A system and method for detecting current of high-voltage box of urban rail train

Abstract

This invention relates to the field of electrical measurement technology, specifically to a current detection system and method for a high-voltage box in an urban rail train. The system includes: a signal acquisition module for acquiring the original current measurement signal; a signal feature quantization module for inputting the original current measurement signal into a current reconstruction model to generate a reconstructed current measurement signal and calculating the signal quality index; a data correlation analysis module for analyzing the signal quality index and the original current measurement signal to construct a correlation strength matrix and a causal relationship graph; a numerical correction module for identifying and correcting outlier data points in the correlation strength matrix and the causal relationship graph, and outputting the corrected current value; and a current state judgment module for predicting a probabilistically reasonable current threshold range based on the multi-dimensional operating condition vector of the urban rail train using a Gaussian process regression model, comparing it with the corrected current value, generating an operating state judgment result, and outputting it along with the predicted value. This system improves the diagnostic depth and environmental adaptability of current detection, ensuring the accuracy and reliability of the measurement results.

Description

Technical Field

[0001] This invention relates to the field of electrical measurement technology, specifically to a current detection system and method for a high-voltage box in an urban rail train. Background Technology

[0002] The high-voltage system of urban rail transit trains is the energy source for its core subsystems such as traction and auxiliary power supply. As a critical node in this system, the large current flowing through the high-voltage box is a core physical quantity reflecting the system's power transmission status. Therefore, obtaining the operating current value of the high-voltage box in real time and accurately is a fundamental prerequisite for electrical system status assessment and safety monitoring. However, in existing technologies, accurately and reliably measuring this current value faces several challenges, specifically in the following aspects:

[0003] Traditional measurement schemes typically employ only one current sensor in each high-voltage box. Under harsh operating conditions characterized by severe train vibrations, wide temperature variations, and strong electromagnetic interference, this single sensor is highly susceptible to physical degradation, such as zero-point drift, decreased linearity, or output signal distortion, and may even fail completely. In such cases, the system's output measurement value will deviate significantly from the true value.

[0004] To improve reliability, some current detection schemes employ multiple sensors for redundant measurements. However, existing data processing methods are relatively rudimentary. While these methods can handle sudden, large deviations from a single sensor, they struggle to effectively identify and correct for simultaneous, small-amplitude, unidirectional, and gradual drifts from multiple sensors, or more subtle anomalies at the signal waveform level. More importantly, existing methods fail to fully utilize the inherent coupling relationships and causal logic between multiple measurement signals to cross-validate and refine the data. Consequently, the accuracy and reliability of the final measurement value output after fusion cannot be fundamentally guaranteed.

[0005] When interpreting the validity of test results, existing technologies mostly rely on fixed, reasonable ranges. However, the normal value of train current can vary drastically depending on its operating conditions. Using a fixed range to evaluate dynamically changing measurements can lead to situations where even accurate measurements may be misjudged as abnormal under certain conditions, and vice versa.

[0006] Therefore, a current detection system and method for high-voltage boxes in urban rail trains are proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a current detection system and method for high-voltage boxes in urban rail trains, used for current detection in high-voltage boxes of urban rail trains. To address the problems existing in the prior art, this invention first uses a signal feature quantization module to perform quality assessment on the original current measurement signal using a current reconstruction model, outputting a signal quality index. Second, a data correlation analysis module constructs a correlation strength matrix and a causal relationship graph. Then, a numerical correction module identifies and intelligently corrects outlier measurement values ​​caused by correlation failures, outputting the corrected current value. Finally, a current state interpretation module, based on the real-time operating conditions of the train, calls a Gaussian process regression model to predict a dynamic and reasonable current threshold range, and performs state interpretation on the corrected current value before outputting it. This invention, through a progressively layered data processing flow, significantly improves the reliability, diagnostic depth, and environmental adaptability of current detection.

[0008] To achieve the above objectives, the present invention provides a high-voltage box current detection system for urban rail trains, specifically comprising the following modules:

[0009] Signal acquisition module: Real-time acquisition of the current in the high-voltage box of the urban rail train to obtain the raw current measurement signal;

[0010] Signal feature quantization module: Inputs the original current measurement signal into the current reconstruction model to generate a reconstructed current measurement signal, and calculates the signal quality index of the original current measurement signal through the quality index generator;

[0011] Data correlation analysis module: Analyzes the signal quality index and the original current measurement signal, infers and constructs a correlation strength matrix and a causal relationship diagram;

[0012] Numerical correction module: Identifies outlier data points in the correlation strength matrix and the causal relationship graph, identifies and corrects outlier measurement values, and outputs the corrected current value;

[0013] Current status judgment module: acquires the multi-dimensional operating condition vector of the urban rail train in real time, calls the Gaussian process regression model to predict the probabilistic reasonable current threshold range; compares the corrected current value with the probabilistic reasonable current threshold range, generates the operating status judgment result and outputs it together with the corrected current value.

[0014] Preferably, the process by which the signal acquisition module acquires the original current measurement signal includes: using a sensor installed inside the high-voltage box to convert the current into an analog voltage measurement signal; and using a multi-channel synchronous analog-to-digital converter interface to synchronously sample and digitize the analog voltage measurement signals of all channels to generate the original current measurement signal.

[0015] Preferably, the signal feature quantization module includes a current reconstruction model and a quality index generator; wherein, the current reconstruction model includes an encoder and a decoder; the original current measurement signal is input into the current reconstruction model and processed by the encoder, and compressed into a current waveform eigenvector; the current waveform eigenvector is input into the decoder, decoded, and a reconstructed current measurement signal is generated; the current reconstruction model is trained offline using healthy current data.

[0016] Preferably, the specific processing steps of the quality index generator include: the quality index generator calculates the point-by-point error between the original current measurement signal and the reconstructed current measurement signal, and integrates to generate a reconstruction error energy value; the reconstruction error energy value is input into the error-exponential response curve for normalization, and the signal quality index is calculated; the signal quality index is inversely proportional to the error energy value.

[0017] Preferably, the specific operations of the data correlation analysis module include: filtering according to the signal quality index of each channel, reorganizing the retained original current measurement signals into a synchronous current measurement matrix; calculating the normalized mutual information between any two original current measurement signals in the synchronous current measurement matrix within a sliding time window to construct a correlation strength matrix; and performing Granger causality tests on signal pairs with correlation degrees higher than a threshold in the correlation strength matrix to infer a causal relationship graph.

[0018] Preferably, the specific operations of the numerical correction module include: comparing the correlation strength matrix and the causal relationship graph with a preset benchmark correlation paradigm item by item; when the normalized mutual information value and causal relationship of the original current measurement signal deviate significantly from the benchmark correlation paradigm, determining that the original current measurement signal is an outlier measurement value; when the outlier measurement value is identified, calling a pre-trained multidimensional current sequence coupling model; the multidimensional current sequence coupling model receives the original current measurement signals from multiple healthy channels as input, analyzes nonlinearity and time-series dependence, calculates and outputs the optimal estimated current waveform; replacing the outlier measurement value with the current value of the optimal estimated current waveform to complete the correction, and outputting the corrected current value.

[0019] Preferably, the multi-dimensional operating condition vector includes: real-time train speed, real-time train acceleration, gradient information of the line where the train is located, traction or braking command level output by the driver controller, high-voltage system bus voltage, and ambient temperature outside the train.

[0020] Preferably, the specific operations of the current state judgment module include: acquiring a multi-dimensional operating condition vector in real time through the train network; inputting the multi-dimensional operating condition vector into a Gaussian process regression model to output the expected value of the predicted current and the uncertainty of the predicted current; calculating and constructing a probabilistic reasonable current threshold range by combining the expected value of the predicted current and the uncertainty of the predicted current with a preset confidence level; comparing the input value of the corrected current with the probabilistic reasonable current threshold range to obtain the operating state judgment result, and outputting the operating state judgment result and the value of the corrected current.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] 1. By evaluating the quality of the original current measurement signal, this invention can identify signal distortion caused by sensor malfunction or complex electromagnetic interference online and in real time, thus solving the problem that traditional single-sensor systems cannot prove their reliability.

[0023] 2. This invention introduces mutual information analysis and causal relationship testing. By comparing the real-time multi-channel correlation pattern with the benchmark paradigm, it can identify hidden faults where the individual channel values ​​are normal but the group correlation behavior is abnormal, thereby improving the system's ability to detect connectivity faults and early performance drift.

[0024] 3. This invention employs a Gaussian process regression model to dynamically generate a probabilistically reasonable current range with confidence intervals based on the real-time, complex, and variable operating conditions of the train. This intelligent and flexible discrimination mechanism effectively captures real anomalies while reducing the false alarm rate caused by drastic changes in operating conditions, thus enhancing the robustness and availability of the detection system.

[0025] 4. This invention, through three progressive processing steps—signal quality assessment, multi-channel consistency verification, and adaptive operating condition discrimination—transforms the original, uncertain measurement signals into highly reliable measurement results that have undergone internal quality verification, external cross-verification, and scenario rationality verification, providing high-quality data support for the safe operation and intelligent maintenance of trains. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the structure of a high-voltage box current detection system for urban rail trains provided in an embodiment of the present invention;

[0027] Figure 2 A flowchart of a method for detecting the current of a high-voltage box in an urban rail train, provided in an embodiment of the present invention;

[0028] Figure 3This is an internal data flow diagram of a high-voltage box current detection system and method for urban rail trains provided in an embodiment of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Please see Figures 1 to 3 This invention provides a current detection system and method for high-voltage boxes in urban rail trains, the technical solution of which is as follows:

[0031] A current detection system and method for high-voltage boxes in urban rail trains, referenced Figure 1 The specific modules of the system proposed in this invention include:

[0032] Signal acquisition module: Real-time acquisition of the current in the high-voltage box of the urban rail train to obtain the raw current measurement signal;

[0033] Signal feature quantization module: Inputs the original current measurement signal into the current reconstruction model to generate a reconstructed current measurement signal, and calculates the signal quality index of the original current measurement signal through the quality index generator;

[0034] Data correlation analysis module: Analyzes the signal quality index and the original current measurement signal, infers and constructs a correlation strength matrix and a causal relationship diagram;

[0035] Numerical correction module: Identifies outlier data points in the correlation strength matrix and the causal relationship graph, identifies and corrects outlier measurement values, and outputs the corrected current value;

[0036] Current status judgment module: acquires the multi-dimensional operating condition vector of the urban rail train in real time, calls the Gaussian process regression model to predict the probabilistic reasonable current threshold range; compares the corrected current value with the probabilistic reasonable current threshold range, generates the operating status judgment result and outputs it together with the corrected current value.

[0037] Example 1

[0038] This embodiment provides a specific application of a high-voltage box current detection system for urban rail trains. A typical application scenario involves an urban rail train containing six high-voltage boxes, where a high-voltage box current detection system is introduced for periodic testing. (Refer to...) Figure 2 , Figure 3This embodiment follows the procedure of the high-voltage box current detection method for urban rail trains provided by the present invention and shows the data flow.

[0039] Furthermore, the current of the high-voltage box of the urban rail train is collected in real time to obtain the raw current measurement signal. The specific process is as follows:

[0040] The signal acquisition module functions through a combination of current sensing devices distributed within six high-voltage boxes and a data synchronization interface with the onboard data processing controller. Each high-voltage box contains a current sensor that linearly converts the large current in the main circuit being measured into a continuously varying analog voltage measurement signal. The onboard data processing controller, through its multi-channel, high-precision analog-to-digital converter interface, performs rigorous synchronous sampling and digitization of these six analog voltage measurement signals at a sampling frequency of 20,000 times per second. The resulting raw digital signal with a sampling rate of 20,000Hz flows through a digital low-pass filter. The cutoff frequency of this filter is set slightly below the Nyquist frequency of the target decimated sampling rate. Subsequently, the low-pass filtered signal passes through a decimator. This decimator operates at a 4:1 decimation ratio, retaining one point every four sampling points, reducing the signal sampling rate from 20,000Hz to 5,000Hz, thus obtaining a smoother and cleaner pre-processed raw current measurement signal with a sampling rate of 5,000Hz. By reducing the sampling rate from 20000Hz to 5000Hz, the computational complexity and resource consumption of the subsequent system are significantly reduced, the signal quality and signal-to-noise ratio are significantly improved, and the effectiveness and stability of subsequent analysis are ensured.

[0041] The signal acquisition process enables the current data between different high-voltage boxes to be aligned in time with precision. After digital conversion, it forms a six-channel parallel raw current measurement signal data stream that can accurately characterize the instantaneous changes in current.

[0042] Further, the original current measurement signal is input into the current reconstruction model to generate a reconstructed current measurement signal, and the quality index generator calculates the signal quality index of the original current measurement signal. The specific process is as follows:

[0043] The raw current measurement signal from each channel is sent to the signal feature quantization module. This module incorporates a current reconstruction model, which consists of an encoder and a decoder. The current reconstruction model is trained offline using a large amount of healthy current data, covering all normal operating conditions of the train, including stationary conditions, traction at different acceleration levels, braking at different power levels, and healthy high-voltage box current data under auxiliary loads in various seasons and passenger densities.

[0044] The raw current measurement signal first flows through an encoder, which consists of five one-dimensional convolutional layers and three max-pooling layers interleaved. As the signal passes through this encoder, its high-dimensional time-domain features are extracted and compressed layer by layer, reducing the complex waveform information into a low-dimensional current waveform eigenvector. This vector can be understood as a condensation of the most essential and core information of the original waveform. Subsequently, this current waveform eigenvector is input to a decoder, which consists of five one-dimensional deconvolutional layers and three upsampling layers. Through progressive decoding and dimensionality upsampling, the decoder reconstructs the current waveform eigenvector into a reconstructed current measurement signal with the same dimension as the original input.

[0045] The current reconstruction model enables the system not only to evaluate signal quality but also to extract eigenvectors that characterize the core dynamics of the current. This provides high-quality, highly condensed input for subsequent deeper diagnosis and state classification of the current, greatly improving efficiency and accuracy.

[0046] Simultaneously, the quality index generator begins operation, receiving both the original and reconstructed current measurement signals. It calculates the root mean square error (RMSE) between the original and reconstructed current measurement signals point-by-point, obtaining the reconstruction error energy value. This reconstruction error energy value is then input into a preset, monotonically decreasing error-exponential response curve for normalization. This function is an inverse S-shaped curve, smoothly and non-linearly mapping the reconstruction error energy value to a signal quality index between 0 and 1. The error-exponential response curve is a dynamically adaptive curve, its shape dynamically adjusted according to the RMS value of the original current measurement signal within the current time window. The system first calculates the RMS value of the original current measurement signal within the current time window as a quantitative indicator of the current energy level of the original current measurement signal. Then, it selects a response curve from a preset family of functions based on the RMS value. This implementation ensures that the signal quality index evaluation criteria can smoothly and automatically adapt to continuous changes in signal load, thus providing accurate and reliable signal quality assessments under all operating conditions. The signal quality index is strictly inversely proportional to the reconstruction error energy value. The higher the signal quality index, the more reliable the original signal quality.

[0047] The process of generating the signal quality index is concretized into calculating the reconstruction error and normalizing it using the response curve, providing a quantitative evaluation method that gives the signal quality of different sensors and at different times a unified and comparable benchmark.

[0048] Furthermore, the signal quality index and the original current measurement signal are analyzed to infer and construct a correlation strength matrix and a causal relationship graph. The specific process is as follows:

[0049] The data correlation analysis module receives all raw current measurement signals and their corresponding signal quality indices. It first filters the signals based on their quality indices, retaining only channels with indices higher than 0.9 for subsequent collaborative analysis to ensure the reliability of the analysis basis. The filtered signals are reassembled into a six-column synchronous current measurement matrix. Within a sliding time window of one second, the module continuously performs in-depth analysis on this matrix. It calculates the normalized mutual information between any two high-voltage box current signals, constructing a real-time correlation strength matrix. This correlation strength matrix quantifies the nonlinear coupling strength between all channels.

[0050] For strongly correlated channel pairs with mutual information values ​​higher than 0.6, the data association analysis module initiates a parallel dynamic relationship analysis process, performing Granger causality tests on these channel pairs. Based on all test results, the data association analysis module generates and updates a directed real-time causal relationship graph. A time delay calculator calculates the cross-correlation function of the strongly correlated signal pairs in parallel and identifies the time delay corresponding to the peak point to accurately quantify the signal transmission delay τ between the two. The data association analysis module then fuses and mutually corroborates the causal direction inferred from the real-time causal relationship graph with the signal transmission delay τ.

[0051] This real-time dynamic relationship description, which includes direction and delay information, will be used for a deeper comparison with the baseline correlation paradigm, thereby enabling the identification of more subtle system failures that not only exhibit abnormal correlation strength but also changes in transmission delay or causal logic.

[0052] By utilizing mutual information and Granger causality tests to construct correlations, the system can go beyond simple linear correlation analysis and uncover the complex nonlinear dependencies and "leader-follower" transmission modes among the currents in each high-voltage box. This enables the earlier and more accurate detection of system-level hidden faults, facilitating precise current detection.

[0053] Furthermore, outlier data points are identified in the correlation strength matrix and the causal relationship graph, and outlier measured values ​​are identified and corrected, with the corrected current value output. The specific process is as follows:

[0054] The numerical correction module then compares the correlation strength matrix and real-time causal relationship graph with a preset benchmark correlation paradigm stored locally for the current train operation mode. When the system runs stably under a specific operating condition, if the signal quality index of all channels remains high and no outlier data points are identified, the numerical correction module activates the paradigm learner. The paradigm learner collects the correlation strength matrix and causal relationship graph during the stable operating period and calculates real-time statistical characteristics. If these real-time statistical characteristics have a small but stable systematic deviation from the existing benchmark correlation paradigm, the learner calculates correction coefficients. After obtaining authorization, the numerical correction module uses the correction coefficients to fine-tune and update the stored benchmark correlation paradigm. Through dynamic fine-tuning of the benchmark correlation paradigm, the long-term accumulation of misjudgments caused by benchmark solidification is avoided, reflecting the reliability of the detection system. The adaptive update mechanism means that the preset benchmark paradigm does not need to be perfect in the initial stage, which reduces the stringent requirements on the initial model training data and enhances the model's generalization ability to different application scenarios.

[0055] During the comparison process, if the numerical correction module detects a significant deviation from the baseline in either the mutual information or causal relationship of a channel, it determines that the channel is an outlier measurement. In this case, the numerical correction module invokes a pre-trained multidimensional current sequence coupling model. This model takes the original current measurement signals from multiple healthy channels as input, analyzes the complex nonlinear and temporal dependencies between them, and calculates and outputs the optimal estimated current waveform for the outlier channel. The numerical correction module then uses the current value of this optimal estimated waveform to replace the original outlier measurement, thus completing the correction and outputting the final corrected current value.

[0056] By proposing an outlier correction method based on a multidimensional current sequence coupling model, the problem of how to handle bad data after it is identified is solved. By using information from the health channel to repair the data, the system can still output a high-quality data stream without interruption, even if some sensors experience momentary failures, thus enhancing the fault tolerance and data integrity of the entire detection system.

[0057] Furthermore, the multi-dimensional operating condition vector of the urban rail train is acquired in real time, and a Gaussian process regression model is called to predict a probabilistically reasonable current threshold range. The corrected current value is compared with the probabilistically reasonable current threshold range to generate an operating state discrimination result, which is output together with the corrected current value. The specific process is as follows:

[0058] The current state determination module is the final decision-making and output unit of this system. Through the train's high-speed data network, the current state determination module acquires in real time the various parameters constituting the multi-dimensional operating condition vector. These parameters include: the train's real-time speed of 65 km / h and its real-time acceleration of 1 m / s². 2The information includes the gradient of the track where the train is located (5 / 1000 uphill), the traction command level output by the driver controller (Level 3), the high-voltage system bus voltage (760 volts), and the ambient temperature outside the train (28 degrees Celsius).

[0059] Listing the specific physical quantities that constitute the multidimensional operating condition vector provides a solid physical foundation for subsequent adaptive operating condition discrimination; it enables the model to fully consider the key external factors affecting current changes, ensuring the accuracy and reliability of dynamic threshold determination.

[0060] The current status interpretation module inputs a real-time multi-dimensional operating condition vector into a Gaussian process regression model. The model predicts an expected current of 60A for each high-voltage box under the current specific operating condition, and provides the prediction uncertainty with a standard deviation of 4A. Based on this, the system constructs a probabilistic reasonable current threshold range with a 99.7% confidence level, i.e., a dynamic confidence interval from 48A to 72A. The current status interpretation module compares all values ​​in the received corrected current value vector with the probabilistic reasonable current threshold range. If all values ​​fall within this range, the module generates a "normal" operating status determination result. Finally, the current status interpretation module integrates the six-dimensional corrected current value vector and its accompanying "normal" status label into a standard data packet, which is then published to the train's main monitoring system for real-time display and long-term health record recording.

[0061] By streamlining the specific operational process of current state interpretation, alarm decision-making is no longer a simple threshold comparison. This makes the system more robust when dealing with ambiguous states, effectively suppressing false alarms caused by minor signal fluctuations and improving the accuracy of current measurement.

[0062] This invention improves the reliability, accuracy, and depth of current measurement results by constructing an innovative, progressive data processing flow from single-channel confidence assessment and multi-channel coupling analysis to adaptive interpretation of operating conditions. While enhancing the system's fault tolerance, it provides technical means for predictive maintenance and lean operation and maintenance of high-voltage systems in rail transit vehicles, and provides technical support for ensuring the safe and efficient operation of trains.

[0063] Example 2

[0064] In this embodiment, a high-voltage box current detection system and method for urban rail trains are applied to the high-voltage box current detection of urban rail train B, focusing on the quality assessment of the original current signal and the process of multi-channel data fusion analysis.

[0065] First, the system's signal acquisition module starts working. Current sensors installed inside each high-voltage box convert real-time current into corresponding analog voltage measurement signals. A multi-channel synchronous analog-to-digital converter interface synchronously samples and digitizes the analog voltage measurement signals of all channels at a preset high sampling frequency, generating multiple parallel raw current measurement signals that can characterize the continuous changes in the current waveform, and then transmits them to the subsequent processing unit.

[0066] The signal acquisition process is concretized into multi-channel synchronous analog-to-digital conversion, which aligns the original current measurement signals of all high-voltage boxes on the time base. This time synchronization is a prerequisite for subsequent multi-channel mutual information analysis, causal relationship testing and correlation collaborative diagnosis, and provides a solid data foundation to ensure the accuracy of the results.

[0067] Next, the signal feature quantization module receives these raw current measurement signals. The core of this module is an offline-trained current reconstruction model, which logically includes an encoder and a decoder. When a raw current measurement signal is input, the encoder first processes it, extracting its deep-seated timing features and compressing it into a low-dimensional current waveform eigenvector. This eigenvector is then fed into the decoder, which decodes it and attempts to reconstruct the original waveform, generating the reconstructed current measurement signal.

[0068] The current reconstruction model enables the system not only to evaluate signal quality but also to extract eigenvectors that characterize the core dynamics of the current. This provides high-quality, highly condensed input for subsequent fault diagnosis and state classification, improving the accuracy of the results.

[0069] Meanwhile, the quality index generator begins operation. It receives both the original and reconstructed current measurement signals, calculates the point-by-point error between them, and integrates it to obtain a scalarized reconstruction error energy value. This energy value is then normalized using a pre-defined, monotonically decreasing error-exponential response curve, ultimately calculating a signal quality index inversely proportional to the error energy value. This index intuitively quantifies the degree to which the original signal deviates from the healthy paradigm, i.e., its inherent reliability.

[0070] The process of generating the signal quality index is concretized into calculating the reconstruction error and normalizing it using the response curve, providing a quantitative evaluation method that gives the signal quality of different sensors and at different times a unified and comparable benchmark.

[0071] Subsequently, the data correlation analysis module began collaborative analysis of the data from all channels. It first filtered the data based on the signal quality index of each channel, retaining only those channels whose indices exceeded a preset reliability threshold, and then reassembled them into a synchronous current measurement matrix. Within a continuously scrolling sliding time window, the module performed in-depth computation on the data in the matrix: first, by calculating the normalized mutual information between any two channels, it constructed a real-time correlation strength matrix; second, for strongly correlated channel pairs, it performed Granger causality tests to infer the potential leader-follower relationship between them, ultimately forming a causal relationship graph. These two results together comprehensively characterize the complex coupling state between the high-voltage box currents of the entire train at the current moment.

[0072] By utilizing mutual information and Granger causality tests to construct correlations, the system can go beyond simple linear correlation analysis and uncover the complex nonlinear dependencies and "leader-follower" transmission modes among the currents in each high-voltage box. This enables the earlier and more accurate detection of system-level hidden faults, facilitating precise current detection.

[0073] Finally, the numerical correction module performs data optimization and correction. It compares the real-time generated correlation strength matrix and causal relationship graph with a locally stored benchmark correlation paradigm representing the train's healthy operating status. When the mutual information value or causal relationship of a certain channel deviates significantly from the benchmark paradigm, the original current measurement signal of that channel is identified as an outlier. Once an outlier is identified, the system does not simply remove it, but instead calls a pre-trained multidimensional current sequence coupling model. This model takes the original current measurement signals of several other healthy channels as input and calculates an optimal estimated current waveform for the outlier channel by analyzing the complex nonlinear and temporal dependencies between them. Finally, the module uses the current instantaneous value of this optimal estimated waveform to replace the original outlier measurement value, completes the correction, and outputs a corrected current value that is internally consistent and reliable.

[0074] Example 3

[0075] In this embodiment, a high-voltage box current detection system and method for urban rail trains are applied to the high-voltage box current detection of urban rail train C, focusing on the intelligent correction of electrical outliers and the adaptive state interpretation process based on operating conditions.

[0076] First, the system performs current verification. The numerical correction module receives the correlation strength matrix and causal relationship graph from the previous module. The numerical correction module internally stores a preset benchmark correlation paradigm, which is obtained through offline learning of a large amount of data under normal operating conditions and includes the ideal correlation strength and causal relationship between different sensor channels.

[0077] The module compares the real-time correlation strength matrix and causal relationship graph with the benchmark correlation paradigm item by item. If the normalized mutual information value and causal relationship of a certain channel's signal significantly deviate from the benchmark paradigm, the system determines that the current measurement value of that channel is an outlier. Once an outlier is identified, the system does not directly remove it, but instead calls a pre-trained multidimensional current sequence coupling model for intelligent correction. This model receives the original current measurement signal sequences of all other channels determined to be healthy as input, analyzes the nonlinearity and time-series dependencies between these healthy signals, and calculates and outputs an optimal estimated current waveform for the outlier channel in real time. Then, the system uses the current value of this optimal estimated current waveform to replace the original outlier measurement value, thus completing the data correction. Finally, the module outputs the verified corrected current value.

[0078] Outlier correction methods solve the problem of how to handle bad data after it is identified; by using information from the health channel to repair the data, the system can still output high-quality data streams without interruption even when some sensors experience momentary failures, thus enhancing the fault tolerance and data integrity of the entire detection system.

[0079] Secondly, the system performs current status analysis. The current status analysis module is responsible for the final judgment on the rationality of the corrected current value. This module obtains a multi-dimensional operating condition vector in real time through the train network. In this embodiment, the vector includes: real-time train speed, real-time acceleration, gradient information of the track where the train is located, traction or braking command level output by the driver controller, high-voltage system bus voltage, and ambient temperature outside the train.

[0080] By clearly defining the specific physical quantities that constitute the multidimensional operating condition vector, the model can fully consider the key external factors affecting current changes, thus providing a solid physical foundation for subsequent adaptive operating condition discrimination and ensuring the accuracy and reliability of dynamic threshold determination.

[0081] The multidimensional operating condition vector is input into a Gaussian process regression model. Based on the current operating conditions, the model outputs the expected value of the predicted current μ and the uncertainty of the predicted current σ in real time. These two values ​​describe the most likely value of the high-voltage box current and its reasonable fluctuation range under the current operating conditions. Subsequently, combined with a preset confidence level, the system calculates and constructs a dynamic probabilistic reasonable current threshold range, which is [μ-3σ, μ+3σ].

[0082] Finally, the current status judgment module compares the corrected current value from the previous module with this dynamic threshold range. If the value falls within the threshold range, the operating status is judged as normal; if it significantly exceeds the threshold range, the operating status is judged as abnormal. Ultimately, the system outputs this judgment result along with the corrected current value for display and recording by the train monitoring system.

[0083] The standardized operational procedures for current state interpretation make alarm decisions no longer a simple comparison; this makes the system more robust when dealing with ambiguous states, effectively suppressing false alarms caused by minor signal fluctuations and improving the accuracy of current measurement.

[0084] Through the process described in this embodiment, the system can not only intelligently repair the data generated by the faulty sensor, but also dynamically and intelligently judge the rationality of the current in combination with the complex operating conditions of the train, which greatly improves the accuracy and reliability of detection and effectively avoids false alarms and missed alarms caused by traditional fixed threshold alarm methods.

[0085] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A current detection system for a high-voltage box of an urban rail train, characterized in that, include: Signal acquisition module: used to acquire the current of the high-voltage box of the urban rail train in real time and obtain the raw current measurement signal; Signal feature quantization module: The signal feature quantization module includes a current reconstruction model and a quality index generator; the current reconstruction model includes an encoder and a decoder, and the current reconstruction model is trained offline from healthy current data; The encoder is used to process the original current measurement signal into the current reconstruction model, which is then compressed into the eigenvectors of the current waveform. The eigenvector of the current waveform is input to the decoder, decoded and used to generate a reconstructed current measurement signal. The quality index generator calculates the point-by-point error between the original current measurement signal and the reconstructed current measurement signal, and integrates it to generate a reconstruction error energy value. The reconstruction error energy value is input into the error-exponential response curve for normalization, and the signal quality index is calculated; the signal quality index is inversely proportional to the error energy value. Data correlation analysis module: Based on the signal quality index of each channel, the original current measurement signals are filtered and analyzed to reassemble them into a synchronous current measurement matrix; Within a sliding time window, the normalized mutual information between any two original current measurement signals in the synchronous current measurement matrix is ​​calculated to infer and construct the correlation strength matrix; Granger causality test is performed on signal pairs in the correlation strength matrix with a correlation degree higher than a threshold to infer the causal relationship graph. Numerical correction module: used to identify outlier data points in the correlation strength matrix and the causal relationship graph, identify and correct outlier measurement values, and output corrected current values; Current status judgment module: used to acquire the multi-dimensional operating condition vector of the urban rail train in real time, call the Gaussian process regression model to predict the probabilistic reasonable current threshold range; compare the corrected current value with the probabilistic reasonable current threshold range, generate the operating status judgment result and output it together with the corrected current value.

2. The high-voltage box current detection system for urban rail trains according to claim 1, characterized in that, The process by which the signal acquisition module acquires the original current measurement signal includes: Using a sensor installed inside the high-voltage box, the current is converted into an analog voltage measurement signal; the analog voltage measurement signals of all channels are synchronously sampled and digitized using a multi-channel synchronous analog-to-digital converter interface to generate the original current measurement signal.

3. The high-voltage box current detection system for urban rail trains according to claim 1, characterized in that, The specific operations of the numerical correction module include: The correlation strength matrix and the causal relationship graph are compared item by item with a preset benchmark correlation paradigm. When the normalized mutual information value and causal relationship of the original current measurement signal deviate significantly from the benchmark correlation paradigm, the original current measurement signal is determined to be an outlier measurement value. When the outlier measurement value is identified, a pre-trained multidimensional current sequence coupling model is invoked. The multidimensional current sequence coupling model receives the original current measurement signals from multiple healthy channels as input, analyzes nonlinearity and time-series dependence, calculates and outputs the optimal estimated current waveform. The current value of the optimal estimated current waveform is used to replace the outlier measurement value to complete the correction, and the corrected current value is output.

4. The high-voltage box current detection system for urban rail trains according to claim 1, characterized in that, The multi-dimensional operating condition vector includes: real-time train speed, real-time train acceleration, gradient information of the track where the train is located, traction or braking command level output by the driver controller, high-voltage system bus voltage, and ambient temperature outside the train.

5. The high-voltage box current detection system for urban rail trains according to claim 1, characterized in that, The specific operations of the current state determination module include: A multi-dimensional operating condition vector is acquired in real time through the train network; the multi-dimensional operating condition vector is input into a Gaussian process regression model, and the expected value of the predicted current and the uncertainty of the predicted current are output; the expected value of the predicted current and the uncertainty of the predicted current are combined with a preset confidence level to calculate and construct a probabilistic reasonable current threshold range; the value of the input corrected current value is compared with the probabilistic reasonable current threshold range to obtain the operating state discrimination result, and the operating state discrimination result and the value of the corrected current value are output.

6. A method for detecting the current in a high-voltage box of an urban rail train, characterized in that, include: Real-time acquisition of the current in the high-voltage box of the urban rail train to obtain the raw current measurement signal; The original current measurement signal is input into the current reconstruction model and processed by the encoder, and then compressed into the eigenvectors of the current waveform. The eigenvector of the current waveform is input to the decoder, decoded and used to generate a reconstructed current measurement signal. The quality index generator calculates the point-by-point error between the original current measurement signal and the reconstructed current measurement signal, and integrates it to generate a reconstruction error energy value. The reconstruction error energy value is input into the error-exponential response curve for normalization, and the signal quality index is calculated. The current reconstruction model includes an encoder and a decoder, and the current reconstruction model is trained offline using healthy current data; the signal quality index is inversely proportional to the error energy value. The original current measurement signals will be filtered according to the signal quality index of each channel, and the retained signals will be analyzed and reassembled into a synchronous current measurement matrix. Within a sliding time window, the normalized mutual information between any two original current measurement signals in the synchronous current measurement matrix is ​​calculated to infer and construct the correlation strength matrix; Granger causality test is performed on signal pairs in the correlation strength matrix with a correlation degree higher than a threshold to infer the causal relationship graph. Outlier data points are identified in the correlation strength matrix and the causal relationship graph, and outlier measurement values ​​are identified and corrected, and the corrected current value is output. The multi-dimensional operating condition vector of the urban rail train is acquired in real time, and a Gaussian process regression model is called to predict a probabilistic reasonable current threshold range. The corrected current value is compared with the probabilistic reasonable current threshold range to generate an operating state discrimination result, which is output together with the corrected current value.

Images