global monitoring dataset is collected based on a preset global monitoring index set, which includes vibration monitoring indexes, temperature monitoring indexes, strain monitoring indexes, and operating condition monitoring indexes. The vibration monitoring indexes include triaxial vibration acceleration at key measuring points, effective values ​​of vibration velocity at key measuring points, and peak values ​​of vibration acceleration at key measuring points. The temperature monitoring indexes include the temperatures at multiple measuring points on the box surface and the lubricating oil temperature. The strain monitoring indexes include the strain values ​​at measuring points near key welds. The operating condition monitoring indexes include tamping depth, clamping force, operating speed, and operating mode.

[0087] In one embodiment, the feature vector acquisition module 11 is further configured to: Several global monitoring datasets were collected, and the proportion of welding defect abnormal events in each key part of the tamping device box under the same working conditions was calculated for each global monitoring dataset. The proportion of abnormal events was recorded as the sample defect probability corresponding to the key part, and several sample defect probabilities were obtained. The key parts of the tamping device box include the weld connecting the vibrator mounting base and the box side plate, the weld at the intersection of the box stiffener plate and the bottom plate, the radial weld around the bearing seat, the fillet weld connecting the front and rear end plates and the side plate of the box, and the weld near the lubricating oil passage. The sample training dataset is constructed based on the aforementioned global monitoring dataset and the corresponding sample defect probabilities. The deep neural network is trained under supervision using the sample training dataset until the preset convergence condition is met, thereby generating a welding defect anomaly recognition model.

[0088] The process of reducing the dimensionality of the global monitoring dataset to generate local monitoring feature vectors includes: Using the global monitoring dataset of the sample training dataset as the training sample, the preset data compression ratio as the constraint condition of the encoder output dimension, and minimizing the difference between the global monitoring dataset of the sample and the reconstructed dataset after encoder-decoder reconstruction as the optimization objective, the encoder and decoder are trained. During training, the encoder compresses the high-dimensional global monitoring dataset of samples into a low-dimensional feature representation that conforms to the preset data compression ratio, and the decoder reconstructs the low-dimensional feature representation into a reconstructed dataset with the same dimension as the original input. The trained encoder is used as a data dimensionality reduction plugin to reduce the dimensionality of the real-time collected global monitoring dataset and generate local monitoring feature vectors.

[0089] The preset data compression ratio is determined through the following steps: Principal component analysis is performed on each monitoring indicator in the preset global monitoring indicator set to calculate the eigenvalues ​​and variance contribution rate of each principal component, and the variance contribution rates are sorted from high to low. The number of principal components required to accumulate the variance contribution rate to reach the preset contribution threshold is determined as the basic compression dimension, and the range of values ​​for the candidate compression ratio is determined based on the basic compression dimension. Within the range of values, for each candidate compression ratio, the data compression performance at that compression ratio is evaluated. The data compression performance is measured by the reconstruction error of the original data by the dimensionality-reduced data. The smaller the reconstruction error, the higher the data compression performance. Among all candidate compression ratios that satisfy the condition that the reconstruction error is less than the preset reconstruction error threshold, the candidate compression ratio with the largest value is selected as the most suitable data compression ratio, so as to minimize the number of monitoring indicators while ensuring that the data compression performance meets the standard. The optimal data compression ratio is used as the preset data compression ratio.

[0090] The construction steps of the welding defect anomaly detector include: The data dimensionality reduction plugin is used to reduce the dimensionality of several global monitoring datasets of samples in the sample training dataset to obtain several local monitoring feature vectors of samples. A first defect probability threshold is set, and the defect probability distribution of the plurality of samples is classified into several samples by binary classification using the first defect probability threshold to obtain several defect discrimination results, wherein the defect discrimination results are welding defects or non-welding defects. Using the local monitoring feature vectors of several samples and the defect discrimination results of several samples as training data, a regression model is trained to generate a welding defect anomaly discriminator.

[0091] In one embodiment, the defect / anomaly detection module 12 is used for: For any sample defect probability distribution, if the sample defect probability of any key part in the sample defect probability distribution is greater than the first defect probability threshold, then the discrimination result corresponding to the sample defect probability distribution is marked as a welding defect. If the sample defect probability of all key parts in the sample defect probability distribution is less than or equal to the first defect probability threshold, then the discrimination result corresponding to the sample defect probability distribution is marked as a non-welding defect.

[0092] In one embodiment, the defect probability prediction module 13 is used for: If the welding defect anomaly discriminator outputs a judgment result indicating the presence of a defect anomaly at K consecutive monitoring nodes, an anomaly warning mechanism is triggered, where K is a positive integer.

[0093] In one embodiment, the defect detection mechanism configuration module 14 is used for: A second defect probability threshold is set, wherein the second defect probability threshold is greater than the first defect probability threshold; Traverse the predicted defect probability distribution and mark key parts with predicted defect probabilities greater than the second defect probability threshold as high-risk defect parts; A mapping relationship is constructed between the predicted defect probability and the defect detection scheme, wherein the detection depth of the defect detection scheme and the predicted defect probability are positively correlated in the mapping relationship; Based on the predicted defect probability corresponding to the high-risk defect location, the mapping relationship is queried to determine the matching defect detection scheme, and welding defect identification is performed on the high-risk defect location.

[0094] Compared to existing technologies, this application first establishes a multi-data fusion monitoring dataset by pre-setting a global monitoring index set, laying a high-quality data foundation for subsequent analysis. Then, it utilizes a deep neural network to construct and supervise the training of a welding defect anomaly identification model, enabling the labeling of defect probabilities in key areas. An encoder-decoder is trained with a pre-set data compression ratio as a constraint and minimizing reconstruction error as the optimization objective. The trained encoder is used as a data dimensionality reduction plugin to achieve nonlinear feature extraction and information condensation from high-dimensional monitoring data, effectively removing redundancy and noise while retaining core state features, providing standardized low-dimensional input for subsequent rapid discrimination. Finally, the sample defect probability distribution is binary-classified to obtain several defect discrimination results. A training dataset is constructed based on the defect discrimination results and the sample local monitoring feature vectors, and a welding defect anomaly discriminator is built and trained, achieving maximum data compression while ensuring information retention quality.

[0095] Secondly, through the abnormal early warning triggering mechanism, when the welding defect anomaly discriminator outputs the presence of defect anomalies under K consecutive monitoring nodes, an early warning is triggered, effectively filtering out single misjudgments caused by accidental factors. Only continuous and stable abnormal patterns can trigger the subsequent in-depth diagnostic process, reducing the false alarm rate while achieving a balance between early warning sensitivity and stability, making the early warning results more reliable.

[0096] Furthermore, by collecting global monitoring data through a dynamic monitoring window, the model can make judgments based on more complete dynamic process characteristics. Through a pre-built welding defect anomaly identification model, the risk of defects in multiple locations is predicted in parallel. The predicted defect probability of key locations provides a scientific basis for subsequent identification of high-risk locations and the setting of appropriate detection schemes, thereby improving the operability of defect identification.

[0097] Finally, by setting a probability threshold higher than the second defect probability threshold, high-risk defect locations exceeding the second defect probability threshold are identified, enabling accurate screening of high-risk locations. By constructing a mapping relationship between predicted defect probabilities and defect detection schemes, a correlation between risk levels and detection strategies is established, achieving optimized allocation of detection resources and improved cost-effectiveness. Ultimately, based on the predicted probability of high-risk locations, a matching detection scheme is determined to ensure that high-risk locations receive in-depth detection commensurate with their risk level, thereby improving the accuracy and efficiency of welding defect identification.

Claims

1. A method for identifying welding defects in sealing components of engineering vehicles, characterized in that the method... include: Collect a global monitoring dataset of the tamping device housing during the operation of the railway tamping machine, and perform dimensionality reduction on the global monitoring dataset to generate local monitoring feature vectors; The local monitoring feature vector is input into a pre-constructed welding defect anomaly discriminator to determine whether there is a defect anomaly. If the abnormal warning mechanism is triggered, the global monitoring dataset in the dynamic monitoring window is collected, the pre-built welding defect abnormality identification model is input, the degree of welding defect abnormality of several key parts of the tamping device box is predicted, and the predicted defect probability distribution is output. The predicted defect probability distribution includes several predicted defect probabilities of several key parts. Based on the predicted defect probability distribution, high-risk defect locations are identified, and an adaptive defect detection mechanism is set according to the predicted defect probability corresponding to the high-risk defect locations to identify welding defects in the high-risk defect locations.

2. The method for identifying welding defects in the sealing components of an engineering vehicle according to claim 1, characterized in that, The global monitoring dataset is collected based on a preset global monitoring index set, which includes vibration monitoring indexes, temperature monitoring indexes, strain monitoring indexes, and operating condition monitoring indexes. The vibration monitoring indexes include triaxial vibration acceleration at key measuring points, effective values ​​of vibration velocity at key measuring points, and peak values ​​of vibration acceleration at key measuring points. The temperature monitoring indexes include the temperatures at multiple measuring points on the box surface and the lubricating oil temperature. The strain monitoring indexes include the strain values ​​at measuring points near key welds. The operating condition monitoring indexes include tamping depth, clamping force, operating speed, and operating mode.

3. The method for identifying welding defects in the sealing components of an engineering vehicle according to claim 1, characterized in that, The steps for constructing the welding defect anomaly identification model include: Several global monitoring datasets were collected, and the proportion of welding defect abnormal events in each key part of the tamping device box under the same working conditions was calculated for each global monitoring dataset. The proportion of abnormal events was recorded as the sample defect probability corresponding to the key part, and several sample defect probabilities were obtained. The key parts of the tamping device box include the weld connecting the vibrator mounting base and the box side plate, the weld at the intersection of the box stiffener plate and the bottom plate, the radial weld around the bearing seat, the fillet weld connecting the front and rear end plates and the side plate of the box, and the weld near the lubricating oil passage. The sample training dataset is constructed based on the aforementioned global monitoring dataset and the corresponding sample defect probabilities. The deep neural network is trained under supervision using the sample training dataset until the preset convergence condition is met, thereby generating a welding defect anomaly recognition model.

4. The method for identifying welding defects in the sealing components of an engineering vehicle according to claim 3, characterized in that, The global monitoring dataset is dimensionality reduced to generate local monitoring feature vectors, including: Using the global monitoring dataset of the sample training dataset as the training sample, the preset data compression ratio as the constraint condition of the encoder output dimension, and minimizing the difference between the global monitoring dataset of the sample and the reconstructed dataset after encoder-decoder reconstruction as the optimization objective, the encoder and decoder are trained. During training, the encoder compresses the high-dimensional global monitoring dataset of samples into a low-dimensional feature representation that conforms to the preset data compression ratio, and the decoder reconstructs the low-dimensional feature representation into a reconstructed dataset with the same dimension as the original input. The trained encoder is used as a data dimensionality reduction plugin to reduce the dimensionality of the real-time collected global monitoring dataset and generate local monitoring feature vectors.

5. The method for identifying welding defects in the sealing components of an engineering vehicle according to claim 4, characterized in that, The preset data compression ratio is determined through the following steps: Principal component analysis is performed on each monitoring indicator in the preset global monitoring indicator set to calculate the eigenvalues ​​and variance contribution rate of each principal component, and the variance contribution rates are sorted from high to low. The number of principal components required to accumulate the variance contribution rate to reach the preset contribution threshold is determined as the basic compression dimension, and the range of values ​​for the candidate compression ratio is determined based on the basic compression dimension. Within the range of values, for each candidate compression ratio, the data compression performance at that compression ratio is evaluated. The data compression performance is measured by the reconstruction error of the original data by the dimensionality-reduced data. The smaller the reconstruction error, the higher the data compression performance. Among all candidate compression ratios that satisfy the condition that the reconstruction error is less than the preset reconstruction error threshold, the candidate compression ratio with the largest value is selected as the most suitable data compression ratio, so as to minimize the number of monitoring indicators while ensuring that the data compression performance meets the standard. The optimal data compression ratio is used as the preset data compression ratio.

6. The method for identifying welding defects in the sealing components of an engineering vehicle according to claim 4, characterized in that, The construction steps of the welding defect anomaly discriminator include: The data dimensionality reduction plugin is used to reduce the dimensionality of several global monitoring datasets of samples in the sample training dataset to obtain several local monitoring feature vectors of samples. A first defect probability threshold is set, and the defect probability distribution of the plurality of samples is classified into several samples by binary classification using the first defect probability threshold to obtain several defect discrimination results, wherein the defect discrimination results are welding defects or non-welding defects. Using the local monitoring feature vectors of several samples and the defect discrimination results of several samples as training data, a regression model is trained to generate a welding defect anomaly discriminator.

7. The method for identifying welding defects in the sealing components of an engineering vehicle according to claim 6, characterized in that, Using the first defect probability threshold, perform binary classification on the defect probability distributions of the plurality of samples, including: For any sample defect probability distribution, if the sample defect probability of any key part in the sample defect probability distribution is greater than the first defect probability threshold, then the discrimination result corresponding to the sample defect probability distribution is marked as a welding defect. If the sample defect probability of all key parts in the sample defect probability distribution is less than or equal to the first defect probability threshold, then the discrimination result corresponding to the sample defect probability distribution is marked as a non-welding defect.

8. The method for identifying welding defects in the sealing components of an engineering vehicle according to claim 6, characterized in that, If the welding defect anomaly discriminator outputs a judgment result indicating the presence of a defect anomaly at K consecutive monitoring nodes, an anomaly warning mechanism is triggered, where K is a positive integer.

9. The method for identifying welding defects in the sealing components of an engineering vehicle according to claim 6, characterized in that, Based on the predicted defect probability distribution, high-risk defect locations are identified, and an adaptive defect detection mechanism is set according to the predicted defect probability corresponding to the high-risk defect locations, including: A second defect probability threshold is set, wherein the second defect probability threshold is greater than the first defect probability threshold; Traverse the predicted defect probability distribution and mark key parts with predicted defect probabilities greater than the second defect probability threshold as high-risk defect parts; A mapping relationship is constructed between the predicted defect probability and the defect detection scheme, wherein the detection depth of the defect detection scheme and the predicted defect probability are positively correlated in the mapping relationship; Based on the predicted defect probability corresponding to the high-risk defect location, the mapping relationship is queried to determine the matching defect detection scheme, and welding defect identification is performed on the high-risk defect location.

10. A welding defect identification system for sealing components of engineering vehicles, characterized in that, A method for identifying welding defects in a sealing component of an engineering vehicle according to any one of claims 1-9, the system comprising: The feature vector acquisition module is used to collect the global monitoring dataset of the tamping device box when the railway tamping machine is working, and to reduce the dimensionality of the global monitoring dataset to generate local monitoring feature vectors. The defect anomaly discrimination module is used to input the local monitoring feature vector into a pre-constructed welding defect anomaly discriminator to determine whether there is a defect anomaly. The defect probability prediction module is used to collect the global monitoring dataset in the dynamic monitoring window if the abnormal warning mechanism is triggered, input the pre-built welding defect abnormality identification model, predict the degree of welding defect abnormality of several key parts of the tamping device box, and output the predicted defect probability distribution, wherein the predicted defect probability distribution includes several predicted defect probabilities of several key parts. The defect detection mechanism configuration module is used to identify and determine high-risk defect locations based on the predicted defect probability distribution, and to set an adaptive defect detection mechanism according to the predicted defect probability corresponding to the high-risk defect locations to identify welding defects in the high-risk defect locations.

Images