Fault early warning method and system for hydrogen power system of hydrogen energy high-speed train

Abstract

The invention provides a fault early warning method and system for a hydrogen power system of a hydrogen energy high-speed train, and belongs to the technical field of train fault monitoring. And a pre-trained anomaly detection model is utilized to process the obtained key component operation state data, and an anomaly detection result of the hydrogen power system of the hydrogen energy high-speed train is obtained. According to the method, the limitation of a fixed threshold value method is overcome, and misjudgment and missed judgment are effectively reduced; early warning can be carried out in advance when a potential fault is not obvious, so that the safety of train operation is greatly improved; and by utilizing distributed calculation and a parallel algorithm, the processing efficiency of large-scale real-time data is remarkably improved, and a powerful technical guarantee is provided for efficient and safe operation of the train.

Description

Technical Field

[0001] This invention relates to the field of train fault monitoring technology, specifically to a fault early warning method and system for hydrogen power systems of hydrogen-powered high-speed trains. Background Technology

[0002] In today's rail transit field, with the continuous exploration and application of clean energy, hydrogen-powered high-speed trains have gradually become a research and development hotspot. Currently, some monitoring and fault diagnosis technologies related to hydrogen-powered trains exist. For example, some existing technologies use sensors to collect certain key parameters of hydrogen-powered high-speed trains in real time, such as hydrogen pressure, temperature, and fuel cell output voltage, and then determine whether the train system is in a normal state based on preset thresholds. This method can, to some extent, perform simple monitoring and judgment of the train's operating status. However, it has obvious problems and shortcomings. On the one hand, relying solely on preset thresholds is too simplistic and mechanical, and cannot adapt to complex and changing train operating conditions. For example, the range of change of various parameters differs under different operating states such as train acceleration, deceleration, and climbing. Fixed thresholds cannot accurately reflect the true state of the system, easily leading to misjudgments or missed judgments. On the other hand, this monitoring method lacks the ability to warn of potential faults. It often only issues an alarm after the fault is already quite obvious and the parameters exceed the threshold range, failing to detect potential fault hazards in advance, thus posing a significant risk to the safe operation of the train.

[0003] While some existing fault diagnosis systems for hydrogen-powered high-speed trains employ data analysis methods, these methods are mostly based on simple statistical analysis or linear models. For complex systems like hydrogen-powered high-speed trains, the interrelationships between their internal components are extremely intricate. Simple data analysis methods struggle to accurately uncover the deep-seated correlations between parameters and potential fault modes. For example, there are complex energy conversion and interaction relationships between the fuel cell system, hydrogen storage system, and power transmission system. Existing technologies cannot comprehensively and accurately analyze these relationships, leading to difficulties in accurately locating the fault source during fault diagnosis. Furthermore, existing fault diagnosis systems lack real-time data processing capabilities, failing to quickly process and analyze large amounts of real-time monitoring data. This can result in the inability to detect faults in a timely manner during train operation, delaying optimal fault handling and impacting train operational efficiency and safety. These problems urgently require a more advanced, accurate, and real-time fault early warning method for hydrogen-powered high-speed train systems.

[0004] In summary, the shortcomings of existing fault monitoring methods for hydrogen-powered high-speed trains are mainly reflected in three aspects. First, monitoring methods relying on fixed thresholds are too simplistic and cannot adapt to complex operating conditions, leading to frequent misjudgments or missed detections under different operating states, thus affecting the normal operation of the system. Second, these methods lack the ability to warn of potential faults, often issuing alarms only after the fault is obvious or the parameters exceed the threshold, failing to prevent potential hazards in advance. Finally, existing fault diagnosis technologies are mostly based on simple statistical analysis or linear models, failing to delve into the complex correlations within the system, resulting in inaccurate fault source localization. Summary of the Invention

[0005] The purpose of this invention is to provide a fault early warning method and system for hydrogen power systems in high-speed hydrogen-powered trains, thereby solving at least one of the technical problems existing in the background art. By employing a more intelligent anomaly detection algorithm and efficient data processing technology, this invention can not only dynamically adapt to complex operating conditions but also issue early warnings before faults occur. Simultaneously, it achieves accurate analysis and real-time monitoring of the health status of key components, thus significantly improving the safety and reliability of train operation.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for early warning of faults in the hydrogen power system of a hydrogen-powered high-speed train, comprising:

[0008] Acquire operational status data of key components of the hydrogen power system in hydrogen-powered high-speed trains;

[0009] A pre-trained anomaly detection model is used to process the acquired operational status data of key components to obtain anomaly detection results for the hydrogen power system of a hydrogen-powered high-speed train. The anomaly detection model includes a feature extraction unit, an association unit, a computation unit, and an output unit. The feature extraction unit consists of an input layer, a convolutional layer, and a position encoding layer. The input layer receives time-series data from multiple sensors, the convolutional layer extracts local features, the position encoding layer adds time and position information to the data, and the output is a sequence of feature vectors. The association unit includes parallel prior association subunits and sequence association subunits. The prior association subunit generates a prior association matrix using a learnable Gaussian kernel function, where each element represents the prior association strength between two time points. The sequence association subunit calculates the sequence association matrix using a multi-head self-attention mechanism, which captures global dependencies throughout the time series. The computation unit calculates the association difference between the prior association matrix and the sequence association matrix. The output unit outputs the calculated association difference as the anomaly score for each time point.

[0010] As a further limitation of the first aspect of the present invention, training the anomaly detection model includes: collecting multi-sensor time-series data of a hydrogen-powered high-speed train under normal operating conditions; performing data augmentation; filling in missing values ​​using linear interpolation; denoising and normalizing using a denoising method combining time and frequency domains to finally obtain multi-sensor time-series data; and designing a loss function using a minimax strategy for model training, which includes correlation difference loss, i.e., maximizing the correlation difference of normal samples, so that the model can generate a large correlation difference on normal data.

[0011] As a further limitation of the first aspect of the present invention, the noise removal algorithm in the abnormal data processing process of the hydrogen power system of the hydrogen-powered high-speed train includes: given a time-domain signal, the signal is divided into frames by short-time Fourier transform, and Fourier transform is performed on each frame; the amplitude spectrum is used as a feature while the original phase is retained to extract the frequency domain feature; the frequency domain feature is denoised using a frequency domain denoising model; and the denoised complex spectrum is converted into a time-domain signal by inverse Fourier transform.

[0012] As a further limitation of the first aspect of the present invention, the frequency domain denoising model uses an LSTM network to learn a mapping from a noisy amplitude spectrum to a clean amplitude spectrum. Its training includes: organizing consecutive noisy amplitude spectrum frames into time series samples; using a sliding window to divide consecutive frames into overlapping segments; constructing an LSTM network: an input layer to accept T time steps; LSTM layers, each containing several hidden units; and fully connected layers to output the clean amplitude spectrum of the current frame. The training objective is to minimize the mean square error between the predicted clean amplitude spectrum and the true clean amplitude spectrum.

[0013] As a further limitation of the first aspect of the present invention, the frequency domain data is processed using a denoising model, including: calculating the Fourier transform of the noisy signal, extracting the amplitude spectrum and preprocessing it; organizing it into a time series to obtain the denoised amplitude spectrum; performing inverse normalization and exponential transformation on the denoised amplitude spectrum; and combining it with the phase spectrum of the original noisy signal to form the denoised complex spectrum.

[0014] As a further limitation of the first aspect of the present invention, the inverse Fourier transform includes: converting the denoised complex spectrum back to a time-domain signal via an inverse STFT:

[0015]

[0016] Where A is the normalization constant of the overlapping window, calculated as:

[0017]

[0018] Obtain the denoised time-domain signal x denoised (n).

[0019] Secondly, the present invention provides a fault early warning system for a hydrogen power system of a hydrogen-powered high-speed train, comprising:

[0020] The acquisition module is used to acquire operational status data of key components of the hydrogen power system of hydrogen-powered high-speed trains;

[0021] The processing module utilizes a pre-trained anomaly detection model to process the acquired operational status data of key components, obtaining anomaly detection results for the hydrogen power system of the hydrogen-powered high-speed train. The anomaly detection model includes a feature extraction unit, an association unit, a computation unit, and an output unit. The feature extraction unit consists of an input layer, a convolutional layer, and a position encoding layer. The input layer receives time-series data from multiple sensors, the convolutional layer extracts local features, the position encoding layer adds temporal and positional information to the data, and outputs a sequence of feature vectors. The association unit includes parallel prior association subunits and sequence association subunits. The prior association subunit generates a prior association matrix using a learnable Gaussian kernel function, where each element represents the prior association strength between two time points. The sequence association subunit calculates the sequence association matrix using a multi-head self-attention mechanism, capturing global dependencies throughout the time series. The computation unit calculates the association difference between the prior association matrix and the sequence association matrix. The output unit outputs the calculated association difference as the anomaly score for each time point.

[0022] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the fault early warning method for the hydrogen power system of a hydrogen-powered high-speed train as described in the first aspect.

[0023] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the fault early warning method for hydrogen power system of hydrogen-powered high-speed train as described in the first aspect.

[0024] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the fault early warning method for the hydrogen power system of a hydrogen-powered high-speed train as described in the first aspect.

[0025] The beneficial effects of this invention are: overcoming the limitations of the fixed threshold method and effectively reducing the occurrence of false positives and false negatives; enabling early warnings when potential faults are not yet obvious, greatly improving the safety of train operation; and significantly improving the processing efficiency of large-scale real-time data by utilizing distributed computing and parallel algorithms, providing strong technical support for the efficient and safe operation of trains.

[0026] The advantages of additional aspects of the invention will be set forth more clearly in the following description or will be learned by practice of the invention. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart illustrating the fault early warning process of the hydrogen-powered high-speed train power system according to an embodiment of the present invention.

[0029] Figure 2 This is a topology diagram of the hydrogen power system for a high-speed train according to an embodiment of the present invention.

[0030] Figure 3 This is a schematic diagram of data acquisition and transmission for a high-speed train hydrogen power system according to an embodiment of the present invention. Detailed Implementation

[0031] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0032] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0033] It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as here.

[0034] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, and / or groups thereof.

[0035] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0036] To facilitate understanding of the present invention, the present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0037] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0038] This invention provides a fault early warning method for hydrogen-powered train systems. The method first decomposes the key components of the hydrogen-powered train system to establish a physical architecture. Then, sensors are deployed at key locations to collect data from these key components. The data is collected through a data acquisition and transmission system, processed, and monitored for anomalies to ultimately provide an early warning. This fault early warning method can detect potential faults in key components of hydrogen-powered high-speed trains in advance, providing sufficient time for maintenance and repair, preventing faults from occurring, or reducing their impact. Through real-time analysis and prediction of monitoring data, early warnings can be issued before faults become apparent, improving the safety and reliability of the train.

[0039] Example 1

[0040] In this embodiment, a fault early warning system for the hydrogen power system of a high-speed hydrogen train is first provided, including: an acquisition module for acquiring operating status data of key components of the hydrogen power system of the high-speed hydrogen train; and a processing module for processing the acquired operating status data of key components using a pre-trained anomaly detection model to obtain anomaly detection results of the hydrogen power system of the high-speed hydrogen train.

[0041] In this embodiment, the above-described system is used to implement a fault early warning method for the hydrogen power system of a hydrogen-powered high-speed train, including: using an acquisition module to acquire the operating status data of key components of the hydrogen power system of the hydrogen-powered high-speed train; and using a processing module to process the acquired operating status data of key components using a pre-trained anomaly detection model to obtain anomaly detection results of the hydrogen power system of the hydrogen-powered high-speed train.

[0042] In this embodiment, the anomaly detection model includes a feature extraction unit, an association unit, a computation unit, and an output unit. The feature extraction unit consists of an input layer, a convolutional layer, and a positional encoding layer. The input layer receives time-series data from multiple sensors, the convolutional layer extracts local features, the positional encoding layer adds temporal and positional information to the data, and the output feature vector sequence is output. The association unit includes parallel prior association subunits and sequence association subunits. The prior association subunit uses a learnable Gaussian kernel function to generate a prior association matrix, the elements of which represent the prior association strength between two time points. The sequence association subunit calculates the sequence association matrix through a multi-head self-attention mechanism, which captures the global dependencies in the entire time series. The computation unit calculates the association difference between the prior association matrix and the sequence association matrix. The output unit outputs the calculated association difference as the anomaly score for each time point.

[0043] Feature Extraction Unit: This unit consists of an input layer, a convolutional layer, and a positional encoding layer. The input layer receives time-series data from multiple sensors (such as hydrogen flow rate, temperature, pressure, etc.), the convolutional layer extracts local features, and the positional encoding layer adds temporal and positional information to the data, outputting a feature vector sequence H = [h1, h2, ..., h...]. T ], where T is the time step, Let d represent the feature vector at time step t, where d is the feature dimension.

[0044] Association Unit: This unit contains two parallel sub-units, which compute prior association and sequence association respectively.

[0045] Prior correlation subunit: A prior correlation matrix is ​​generated using a learnable Gaussian kernel function. The element P of the matrix i,j The prior association strength between time point i and time point j is expressed by the following formula:

[0046]

[0047] Where, σ iis a learnable scaling parameter for time point i, controlling the width of the Gaussian kernel (i.e., the range of the neighborhood of interest). This design allows each time point to primarily focus on its neighboring time points, exhibiting localized focusing characteristics.

[0048] Sequence association subunit: The sequence association matrix is ​​calculated using a multi-head self-attention mechanism. The calculation formula is as follows:

[0049]

[0050] Where Q and K are the query matrix and key matrix obtained by linear transformation of the input feature sequence H, respectively, and d k This is the dimension of the query and key vectors. This matrix adaptively captures global dependencies throughout the time series.

[0051] Computational Unit: This unit calculates the association difference between the prior association matrix P and the sequential association matrix S. For each time point t, it calculates its prior association distribution P. t (i.e., the t-th row of matrix P) and the sequence correlation distribution S t The symmetric KL divergence (Kullback-Leibler Divergence) between the t-th rows of matrix S is used as the outlier score:

[0052] D t =KL(S t ||P t )+KL(P t ||S t )

[0053] The formula for calculating the KL divergence is:

[0054]

[0055] Association Differences D t The larger the value, the higher the degree of anomaly at time point t.

[0056] Output unit: The calculated correlation difference D t Output the outlier score at each time point.

[0057] Multi-sensor time-series data of a hydrogen-powered high-speed train under normal operating conditions are collected. This data includes, but is not limited to, hydrogen flow rate, temperature, pressure, valve status, cooling water parameters, and vibration frequency. The training set contains only normal data and no fault data. Data augmentation and preprocessing are performed on the training set before training the model to obtain a trained anomaly detection model.

[0058] In this embodiment, the steps for training the anomaly detection model are as follows:

[0059] Training set construction:

[0060] Data source: Multi-sensor time-series data collected from hydrogen-powered high-speed trains under normal operating conditions. This data includes, but is not limited to: hydrogen flow rate, temperature, pressure, valve status, cooling water parameters, vibration frequency, etc. The training set contains only normal data and does not include any fault data.

[0061] Data augmentation: To enhance the robustness of the model, the following augmentation operations are performed on the training data:

[0062] Add Gaussian noise: Add Gaussian noise with a mean of 0 and a standard deviation of 5% of the original data standard deviation to each sensor data.

[0063] Time warping: Shifting (e.g., ±5% of the time length) and scaling (e.g., 0.95 to 1.05 times) time series data by random time windows.

[0064] Step 2. Data Preprocessing:

[0065] 1. Missing value handling: Missing values ​​are filled using linear interpolation.

[0066] 2. Noise Reduction: A noise reduction method combining time and frequency domains is employed. The specific steps are as follows:

[0067] 1) For each sensor time series data X = [x1, x2, ..., x...] N Perform a short-time Fourier transform (STFT) to obtain the frequency domain representation. Where M is the number of frequency points and K is the number of time frames.

[0068] 2) Use the amplitude spectrum of X as a feature and input it into the LSTM network to train a denoising model to predict a clean amplitude spectrum.

[0069] 3) Reconstruct the time-domain signal using the predicted amplitude spectrum and the original phase spectrum via inverse STFT.

[0070] 3. Normalization: Perform min-max normalization on each sensor data point, scaling it to the [0,1] interval:

[0071]

[0072] Where, x min and x max These are the minimum and maximum values ​​of the sensor data in the training set, respectively.

[0073] Step 3. Model Training:

[0074] 1. Input: Preprocessed multi-sensor time series data, with shape (T,d), where T is the time step and d is the number of sensors (feature dimension).

[0075] 2. Loss Function: The loss function is designed using a minimax strategy and consists of two parts:

[0076] Association Difference Loss: Maximize the association difference of normal samples so that the model can generate a large association difference on normal data (because sequence association in normal data focuses on the global, while prior association focuses on the local, and the two are very different).

[0077] Regularization term: Prevents model overfitting.

[0078] The specific formula is as follows:

[0079]

[0080] Among them, D t θ represents the correlation difference at time point t, θ represents all trainable parameters of the model (including convolutional kernel weights, attention parameters, Gaussian kernel scale parameters, etc.), and λ is the regularization coefficient (e.g., set to 0.001).

[0081] 3. Optimizer: The Adam optimizer is used with an initial learning rate of 0.001 and a learning rate decay strategy is adopted (e.g., decaying to 0.9 every 10 epochs).

[0082] 4. Training Process: Divide the training set into a training set and a validation set (e.g., 80%:20%). During training, input one batch of data (batch size = 32) for each iteration, calculate the loss, and backpropagate to update the parameters. Use an early stopping strategy: stop training when the validation set loss has not decreased for 10 consecutive epochs.

[0083] Model output: After training is complete, save the model parameters to obtain the trained anomaly detection model.

[0084] Note: During training, since the training set only contains normal data, the model learns the patterns of normal data. In subsequent anomaly detection, if the input data deviates from the normal pattern, the correlation difference D will be used. t The difference will decrease (when anomalies occur, sequence association also focuses on local factors, similar to prior association, resulting in small differences), but according to the design of the Anomaly Transformer, the association differences for normal data are larger, while the association differences for anomaly data are smaller. Therefore, in practical applications, we will consider the association difference D as... t Take the negative or reciprocal, or use a dynamic threshold to determine anomalies (i.e., set a threshold, when D...). t(If the value is below this threshold, it is considered abnormal). The dynamic threshold can be determined based on the distribution of association differences in normal data on the validation set (for example, taking the 1st percentile of association differences on the validation set as the threshold).

[0085] Example 2

[0086] In this embodiment 2, a method for anomaly detection in the hydrogen power system of a hydrogen-powered high-speed train is provided. The method includes: first, acquiring operational status data of key components of the hydrogen power system of the hydrogen-powered high-speed train; then, processing the acquired operational status data of the key components using a pre-trained anomaly detection model to obtain the anomaly detection result of the hydrogen power system of the hydrogen-powered high-speed train. The anomaly detection model is the model trained using the training method provided in embodiment 1.

[0087] Specifically, in this embodiment, the acquisition of operational status data for key components of the hydrogen power system of a hydrogen-powered high-speed train is described in detail below:

[0088] The hydrogen-powered high-speed train's hydrogen propulsion system comprises multiple components, including hydrogen cylinders, pressure reducing valves, proportional valves, solenoid valves, fuel cell stacks, gas-water separators, hydrogen discharge valves, drain valves, hydrogen circulation pumps, radiators, water tanks, circulating water pumps, air filters, air compressors, intercoolers, humidifiers, and back pressure valves. Hydrogen gas originates from the cylinders, passes through pressure reducing valves, hydrogen flow sensors, proportional valves, and solenoid valves, and enters the fuel cell stack. After the reaction, the hydrogen is separated by the gas-water separator and returned to the stack via the circulation pump for further reaction. The remaining gases and water are discharged through the hydrogen discharge valve and drain valve, respectively. Air enters through the air filter, passes through the air compressor, intercooler, and humidifier, and enters the fuel cell stack to provide oxygen for the chemical reactions within the stack. The reacted gases are discharged through the humidifier and back pressure valve. The heat generated by the fuel cell stack is removed by a water circulation system containing radiators, a PWM controller, a water tank, and a circulating water pump, ensuring normal system operation. The entire system achieves hydrogen and air supply, electrochemical reactions, power output, and heat management through the coordinated operation of its components.

[0089] In this embodiment, sensors installed in the system can acquire monitoring data of key components, including hydrogen flow rate at the hydrogen cylinder outlet, temperature and pressure at the hydrogen inlet of the fuel cell stack, air flow rate of the air filter, temperature and pressure at the air inlet and outlet of the fuel cell stack, status of solenoid valves, pressure reducing valves, proportional valves, back pressure valves, hydrogen discharge valves and drain valves, temperature, pressure and flow rate of cooling water in the cooling system, hydrogen concentration in the surrounding environment of the system, as well as the acceleration, vibration angle, vibration frequency of the train, temperature and air pressure in the environment, etc.

[0090] like Figure 2The diagram shows the topology of the hydrogen power system and related systems in a hydrogen-powered high-speed train. The fuel cell system, the core of the entire system, is located slightly to the left of the center. To its left is the hydrogen storage system, responsible for supplying hydrogen; to the upper left is the cooling system, used to dissipate the heat generated by the fuel cell. The fuel cell system is also connected to the ECU (Electronic Control Unit) via electrical energy, control signals, hydrogen, and coolant. The ECU connects to the entire vehicle via CAN-to-Ethernet. Furthermore, the fuel cell system is connected to the energy storage system to its right. The energy storage system outputs AC400V and is connected to the traction inverter and charger. The traction inverter and charger are further connected to the auxiliary inverter and charger, outputting DC110V and AC380V respectively to power the traction motor, battery, and AC380V and DC110V loads.

[0091] The entire system powers the vehicle using hydrogen energy. A hydrogen storage system ensures a hydrogen supply, while the fuel cell system reacts hydrogen and oxygen to generate electricity. A cooling system ensures the fuel cell operates at a suitable temperature. A portion of the generated electricity is stored in the energy storage system, and the remainder is converted into voltage by a traction converter and charger, and an auxiliary inverter and charger, providing appropriate voltage to different vehicle components. For example, the traction motor drives the vehicle, the battery stores electrical energy for auxiliary power supply when necessary, and it also powers various loads on the vehicle. The ECU, as the control center, coordinates the operation of each component through control signals, ensuring the efficient and stable operation of the entire hydrogen power system.

[0092] In this embodiment, data preprocessing includes, but is not limited to, missing value handling, outlier detection and handling, data consistency checking, data correction, noise removal, and data standardization or normalization. A noise removal algorithm combining time and frequency domains is employed. First, Fourier Transform (STFT) is used to transform the time-domain data to the frequency domain, extracting frequency domain features. Then, the frequency domain features are input into an LSTM network to learn a frequency-domain denoising model. This model is used to process the frequency-domain data, and finally, an inverse Fourier transform is used to obtain the time-domain data.

[0093] Step 1. Short-Time Fourier Transform (STFT)

[0094] Given a time-domain signal x(n), where n represents the discrete-time index, the STFT divides the signal into frames (typically using overlapping windows) and performs a Fourier transform on each frame.

[0095]

[0096] Where: m is the frame index (integer); k is the frequency index (integer, k = 0, 1, ..., N-1); N is the number of Fourier transform points; ω(n) is the window function (such as a Hamming window) with a length of N. ωR is the frame shift (in the number of sample points), which is usually 1 / 4 or 1 / 2 of the window length.

[0097] In actual calculations, only the positive frequency portion (i.e., k = 0, 1, ..., N / 2) is taken because the spectrum of a real signal is conjugate symmetric.

[0098] The result of STFT is expressed in complex form:

[0099]

[0100] Where |X(m,k)| is the amplitude spectrum. It is the phase spectrum.

[0101] Step 2. Frequency Domain Feature Extraction

[0102] In denoising tasks, the amplitude spectrum is used as a feature while the original phase is preserved. This is because phase information is crucial for signal reconstruction, but directly denoising the phase is difficult. The denoising model only processes the amplitude spectrum and uses the original phase for reconstruction.

[0103] Feature extraction steps:

[0104] Input: The complex matrix X(m,k) obtained by STFT.

[0105] Extracting the amplitude spectrum: |X(m,k)|

[0106] Typically, a logarithmic transformation is performed on the amplitude spectrum to compress the dynamic range: log(1+|X(m,k)|), followed by normalization (such as mean-variance normalization).

[0107] Step 3. LSTM Frequency Domain Denoising Model

[0108] Use an LSTM network to learn a mapping from a noisy amplitude spectrum to a clean amplitude spectrum.

[0109] 1. Model input: Amplitude spectrum features of the noisy signal (may contain multiple frames to utilize contextual information).

[0110] 2. Model Output: Clean amplitude spectrum of the same time frame

[0111] Specific steps:

[0112] a. Data Preparation: Organize consecutive noisy amplitude spectrum frames into time series samples. Each sample represents a time step, with a feature dimension of [missing value]. (Number of frequency bins)

[0113] b. Use a sliding window to divide consecutive frames into overlapping segments, each segment containing T frames (e.g., T=5), that is, the current frame and the two frames before and after it. Thus, the input shape for each sample is...

[0114] c. Constructing an LSTM network:

[0115] Input layer: Accepts T time steps, with the number of features at each time step being...

[0116] LSTM layers: There can be multiple layers, each containing a number of hidden units (e.g., 128).

[0117] Output layer: Fully connected layer, output dimension is (i.e., the clean amplitude spectrum of the current frame).

[0118] 3. Training objective: Minimize the mean square error (MSE) between the predicted clean amplitude spectrum and the true clean amplitude spectrum.

[0119] Loss function:

[0120]

[0121] in: Y(m) is the clean amplitude spectrum output by the model for the mm-th frame. Y(m) is the true clean amplitude spectrum (target value) for the mm-th frame. M is the number of training samples.

[0122] During training, noisy-clean signal pairs are used as training data. These data can be generated by adding simulated noise (such as Gaussian white noise, impulse noise, etc.) to the clean signal.

[0123] Step 4: Process frequency domain data using a denoising model.

[0124] After the model is trained, denoising is performed on the new noisy signal:

[0125] a. Calculate the STFT of the noisy signal to obtain X. noisy (m,k).

[0126] b. Extract the amplitude spectrum |X noisy (m,k)| and perform the same preprocessing as during training (such as logarithmic transformation and normalization).

[0127] c. Organize the data into a time series (using a sliding window, with the same window size as during training), input it into the LSTM model, and obtain the denoised amplitude spectrum.

[0128] d. To Perform inverse normalization and exponential transformation (if a logarithmic transformation was used previously, then calculate...) Obtain the denoised amplitude spectrum |X denoised (m,k)|.

[0129] e. Combining the phase spectrum of the original noisy signal with the denoised complex spectrum:

[0130]

[0131] in It is the phase of the noisy signal.

[0132] Step5. Inverse STFT (ISTFT)

[0133] The denoised complex spectrum X denoised (m,k) is converted back to a time-domain signal via an inverse STFT.

[0134] ISTFT formula:

[0135]

[0136] Where A is the normalization constant of the overlapping window, calculated as:

[0137]

[0138] Finally, the denoised time-domain signal x is obtained. denoised (n).

[0139] Fault warning can be considered a type of anomaly detection, that is, judging whether there is abnormal data (or a trend of becoming abnormal data) in normal data. In this embodiment, anomaly detection is implemented using Anomaly Transformer, which can simultaneously model prior association and series association to calculate association discrepancy, and uses a Minimax Strategy to amplify the normal-abnormal discrepancy of association discrepancies. Prior association: A learnable Gaussian kernel function is used, with its center at the index of the corresponding time point. This design can take advantage of the unimodal characteristic of Gaussian distribution to make the data focus more on neighboring points. At the same time, in order to make prior association adaptable to different time series patterns, the Gaussian kernel function includes a learnable scaling parameter. Series association: It is obtained by the attention calculation in standard Transformer. The series association of a point is the attention weight distribution of the corresponding row in the attention matrix. This branch is to mine the associations in the original sequence, allowing the model to adaptively capture the most effective associations. These two associations serve to support subsequent association differences, which are a standard for anomaly measurement. Prior associations employ a normal distribution with unimodal characteristics. When an anomaly occurs, neighboring points around the current point are highly likely to be affected. Therefore, both prior and sequence associations focus on local information, resulting in a small difference between the two associations. However, in normal sequences, the attention mechanism exhibits global attention, distributing its focus across non-neighboring points, such as in periodic sequences. This contrasts sharply with the clearly unimodal prior association, leading to a significant difference between the two associations. Therefore, association differences can be used for anomaly detection.

[0140] Example 3

[0141] In this embodiment 3, a fault early warning system for the hydrogen power system of a hydrogen-powered high-speed train is provided. This system includes: an acquisition module for acquiring operational status data of key components of the hydrogen power system of the hydrogen-powered high-speed train; and a processing module for processing the acquired operational status data of the key components using a pre-trained anomaly detection model to obtain anomaly detection results for the hydrogen power system of the hydrogen-powered high-speed train.

[0142] In this embodiment, multiple sensors first monitor the operating status data of key components of the hydrogen power system in the high-speed hydrogen train. This data is then transmitted via network to the data receiving unit in the processing module of the hydrogen power system fault early warning system for data analysis. During the data analysis phase, this invention employs an anomaly detection method based on AnomalyTransformer. This method calculates correlation differences by simultaneously modeling prior correlations and sequence correlations, and combines a minimax strategy to amplify the distinguishability between normal and abnormal conditions, thereby achieving high-precision fault early warning. The entire process is based on distributed computing and parallel algorithms, ensuring the high efficiency of real-time data processing. Furthermore, through repeatedly optimized data analysis models, this invention can dynamically capture the complex correlations between the fuel cell system, hydrogen storage system, and power transmission system, achieving precise location of the fault source.

[0143] Specifically, in this embodiment, the methods for acquiring health data of the hydrogen-powered high-speed train include several key elements. First, various types of sensors are deployed at key locations on the hydrogen-powered high-speed train. Temperature sensors, voltage sensors, and current sensors are installed at various key parts of the hydrogen fuel cell stack to accurately acquire the electrical performance parameters and thermal state parameters of the fuel cell during operation. In the hydrogen storage and transportation system, high-precision pressure sensors and hydrogen leak detection sensors are installed to monitor the hydrogen pressure and detect any leaks in real time. Simultaneously, vibration sensors and torque sensors are installed in the motors, gearboxes, and other components of the power transmission system to acquire mechanical state data during power transmission. Furthermore, displacement sensors and acceleration sensors are installed in the train's running gear to detect changes in displacement and acceleration during train operation, thereby comprehensively collecting data reflecting the train's health status.

[0144] In this embodiment, the data acquisition system collects data from each sensor at a specific sampling frequency and assigns a precise timestamp to each data point to ensure the timeliness of the data. The collected data is transmitted to the train's central data processing unit via a high-speed, stable, and interference-resistant communication network. During data transmission, encryption and verification algorithms are employed to ensure data integrity and security. Furthermore, the data acquisition system is scalable, allowing for easy integration of new sensors or upgrading of existing sensors to adapt to the evolving technological needs and health data acquisition requirements of hydrogen-powered high-speed trains.

[0145] Sensor data transmission methods such as Figure 3As shown, in the hydrogen power system of high-speed trains, the flow rate, temperature, pressure, valve status, and hydrogen concentration, temperature, pressure, acceleration, vibration frequency and other data collected by sensors in the internal pipelines and components are transmitted to the microcontroller MCU via CAN bus or 485 communication. After being processed by the microcontroller MCU, the data is transmitted to the background TCP server via network cable.

[0146] Example 4

[0147] This embodiment 4 provides a non-transitory computer-readable storage medium for storing computer instructions. When these computer instructions are executed by a processor, they implement the fault early warning method for the hydrogen power system of a high-speed hydrogen-powered train as described above. The method includes:

[0148] Acquire operational status data of key components of the hydrogen power system in hydrogen-powered high-speed trains;

[0149] A pre-trained anomaly detection model is used to process the acquired operational status data of key components to obtain anomaly detection results for the hydrogen power system of a hydrogen-powered high-speed train. The anomaly detection model includes a feature extraction unit, an association unit, a computation unit, and an output unit. The feature extraction unit consists of an input layer, a convolutional layer, and a position encoding layer. The input layer receives time-series data from multiple sensors, the convolutional layer extracts local features, the position encoding layer adds time and position information to the data, and the output is a sequence of feature vectors. The association unit includes parallel prior association subunits and sequence association subunits. The prior association subunit generates a prior association matrix using a learnable Gaussian kernel function, where each element represents the prior association strength between two time points. The sequence association subunit calculates the sequence association matrix using a multi-head self-attention mechanism, which captures global dependencies throughout the time series. The computation unit calculates the association difference between the prior association matrix and the sequence association matrix. The output unit outputs the calculated association difference as the anomaly score for each time point.

[0150] Example 5

[0151] This embodiment 5 provides a computer device, including a memory and a processor, wherein the processor and the memory communicate with each other, and the memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute the fault early warning method for the hydrogen power system of a hydrogen-powered high-speed train as described above. The method includes:

[0152] Acquire operational status data of key components of the hydrogen power system in hydrogen-powered high-speed trains;

[0153] A pre-trained anomaly detection model is used to process the acquired operational status data of key components to obtain anomaly detection results for the hydrogen power system of a hydrogen-powered high-speed train. The anomaly detection model includes a feature extraction unit, an association unit, a computation unit, and an output unit. The feature extraction unit consists of an input layer, a convolutional layer, and a position encoding layer. The input layer receives time-series data from multiple sensors, the convolutional layer extracts local features, the position encoding layer adds time and position information to the data, and the output is a sequence of feature vectors. The association unit includes parallel prior association subunits and sequence association subunits. The prior association subunit generates a prior association matrix using a learnable Gaussian kernel function, where each element represents the prior association strength between two time points. The sequence association subunit calculates the sequence association matrix using a multi-head self-attention mechanism, which captures global dependencies throughout the time series. The computation unit calculates the association difference between the prior association matrix and the sequence association matrix. The output unit outputs the calculated association difference as the anomaly score for each time point.

[0154] Example 6

[0155] This embodiment 6 provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions to implement the fault early warning method for the hydrogen power system of a hydrogen-powered high-speed train as described above. The method includes:

[0156] Acquire operational status data of key components of the hydrogen power system in hydrogen-powered high-speed trains;

[0157] A pre-trained anomaly detection model is used to process the acquired operational status data of key components to obtain anomaly detection results for the hydrogen power system of a hydrogen-powered high-speed train. The anomaly detection model includes a feature extraction unit, an association unit, a computation unit, and an output unit. The feature extraction unit consists of an input layer, a convolutional layer, and a position encoding layer. The input layer receives time-series data from multiple sensors, the convolutional layer extracts local features, the position encoding layer adds time and position information to the data, and the output is a sequence of feature vectors. The association unit includes parallel prior association subunits and sequence association subunits. The prior association subunit generates a prior association matrix using a learnable Gaussian kernel function, where each element represents the prior association strength between two time points. The sequence association subunit calculates the sequence association matrix using a multi-head self-attention mechanism, which captures global dependencies throughout the time series. The computation unit calculates the association difference between the prior association matrix and the sequence association matrix. The output unit outputs the calculated association difference as the anomaly score for each time point.

[0158] In summary, the key aspects of the fault early warning method and system for hydrogen-powered high-speed trains described in this invention mainly include the following: First, an efficient sensor data acquisition and transmission system that supports parallel operation of multiple sensors, ensuring data integrity and real-time performance; second, an anomaly detection algorithm based on Anomaly Transformer, which achieves high-precision early warning of potential faults in key components by calculating correlation differences; and finally, a denoising algorithm combining time-domain and frequency-domain features, effectively improving the accuracy and robustness of data preprocessing through LSTM networks and Fourier transform models. These technical points together constitute the complete fault early warning scheme of this invention and are the core content of technical protection.

[0159] Compared with existing technologies, this invention has significant advantages. It proposes a novel method for early warning of anomalies in hydrogen-powered train systems and a method for processing sensor data from key components of hydrogen-powered train systems. By dynamically adapting to complex operating conditions through intelligent algorithms, this invention overcomes the limitations of fixed threshold methods, effectively reducing false positives and false negatives. Furthermore, by utilizing the AnomalyTransformer model, this invention can provide early warnings before potential faults become apparent, significantly improving the safety of train operation. Finally, by employing distributed computing and parallel algorithms, this invention significantly improves the processing efficiency of large-scale real-time data, providing strong technical support for the efficient and safe operation of trains.

[0160] Regarding the key technologies of this invention, some implementation methods can be replaced by other means. For example, in the anomaly detection algorithm, the Anomaly Transformer can be replaced by other deep learning models such as graph neural network (GNN) based anomaly detection algorithms to model the complex relationships between multiple components. In terms of data denoising, the LSTM network can also be replaced by a combination of convolutional neural network (CNN) and autoencoder to improve the extraction effect of frequency domain features. In addition, the data transmission method can be further improved by adopting 5G communication technology to improve transmission speed and anti-interference capability. These alternative solutions can, to some extent, expand the scope of patent protection.

[0161] During implementation, the type and number of sensors can be adjusted according to the actual operating environment of the train. For example, temperature compensation sensors can be added in low-temperature areas to cope with the effects of special climatic conditions. Furthermore, the distributed computing-based processing framework supports modular design, allowing for direct integration of new algorithm models during system upgrades, thus improving the system's scalability and future applicability. These design details contribute to a better understanding and implementation of the core technical solutions of this invention.

[0162] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0163] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0164] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0165] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0166] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A method for early warning of faults in the hydrogen power system of a hydrogen-powered high-speed train, characterized in that, include: Acquire operational status data of key components of the hydrogen power system in hydrogen-powered high-speed trains; A pre-trained anomaly detection model is used to process the acquired operational status data of key components to obtain anomaly detection results for the hydrogen power system of a hydrogen-powered high-speed train. The anomaly detection model includes a feature extraction unit, an association unit, a computation unit, and an output unit. The feature extraction unit consists of an input layer, a convolutional layer, and a position encoding layer. The input layer receives time-series data from multiple sensors, the convolutional layer extracts local features, the position encoding layer adds time and position information to the data, and the output is a sequence of feature vectors. The association unit includes parallel prior association subunits and sequence association subunits. The prior association subunit uses a learnable Gaussian kernel function to generate a prior association matrix, whose elements represent the prior association strength between two time points. The sequence association subunit calculates the sequence association matrix through a multi-head self-attention mechanism, which captures the global dependencies throughout the time series. The calculation unit calculates the correlation difference between the prior correlation matrix and the sequential correlation matrix; the output unit outputs the calculated correlation difference as the anomaly score for each time point.

2. The fault early warning method for hydrogen power system of hydrogen-powered high-speed train according to claim 1, characterized in that, Training the anomaly detection model includes: collecting multi-sensor time-series data of a hydrogen-powered high-speed train under normal operating conditions; performing data augmentation; filling in missing values ​​using linear interpolation; denoising and normalizing using a combination of time and frequency domain denoising methods to obtain the final multi-sensor time-series data; and designing a loss function using a minimax strategy for model training, which includes correlation difference loss, i.e., maximizing the correlation difference of normal samples, so that the model can generate a large correlation difference on normal data.

3. The fault early warning method for hydrogen power system of hydrogen-powered high-speed train according to claim 2, characterized in that, The noise removal algorithm in the abnormal data processing of the hydrogen power system of the hydrogen-powered high-speed train includes: given a time-domain signal, the signal is divided into frames by short-time Fourier transform, and Fourier transform is performed on each frame; the amplitude spectrum is used as a feature while the original phase is preserved to extract the frequency domain feature; the frequency domain feature is denoised using a frequency domain denoising model; and the denoised complex spectrum is converted into a time-domain signal by inverse Fourier transform.

4. The fault early warning method for hydrogen power system of hydrogen-powered high-speed train according to claim 3, characterized in that, The frequency domain denoising model uses an LSTM network to learn a mapping from noisy amplitude spectrum to clean amplitude spectrum. Its training includes: organizing consecutive noisy amplitude spectrum frames into time series samples; using a sliding window to divide consecutive frames into overlapping segments; constructing an LSTM network: the input layer accepts T time steps; LSTM layers, each containing several hidden units; and fully connected layers that output the clean amplitude spectrum of the current frame. The training objective is to minimize the mean squared error between the predicted clean amplitude spectrum and the true clean amplitude spectrum.

5. The fault early warning method for hydrogen power system of hydrogen-powered high-speed train according to claim 4, characterized in that, The process of processing frequency domain data using a denoising model includes: calculating the Fourier transform of the noisy signal, extracting the amplitude spectrum and performing preprocessing; organizing it into a time series to obtain the denoised amplitude spectrum; performing inverse normalization and exponential transform on the denoised amplitude spectrum; and combining it with the phase spectrum of the original noisy signal to construct the denoised complex spectrum.

6. The fault early warning method for hydrogen power system of hydrogen-powered high-speed train according to claim 5, characterized in that, The inverse Fourier transform includes converting the denoised complex spectrum back to the time domain signal via an inverse STFT: Where A is the normalization constant of the overlapping window, calculated as: Obtain the denoised time-domain signal x denoised (n).

7. A fault early warning system for a hydrogen-powered high-speed train's hydrogen propulsion system, characterized in that, include: The acquisition module is used to acquire operational status data of key components of the hydrogen power system of hydrogen-powered high-speed trains; The processing module is used to process the acquired operational status data of key components using a pre-trained anomaly detection model to obtain anomaly detection results for the hydrogen power system of the hydrogen-powered high-speed train. The anomaly detection model includes a feature extraction unit, an association unit, a calculation unit, and an output unit. The feature extraction unit consists of an input layer, a convolutional layer, and a position encoding layer. The input layer receives time-series data from multiple sensors, the convolutional layer extracts local features, the position encoding layer adds time and position information to the data, and the output is a sequence of feature vectors. The association unit includes parallel prior association subunits and sequence association subunits. The prior association subunit uses a learnable Gaussian kernel function to generate a prior association matrix, the elements of which represent the prior association strength between two time points. The sequence association subunit calculates the sequence association matrix through a multi-head self-attention mechanism, which captures the global dependencies throughout the time series. The calculation unit calculates the correlation difference between the prior correlation matrix and the sequential correlation matrix; the output unit outputs the calculated correlation difference as the anomaly score for each time point.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the fault early warning method for hydrogen power system of hydrogen-powered high-speed train as described in any one of claims 1-6.

9. A computer device, characterized in that, The system includes a memory and a processor, which communicate with each other. The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the fault early warning method for hydrogen power system of hydrogen-powered high-speed train as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions to implement the fault early warning method for the hydrogen power system of a hydrogen-powered high-speed train as described in any one of claims 1-6.

Images