Rail transit nondestructive testing equipment and testing method thereof

Abstract

The invention provides rail transit nondestructive testing equipment and a testing method thereof. The method comprises the following steps: acquiring testing data for rail transit equipment; performing data preprocessing on the detection data to obtain preprocessed data, and extracting a feature vector based on the preprocessed data; based on the feature vector and the preprocessed data, performing data enhancement on the detection data through the few-step diffusion model to obtain synthetic data; constructing a jump data diffusion model based on the detection model, and performing robustness analysis on the detection data based on the jump data diffusion model to obtain a robustness analysis result; and based on the robustness analysis result, using a preset implicit neural representation network to extract time continuity features in the detection data, and performing time sequence anomaly detection on the time continuity features to obtain a time sequence anomaly detection result.

Description

Technical Field

[0001] The present application relates to the field of rail transit technology, and in particular to a rail transit nondestructive testing device and a testing method thereof. Background Art

[0002] As an important mode of transportation, rail transit's safety and reliability are directly related to the safety of people's lives and property. Existing rail transit equipment inspection technologies primarily include ultrasonic testing, temperature monitoring, and vibration monitoring. Ultrasonic testing can detect internal defects in equipment, temperature monitoring can promptly identify abnormal heating during operation, and vibration monitoring can detect abnormal vibrations during operation.

[0003] However, existing technologies have the following problems: first, the reliability and accuracy of a single detection method are limited, making it difficult to fully reflect the operating status of the equipment; second, the fusion and analysis capabilities of multi-source detection data are insufficient, making it impossible to fully utilize the data collected by various sensors.

[0004] Therefore, the existing rail transit equipment detection technology has certain limitations and low detection accuracy. Summary of the Invention

[0005] In view of this, the present application provides a rail transit non-destructive testing device and a testing method thereof, which solves the problem that the rail transit equipment testing technology in the prior art has limitations and low detection accuracy.

[0006] In a first aspect, an embodiment of the present application provides a rail transit nondestructive testing method, comprising: Acquiring detection data for rail transit equipment, the detection data including ultrasonic reflection signals from a multi-channel ultrasonic array detection system, and temperature data and vibration data from a multi-source sensor network system; performing data preprocessing on the detection data to obtain preprocessed data, and extracting a feature vector based on the preprocessed data; constructing a few-step diffusion model based on the feature vector and the preprocessed data, and performing data enhancement on the detection data using the few-step diffusion model to obtain synthetic data; Based on the synthetic data, a dense reward difference learning framework is constructed, and the dense reward difference learning framework is optimized to obtain a detection model; Based on the detection model, constructing a jump data diffusion model, and performing robustness analysis on the detection data based on the jump data diffusion model to obtain a robustness analysis result; Based on the robustness analysis result, a preset implicit neural representation network is used to extract the time continuity features in the detection data, and a time series anomaly detection is performed on the time series anomaly features to obtain a time series anomaly detection result; Multi-model data fusion is performed on the abnormal data detection result, the robustness analysis result and the time series anomaly detection result to obtain a detection warning result, wherein the abnormal data detection result is obtained after the detection model performs anomaly detection on the detection data.

[0007] In one possible implementation, preprocessing the detection data to obtain preprocessed data includes: Performing wavelet transform processing on the ultrasonic reflection signal, processing the high-frequency coefficients obtained after the wavelet transform, and reconstructing the signal based on the processed high-frequency coefficients; Performing median filtering on the temperature data to eliminate mutation points in the temperature data, and performing low-pass filtering on the temperature data to eliminate high-frequency noise in the temperature data; performing bandpass filtering on the vibration data and processing abnormal value points in the vibration data; Based on the timestamp information, the processed ultrasonic reflection signal, the processed temperature data, and the processed vibration data are aligned and then standardized to obtain preprocessed data.

[0008] In one possible implementation, the preprocessed data includes a preprocessed ultrasonic reflection signal, preprocessed temperature data, and preprocessed vibration data, and extracting a feature vector based on the preprocessed data includes: Performing time domain feature extraction and frequency domain feature extraction on the preprocessed ultrasonic reflection signal to obtain ultrasonic reflection signal feature data; Extracting temperature characteristic data of the preprocessed temperature data, wherein the temperature characteristic data includes temperature gradient, temperature field distribution characteristics and temperature fluctuation characteristics; Extracting vibration feature data of the preprocessed vibration data, wherein the vibration feature data includes time domain statistical features, frequency domain features, and time domain joint features; The ultrasonic reflection signal characteristic data, the temperature characteristic data, and the vibration characteristic data are screened and evaluated to obtain a characteristic vector.

[0009] In one possible implementation, constructing a dense reward difference learning framework based on the synthetic data, and optimizing the dense reward difference learning framework to obtain a detection model include: Based on the synthesized data, constructing a local feature reward function and a global feature reward function, wherein the local feature reward function is used to indicate a reward for the similarity of local features, and the global feature reward function is used to indicate a reward for the consistency of global features; Based on the local feature reward function and the global feature reward function, constructing a dense reward difference learning framework; Calculating sample weights based on the dense reward differential learning framework and performing differential optimization based on contrastive learning on the synthetic data; Optimizing parameters of the dense reward difference learning framework based on the sample weights; The dense reward difference learning framework after parameter optimization is evaluated and optimized to obtain the detection model.

[0010] In one possible implementation, constructing a jump data diffusion model based on the detection model, and performing robustness analysis on the detection data based on the jump data diffusion model to obtain a robustness analysis result includes: Based on the detection model, a multi-scale feature extraction network is constructed, wherein the multi-scale feature extraction network includes a skip connection structure in which a plurality of feature scale layers are arranged in cascade and adjacent feature scale layers are arranged; Constructing a feature fusion mechanism, wherein the feature fusion mechanism is used to instruct to use a preset channel attention mechanism and a spatial attention mechanism to perform feature fusion on the detection data; Constructing an anti-interference network, wherein the anti-interference network is used to instruct a preset spatial transformation network to perform feature alignment on the detection data; Constructing a jump data diffusion model based on the multi-scale feature extraction network, the feature fusion mechanism and the anti-interference network; Performing system performance evaluation and optimization on the jump data diffusion model; A robustness analysis is performed on the detection data based on the optimized jump data diffusion model to obtain a robustness analysis result.

[0011] In one possible implementation, based on the robustness analysis result, a preset implicit neural representation network is used to extract the time continuity feature in the detection data, and time series anomaly detection is performed on the time series anomaly feature to obtain the time series anomaly detection result, including: Based on the robustness analysis results, constructing an implicit neural representation network, wherein the implicit neural representation network includes a time continuity encoder, a temporal dependency model, and an anomaly detection discriminator; extracting a temporal continuity feature from the detection data based on the implicit neural representation network, and calculating an anomaly score of the temporal continuity feature; The anomaly score of the time continuity feature is judged based on a preset anomaly threshold to obtain a corresponding time series anomaly detection result.

[0012] In one possible implementation, performing multi-model data fusion on the abnormal data detection result, the robustness analysis result, and the time series anomaly detection result to obtain the detection warning result includes: Assigning corresponding weight coefficients to the abnormal data detection result, the robustness analysis result, and the time series anomaly detection result; Performing weighted fusion on the abnormal data detection result, the robustness analysis result, and the time series anomaly detection result according to the weight coefficient; Output corresponding detection and warning results based on the weighted fusion results.

[0013] In a possible implementation, before acquiring the detection data for the rail transit equipment, the method further includes: Obtain ultrasonic testing requirements, environmental monitoring requirements and structural characteristics of the rail transit equipment; Determine the structural strategy and ultrasonic excitation parameters of the multi-channel ultrasonic array based on the ultrasonic detection requirements and the structural characteristics of the rail transit equipment, wherein the structural strategy includes the number, spacing, and arrangement of array elements; Configuring ultrasonic signal acquisition parameters of the multi-channel ultrasonic array, wherein the ultrasonic signal acquisition parameters include ultrasonic signal sampling frequency and quantization accuracy; Based on the structural strategy of the multi-channel ultrasonic array, the ultrasonic signal acquisition parameters, and the ultrasonic excitation parameters, an ultrasonic array arrangement strategy is formed and sent to the multi-channel ultrasonic array detection system, so that the multi-channel ultrasonic array detection system is arranged based on the ultrasonic array arrangement strategy; Determining a layout strategy for a temperature sensor array and a deployment strategy for vibration sensors based on the environmental monitoring requirements and the structural characteristics of the rail transit equipment; Configuring sensor data acquisition parameters of the temperature sensor array and the vibration sensor, wherein the sensor data acquisition parameters include sensor data sampling frequency and triggering mode; Based on the layout strategy of the temperature sensor array, the deployment strategy of the vibration sensor and the sensor data acquisition parameters, a sensor network layout strategy is formed and sent to the multi-source sensor network system, so that the multi-source sensor network system is deployed based on the sensor network layout strategy.

[0014] In one possible implementation, constructing a few-step diffusion model based on the feature vector and the preprocessed data includes: Based on the feature vector, constructing a few-step diffusion initial model, the few-step diffusion initial model including a basic model architecture, a noise scheduling strategy, a sampling mechanism, a loss function, and an optimization strategy; Dividing the feature vector and the preprocessed data into a training set, a validation set, and a test set; Training the few-step diffusion initial model using the training set; Using the validation set to evaluate and optimize the trained initial few-step diffusion model; The optimized few-step diffusion initial model is tested using the test set to obtain the few-step diffusion model.

[0015] In a second aspect, an embodiment of the present application further provides a rail transit nondestructive testing device, comprising: processor; memory for storing computer programs; Wherein, when the processor executes the computer program, the steps of the rail transit non-destructive testing method as described in the above-mentioned first aspect or any possible implementation method of the first aspect are implemented.

[0016] The rail transit nondestructive testing method and equipment provided in this application solves the problem of a single detection method by acquiring ultrasonic reflection signals from a multi-channel ultrasonic array detection system for rail transit equipment, as well as temperature data and vibration data from a multi-source sensor network system for rail transit equipment. This achieves the collection of multi-source detection data, improving the comprehensiveness and accuracy of detection. Subsequently, by preprocessing and extracting features from the multi-source detection data and using a few-step diffusion model for data enhancement, the fusion capability of the multi-source detection data is improved, which can meet the training requirements of subsequent models. Furthermore, by constructing a few-step diffusion model and a dense reward difference learning framework, the accuracy of abnormal data detection is improved. By constructing a jump data diffusion model, the robustness of the detection results of the detection model is improved, and it can be applied to different working conditions and environmental conditions. By extracting temporal continuity features through an implicit neural representation network for time series anomaly detection, in-depth analysis of the time series characteristics of the detection data is achieved, thereby performing multi-model data fusion on the abnormal data detection results, the robustness analysis results, and the time series anomaly detection results, improving the detection accuracy and the reliability of the detection warning results. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the disclosed embodiments in the embodiments of the present application, the following will briefly introduce the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without any creative work.

[0018] Figure 1 This is a flow chart of a rail transit nondestructive testing method provided in an embodiment of the present application; Figure 2 A schematic structural diagram of a rail transit nondestructive testing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0019] In order to make the objectives, technical solutions and advantages of this application more clear, the implementation methods of this application will be further described in detail with reference to the accompanying drawings.

[0020] Figure 1 This is a flow chart of a rail transit non-destructive testing method provided in an embodiment of the present application. Based on this example flow, the embodiment of the present application provides a rail transit non-destructive testing method, including the following steps S101 to S107.

[0021] S101 : Acquire detection data for rail transit equipment, where the detection data includes ultrasonic reflection signals from a multi-channel ultrasonic array detection system, and temperature data and vibration data from a multi-source sensor network system.

[0022] In some embodiments, before step S101, the method further includes: Obtain ultrasonic testing requirements, environmental monitoring requirements and structural characteristics of the rail transit equipment; Determine the structural strategy and ultrasonic excitation parameters of the multi-channel ultrasonic array based on the ultrasonic detection requirements and the structural characteristics of the rail transit equipment, wherein the structural strategy includes the number, spacing, and arrangement of array elements; Configuring ultrasonic signal acquisition parameters of the multi-channel ultrasonic array, wherein the ultrasonic signal acquisition parameters include ultrasonic signal sampling frequency and quantization accuracy; Based on the structural strategy of the multi-channel ultrasonic array, the ultrasonic signal acquisition parameters, and the ultrasonic excitation parameters, an ultrasonic array arrangement strategy is formed and sent to the multi-channel ultrasonic array detection system, so that the multi-channel ultrasonic array detection system is arranged based on the ultrasonic array arrangement strategy; Determining a layout strategy for a temperature sensor array and a deployment strategy for vibration sensors based on the environmental monitoring requirements and the structural characteristics of the rail transit equipment; Configuring sensor data acquisition parameters of the temperature sensor array and the vibration sensor, wherein the sensor data acquisition parameters include sensor data sampling frequency and triggering mode; Based on the layout strategy of the temperature sensor array, the deployment strategy of the vibration sensor and the sensor data acquisition parameters, a sensor network layout strategy is formed and sent to the multi-source sensor network system, so that the multi-source sensor network system is deployed based on the sensor network layout strategy.

[0023] In this embodiment, to ensure the quality and reliability of detection data, a multi-channel ultrasonic array detection system and a multi-source sensor network system must be properly arranged. Specifically, ultrasonic detection requirements must be obtained, including key information such as the detection area, detection accuracy requirements, and detection frequency. Furthermore, environmental monitoring requirements must be obtained, including information such as the temperature monitoring range and vibration monitoring frequency band, as well as the structural characteristics of rail transit equipment, such as its geometric dimensions and material properties.

[0024] When determining the structural strategy for a multi-channel ultrasonic array, it's important to consider both the coverage requirements and spatial constraints of the inspection area. For example, for rail transit wheel inspection, a linear array layout can be employed, with 16, 32, or 64 elements. The element spacing is typically set between half a wavelength and one wavelength (e.g., less than 0.5 mm), and the element size can be λ / 2) to achieve optimal acoustic field coverage. Ultrasonic excitation parameters include center frequency (e.g., 2-5 MHz), pulse width, transmitted sound pressure (e.g., 5 MPa), and repetition rate. These parameters must be selected in a balanced manner with inspection depth and resolution requirements to ensure signal penetration and resolution.

[0025] In addition, the configuration of the ultrasonic signal acquisition parameters is also crucial. In an embodiment of the present application, the ultrasonic signal sampling frequency can be set to 5-10 times the center frequency to meet the requirements of the Nyquist sampling theorem. The quantization accuracy is generally selected to be 14 bits or 16 bits to ensure that the dynamic range of the acquired signal meets the detection requirements. For example, when the ultrasonic center frequency is 5MHz, the sampling frequency can be set to 30MHz and the quantization accuracy can be set to 14 bits. The ultrasonic signal acquisition parameters also include data cache size, etc. Optionally, this embodiment establishes a real-time monitoring mechanism to ensure the reliability of data acquisition by setting a signal-to-noise ratio threshold (not less than 20dB) and waveform quality indicators.

[0026] Furthermore, based on the structural strategy, ultrasonic signal acquisition parameters and ultrasonic excitation parameters of the multi-channel ultrasonic array, an ultrasonic array layout strategy is formed and sent to the multi-channel ultrasonic array detection system, so that the multi-channel ultrasonic array detection system is arranged based on the ultrasonic array layout strategy to achieve the collection of high-quality and high-precision ultrasonic reflection signals.

[0027] It should be noted that the layout strategy of the temperature sensor array needs to take into account the hot spot distribution characteristics of rail transit equipment. In this embodiment, the temperature sensor array can be arranged in key parts of the equipment, such as bearings, motors and other parts that are prone to heat. Exemplarily, the temperature sensor can be a sensor with a temperature measurement range of -40°C to 85°C and an accuracy better than ±0.5°C. For the deployment of vibration sensors, it is necessary to consider the vibration transmission path and modal characteristics of the equipment, and the sensors are usually arranged in positions where the vibration response is obvious. For example, for the monitoring of rail transit vehicle bearings, vibration sensors can be arranged in the radial and axial directions of the bearing seat. Exemplarily, the vibration sensor can be an acceleration sensor with a measurement range of ±50g and a frequency response of 0-5kHz.

[0028] When configuring sensor data acquisition parameters, the sampling frequency for temperature data can be relatively low, such as 1Hz-10Hz, while the sampling frequency for vibration data requires a higher frequency, typically between a few kHz and tens of kHz. Triggering can be either timed or conditional, depending on actual needs. For example, data acquisition can be triggered when the device speed reaches a certain threshold. Sensor data acquisition parameters also include data packaging formats, enabling synchronous collection of environmental data.

[0029] Optionally, this embodiment can implement self-diagnosis capabilities for the multi-source sensor network system, ensuring the stability and reliability of data collection through periodic testing and status monitoring. Furthermore, a data preprocessing module is constructed to perform preliminary filtering and calibration on the collected environmental data, providing a high-quality data foundation for subsequent analysis.

[0030] Therefore, to ensure data quality, this application also requires real-time monitoring of the operating status of each sensor, timely diagnosis and treatment of possible sensor failures, and ensuring the reliability and continuity of data collection. After completing the above steps, a complete multi-source detection data stream is formed, which serves as the data foundation for subsequent processing and analysis. This solves the problem of a single detection method, realizes the collection of multi-source detection data, and improves the comprehensiveness and accuracy of detection.

[0031] S102 : performing data preprocessing on the detection data to obtain preprocessed data, and extracting a feature vector based on the preprocessed data.

[0032] In this application, data cleaning, noise reduction and standardization are achieved by performing data preprocessing and feature extraction on the detection data.

[0033] S103 : constructing a few-step diffusion model based on the feature vector and the preprocessed data, and performing data enhancement on the detection data using the few-step diffusion model to obtain synthetic data.

[0034] S104: Based on the synthetic data, construct a dense reward difference learning framework, optimize the dense reward difference learning framework, and obtain a detection model.

[0035] S105 : Construct a jump data diffusion model based on the detection model, and perform robustness analysis on the detection data based on the jump data diffusion model to obtain a robustness analysis result.

[0036] S106. Based on the robustness analysis result, a preset implicit neural representation network is used to extract the time continuity features in the detection data, and a time series anomaly detection is performed on the time series anomaly features to obtain a time series anomaly detection result.

[0037] S107. Perform multi-model data fusion on the abnormal data detection result, the robustness analysis result, and the time series anomaly detection result to obtain a detection warning result, wherein the abnormal data detection result is obtained after the detection model performs anomaly detection on the detection data.

[0038] The rail transit nondestructive testing method provided in this embodiment solves the problem of a single detection method by acquiring ultrasonic reflection signals from a multi-channel ultrasonic array detection system for rail transit equipment, as well as temperature data and vibration data from a multi-source sensor network system for rail transit equipment. It achieves the collection of multi-source detection data, improving the comprehensiveness and accuracy of detection. Subsequently, by preprocessing and extracting features from the multi-source detection data and using a few-step diffusion model for data enhancement, the fusion capability of the multi-source detection data is improved, which can meet the training requirements of subsequent models. Furthermore, by constructing a few-step diffusion model and a dense reward difference learning framework, the accuracy of abnormal data detection is improved. By constructing a jump data diffusion model, the robustness of the detection model's detection results is improved, and it can be applied to different working conditions and environmental conditions. By extracting temporal continuity features through an implicit neural representation network for time series anomaly detection, it achieves in-depth analysis of the time series characteristics of the detection data, thereby performing multi-model data fusion on the abnormal data detection results, the robustness analysis results, and the time series anomaly detection results, improving the detection accuracy and the reliability of the detection warning results.

[0039] In some embodiments, step S102 includes: Performing wavelet transform processing on the ultrasonic reflection signal, processing the high-frequency coefficients obtained after the wavelet transform, and reconstructing the signal based on the processed high-frequency coefficients; Performing median filtering on the temperature data to eliminate mutation points in the temperature data, and performing low-pass filtering on the temperature data to eliminate high-frequency noise in the temperature data; performing bandpass filtering on the vibration data and processing abnormal value points in the vibration data; Based on the timestamp information, the processed ultrasonic reflection signal, the processed temperature data, and the processed vibration data are aligned and then standardized to obtain preprocessed data.

[0040] In this embodiment, the ultrasonic reflection signal is processed using a wavelet transform method, specifically, Daubechies wavelet or Symlet wavelet, among others. Wavelet transform can decompose the signal into sub-signals of different frequency bands, and threshold processing can be performed on the high-frequency coefficients to remove noise, followed by signal reconstruction. For example, a soft thresholding method can be used to process the high-frequency coefficients, with the threshold determined based on a noise variance estimate. For example, the db4 wavelet basis function is selected, a four-layer wavelet decomposition is performed, and the high-frequency coefficients are processed using a soft thresholding method. The reconstructed signal improves the signal-to-noise ratio by at least 10 dB.

[0041] For temperature data, the median filter window size is typically 3-5 data points to effectively remove sudden changes without oversmoothing the valid signal. Low-pass filtering can use a Butterworth filter, with the cutoff frequency set based on the actual temperature variation characteristics, typically between 0.1Hz and 1Hz. For example, a 5-point median filter is used to eliminate sudden changes, while a Butterworth low-pass filter (cutoff frequency 0.1Hz) is used to remove high-frequency noise and ensure smooth temperature data.

[0042] For vibration data, the bandpass filter should be designed based on the characteristic frequency of the equipment. For example, for bearing vibration signals, the passband of the bandpass filter can be set within 0.5-2 times the fault characteristic frequency. For outlier processing, the 3σ criterion can be used for judgment and processing. For example, bandpass filtering (with a passband of 10Hz-2kHz) is applied to the vibration data to eliminate low-frequency drift and high-frequency interference, and an adaptive threshold method is used to identify and process outliers.

[0043] Furthermore, optionally, data alignment and synchronization are performed on the processed ultrasonic reflection signals, processed temperature data, and processed vibration data. Data from different sources are aligned to a unified time base based on timestamp information (with an accuracy better than 1ms). Subsequently, data normalization is performed using the Z-score method to normalize all data to a mean of 0 and a standard deviation of 1, facilitating subsequent analysis. Furthermore, a data quality assessment mechanism is established to evaluate the quality of preprocessing results by calculating metrics such as signal-to-noise ratio and effective data rate to ensure data availability.

[0044] Therefore, this embodiment performs data preprocessing through a multi-stage signal processing method to achieve noise reduction and standardization of the original detection data, thereby improving the quality of the detection data.

[0045] In some embodiments, step S102 further includes: Performing time domain feature extraction and frequency domain feature extraction on the preprocessed ultrasonic reflection signal to obtain ultrasonic reflection signal feature data; Extracting temperature characteristic data of the preprocessed temperature data, wherein the temperature characteristic data includes temperature gradient, temperature field distribution characteristics and temperature fluctuation characteristics; Extracting vibration feature data of the preprocessed vibration data, wherein the vibration feature data includes time domain statistical features, frequency domain features, and time domain joint features; The ultrasonic reflection signal characteristic data, the temperature characteristic data, and the vibration characteristic data are screened and evaluated to obtain a characteristic vector.

[0046] In this embodiment, the time-domain features of the preprocessed ultrasonic reflection signal corresponding to the ultrasonic reflection signal include statistical features such as peak value, mean, variance, skewness, and kurtosis. They also include features such as zero-crossing rate, envelope curve characteristics, and energy distribution characteristics of the time-domain waveform. Frequency-domain features include spectral energy distribution, dominant frequency component, spectral center of gravity, bandwidth, energy proportion of each frequency band (dividing the spectrum into 10 equally divided frequency bands), and frequency-domain entropy. Specifically, the spectrum can be obtained through a fast Fourier transform, and then features such as the frequency band energy ratio and spectral center of gravity can be extracted.

[0047] For temperature feature data, its extraction needs to consider both spatial and temporal dimensions. For example, temperature gradient features can be obtained by calculating the temperature difference between adjacent measurement points, temperature field distribution features can be extracted using principal component analysis, and temperature fluctuation features can be characterized by calculating the temperature change rate and fluctuation period.

[0048] For vibration characteristic data, time domain statistical features include root mean square value, peak factor, margin factor, etc., frequency domain features include frequency band energy, harmonic ratio, etc., and time-frequency joint features can be obtained through methods such as wavelet packet decomposition or Hilbert-Huang transform.

[0049] When screening and evaluating features, a combination of correlation analysis and importance assessment can be used. First, the correlation coefficient between features is calculated to eliminate highly correlated, redundant features. Feature importance is then assessed using methods such as random forests, retaining features that contribute significantly to the detection results. For example, a feature selection mechanism is constructed. Based on the mutual information criterion and Fisher score, the most discriminative feature subset (with a feature dimension limited to 20-30) is selected from the initial feature set to generate the final feature vector. A feature evaluation system is then established to evaluate the quality of feature extraction results by calculating feature stability and discrimination indices to ensure feature validity and reliability.

[0050] In this embodiment, the feature extraction link is of great significance for improving detection accuracy and reliability. First, feature extraction is performed on the pre-processed ultrasonic reflection signal. In terms of time domain feature extraction, it mainly includes the statistical characteristics of the signal, such as the peak amplitude can reflect the size of the defect, the root mean square value can characterize the energy level of the signal, the peak factor and margin factor can reflect the waveform characteristics of the signal, and the skewness and kurtosis can describe the degree of deviation of the signal amplitude distribution. In addition, it also includes the envelope characteristics of the signal. The instantaneous amplitude and phase information of the signal can be obtained through the Hilbert transform, which plays an important role in identifying the reflection characteristics of ultrasonic waves at the interfaces of different media. In terms of frequency domain feature extraction, the spectrum of the signal is first obtained through fast Fourier transform, and then the characteristics such as the band energy ratio, the main frequency component, and the center of gravity of the spectrum are extracted. These characteristics can reflect the frequency composition and energy distribution characteristics of the signal.

[0051] To preprocess temperature data, features must be extracted from both spatial and temporal dimensions. Temperature gradient features are derived by calculating the temperature difference between adjacent measurement points and can reflect the changing trends and heat conduction characteristics of the temperature field. Temperature field distribution features utilize principal component analysis to map high-dimensional temperature field data into a low-dimensional feature space and extract the primary temperature distribution patterns. Temperature fluctuation features include the temperature change rate, fluctuation period, and fluctuation amplitude. These features can reflect the dynamic temperature changes during equipment operation.

[0052] To preprocess vibration data, it is necessary to extract time-domain statistical features, frequency-domain features, and joint time-frequency features. Time-domain statistical features include root mean square value, crest factor, margin factor, and waveform factor. These features characterize the intensity and waveform characteristics of the vibration signal. Frequency-domain features are obtained by performing spectral analysis on the vibration signal and include the energy distribution of each frequency band, harmonic ratio, and sideband ratio. These features are important for identifying equipment fault types. Joint time-frequency features are obtained through methods such as wavelet packet decomposition or the Hilbert-Huang transform and can describe the local characteristics of the vibration signal in both the time and frequency dimensions.

[0053] After completing the above feature extraction, the feature data needs to be screened and evaluated to reduce feature redundancy and improve its contribution to the detection results. First, by calculating the Pearson correlation coefficient between features, highly correlated feature pairs are identified, and the feature with the better physical interpretation is retained. Then, methods such as random forests are used to assess the importance of features and calculate the contribution of each feature to model performance. Specifically, features can be ranked by feature importance score, and a subset of features with high scores can be selected. In addition, domain knowledge can be incorporated to retain features that have clear physical meaning and are important indicators for fault diagnosis. Through this feature screening and evaluation mechanism, a feature vector with moderate dimensionality and rich information is ultimately obtained, providing a reliable data foundation for subsequent fault detection and diagnosis.

[0054] It should be noted that the feature extraction and screening process needs to consider the characteristics of the actual application scenario. For example, for high-speed rail transit equipment, it may be necessary to pay more attention to the high-frequency characteristics of the vibration signal; for components prone to overheating failures, it is necessary to focus on extracting abnormal characteristics of the temperature field. Therefore, in actual applications, the feature extraction strategy can be appropriately adjusted according to specific needs to obtain the optimal detection effect. In this embodiment, feature extraction and selection are performed on multi-source detection data to achieve data dimensionality reduction and information extraction, facilitating multi-source data fusion.

[0055] In some embodiments, constructing a few-step diffusion model based on the feature vector and the preprocessed data includes: Based on the feature vector, constructing a few-step diffusion initial model, the few-step diffusion initial model including a basic model architecture, a noise scheduling strategy, a sampling mechanism, a loss function, and an optimization strategy; Dividing the feature vector and the preprocessed data into a training set, a validation set, and a test set; Using the training set to train the few-step diffusion initial model; Using the validation set to evaluate and optimize the trained initial few-step diffusion model; The optimized few-step diffusion initial model is tested using the test set to obtain the few-step diffusion model.

[0056] In this embodiment, in order to solve the problem of insufficient model training data, the detection data is enhanced by constructing a few-step diffusion model to achieve the generation of high-quality samples and the optimization of data distribution. First, a few-step diffusion initial model is constructed based on the obtained feature vector. The basic model architecture of the few-step diffusion initial model uses U-Net as the backbone network, including two main parts: an encoder and a decoder. Among them, the encoder is composed of multiple downsampling blocks, each of which contains two convolutional layers, a batch normalization layer and a ReLU activation function, and extracts multi-scale features by gradually reducing the spatial resolution of the feature map. The decoder contains corresponding upsampling blocks, which gradually restore the spatial resolution of the feature map through transposed convolution or upsampling operations. It should be noted that a jump connection is set between the encoder and the decoder, which can effectively retain detail information and prevent information from being lost during multiple downsampling processes.

[0057] For example, the basic architecture of the few-step diffusion initialization model uses a U-Net-based backbone network, consisting of a four-layer encoder and a four-layer decoder. Each encoder layer contains two 3×3 convolutional layers (with 64, 128, 256, and 512 channels, respectively) and a maximum pooling layer. Each decoder layer contains a transposed convolutional layer and two 3×3 convolutional layers. Skip connections are added between layers to ensure the preservation of detailed information. Furthermore, the dimensions of the input and output layers match the dimensions of the feature vectors.

[0058] In addition, the noise scheduling strategy for the few-step diffusion initialization model adopts cosine scheduling. Specifically, during training, the noise level gradually decreases from high to low, and the decreasing curve follows the cosine function. This scheduling method can improve computational efficiency while ensuring sampling quality. For example, the progressive noise addition mechanism sets the number of diffusion steps to 10, and the noise scale increases linearly in the range of β = 0.1 to 0.9 at each step, achieving adaptive noise intensity adjustment. The noise parameters are dynamically adjusted according to data characteristics, and noise impact evaluation indicators are established, including signal-to-noise ratio and data fidelity.

[0059] In this embodiment, the sampling mechanism for the few-step diffusion initialization model uses a predictor-corrector sampler, which gradually transforms noisy data into meaningful samples through multi-step prediction and correction operations. Specifically, the sampling mechanism implements a predictor-corrector sampling strategy, employs importance sampling techniques to accelerate convergence, designs a batch parallel sampling mechanism to improve generation efficiency, and constructs a sampling quality assessment indicator system.

[0060] In this embodiment, the loss function for the few-step diffusion initialization model uses a combination of a simple loss and a variational lower bound. The simple loss is used to measure the reconstruction error, while the variational lower bound is used to constrain the distribution of the latent space. For example, a mean squared error loss and a knowledge distillation loss are combined, a regularization term is added to control model complexity, and a contrastive learning loss is introduced to enhance feature representation, with weight coefficients set to 0.5, 0.3, and 0.2, respectively.

[0061] In this example, the Adam optimizer is used for the optimization strategy of the initial model of the few-step diffusion, and the learning rate is scheduled using cosine annealing to achieve better convergence. For example, the Adam optimizer is used, the initial learning rate is set to 0.001, cosine annealing is used, the cycle is set to 10 epochs, an early stopping mechanism is implemented, the tolerance is set to 5 epochs, gradient clipping is implemented, and the threshold is set to 5.0.

[0062] Before model training, it is necessary to reasonably divide the feature vectors and preprocessed data. In the embodiment of the present application, the data set is divided into a training set, a validation set, and a test set in a ratio of 6:2:2. The training set is used for parameter learning of the model, the validation set is used for performance evaluation and hyperparameter tuning of the model, and the test set is used to evaluate the generalization ability of the model. It should be noted that the consistency of the data distribution should be maintained when dividing the data. Stratified sampling can be used for division to ensure that the proportions of samples of different categories in each subset are similar.

[0063] During the model training phase, small batch stochastic gradient descent is used for optimization. Specifically, the training data is divided into multiple batches, and the batch size can be set to 32 or 64. The specific value needs to be determined according to the GPU memory capacity and training efficiency. During the training process, the model parameters are continuously updated through the back propagation algorithm until the preset number of training rounds is reached or the convergence condition is met. For example, it can be set that when the loss on the validation set changes by less than 0.1% in 5 consecutive rounds, the model is considered to have reached convergence. Exemplarily, this embodiment performs progressive training, adopts a curriculum learning strategy, and gradually increases the noise level. The noise scale is set to 0.1 in the initial training phase, and increases by 0.1 every 5 epochs. At least 100 iterations are performed at each noise level, and the quality indicators of the generated samples are monitored, including the FID score and the IS score.

[0064] During the model evaluation and optimization phase, the primary focus is on performance metrics on the validation set, including reconstruction error and quality scores of generated samples. Based on these evaluation results, model hyperparameters can be adjusted, such as changing the number of network layers, adjusting the learning rate, and modifying the batch size. Furthermore, an early stopping strategy can be employed, terminating training when performance metrics on the validation set no longer improve to prevent model overfitting.

[0065] Finally, the optimized model is tested on the test set to evaluate its generalization performance. The testing phase focuses on the model's performance on unseen data, including the quality and diversity of the generated samples, as well as their consistency with the real-world data distribution. It is important to note that the test set should only be used once, and further adjustments to the model based on the test results should not be made to ensure the objectivity of the evaluation results.

[0066] Furthermore, after completing model training and testing, the trained few-step diffusion model can be used to perform data augmentation on the test data. Specifically, the test data is first input into the model, and a sampling process is used to generate synthetic data with similar characteristics to the original data. This synthetic data retains the statistical properties of the original data while also incorporating a certain amount of random variation, effectively expanding the number and diversity of training samples. It is important to note that the generated synthetic data should undergo a quality assessment to remove obviously unreasonable or low-quality samples to ensure the reliability of the augmented dataset.

[0067] Therefore, this embodiment enhances the detection data by constructing a few-step diffusion model, generating high-quality samples and optimizing data distribution. Specifically, by optimizing the training of the initial few-step diffusion model, the model performance is optimized and the generation quality is improved.

[0068] In some embodiments, step S104 includes: Based on the synthesized data, constructing a local feature reward function and a global feature reward function, wherein the local feature reward function is used to indicate a reward for the similarity of local features, and the global feature reward function is used to indicate a reward for the consistency of global features; Based on the local feature reward function and the global feature reward function, constructing a dense reward difference learning framework; Calculating sample weights based on the dense reward differential learning framework and performing differential optimization based on contrastive learning on the synthetic data; Optimizing parameters of the dense reward difference learning framework based on the sample weights; The dense reward difference learning framework after parameter optimization is evaluated and optimized to obtain the detection model.

[0069] Optionally, the local feature reward function Rlocal can be defined as: ; Where fi represents the feature representation of the i-th sample, N(i) represents the neighbor set of the i-th sample, fj represents the feature representation of the j-th sample, and sim() represents the cosine similarity. The samples in the local feature reward function are derived from synthetic data.

[0070] The global feature reward function Rglobal can be defined as: ; in, Represents the mean of all sample features.

[0071] In calculating sample weights When , it can be calculated based on the sparsity and representativeness of the sample: ; in, is a balancing factor, usually 0.5, sparsity(fi) represents the sparsity of the feature representation of the i-th sample, and representativeness(fi) represents the representativeness of the feature representation of the i-th sample.

[0072] In this embodiment, the construction of a dense reward difference learning framework is of great significance for improving the performance of the detection model. First, a local feature reward function and a global feature reward function are constructed. Among them, the local feature reward function is mainly used to measure the similarity of samples in the local feature space, and the reward value is obtained by calculating the feature similarity between each sample and its neighboring samples. This local feature similarity calculation can help the model better understand the local structural characteristics of the data. The global feature reward function focuses on the distribution of samples in the overall feature space, and obtains the reward value by calculating the difference between the sample features and the overall feature mean, which helps to maintain the diversity of sample features and prevent the model from falling into a local optimal solution.

[0073] When building a dense reward differential learning framework, it is necessary to organically combine local feature reward functions with global feature reward functions. Specifically, a basic deep neural network is first designed as a feature extractor to extract valid features from the input data. Then, local and global reward values ​​are calculated based on the extracted features, and these two rewards are combined in a weighted manner to form the final reward signal. Furthermore, a corresponding loss function is designed to guide the model training process, allowing it to simultaneously focus on the similarity of local features and the consistency of global features.

[0074] When calculating sample weights and performing variance optimization, the importance of each sample must first be assessed. Specifically, weights can be calculated based on the sample's sparsity and representativeness. Sparsity reflects the uniqueness of a sample in the feature space, while representativeness reflects its contribution to the overall data distribution. After obtaining sample weights, variance optimization is performed, using comparative learning to enable the model to learn subtle differences between samples. This process prioritizes samples with higher weights, thereby improving the model's ability to identify key samples.

[0075] Based on the obtained sample weights, the dense reward difference learning framework is optimized. During this phase, model parameters are updated in a weighted manner, so that the model pays more attention to samples with higher weights during training. Specifically, targeted parameter optimization can be achieved by adjusting the contribution ratio of different samples in the loss function or by weighting the gradients during backpropagation. It is important to note that the parameter optimization process should adopt an appropriate learning rate scheduling strategy to ensure stable model convergence.

[0076] Finally, the optimized framework is systematically evaluated and optimized. The evaluation process focuses on the model's performance metrics on the validation set, including detection accuracy, recall, and precision. Based on the evaluation results, further adjustments can be made to the various components of the framework, such as modifying the network structure, adjusting the weights of the reward function, and optimizing the calculation of sample weights. This process requires multiple iterations until the model performance reaches the desired level. It is important to note that to ensure model robustness, the evaluation data should encompass a wide range of possible operating conditions and fault types to ensure that the resulting detection model has good generalization capabilities.

[0077] Furthermore, after obtaining the final detection model, model compression and acceleration optimization are required to meet the needs of actual deployment. For example, techniques such as knowledge distillation, model pruning, or quantization can be used to reduce the model's computational complexity and storage overhead while maintaining detection performance, enabling efficient operation in resource-constrained environments.

[0078] Therefore, this embodiment constructs a dense reward difference learning framework by constructing a local feature reward function and a global feature reward function, thereby constructing a detection model, and realizing the accurate identification and evaluation of abnormal samples in the detection data using the detection model, thereby improving the accuracy of abnormal identification of the detection data.

[0079] In some embodiments, step S105 includes: Based on the detection model, a multi-scale feature extraction network is constructed, wherein the multi-scale feature extraction network includes a skip connection structure in which a plurality of feature scale layers are arranged in cascade and adjacent feature scale layers are arranged; Constructing a feature fusion mechanism, wherein the feature fusion mechanism is used to instruct to use a preset channel attention mechanism and a spatial attention mechanism to perform feature fusion on the detection data; Constructing an anti-interference network, wherein the anti-interference network is used to instruct a preset spatial transformation network to perform feature alignment on the detection data; Constructing a jump data diffusion model based on the multi-scale feature extraction network, the feature fusion mechanism and the anti-interference network; Performing system performance evaluation and optimization on the jump data diffusion model; A robustness analysis is performed on the detection data based on the optimized jump data diffusion model to obtain a robustness analysis result.

[0080] It should be noted that the multi-scale feature extraction network of the jump data diffusion model adopts a pyramid structure, consisting of 3-5 feature scale layers, with the feature map size of each layer being halved. The skip connections between adjacent scale layers use residual connections, which effectively avoid the vanishing gradient problem.

[0081] The channel attention in the feature fusion mechanism uses a squeeze-and-excitation structure, while the spatial attention uses a self-attention mechanism. Specifically, the channel attention module first performs global average pooling on the feature map and then learns channel weights through two fully connected layers. The spatial attention module calculates the correlation between different locations in the feature map to obtain spatial weights.

[0082] The interference-rejection network adopts the concept of a spatial transformer network and consists of three main components: a localization network, a grid generator, and a sampler. The localization network predicts the spatial transformation parameters, the grid generator generates a sampling grid based on these parameters, and the sampler resamples the feature map according to the sampling grid. This mechanism adaptively aligns and corrects features, improving the model's robustness to pose variations.

[0083] In this embodiment, the construction of a jump data diffusion model is a key link in achieving robustness detection. First, a multi-scale feature extraction network is constructed, which adopts a pyramid structure and contains 3 to 5 feature scale layers. Each feature scale layer gradually reduces the spatial resolution of the feature map through convolution operations and downsampling operations, while increasing the number of feature channels to obtain feature representations of different scales. Skip connections are set between adjacent feature scale layers in the form of residual connections. This design can effectively alleviate the gradient vanishing problem and can retain more detailed information during the feature extraction process. It should be noted that each feature scale layer contains multiple residual blocks, each residual block consists of a convolution layer, a batch normalization layer and an activation function. This deep residual structure can improve the feature extraction capability of the network.

[0084] The feature fusion mechanism consists of two key components: channel attention and spatial attention. The channel attention mechanism employs a squeeze-and-excitation architecture. It first obtains statistical features for each channel through global average pooling. It then learns the inter-channel relationships through a two-layer fully connected network, ultimately generating channel weights. This mechanism helps the model identify and enhance important feature channels. The spatial attention mechanism, on the other hand, employs a self-attention architecture, generating spatial weights by calculating correlations between different locations in the feature map. This allows the model to adaptively focus on important spatial regions, improving the targeted nature of feature extraction. Furthermore, the feature fusion process employs a multi-level fusion strategy, integrating features across different feature scales to fully leverage multi-scale feature information.

[0085] The anti-interference network is constructed based on the concept of a spatial transformer network. First, a localization network predicts the spatial transformation parameters. These parameters describe the required geometric transformations of the feature map, such as translation, rotation, and scaling. Then, a grid generator generates a sampling grid based on the predicted transformation parameters. This grid defines the sampling locations on the feature map. Finally, a sampler resamples the feature map according to the sampling grid to obtain a spatially transformed feature map. This mechanism enables the model to adaptively align and correct features, effectively reducing interference caused by pose variations. It is important to note that the spatial transformation parameters are learned end-to-end, requiring no additional supervision.

[0086] Based on the three core components described above, a complete skip data diffusion model is constructed. This model employs an encoder-decoder architecture, where the encoder includes a multi-scale feature extraction network to obtain multi-scale feature representations. The decoder reconstructs a representation of the input data through progressive upsampling and feature fusion. Multiple layers of skip connections are established between the encoder and decoder, enabling the decoder to better utilize the feature information extracted by the encoder. The model is trained end-to-end, with the loss function consisting of a reconstruction loss to ensure the accuracy of the reconstruction results and a regularization term to constrain the model's complexity.

[0087] During the system performance evaluation and optimization phase, the model's performance under various operating conditions, including varying noise levels, interference types, and failure modes, is first assessed. Evaluation metrics include reconstruction error, feature extraction capability, and interference immunity. Based on the evaluation results, targeted optimization of the model's components can be performed, such as adjusting the network structure, modifying the attention mechanism, and optimizing the loss function. Furthermore, the model's computational efficiency and resource consumption must be considered. Techniques such as model compression or knowledge distillation can be employed as necessary.

[0088] Finally, the optimized model is used to perform a robustness analysis on the detection data. Specifically, different types of interference data are first constructed, such as adding noise, performing geometric transformations, and changing lighting conditions. The model's performance on this interference data is then evaluated, including the stability of feature extraction and the consistency of detection results. This systematic robustness analysis allows for a comprehensive assessment of the model's anti-interference capabilities and identifies potential performance bottlenecks. It is important to note that the results of this robustness analysis can serve as an important basis for further model optimization, helping to improve the model's reliability in practical applications.

[0089] Therefore, this embodiment constructs a jump data diffusion model through a multi-scale feature extraction network, a feature fusion mechanism and an anti-interference network, which can be used to perform data enhancement and feature extraction on the detection data, thereby meeting the robustness analysis requirements of the detection data and realizing the robustness of the analysis of the detection data. It solves the problem that the existing detection model is insufficient in robustness and makes it difficult to adapt to different working conditions and environmental conditions, realizes the enhanced adaptability and stability of data under different working conditions, and improves the comprehensiveness and accuracy of the analysis of the detection data.

[0090] In some embodiments, step S106 includes: Based on the robustness analysis results, constructing an implicit neural representation network, wherein the implicit neural representation network includes a time continuity encoder, a temporal dependency model, and an anomaly detection discriminator; extracting a temporal continuity feature from the detection data based on the implicit neural representation network, and calculating an anomaly score of the temporal continuity feature; The anomaly score of the time continuity feature is judged based on a preset anomaly threshold to obtain a corresponding time series anomaly detection result.

[0091] In this embodiment, the temporal continuity encoder of the implicit neural representation network uses a multilayer perceptron structure. The number of neurons in the input layer is equal to the feature dimension, and the hidden layer uses the ReLU activation function. The temporal dependency model uses a long short-term memory (LSTM) network structure based on an attention mechanism to capture both long-term and short-term temporal dependencies. The anomaly detection discriminator uses a reconstruction error-based approach, first reconstructing the input features and then calculating the reconstruction error as an anomaly score.

[0092] When determining the anomaly threshold, it can be set based on the anomaly score distribution of normal samples. Then, when the anomaly score is detected to exceed the anomaly threshold, it is considered that a time series anomaly exists, and the corresponding time series anomaly detection result is output.

[0093] In this embodiment, time series anomaly detection is implemented based on an implicit neural representation network. First, an implicit neural representation network is constructed, which includes three main components: a time continuity encoder, a time dependency model, and an anomaly detection discriminator. Among them, the time continuity encoder adopts a multi-layer perceptron structure, which maps the input features to an implicit feature space through multiple fully connected layers and nonlinear activation functions. It should be noted that the design of the encoder needs to consider the time continuity of the features, so a time embedding layer is added to the network structure to capture the time series information of the features. The time dependency model adopts a long short-term memory network structure based on the attention mechanism, which can effectively model long-term and short-term time series dependencies. Specifically, through the gating mechanism and memory units, the model can selectively remember and forget historical information, thereby capturing important patterns in the data sequence. The anomaly detection discriminator adopts a reconstruction-based approach to achieve anomaly detection by reconstructing the input features and calculating the reconstruction error.

[0094] During the temporal continuity feature extraction and anomaly score calculation phase, the detection data is first organized into time series and fed into an implicit neural representation network for processing. The temporal continuity encoder converts the input features into implicit representations, preserving the temporal continuity information of the features. A temporal dependency model then models the implicit representations, capturing the temporal dependencies within the feature sequence. By comparing the features at the current moment with the features predicted by the model, potential anomaly patterns can be identified. Furthermore, the multi-scale nature of the features needs to be considered. Therefore, a multi-scale analysis approach is employed during feature extraction, extracting features and calculating anomaly scores at different time scales. It is important to note that the calculation of the anomaly score considers not only the reconstruction error but also the temporal continuity and prediction bias of the features, enabling a more comprehensive assessment of the degree of anomaly in the data.

[0095] Finally, the anomaly score is judged based on the preset anomaly threshold to obtain the time series anomaly detection result. The setting of the threshold is a key link, and a balance needs to be struck between the false alarm rate and the missed alarm rate. In the embodiment of the present application, a method based on the statistical characteristics of normal samples is used to determine the threshold. Specifically, a reasonable threshold can be set by analyzing the distribution characteristics of the anomaly score under normal operating conditions. In addition, considering that the characteristics of different types of anomalies may be different, different thresholds can be set for different feature dimensions. This adaptive threshold setting method can improve the accuracy of detection. It should be noted that in actual applications, the setting of the threshold also needs to take into account the specific application scenarios and safety requirements, and can be dynamically adjusted according to the actual operating conditions.

[0096] In addition, to improve the reliability of the detection results, the embodiments of the present application also introduce a time persistence judgment mechanism. Specifically, the abnormal detection result is triggered only when the abnormal state persists for a certain period of time, which can effectively filter out false alarms caused by instantaneous fluctuations. At the same time, different duration requirements are set for abnormalities of different severity. This hierarchical judgment mechanism can better guide actual maintenance decisions. It should be noted that the judgment threshold of time persistence can also be adjusted according to actual operating experience to adapt to different application requirements.

[0097] Finally, the system records and analyzes historical anomaly detection results. By statistically analyzing the patterns and characteristics of anomaly occurrences, it continuously optimizes detection strategies and parameter settings. For example, it can analyze the temporal and spatial distribution of anomalies and their correlation with equipment operating status. These results provide important insights for predictive maintenance. Furthermore, anomaly detection models can be regularly evaluated and updated based on historical data to ensure the continued effectiveness of detection performance.

[0098] Therefore, this embodiment extracts the time continuity features in the detection data through the implicit neural representation network to analyze whether there are time series anomalies in the detection data, realizes in-depth analysis of the time series features of the detection data, and improves the accuracy of the detection analysis.

[0099] In some embodiments, step S107 includes: Assigning corresponding weight coefficients to the abnormal data detection result, the robustness analysis result, and the time series anomaly detection result; Performing weighted fusion on the abnormal data detection result, the robustness analysis result, and the time series anomaly detection result according to the weight coefficient; Output corresponding detection and warning results based on the weighted fusion results.

[0100] In this embodiment, the weight coefficients corresponding to the abnormal data detection results, robustness analysis results, and time series anomaly detection results are determined respectively, and then weighted calculations are performed based on them to output the corresponding detection and warning results. It should be noted that the output of the detection and warning results needs to take into account the needs of the actual application scenario. In the embodiment of the present application, the warning results can be divided into multiple levels, such as normal, slight abnormality, severe abnormality, etc. Each level corresponds to a different processing strategy, such as continued monitoring, increased detection frequency, immediate shutdown and inspection, etc.

[0101] Furthermore, to improve the reliability of early warning results, a time persistence mechanism can be introduced. For example, an early warning of the corresponding level will only be triggered when the abnormal state persists for a certain period of time (e.g., three consecutive detection cycles). This can effectively avoid false alarms caused by instantaneous fluctuations.

[0102] Therefore, this embodiment performs multi-model data fusion on the abnormal data detection results, robustness analysis results and time series anomaly detection results, thereby improving the detection accuracy and the reliability of the detection warning results.

[0103] An embodiment of the present application provides a rail transit non-destructive testing device, including a processor; a memory for storing a computer program; wherein, when the processor executes the computer program, the steps of the rail transit non-destructive testing method described in the above embodiment are implemented.

[0104] In an optional embodiment, a nondestructive testing device for rail transportation is provided, such as Figure 2 As shown, Figure 2 The illustrated rail transit nondestructive testing device 200 includes: a processor 201 and a memory 203. The processor 201 and the memory 203 are connected, for example, via a bus 202. Optionally, the rail transit nondestructive testing device 200 may further include a transceiver 204, which may be used for data exchange between the rail transit nondestructive testing device and other devices, such as data transmission and / or data reception. It should be noted that in actual applications, the number of transceivers 204 is not limited to one, and the structure of the rail transit nondestructive testing device 200 does not constitute a limitation on the embodiments of the present application.

[0105] Processor 201 can be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 201 can also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0106] The bus 202 may include a path for transmitting information between the above components. The bus 202 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. The bus 202 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 2Only one thick line is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0107] The memory 203 may be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media, other magnetic storage devices, or any other medium that can be used to carry or store computer programs and can be read by a computer, without limitation herein.

[0108] The memory 203 is used to store the computer program for executing the embodiments of the present application, and the execution is controlled by the processor 201. The processor 201 is used to execute the computer program stored in the memory 203 to implement the steps shown in the above method embodiments.

[0109] The above is a detailed description of the specific implementation of the embodiment of the present application. Through the implementation of the above technical solution, efficient, accurate and reliable non-destructive testing of rail transit equipment can be achieved, providing a strong guarantee for the safe operation of rail transit.

Claims

1. A rail transit nondestructive testing method, characterized in that: include: Acquire detection data for rail transit equipment, the detection data including ultrasonic reflection signals from a multi-channel ultrasonic array detection system, and temperature data and vibration data from a multi-source sensor network system; Performing data preprocessing on the detection data to obtain preprocessed data, and extracting a feature vector based on the preprocessed data; Based on the feature vector and the preprocessed data, a few-step diffusion model is constructed, and data enhancement is performed on the detection data through the few-step diffusion model to obtain synthetic data; Based on the synthetic data, a dense reward difference learning framework is constructed, and the dense reward difference learning framework is optimized to obtain a detection model; Based on the detection model, a jump data diffusion model is constructed, and based on the jump data diffusion model, a robustness analysis is performed on the detection data to obtain a robustness analysis result; Based on the robustness analysis result, a preset implicit neural representation network is used to extract the time continuity feature in the detection data, and a time series anomaly detection is performed on the time series anomaly feature to obtain a time series anomaly detection result; Multi-model data fusion is performed on the abnormal data detection result, the robustness analysis result and the time series anomaly detection result to obtain a detection warning result, wherein the abnormal data detection result is obtained after the detection model performs anomaly detection on the detection data.

2. The method according to claim 1, characterized in that The performing data preprocessing on the detection data to obtain preprocessed data includes: Performing wavelet transform processing on the ultrasonic reflection signal, processing the high-frequency coefficients obtained after the wavelet transform, and reconstructing the signal based on the processed high-frequency coefficients; Performing median filtering on the temperature data to eliminate mutation points in the temperature data, and performing low-pass filtering on the temperature data to eliminate high-frequency noise in the temperature data; Performing bandpass filtering on the vibration data and processing abnormal value points in the vibration data; Based on the timestamp information, the processed ultrasonic reflection signal, the processed temperature data and the processed vibration data are aligned and standardized to obtain preprocessed data.

3. The method according to claim 2, characterized in that The preprocessed data includes preprocessed ultrasonic reflection signals, preprocessed temperature data, and preprocessed vibration data. The extracting of feature vectors based on the preprocessed data includes: Performing time domain feature extraction and frequency domain feature extraction on the preprocessed ultrasonic reflection signal to obtain ultrasonic reflection signal feature data; Extracting temperature characteristic data of the preprocessed temperature data, wherein the temperature characteristic data includes temperature gradient, temperature field distribution characteristics and temperature fluctuation characteristics; Extracting vibration feature data of the preprocessed vibration data, wherein the vibration feature data includes time domain statistical features, frequency domain features, and time domain joint features; The ultrasonic reflection signal characteristic data, the temperature characteristic data and the vibration characteristic data are screened and evaluated to obtain a characteristic vector.

4. The method according to claim 1, characterized in that: The step of constructing a dense reward difference learning framework based on the synthetic data, optimizing the dense reward difference learning framework, and obtaining a detection model includes: Based on the synthetic data, construct a local feature reward function and a global feature reward function, wherein the local feature reward function is used to indicate a reward for the similarity of local features, and the global feature reward function is used to indicate a reward for the consistency of global features; Based on the local feature reward function and the global feature reward function, construct a dense reward difference learning framework; Based on the dense reward difference learning framework, calculating sample weights, and performing difference optimization based on contrastive learning of the synthetic data; Optimizing parameters of the dense reward difference learning framework based on the sample weights; The dense reward difference learning framework with optimized parameters is evaluated and optimized to obtain the detection model.

5. The method according to claim 1, characterized in that The step of constructing a jump data diffusion model based on the detection model, and performing robustness analysis on the detection data based on the jump data diffusion model to obtain a robustness analysis result includes: Based on the detection model, a multi-scale feature extraction network is constructed, wherein the multi-scale feature extraction network includes a skip connection structure in which a plurality of feature scale layers arranged in cascade are arranged between adjacent feature scale layers; Constructing a feature fusion mechanism, wherein the feature fusion mechanism is used to indicate that a preset channel attention mechanism and a spatial attention mechanism are used to perform feature fusion on the detection data; Constructing an anti-interference network, wherein the anti-interference network is used to instruct to perform feature alignment on the detection data using a preset spatial transformation network; Based on the multi-scale feature extraction network, the feature fusion mechanism and the anti-interference network, a jump data diffusion model is constructed; Performing system performance evaluation and optimization on the jump data diffusion model; Based on the optimized jump data diffusion model, a robustness analysis is performed on the detection data to obtain a robustness analysis result.

6. The method according to claim 1, characterized in that Based on the robustness analysis result, the preset implicit neural representation network is used to extract the time continuity feature in the detection data, and the time continuity feature is subjected to time series anomaly detection to obtain the time series anomaly detection result, including: Based on the robustness analysis result, constructing an implicit neural representation network, wherein the implicit neural representation network includes a time continuity encoder, a temporal dependency model and an anomaly detection discriminator; Extracting a temporal continuity feature in the detection data based on the implicit neural representation network, and calculating an abnormality score of the temporal continuity feature; The anomaly score of the time continuity feature is judged based on a preset anomaly threshold to obtain a corresponding time series anomaly detection result.

7. The method according to claim 1, characterized in that The performing multi-model data fusion on the abnormal data detection result, the robustness analysis result and the time series abnormality detection result to obtain the detection warning result includes: Assigning corresponding weight coefficients to the abnormal data detection result, the robustness analysis result and the timing anomaly detection result; Performing weighted fusion on the abnormal data detection result, the robustness analysis result and the time series abnormality detection result according to the weight coefficient; Output the corresponding detection and warning results based on the weighted fusion results.

8. The method according to claim 1, characterized in that Before obtaining the detection data for the rail transit equipment, the method further includes: Obtaining ultrasonic testing requirements, environmental monitoring requirements and structural characteristics of the rail transit equipment; Based on the ultrasonic detection requirements and the structural characteristics of the rail transit equipment, determine the structural strategy and ultrasonic excitation parameters of the multi-channel ultrasonic array, wherein the structural strategy includes the number, spacing and arrangement of array elements; Configuring ultrasonic signal acquisition parameters of the multi-channel ultrasonic array, wherein the ultrasonic signal acquisition parameters include ultrasonic signal sampling frequency and quantization accuracy; Based on the structural strategy of the multi-channel ultrasonic array, the ultrasonic signal acquisition parameters and the ultrasonic excitation parameters, an ultrasonic array arrangement strategy is formed and sent to the multi-channel ultrasonic array detection system, so that the multi-channel ultrasonic array detection system is arranged based on the ultrasonic array arrangement strategy; Based on the environmental monitoring requirements and the structural characteristics of the rail transit equipment, determine the layout strategy of the temperature sensor array and the deployment strategy of the vibration sensor; Configuring sensor data acquisition parameters of the temperature sensor array and the vibration sensor, wherein the sensor data acquisition parameters include sensor data sampling frequency and triggering mode; Based on the layout strategy of the temperature sensor array, the deployment strategy of the vibration sensor and the sensor data acquisition parameters, a sensor network deployment strategy is formed and sent to the multi-source sensor network system, so that the multi-source sensor network system is deployed based on the sensor network deployment strategy.

9. The method according to claim 1, characterized in that: The constructing of a few-step diffusion model based on the feature vector and the preprocessed data includes: Based on the feature vector, construct a few-step diffusion initial model, wherein the few-step diffusion initial model includes a basic model architecture, a noise scheduling strategy, a sampling mechanism, a loss function, and an optimization strategy; Dividing the feature vector and the preprocessed data into a training set, a validation set and a test set; Using the training set to train the few-step diffusion initial model; Using the validation set to evaluate and optimize the trained initial model of few-step diffusion; The optimized few-step diffusion initial model is tested using the test set to obtain the few-step diffusion model.

10. A rail transit nondestructive testing device, characterized in that: include: processor; memory for storing computer programs; Wherein, when the processor executes the computer program, the steps of the method according to any one of claims 1 to 9 are implemented.

Images