High-speed maglev train boost chopper fault diagnosis method based on depth time sequence modeling

Abstract

The invention provides a high-speed maglev train boost chopper fault diagnosis method based on depth time sequence modeling. The method comprises the following steps: organizing original sampling data of a multivariable time sequence of an actual operation platform of a boost chopper system of the high-speed maglev train by adopting a sliding time window mechanism, and generating a category equalization time sequence sample set; performing periodic pattern analysis on the time sequence sample to obtain a dominant periodic sequence; the method comprises the following steps: performing nonlinear transformation and residual connection processing on a time sequence sample by using a two-dimensional remolding mechanism through a TimesNet module to obtain a deep feature sequence, and predicting category probability distribution of an operation state of a boost chopper system of a high-speed maglev train through nonlinear activation, regularization, flattening operation and full-connection mapping processing. According to the invention, fault diagnosis of the boost chopper can be realized under complex operation conditions, and an efficient and reliable technical scheme is provided for state monitoring and safe operation of the high-speed maglev train boost chopper.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for maglev trains, and in particular to a fault diagnosis method for the boost chopper of a high-speed maglev train based on deep time-series modeling. Background Technology

[0002] The 600 km / h high-speed maglev train, with its ultra-high speed, excellent operational stability, and low mechanical loss characteristics, is leading the development direction of future rail transit equipment. As a core electrical component of the onboard power grid system of the high-speed maglev train, the boost chopper's operational stability directly affects the overall power transmission and system safety of the vehicle. During long-term operation, the boost chopper is susceptible to factors such as temperature cycling, current surges, and device aging, leading to faults such as open circuits, short circuits, and parameter degradation in the IGBTs (Insulated Gate Bipolar Transistors) or diodes.

[0003] Currently, existing technologies for fault diagnosis of boost choppers in high-speed maglev trains mainly rely on current and voltage sensors to collect signals, combined with threshold comparisons or simple logical criteria for judgment. However, as the switching frequency of boost chopper systems in high-speed maglev trains increases and operating conditions become increasingly complex, traditional diagnostic methods are insufficient in terms of feature extraction capabilities, noise immunity, and early fault identification sensitivity, making it difficult to meet the high reliability and adaptive diagnostic requirements of intelligent operation and maintenance. Summary of the Invention

[0004] The embodiments of the present invention provide a fault diagnosis method for the boost chopper of a high-speed maglev train based on deep time series modeling, so as to effectively diagnose faults in the boost chopper of a high-speed maglev train.

[0005] To achieve the above objectives, the present invention adopts the following technical solution.

[0006] A fault diagnosis method for a high-speed maglev train boost chopper based on deep time-series modeling includes:

[0007] The fault state and normal operation state of the boost chopper system of the high-speed maglev train in actual operation are determined. The original sampling data of the multivariate time series of the actual operation platform of the boost chopper system of the high-speed maglev train are obtained. The original sampling data are organized by the sliding time window mechanism to generate a class-balanced time series sample set.

[0008] Periodic pattern analysis is performed on the time series samples in the aforementioned category-balanced time series sample set to obtain the dominant periodic sequence;

[0009] Using the dominant periodic sequence as a structural prior, the TimeNet module performs nonlinear transformation and residual connection processing on the time series samples in the class-balanced time series sample set using a period-driven two-dimensional reshaping mechanism to obtain the deep feature sequence.

[0010] The deep feature sequence is processed by nonlinear activation, regularization, flattening and fully connected mapping to predict the category probability distribution of the operating state of the boost chopper system of the high-speed maglev train. The category with the highest probability in the category probability distribution is taken as the operating state of the boost chopper system of the high-speed maglev train.

[0011] Preferably, determining the fault state and normal operating state of the boost chopper system of the high-speed maglev train during actual operation includes:

[0012] To address common operational anomalies and component failures in the boost chopper system of high-speed maglev trains during actual operation, 13 fault modes and one normal operating state were selected. The 13 fault modes include: abnormal HS equipment status, abnormal external power supply voltage, battery current exceeding the allowable value, battery voltage exceeding the allowable value, average value of HS output current exceeding the allowable value, abnormal 440V battery status, communication failure, air conditioning control signal error, abnormal 440V battery voltage monitoring, timing controller failure, fluctuating signal K1 error, fluctuating signal K2 error, and DC motor speed diagnosis failure.

[0013] Preferably, the acquisition of the original sampling data of the multivariate time series of the actual operating platform of the high-speed maglev train booster chopper system, using a sliding time window mechanism to organize the original sampling data and generate a class-balanced time series sample set, includes:

[0014] The raw sampling data of the multivariate time series of the actual operating platform of the high-speed maglev train boost chopper system is obtained. The raw sampling data is organized by a sliding time window mechanism to obtain multiple continuous samples. The length of the sliding time window is 130 consecutive time steps. Each sample is represented in the form of a two-dimensional tensor with dimensions [130, 6], where 130 is the time step size and 6 is the sensor reading of the six channels. All samples constitute a structured class-balanced time series sample set.

[0015] Based on the aforementioned category-balanced time-series sample set, a corresponding sample set is constructed for each type of fault mode state and a normal operating state. The samples for each fault category include pure fault samples that are completely in a fault state and the state transition process from the initial stage of fault occurrence to the final stage of recovery. For the initial stage of a certain fault state, the first sample contains data from the first 129 normal operating time points and fault data from the 130th time point. Subsequently, the window gradually slides: the second sample contains 128 normal points and 2 consecutive fault points, the third sample contains 127 normal points and 3 fault points, and so on. When the window fully enters the fault duration, several intermediate samples containing only fault data are generated. In the fault end stage, the transition logic is executed in reverse: the last fault-dominant sample contains 129 fault points and 1 data point after recovery, followed by 128 fault points + 2 normal points, until finally a sample containing 1 fault point + 129 normal points is generated.

[0016] Preferably, the step of performing periodic pattern analysis on the time series samples in the class-balanced time series sample set to obtain the dominant periodic sequence includes:

[0017] Time series data were extracted from the class-balanced time series sample set using the TimesNet module. Where T represents the time step and C represents the number of variable channels, for the time series Each variable channel is transformed independently using FFT: After obtaining the complex spectrum, the amplitude is calculated: ;

[0018] The spectral amplitude A is obtained by averaging all amplitudes along the variable dimension using the TimesNet module.

[0019] ;

[0020] This represents the input two-dimensional time series matrix. Indicates to The complex spectrum obtained after performing a fast Fourier transform along the time dimension; This represents the amplitude matrix obtained by taking the modulus of the complex spectrum. This represents the arithmetic mean operation along the variable channel dimension, used to fuse the energy information of multi-channel signals; A is a one-dimensional vector of the output, representing the average energy distribution of all variable channels at each frequency;

[0021] From spectral amplitude The former From the elements, select the k frequency indices with the highest energy:

[0022] ;

[0023] in Indicates taking the first The index of the maximum value.

[0024] frequency Convert to the corresponding period length In the Discrete Fourier Transform, frequency The corresponding period is Period length The calculation formula is as follows:

[0025] ;

[0026] A set of dominant periodic sequences was obtained .

[0027] Preferably, the method of... The module uses the dominant periodic sequence as a structural prior and employs a period-driven two-dimensional reshaping mechanism to perform nonlinear transformation and residual connection processing on the time series samples in the class-balanced time series sample set to obtain a deep feature sequence, including:

[0028] The TimeNet module is stacked with multiple TimeBlock modules. The TimeNet module uses a periodically driven two-dimensional reshaping mechanism to reshape the time series data in the class-balanced time series sample set. Nonlinear transformations and representation enhancements are performed on each TimesBlock module, which is the output sequence of the previous module. As input, each Timesblock module analyzes the input sequence using FFT. spectral amplitude Select the k frequencies with the highest energy. And mapped to the dominant periodic sequence ;

[0029] ;

[0030] Using the TimesNet module to analyze the dominant periodic sequence in the time dimension Perform zero-filling:

[0031] ;

[0032] Length of the padded sequence able to be Divisible;

[0033] The TimesNet module is used to perform the above Perform two-dimensional reshaping processing;

[0034] ;

[0035] Indicates the first One-dimensional time series features output by the layer. Indicates the first In the TimesBlock module of layer 1, based on the first A significant cycle Output to the previous layer The two-dimensional feature matrix obtained after zero-filling and two-dimensional reshaping;

[0036] Let the shared convolution function be... Then the input for the i-th cycle is:

[0037] ;

[0038] in The two-dimensional structure after convolution processing. ,parameter It is fully shared across all k periodic branches;

[0039] Two-dimensional structure Expand into a one-dimensional sequence in row-major order:

[0040] ;

[0041] This indicates a flattening operation along the spatial dimension, expanding a two-dimensional matrix into a one-dimensional vector in row-major order. The length of the flattened one-dimensional sequence. is the hidden state dimension of the model, and represents the length of the feature vector at each time step;

[0042] Remove the redundant parts introduced by zero padding and restore the original length:

[0043] ;

[0044] In the above formula, This indicates a truncation operation, meaning that only the first few bits are retained. The feature vectors at each time step are discarded, with the last one being discarded. The redundant features corresponding to the time points introduced by zero padding are restored to the original sequence length; The final output is a one-dimensional feature sequence. Each time step corresponds to one 3D eigenvectors;

[0045] Obtaining one-dimensional outputs at k different periodic scales Subsequently, TimesBlock employs an adaptive weighting mechanism based on frequency domain energy for fusion processing:

[0046]

[0047] ;

[0048] in, This represents the number of significant cycles detected in the current layer. Indicates the first Significant frequencies obtained during the layer periodicity detection phase The corresponding amplitude value, where Softmax represents the normalization function, is used to convert the original amplitude value into a probability distribution; It is the weight vector obtained after Softmax normalization, which satisfies =1; Indicates the first The transformed output feature sequence obtained after multi-scale fusion in the TimesBlock module;

[0049] The final output is generated through residual connections:

[0050] ;

[0051] In the In the TimesBlock module, the transform output is obtained through multi-scale periodic analysis and two-dimensional convolution extraction. ,Will With the original input Perform element-wise addition and output the result. As input to the next TimesBlock module, it is passed layer by layer, passing through all layers. Composition of deep feature sequences Where B is the batch size, T is the length of the input sequence, and D is the feature dimension obtained after TimeNet feature transformation.

[0052] Preferably, the step of performing nonlinear activation, regularization, flattening, and fully connected mapping on the deep feature sequence to predict the category probability distribution of the operating state of the boost chopper system of the high-speed maglev train, and taking the category with the highest probability in the category probability distribution as the operating state of the boost chopper system of the high-speed maglev train, includes:

[0053] In the deep feature sequence Apply the GELU activation function:

[0054] ;

[0055] The activated features are then regularized using the Dropout mechanism.

[0056] ;

[0057] Dropout randomly sets the output of some neurons to zero with a preset probability, affecting the three-dimensional feature tensor. Flattening operations are performed along the time and feature dimensions:

[0058] ;

[0059] This operation converts the sequence structure into vector form, generating a global feature vector containing complete contextual information. This vector serves as the input to a fully connected layer. The flattened feature is then fed into a fully connected layer, mapping it to the class space. Assuming the total number of classes is K, an unnormalized feature vector is obtained through a linear transformation. :

[0060] ;

[0061] in, This represents the weight matrix of the fully connected layer, where the number of rows equals the dimension of the input feature vector. The number of columns is equal to the total number of category judgments K, and b represents the bias vector of the fully connected layer, whose length is equal to the number of categories K, and is used to adjust the baseline level of each category score; This represents the unnormalized classification score vector output by the model, where each element corresponds to the original confidence value of a class.

[0062] Using the Softmax function After normalization, the predicted probability distributions of the operating states of the boost chopper system of the high-speed maglev train for each category are obtained:

[0063] ;

[0064] The probability distribution for this category is one. dimensional vector ,in Given the preset total number of categories, each component... This indicates that the input sample belongs to the first... The probability of class, and satisfying =1;

[0065] The category with the highest probability is output as the predicted operating state of the boost chopper system of the high-speed maglev train:

[0066] ;

[0067] Represents the probability across all categories Find the category index K that maximizes the probability value; This represents the final classification result output by the model, i.e., the predicted class index of the input sample, and its value ranges from 1 to 2. .

[0068] As can be seen from the technical solutions provided by the embodiments of the present invention above, the present invention can realize fault diagnosis of boost chopper under complex operating conditions, and has good robustness and generalization ability, providing an efficient and reliable technical solution for the condition monitoring and safe operation of boost chopper of high-speed maglev train.

[0069] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0070] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0071] Figure 1 A flowchart illustrating a fault diagnosis method for a high-speed maglev train boost chopper based on deep time-series modeling, provided in an embodiment of the present invention.

[0072] Figure 2 A structural diagram of a TimesBlock module provided in an embodiment of the present invention;

[0073] Figure 3 A structural diagram of a confusion matrix provided in an embodiment of the present invention;

[0074] Figure 4 This is a bar chart of performance indicators provided in an embodiment of the present invention. Detailed Implementation

[0075] 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.

[0076] 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, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0077] 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. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0078] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0079] This invention proposes a fault diagnosis method for high-speed maglev train booster choppers based on deep time series modeling. Addressing the shortcomings of existing fault diagnosis technologies, such as insufficient ability to model the periodic dynamics of multivariate time series of high-speed maglev train booster choppers under complex operating conditions, weak multi-scale fault feature capture capabilities, and lack of predictive reliability quantification, this invention constructs an end-to-end neural network framework. By introducing a TimesNet backbone structure, it utilizes Fast Fourier Transform to adaptively detect the dominant period in the input signal of the high-speed maglev train booster chopper, reconstructing the one-dimensional time series into a two-dimensional structure. Furthermore, it combines parameter-sharing convolutional modules to synchronously capture local changes within the period and the evolution trend during the period, achieving efficient modeling of the complex dynamic characteristics of the high-speed maglev train booster chopper. Deep feature extraction is performed through multi-layer TimesBlock stacking to fully exploit the nonlinear, non-stationary, and multi-scale characteristics of the booster chopper time series during high-speed maglev train operation. Combined with a lightweight classification head, it achieves accurate fault category discrimination.

[0080] The processing flow of a fault diagnosis method for a high-speed maglev train boost chopper based on deep time-series modeling proposed in this embodiment of the invention is as follows: Figure 1 As shown, the processing steps include the following:

[0081] Step S10: Determine the fault state and normal operation state of the boost chopper system of the high-speed maglev train during actual operation. Obtain the original sampling data of the multivariate time series of the actual operating platform of the boost chopper system of the high-speed maglev train, and organize the original sampling data using a sliding time window mechanism to generate a class-balanced time series sample set.

[0082] This invention addresses common operational anomalies and component failures in the boost chopper system of high-speed maglev trains during actual operation. It selects 13 typical fault modes and, combined with normal operating conditions, constructs a multi-category fault diagnosis task covering 14 system states. The 13 typical fault types include: HS equipment status anomaly, external power supply voltage anomaly, battery current exceeding permissible values, battery voltage exceeding permissible values, average HS output current exceeding permissible values, 440V battery status anomaly, communication failure, air conditioning control signal error, 440V battery voltage monitoring anomaly, timing controller failure, fluctuating signal K1 error, fluctuating signal K2 error, and DC motor speed diagnosis failure. Each fault state is precisely defined and time-annotated based on alarm codes and parameter limit records in the actual operation log, ensuring that the fault labels have engineering interpretability and field relevance. All state categories together constitute a complete system behavior classification system, providing fine-grained label support for subsequent model training.

[0083] The original sampling data of the multivariate time series of the actual operating platform of the high-speed maglev train boost chopper system is obtained. The original sampling data comes from the real-time communication messages of the actual operating platform. After the original CAN bus data and communication messages are parsed by the protocol, the key physical quantities are extracted and structured and stored in the high-performance time series database TDengine to form a high-precision, multivariate industrial operation log. Through the SQL query interface, the present invention reads the sampling data in batches from the TDengine parsing database in chronological order. The sampling data includes a number of core electrical parameters for training and evaluation of the fault diagnosis model. The specific parameters include: (1) battery current; (2) battery voltage; (3) average value of boost chopper output current; (4) average value of linear motor current; (5) average value of DC current of external power supply module of boost chopper; (6) UDP voltage value.

[0084] The data sampling period is set to acquire one record every 500ms, corresponding to a sampling frequency of 2Hz. This effectively captures the slow-changing dynamic process of the system while avoiding redundant interference introduced by high-frequency noise. All parameters have undergone unit normalization, outlier filtering, and clock synchronization before being written to the database to ensure that the data quality meets the modeling requirements.

[0085] To support multivariate time series fault diagnosis based on deep learning, this invention employs a sliding time window mechanism to organize the raw sampled data. Combining the dynamic response characteristics and fault evolution patterns of the boost chopper system, the time window length is set to 130 consecutive time steps, meaning each sample covers 65 seconds (130 × 0.5 s) of historical operating data. This ensures a sufficient number of system state change cycles are included, which is beneficial for the model to capture the evolution trend and periodic dynamic characteristics of the fault. Each sample is represented in two-dimensional tensor form, with dimensions of... Where 130 represents the time step and 6 represents the sensor readings of the six channels mentioned above. All samples constitute a structured class-balanced time-series sample set, which serves as the baseline input for subsequent model training and validation. This set is applied to frequency domain analysis in the periodic detection module, multi-scale periodic structure transformation in the two-dimensional reshaping module, deep time-frequency feature extraction and adaptive fusion in the TimesBlock module, and the final classification task output processing and performance evaluation. This provides the model with a high-quality, class-balanced, and engineering-interpretable learning foundation, effectively improving the accuracy and generalization ability of the fault diagnosis system in identifying various states.

[0086] Based on the aforementioned categorically balanced time-series sample set, a corresponding sample set is constructed for each type of system state (including 13 types of faults and 1 type of normal state). All samples are extracted from labeled time-series segments, ensuring that their start and end times are precisely aligned with the state labels. To improve the model's sensitivity to the fault occurrence and recovery process, this invention designs a transition boundary enhanced sliding strategy, which is particularly suitable for data slices in the fault injection or exit phase.

[0087] When constructing samples for fault categories, this invention not only includes pure fault samples that are completely in a fault state, but also specifically considers the state transition process between the initial stage of fault occurrence and the final stage of recovery, in order to enhance the model's ability to identify fault boundaries. Specifically:

[0088] For the initial stage of a certain fault state, the first sample contains data from the first 129 normal operation time points and fault data from the 130th time point;

[0089] The window then slides gradually: the second sample contains 128 normal points and 2 consecutive fault points, the third sample contains 127 normal points and 3 fault points, and so on;

[0090] Once the window has fully entered the fault duration, several intermediate samples containing only fault data are generated.

[0091] During the fault termination phase, the transition logic is executed in reverse: the last fault-dominant sample contains 129 fault points and 1 data point after recovery, followed by 128 fault points + 2 normal points, until finally a sample of 1 fault point + 129 normal points is generated.

[0092] Because the actual occurrence frequencies of different fault categories vary significantly, training data imbalance can occur, affecting the model's ability to identify rare faults. To address this, this invention constructs a class-balanced small training subset. For each system state category, a maximum of 400 samples are extracted and added to this subset; if a category has fewer than 400 available samples, all samples are retained. This strategy ensures that each category has a similar number of samples in the training set, effectively mitigating class imbalance and improving the model's overall generalization performance and fairness. This small subset is then used for model training and validation set partitioning, providing a high-quality, representative learning foundation for subsequent fault diagnosis models.

[0093] Step S20: Perform periodic pattern analysis on the time series samples in the above-mentioned category-balanced time series sample set to extract a set of dominant periodic sequences. .

[0094] The input data for periodic pattern analysis is the time series data from the aforementioned class-balanced time series sample set. Where T represents the time step (series length) and C represents the number of variable channels (e.g., multi-sensor readings). This time series It is a standard two-dimensional matrix, with each row corresponding to a time point and each column corresponding to a variable.

[0095] The Fast Fourier Transform (FFT) can convert a time-domain signal into a frequency-domain signal, decomposing a sequence of length T into T complex frequency components, thus revealing the frequency composition of the signal. For the input time series... First, perform an FFT transformation independently for each variable channel: After obtaining the complex spectrum, calculate its amplitude: .

[0096] Amplitude reflects the energy intensity of each frequency component. Since different variable channels may have different noise levels, to enhance the robustness of period detection,

[0097] The TimesNet module used in this invention is a deep neural network architecture specifically designed for complex time series modeling tasks. Its core idea is to capture the dynamic evolution of a system by mining the multi-periodic characteristics in the data. This model combines the efficiency of convolutional neural networks (CNNs) in local feature extraction with the advantages of the Transformer architecture in modeling long-range dependencies, overcoming the problem of traditional time series models' insufficient ability to represent periodic structures when dealing with non-stationary, multivariate, and long-sequence data. In this embodiment, the TimesNet module first processes the input time series... Frequency domain analysis was performed to identify the dominant periodic component, which then guided the subsequent multi-scale feature modeling process. The TimesNet module was used to average all amplitudes along the variable dimensions to obtain the spectral amplitude A:

[0098] ;

[0099] This represents the input two-dimensional time series matrix. Indicates to The complex spectrum obtained after performing a fast Fourier transform along the time dimension; This represents the amplitude matrix obtained by taking the modulus of the above complex spectrum, reflecting the energy intensity of each channel at different frequencies; This represents the arithmetic mean operation along the variable channel dimension, used to fuse the energy information of multi-channel signals; A is the one-dimensional vector of the output, representing the average energy distribution of all variable channels at each frequency, serving as the basis for subsequent significant periodicity identification.

[0100] After obtaining the spectral amplitude A, it is necessary to extract the k frequencies with the highest energy from the spectral amplitude A. These frequencies correspond to the dominant cycle. The specific steps are as follows:

[0101] Frequency index selection: Due to the conjugate symmetry of the FFT results of real number sequences, and the frequency range is... (Nyquist frequency), therefore only the preceding frequency needs to be considered. Each frequency component. In practical applications, we usually only focus on the low-frequency part, because the high-frequency part often corresponds to noise.

[0102] Top-k frequency selection: Select the k highest-energy frequency indices from the first ⌊T / 2⌋ elements of the spectral amplitude A:

[0103] ;

[0104] in This represents the index of the first k largest values.

[0105] Period length calculation: frequency Convert to the corresponding period length In the Discrete Fourier Transform, the period corresponding to the frequency f is... Therefore, for Frequency, period length The calculation formula is as follows:

[0106] ;

[0107] This yields a set of dominant periodic sequences. .

[0108] Step S30: Using the TimesNet module to control the periodic sequence As a structural prior, a periodically driven two-dimensional reshaping mechanism is used to reshape temporal samples. Nonlinear transformation and residual connection processing are performed to obtain the deep feature sequence. .

[0109] The TimesNet module uses the time-series samples constructed in step S10. As the basic input signal, and simultaneously using a set of dominant periodic sequences extracted through frequency domain analysis in step S20. As a structural prior, it guides the multi-scale feature modeling process. The TimesNet module, through a periodically driven two-dimensional reshaping mechanism, combines convolutional feature extraction, multi-scale fusion, and residual connections to process the input time series data. Nonlinear transformations and representation enhancement processes are performed. Finally, a representation similar to the time-series samples is output. Equal-length one-dimensional feature sequences The sequence It integrates the local dynamics and global evolution features of the original signal at multiple significant periodic scales, and uses this as the final feature representation for classification decision-making.

[0110] Due to the period length The original sequence length T may not be divisible by the time interval, and direct reshaping could result in an incomplete last period. To address this issue, the TimesNet module modifies the dominant periodic sequence along the time dimension. Perform zero-filling:

[0111] ;

[0112] Length of the padded sequence Can be Integer division ensures that all cycles have the same length.

[0113] The reshaping operation formula is as follows:

[0114] ;

[0115] This indicates that zero-padding is performed on the input sequence along the time dimension. This indicates that the padded sequence will be reorganized into a shape according to row priority. Two-dimensional tensor, For the output two-dimensional tensor, This indicates that the reshaping operation applies to each detected dominant cycle. Execute independently to generate K two-dimensional representations at different scales, forming a multi-period feature space.

[0116] The resulting two-dimensional tensor The rows represent different periods (changes within a period), and the columns represent moments within a single period (changes within a period). This structure naturally possesses two types of local dependencies: adjacent elements in the column direction reflect short-term fluctuations within a period, while adjacent elements in the row direction reflect cross-period trend evolution. Therefore, short-range fluctuations and long-term trend evolution can be modeled simultaneously using two-dimensional convolutional kernels, effectively enhancing the model's ability to represent complex time structures. Finally, a set of k two-dimensional tensors is generated. Each two-dimensional tensor corresponds to a two-dimensional representation of time at different scales. By converting a one-dimensional time series into a two-dimensional matrix, the TimesNet module allows traditional image processing techniques (such as 2D convolution) to be directly applied to time series analysis, thereby efficiently capturing these two dependencies.

[0117] The present invention provides a structure for a TimesNet module as follows: Figure 2 As shown, a deep network structure is built by stacking multiple TimesBlock modules to achieve layer-by-layer abstraction and enhancement of time series features. Each TimesBlock module is the output sequence of the previous module. As input, and output a new value through residual connection. As input to the next module, it is passed layer by layer, thereby realizing the progressive evolution and optimization of features in deep networks.

[0118] In the Timesblock module, the mathematical form of periodicity detection is basically the same as described above:

[0119] ;

[0120] Among them, through Analyze the input sequence spectral amplitude Select the k frequencies with the highest energy. And mapped to period length .

[0121] The key difference lies in the fact that this operation is repeated within each TimesBlock module, affecting the feature representation of the current layer, rather than being performed only once at the model input layer. This means that periodicity detection is no longer a static preprocessing step, but a dynamic, learnable process. As network depth increases and features become more abstract, the model can re-evaluate and adjust the periodicity hypothesis in a higher-level semantic space.

[0122] The operational formulas for 2D reshaping and zero-filling are the same as those mentioned above:

[0123] ;

[0124] ;

[0125] Indicates the first One-dimensional time series features output by the layer. Indicates the first In the TimesBlock module of layer 1, based on the first A significant cycle Output to the previous layer The two-dimensional feature matrix obtained after zero-filling and two-dimensional reshaping.

[0126] The input sequence will be padded with zeros to its length. and reshaped into The structure is a two-dimensional structure, where rows represent the number of cycles and columns represent the length of the cycles.

[0127] Style convolution modules from two-dimensional structures Extract local spatiotemporal features.

[0128] Let the shared convolution function be... Then the input for the i-th cycle is:

[0129] ;

[0130] in The two-dimensional structure after convolution processing. .parameter The parameters are fully shared across all k period branches, which is the key to the efficiency of the TimesBlock module: no matter how many periods are detected, the total number of parameters of the convolutional layer is constant, depending only on the kernel size and the number of output channels, and does not increase with the increase of k.

[0131] Two-dimensional structure It needs to be restored to a one-dimensional form in order to perform cross-cycle fusion. This step consists of two parts:

[0132] Flattening: Unfolding a two-dimensional structure into a one-dimensional sequence in row-major order.

[0133] ;

[0134] This indicates a flattening operation along spatial dimensions (rows and columns), expanding a two-dimensional matrix into a one-dimensional vector in row-major order. The length of the flattened one-dimensional sequence. is the hidden state dimension of the model, and represents the length of the feature vector at each time step.

[0135] Truncation: Removes redundant portions introduced by zero padding, restoring the original length.

[0136] ;

[0137] In the above formula, This indicates a truncation operation, meaning that only the first few bits are retained. The feature vectors at each time step are discarded, with the last one being discarded. The redundant features corresponding to the time points introduced by zero padding are restored to the original sequence length; The final output is a one-dimensional feature sequence, with each time step corresponding to one... The feature vectors are 3D in length, with the same length as the input, and can be used for subsequent multi-scale fusion and residual connections.

[0138] Obtaining one-dimensional outputs at k different periodic scales Then, they need to be merged into a unified representation. TimesBlock adopts an adaptive weighting mechanism based on frequency domain energy:

[0139] ;

[0140] ;

[0141] In the above formula, This represents the number of significant cycles detected in the current layer. Indicates the first Significant frequencies obtained during the layer periodicity detection phase The corresponding amplitude value reflects the relative importance of the periodic component in the current input feature; Softmax represents the normalization function, used to transform the original amplitude value into a probability distribution; It is the weight vector obtained after Softmax normalization, which satisfies =1; Indicates the first The transformed output feature sequence obtained after multi-scale fusion in the TimesBlock module is the result of the nonlinear processing of the current layer and is used for subsequent residual connections.

[0142] Finally, the final output is generated through residual connection:

[0143] ;

[0144] In the In the TimesBlock module, the transform output is obtained through multi-scale periodic analysis and two-dimensional convolution extraction. and compare it with the original input. Element-wise addition is performed to implement the residual join mechanism. The output obtained by this operation is... This represents the final feature representation of this layer.

[0145] All layers Composition of deep feature sequences Where B is the batch size, T is the length of the input sequence, and D is the feature dimension obtained after TimeNet feature transformation.

[0146] Residual connections effectively alleviate the vanishing gradient problem in deep networks by directly adding the input features to the output features after nonlinear transformation, making it possible to train deeper networks.

[0147] The 2D transformation framework proposed by the TimesNet module exhibits excellent versatility and scalability. Its core Inception module can be replaced by any advanced 2D vision backbone network, such as ResNet, ConvNeXt, and SwinTransformer. This demonstrates that the framework is not only applicable to current time series modeling tasks but also provides a technical path for integrating more powerful visual representation models, enabling time series analysis to continuously benefit from cutting-edge advancements in computer vision.

[0148] After performing deep time-frequency domain modeling of the input time series using a multi-layered stacked structure of TimesNet modules, the model outputs a one-dimensional serialized representation rich in periodic patterns and local variation features. This one-dimensional serialized representation not only preserves the dynamic evolution characteristics of the original signal at different time scales but also enhances cross-period semantic associations through parameter-efficient convolution operations. However, this type of serialized output still belongs to the implicit representation of the intermediate feature space and cannot be directly used for category discrimination tasks. Therefore, it is necessary to further design a corresponding output processing mechanism to integrate the distributed temporal features into a fixed-dimensional vector with global discriminative capabilities, thereby achieving a semantic transition from "time series modeling" to "pattern classification".

[0149] Step S40: Perform nonlinear activation, regularization, flattening and fully connected mapping on the deep feature sequence to generate a class probability distribution. Finally, output the class with the highest probability in the class probability distribution as the prediction result.

[0150] Step S40 takes this deep feature sequence as input and generates an unnormalized classification through nonlinear activation, regularization, flattening, and fully connected mapping. The probability distribution is then normalized using the Softmax function and transformed into a class probability distribution. Finally, the model outputs the class with the highest probability in this distribution as the prediction result, representing the final decision output of the classification task.

[0151] After completing the feature transformation of the multi-layer TimesBlock, the model outputs a one-dimensional deep feature sequence. Where B is the batch size, T is the input sequence length, and D is the feature dimension obtained after TimesNet feature transformation. To further enhance the model's nonlinear fitting ability, this feature is applied... Activation function:

[0152] ;

[0153] This operation introduces a smooth nonlinear transformation, which helps the network learn more complex decision boundaries while maintaining the stability of gradient propagation.

[0154] To prevent the model from overfitting in scenarios with small samples or high-dimensional features, a Dropout regularization mechanism is introduced for the activated features:

[0155] ;

[0156] Dropout randomly sets the output of some neurons to zero with a preset probability, forcing the network to be independent of specific feature channels, thereby improving generalization performance.

[0157] Classification tasks require integrating variable-length temporal dynamic information into a fixed-dimensional global representation. To this end, a three-dimensional feature tensor is needed. Flattening operations are performed along the time and feature dimensions:

[0158] ;

[0159] This operation converts the sequence structure into a vector form, generating a global feature vector containing complete contextual information, which serves as the input to the fully connected layer.

[0160] The flattened features are input into a fully connected layer, which maps them to the class space. Let the total number of classes be K, then an unnormalized feature is obtained through a linear transformation. :

[0161] ;

[0162] in, This represents the weight matrix of the fully connected layer, where the number of rows equals the dimension of the input feature vector. The number of columns is equal to the total number of category judgments K, which is used to realize the linear projection from the high-dimensional temporal representation to the category space; b represents the bias vector of the fully connected layer, whose length is equal to the number of categories K, and is used to adjust the baseline level of each category score; This represents the unnormalized classification score vector output by the model, where each element corresponds to the original confidence value of a class, and serves as the input to the subsequent Softmax function.

[0163] Subsequently, the Softmax function was used to... After normalization, the predicted probability distributions for each category are obtained:

[0164] ;

[0165] The probability distribution for this category is one. dimensional vector ,in Given the preset total number of categories, each component... This indicates that the input sample belongs to the first... The probability of class, and satisfying =1.

[0166] Ultimately, the model outputs the class with the highest probability as the prediction result:

[0167] ;

[0168] Represents the probability across all categories Find the category index K that maximizes the probability value; This represents the final classification result output by the model, i.e., the predicted class index of the input sample, and its value ranges from 1 to 2. .

[0169] Step S50: Fault diagnosis result evaluation and performance analysis.

[0170] To comprehensively evaluate the classification performance of the multi-state fault diagnosis model for boost choppers proposed in this invention, a standard machine learning evaluation system was used to quantitatively analyze the model's prediction results. The evaluation results are presented through two key visualization charts: a confusion matrix to reveal the misclassification relationships between different categories; and a per-class metrics bar chart to show the specific performance of each fault category in terms of precision, recall, and F1-score. These results together constitute a comprehensive evaluation of the model's overall recognition capability and robustness.

[0171] Figure 3 A structural diagram of a confusion matrix provided in an embodiment of the present invention, such as... Figure 3 As shown, the confusion matrix illustrates the distribution of the model's prediction accuracy across 14 system states (including normal states and 13 typical fault types). The diagonal elements are significantly higher than the off-diagonal elements, indicating that the model has a very strong ability to distinguish between different states. The vast majority of real samples were correctly classified, with very few misclassifications and a low concentration of errors, indicating no systematic misclassification.

[0172] In particular, for key fault types such as "HS equipment status abnormality," "battery current over-limit," "communication failure," and "timing controller failure," the model achieved a near 100% recognition accuracy. Almost all samples of the corresponding categories fell on the main diagonal, demonstrating extremely high discriminative consistency and stability. In addition, edge faults with weaker characteristics, such as "fluctuating signal K1 error" and "DC motor speed diagnosis fault," also showed good separability, indicating that the model can effectively capture weak but regular temporal patterns.

[0173] Figure 4 This is a bar chart of performance indicators provided in an embodiment of the present invention. Figure 4 This demonstrates the performance of each fault category under three key metrics: precision, recall, and F1-score. These three metrics respectively measure the model's predictive accuracy, coverage completeness, and overall balance, and are core standards for evaluating the performance of classification tasks.

[0174] Precision: This refers to the proportion of samples predicted by the model to belong to a specific class, but which actually do not; it reflects the "reliability of the prediction." The formula for calculation is:

[0175] ;

[0176] Recall rate: The proportion of samples that truly belong to a certain class that are correctly identified, reflecting the "comprehensiveness of the detection". The formula is:

[0177] ;

[0178] Precision is the harmonic mean of precision and recall, balancing both and suitable for class imbalance scenarios. The formula is:

[0179] ;

[0180] In summary, the fault diagnosis model for the high-speed maglev train boost chopper constructed in this embodiment of the invention demonstrates extremely high classification accuracy, excellent generalization ability, and outstanding engineering practicality on real operating data. The model can achieve rapid, accurate, and stable identification of typical boost chopper faults, meeting the stringent requirements of intelligent diagnostic technology in fields such as rail transit and power electronics systems. This method is ready for large-scale application and possesses significant technological advantages and promotional value.

[0181] The beneficial effects of this invention are as follows: It introduces a periodically driven two-dimensional modeling mechanism into power electronic system fault diagnosis, overcoming the limitations of traditional one-dimensional sequence modeling in the collaborative extraction of local and global features; through frequency domain prior-guided temporal reshaping, convolutional operations can effectively capture intra-cycle fluctuations and cross-cycle trends, improving sensitivity to weak feature faults; multi-scale fusion and residual structures ensure the training stability of deep networks; the entire method requires no manual feature engineering, achieving end-to-end inference from raw data to fault judgment, significantly reducing deployment complexity. Compared to traditional diagnostic methods relying on expert rules or shallow models, this invention has significant advantages in feature representation capability, anti-interference ability, and engineering practicality. Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of an embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this invention.

[0182] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0183] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0184] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A high-speed maglev train boost chopper fault diagnosis method based on deep time series modeling, characterized by, The application relates to a method for determining the running state of a boost chopper system of a high-speed maglev train. The method comprises the following steps: determining the fault state and normal running state of the boost chopper system of the high-speed maglev train in actual running, obtaining original sampling data of a multivariate time sequence of an actual running platform of the boost chopper system of the high-speed maglev train, organizing the original sampling data by adopting a sliding time window mechanism to generate a category-balanced time sequence sample set, performing periodic pattern analysis on time sequence samples in the category-balanced time sequence sample set to obtain a dominant cycle sequence, performing nonlinear transformation and residual connection processing on the time sequence samples in the category-balanced time sequence sample set by adopting a two-dimensional remodeling mechanism driven by the TimesNet module and taking the dominant cycle sequence as a structure priori to obtain a deep feature sequence, performing nonlinear activation, regularization, flattening operation and full connection mapping processing on the deep feature sequence to predict a category probability distribution of the running state of the boost chopper system of the high-speed maglev train, and taking a category with the highest probability in the category probability distribution as the running state of the boost chopper system of the high-speed maglev train. The method for determining the fault state and normal running state of the boost chopper system of the high-speed maglev train in actual running comprises the following steps: Thirteen fault mode states and one normal running state are selected for common running abnormalities and component faults of the boost chopper system of the high-speed maglev train in actual running, the thirteen fault mode states comprising: HS device state abnormality, external power supply voltage abnormality, battery current exceeding the allowable value, battery voltage exceeding the allowable value, average value of HS output current exceeding the allowable value, 440V battery state abnormality, communication fault, air conditioner control signal error, 440V battery voltage monitoring abnormality, timing controller fault, heave signal K1 error, heave signal K2 error and DC motor speed diagnosis fault. The method for obtaining the original sampling data of the multivariate time sequence of the actual running platform of the boost chopper system of the high-speed maglev train and organizing the original sampling data by adopting the sliding time window mechanism to generate the category-balanced time sequence sample set comprises the following steps:

2. The method of claim 1, wherein, The original sampling data of the multivariate time sequence of the actual running platform of the boost chopper system of the high-speed maglev train is obtained, the original sampling data is organized by adopting the sliding time window mechanism, a plurality of continuous samples are obtained, the length of the sliding time window is 130 continuous time steps, each sample is represented in the form of a two-dimensional tensor, the dimension is [130, 6], wherein 130 is the time step length, 6 is the sensor reading of six channels, and all the samples constitute a structured category-balanced time sequence sample set. ​ 3. The method of claim 1, wherein, ​ ​ Based on the category balanced time sequence sample set, a corresponding sample set is constructed for each fault mode state and a normal operation state, and the samples of the fault category include pure fault samples in a complete fault state and state transition processes at the initial stage of fault occurrence and the final stage of recovery. For the initial stage of a fault state, the first sample includes data of the first 129 normal operation time points and fault data at the 130th time point. The window is then gradually slid: the second sample includes 128 normal points and 2 consecutive fault points, the third sample includes 127 normal points and 3 fault points, and so on. When the window completely enters the fault duration, a plurality of intermediate samples including only fault data are generated. At the end of the fault, the transition logic is performed in reverse: the last fault dominant sample includes 129 fault points and 1 data point after recovery to normal, followed by 128 fault points + 2 normal points, and finally 1 fault point + 129 normal points.

4. The method according to claim 1 or 2 or 3, characterized in that, The periodic pattern analysis on the time sequence samples in the category balanced time sequence sample set to obtain a dominant period sequence includes: extracting time series in the category balanced time series sample set through a TimesNet module where T represents a time step, C represents a variable channel number, and an FFT transform is independently performed on each variable channel of the time series After obtaining a complex spectrum, an amplitude is calculated ;​ Average all amplitude along variables dimensions through TimesNet module to get frequency spectrum amplitude A: ； represents the input two-dimensional time series matrix, represents the operation of taking the modulus of the complex spectrum, represents the complex spectrum after fast Fourier transform along the time dimension; represents the amplitude matrix after taking the modulus operation on the complex spectrum, represents the arithmetic mean operation along the variable channel dimension for fusing the energy information of the multi-channel signal; A is a one-dimensional vector output, representing the average energy distribution of all variable channels at each frequency; Select the highest energy k frequency indexes from the first ⌊T / 2⌋ elements of frequency spectrum amplitude A: ； wherein denotes the indices of the k largest values; frequency Convert to the corresponding period length In the Discrete Fourier Transform, frequency The corresponding period is Period length The calculation formula is as follows: ； obtaining a set of dominant periodic sequences .

5. The method of claim 4, wherein, The non-linear transformation and residual connection processing of the time sequence samples in the category balanced time sequence sample set through the two-dimensional remodeling mechanism driven by the period using the dominant period sequence as the structural prior through the TimesNet module to obtain a deep feature sequence, including: The TimesNet module stacks multiple TimesBlock modules, and the TimesNet module performs a periodic driving two-dimensional reshaping mechanism on the time series in the category balanced time series sample set Nonlinear transformation and feature enhancement are performed, and each TimesBlock module takes the output sequence of a previous module as input As input, each Timesblock module analyzes the input sequence by FFT The spectrum amplitude of the input sequence Select the k frequencies with the highest energy And map to the dominant period sequence ; ； Padding the dominant period sequence in time dimension by the TimesNet module Padding the dominant period sequence in time dimension by the TimesNet module ； filled sequence length can be divisible by performing two-dimensional reshaping processing on the by the TimesNet module; ； represents the output of the layer, represents the output of the layer TimesBlock module, based on the significant periods of the previous layer output after zero padding and two-dimensional reshaping Let the shared convolution function be Then for the input of the i-th cycle: ； wherein is a two-dimensional structure after convolution processing, , parameters is completely shared among all k cycle branches; The two-dimensional structure unfolded into a one-dimensional sequence in row-major order: ； represents a flattening operation along the spatial dimension, expanding a two-dimensional matrix into a one-dimensional vector in row-major order, is the length of the flattened one-dimensional sequence, is the hidden state dimension of the model, representing the length of the feature vector at each time step; Remove the redundant part introduced by zero padding and restore to the original length: ； In the above formula, represents the truncation operation, that is, the feature vectors of the first time steps are retained, and the redundant features corresponding to the last time points filled by zeros are discarded to restore the original sequence length; is a one-dimensional feature sequence output finally, each time step of which corresponds to a dimensional feature vector; The one-dimensional output under k different periodic scales is obtained After that, TimesBlock adopts an adaptive weighting mechanism based on frequency domain energy for fusion processing: ； ； wherein, the number of significant periods detected for the current layer; denotes the each significant frequency obtained in the period detection stage of the corresponding amplitude value, Softmax denotes a normalization function, which is used to change the original amplitude value into a probability distribution; is a weight vector obtained after Softmax normalization, satisfying = 1; denotes the the transformed output feature sequence obtained after multi-scale fusion in the TimesBlock module of the Generate the final output through residual connection: ； In the In the TimesBlock module, the transform output is obtained through multi-scale periodic analysis and two-dimensional convolution extraction. ,Will With the original input Perform element-wise addition and output. As input to the next TimesBlock module, it is passed layer by layer, passing through all layers. Composition of deep feature sequences Where B is the batch size, T is the length of the input sequence, and D is the feature dimension obtained after TimeNet feature transformation.

6. The method of claim 5, wherein, The non-linear activation, regularization, flattening operation and fully connected mapping processing of the deep feature sequence to predict the class probability distribution of the operation state of the boost chopper system of the high-speed maglev train, and the class with the highest probability in the class probability distribution is taken as the operation state of the boost chopper system of the high-speed maglev train, including: In the deep feature sequence The upper application GELU activation function: ； Introduce Dropout regularization mechanism to the activated features: ； where Dropout randomly sets some neuron outputs to zero at a pre-set probability, and the three-dimensional feature tensor Flattening operation in time and feature dimensions: ； The operation converts the sequence structure into a vector form, generates a global feature vector containing complete context information as the input of the full connection layer, and inputs the flattened features into a full connection layer to map them to the category space. Assuming that the total number of categories is K, an unnormalized : ； wherein, represents a weight matrix of the fully connected layer, whose row number is equal to the dimension of the input feature vector , and whose column number is equal to the total number of class judgments K, b represents a bias vector of the fully connected layer, whose length is equal to the number of classes K, and is used to adjust the baseline level of the score of each class; represents an unnormalized classification score vector output by the model, each element of which corresponds to the original confidence value of a class; The Softmax function is used to normalize obtain the prediction probability distribution of each category of the operating state of the boost chopper system of the high-speed maglev train: ； The category probability distribution is a dimensional vector where is a preset total number of categories, each component represents a probability that the input sample belongs to the th category, and satisfies =1. Output the class with the highest probability as the prediction result of the operation state of the boost chopper system of the high-speed maglev train: ； finds the class index K that maximizes the probability value; finds the class index K that maximizes the probability value; represents the final classification result output by the model, i.e., the class index predicted by the input sample, and the value range is .

Images