Traction system fault diagnosis method based on shape primitive and time-frequency mixed prompt

Abstract

The invention discloses a traction system fault diagnosis method based on shape primitives and time-frequency mixed prompt. The method comprises the following steps: firstly, acquiring a monitoring signal sequence of a system; a time-frequency mixed prompt set is constructed according to the monitoring signal sequence; then determining a gating weight according to a preset prompt set and the monitoring signal sequence; then, based on the monitoring signal sequence, the time-frequency mixed prompt set and the gating weight, obtaining a shape primitive; performing reconstruction processing on the signal sequence according to the shape elements to obtain a reconstructed signal sequence; and obtaining a fault diagnosis result according to a preset fault threshold value and the reconstruction signal sequence. According to the method, the robustness of cross-domain fault detection is effectively improved and the abnormal time period and the shape type are accurately positioned by constructing the time-frequency mixed prompt, extracting the shape primitive and combining the synergistic effect of the gating weight, meanwhile, the efficient identification of a single fault and a composite fault is realized, and the false alarm rate and the missing report rate under the complex working condition are effectively reduced.

Description

Technical Field

[0001] The embodiments of this application relate to, but are not limited to, the field of sequence signal modeling and fault detection technology, and in particular to a traction system fault diagnosis method based on shape primitives and time-frequency hybrid prompts. Background Technology

[0002] High-speed train traction systems operate under complex conditions such as acceleration, deceleration, braking, and load changes for extended periods. Internal monitoring signals for voltage, current, and speed are susceptible to environmental disturbances during operational condition transitions and equipment aging. Fault characteristics often manifest as weak amplitude, short duration, and are easily masked by noise. Traditional fault detection methods based on empirical rules or fixed thresholds are ill-suited to multi-condition, high-noise scenarios, leading to false alarms and missed alarms.

[0003] With the development of deep learning, methods for anomaly detection using models such as autoencoders to reconstruct errors perform well in single data domains. However, in practice, traction systems face changes in data distribution across lines and vehicles, and the detection performance of single-domain models deteriorates significantly after migrating to new domains. Existing cross-domain studies, while introducing domain information to alleviate overfitting, mainly focus on overall distribution alignment and residual magnitude determination. They have limited ability to characterize the fine-grained shape and structure of fault signals and cross-frequency energy distribution, and it is difficult to determine the specific time period and related shape type of the anomaly. Furthermore, in complex fault scenarios with multiple superimposed faults, there is a lack of targeted identification methods, making it impossible to effectively distinguish between single faults and complex fault modes. Summary of the Invention

[0004] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0005] This application provides a traction system fault diagnosis method based on shape primitives and time-frequency hybrid prompts, which can effectively improve the robustness of cross-domain fault detection, accurately locate abnormal time periods and shape types, and achieve efficient identification of single faults and compound faults, effectively reducing the false alarm rate and false alarm rate under complex working conditions.

[0006] This application provides a traction system fault diagnosis method based on shape primitives and time-frequency hybrid prompts, comprising: acquiring a monitoring signal sequence of the traction system; constructing a time-frequency hybrid prompt set based on the monitoring signal sequence; determining a gating weight based on a preset prompt set and the monitoring signal sequence; obtaining shape primitives based on the monitoring signal sequence, the time-frequency hybrid prompt set, and the gating weights; reconstructing the monitoring signal sequence based on the shape primitives to obtain a reconstructed signal sequence; and obtaining a fault diagnosis result based on a preset fault threshold and the reconstructed signal sequence.

[0007] In one embodiment of this application, obtaining a shape primitive based on the monitoring signal sequence, the time-frequency hybrid cue set, and the gating weights includes: calculating an attribute marker based on the sequence features of the monitoring signal sequence, combined with the time-frequency hybrid cue set and the gating weights; performing vector quantization on the attribute marker to obtain a discrete code; and performing shape decoding processing on the discrete code to obtain the shape primitive.

[0008] In one embodiment of this application, the step of reconstructing the monitoring signal sequence based on the shape primitive to obtain a reconstructed signal sequence includes: performing attribute encoding based on the shape parameters of the shape primitive and the corresponding discrete code to obtain attribute tags; and performing reconstruction processing on the monitoring signal sequence based on the attribute tags to obtain a reconstructed signal sequence.

[0009] In one embodiment of this application, the step of acquiring the monitoring signal sequence of the traction system includes: collecting a monitoring signal dataset; cleaning, denoising, and standardizing the monitoring signal dataset to obtain a standardized signal set; and dividing the standardized signal set into sample sequences of equal length according to a preset time window to obtain the monitoring signal sequence.

[0010] In one embodiment of this application, constructing a time-frequency hybrid prompt set based on the monitoring signal sequence includes: performing Fourier transform and standardization on the monitoring signal sequence to obtain a spectrum matrix; performing amplitude standardization on the spectrum matrix to obtain spectrum data; performing frequency band division and similarity pooling on the spectrum data to obtain frequency domain prompts; obtaining time domain prompts based on the time characteristics of the monitoring signal sequence; and performing weighted processing on the frequency domain prompts and the time domain prompts according to preset time-frequency weights to obtain a time-frequency hybrid prompt set.

[0011] In one embodiment of this application, obtaining a fault diagnosis result based on a preset fault threshold and the reconstructed signal sequence includes: calculating a signal sequence residual value based on the monitoring signal sequence and the reconstructed signal sequence; and obtaining a fault diagnosis result based on the signal sequence residual value and the preset fault threshold.

[0012] In one embodiment of this application, the step of calculating the signal sequence residual value based on the monitoring signal sequence and the reconstructed signal sequence includes: obtaining a residual sequence based on the monitoring signal sequence and the reconstructed signal sequence; calculating the mean of the signal sequence residuals based on the residual sequence; calculating the standard deviation of the signal sequence residuals based on the mean of the signal sequence residuals; and calculating the signal sequence residual value based on the mean of the signal sequence residuals and the standard deviation of the signal sequence residuals.

[0013] In one embodiment of this application, obtaining a fault diagnosis result based on the signal sequence residual value and a preset fault threshold includes: obtaining a fault diagnosis result indicating that the monitored signal sequence has an abnormal fault when the number of times the signal sequence residual value is continuously greater than the preset fault threshold exceeds a first threshold; and obtaining a fault diagnosis result indicating that the monitored signal sequence does not have an abnormal fault when the number of times the signal sequence residual value is continuously less than or equal to the preset fault threshold exceeds a second threshold.

[0014] In one embodiment of this application, after obtaining the fault diagnosis result, the method further includes: determining the fault frequency band in the monitoring signal sequence based on the fault diagnosis result; and performing fault identification on the fault frequency band based on the time-frequency hybrid cue set and the shape primitive to obtain the fault type.

[0015] In one embodiment of this application, the step of identifying the fault type of the fault frequency band based on the time-frequency hybrid cue set and the shape primitive includes: identifying the category of the fault mode of the fault frequency band based on the time-frequency hybrid cue set to obtain a target fault mode; determining the fault cue corresponding to the target fault mode; splicing and fusing the fault cue to obtain a composite fault cue; obtaining a fault category probability distribution based on the composite fault cue; and interpreting the fault category probability distribution based on the shape primitive to obtain the fault type.

[0016] This application provides a traction system fault diagnosis method based on shape primitives and time-frequency hybrid prompts. First, the monitoring signal sequence of the traction system is acquired. Then, a time-frequency hybrid prompt set is constructed based on the monitoring signal sequence. Next, gating weights are determined based on the preset prompt set and the monitoring signal sequence. Subsequently, shape primitives are obtained based on the monitoring signal sequence, the time-frequency hybrid prompt set, and the gating weights. Then, the signal sequence is reconstructed based on the shape primitives to obtain a reconstructed signal sequence. Finally, the fault diagnosis result is obtained based on a preset fault threshold and the reconstructed signal sequence. This application effectively improves the robustness of cross-domain fault detection by constructing time-frequency hybrid prompts, solving the problem of decreased performance in single-domain model transfer. With the extraction of shape primitives, the time period and related shape type of the anomaly can be accurately located, compensating for the shortcomings of existing methods in characterizing fine-grained features. Simultaneously, through the synergistic effect of the prompt set, gating weights, and shape primitives, effective identification of single and compound faults is achieved, effectively reducing the false alarm rate and false negative rate in complex scenarios. Attached Figure Description

[0017] Figure 1 This is a flowchart of the traction system fault diagnosis method provided in the embodiments of this application; Figure 2 This is provided by the embodiments of this application. Figure 1 The detailed flowchart of step 110; Figure 3 This is provided by the embodiments of this application. Figure 1 The detailed flowchart of step 120; Figure 4 This is provided by the embodiments of this application. Figure 1 The detailed flowchart of step 140; Figure 5 This is a system architecture diagram of a traction system fault diagnosis method provided in one embodiment of this application; Figure 6 This is provided by the embodiments of this application. Figure 1 The detailed flowchart of step 160; Figure 7 This is provided by the embodiments of this application. Figure 6 The detailed flowchart of step 610; Figure 8 This is a flowchart of the fault identification process provided in the embodiments of this application; Figure 9 This is a technical framework diagram of traction system fault diagnosis provided in the embodiments of this application; Figure 10(a) is a schematic diagram of the fault diagnosis results for a single fault scenario provided in the embodiments of this application; Figure 10(b) is a schematic diagram of the fault diagnosis results of the composite fault scenario provided in the embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] It should be noted that although the flowchart shows a logical order, in some cases, the steps shown or described may be performed in a different order than that shown in the flowchart. The terms "first," "second," etc., used in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the structures, proportions, sizes, etc., depicted in the drawings are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the implementation conditions of this application. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, without affecting the effects and purposes achieved by this application, should still fall within the scope of the technical content disclosed in this application. Similarly, the terms such as "upper," "lower," "left," "right," "middle," and "one" used in this specification are only for clarity of description and are not used to limit the scope of implementation of this application. Changes or adjustments in their relative relationships, without substantially altering the technical content, should also be considered within the scope of implementation of this application.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] High-speed train traction systems operate under complex conditions such as acceleration, deceleration, braking, and load changes for extended periods. Internal monitoring signals for voltage, current, and speed are susceptible to environmental disturbances during operational condition transitions and equipment aging. Fault characteristics often manifest as weak amplitude, short duration, and are easily masked by noise. Traditional fault detection methods based on empirical rules or fixed thresholds are ill-suited to multi-condition, high-noise scenarios, leading to false alarms and missed alarms.

[0022] With the development of deep learning, methods for anomaly detection using models such as autoencoders to reconstruct errors perform well in single data domains. However, in practice, traction systems face changes in data distribution across lines and vehicles, and the detection performance of single-domain models deteriorates significantly after migrating to new domains. Existing cross-domain studies, while introducing domain information to alleviate overfitting, mainly focus on overall distribution alignment and residual magnitude determination. They have limited ability to characterize the fine-grained shape and structure of fault signals and cross-frequency energy distribution, and it is difficult to determine the specific time period and related shape type of the anomaly. Furthermore, in complex fault scenarios with multiple superimposed faults, there is a lack of targeted identification methods, making it impossible to effectively distinguish between single faults and complex fault modes.

[0023] In view of this, this application provides a traction system fault diagnosis method based on shape primitives and time-frequency hybrid prompts. First, the monitoring signal sequence of the traction system is acquired; then, a time-frequency hybrid prompt set is constructed based on the monitoring signal sequence; next, gating weights are determined based on the preset prompt set and the monitoring signal sequence. Subsequently, shape primitives are obtained based on the monitoring signal sequence, the time-frequency hybrid prompt set, and the gating weights; then, the signal sequence is reconstructed based on the shape primitives to obtain a reconstructed signal sequence; finally, the fault diagnosis result is obtained based on a preset fault threshold and the reconstructed signal sequence. This application effectively improves the robustness of cross-domain fault detection by constructing time-frequency hybrid prompts, solving the problem of decreased performance in single-domain model transfer; with the extraction of shape primitives, the time period and related shape type of the anomaly can be accurately located, compensating for the shortcomings of existing methods in characterizing fine-grained features; simultaneously, through the synergistic effect of the prompt set, gating weights, and shape primitives, effective identification of single and compound faults is achieved, effectively reducing the false alarm rate and false negative rate in complex scenarios, making the fault diagnosis results more accurate and reliable.

[0024] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0025] Reference Figure 1 , Figure 1 This is a flowchart of a traction system fault diagnosis method provided in an embodiment of this application. The process may specifically include, but is not limited to, steps 110 to 160.

[0026] Step 110: Obtain the monitoring signal sequence of the traction system; Step 120: Construct a time-frequency hybrid cue set based on the monitoring signal sequence; Step 130: Determine the gating weights based on the preset prompt set and monitoring signal sequence; Step 140: Based on the monitoring signal sequence, the time-frequency hybrid cue set, and the gating weights, obtain the shape primitives; Step 150: Reconstruct the monitoring signal sequence based on the shape primitives to obtain the reconstructed signal sequence; Step 160: Obtain the fault diagnosis result based on the preset fault threshold and the reconstructed signal sequence.

[0027] Steps 110 to 160 will be described in detail below.

[0028] In a feasible embodiment, in step 110, the monitoring signal sequence refers to an ordered dataset composed of various physical quantity parameters that are continuously collected at preset time intervals during the operation of the traction system, reflecting the working status of the system. It mainly includes the continuous value sequence of monitoring parameters such as voltage, current, and speed that are directly related to the power output and equipment operation of the traction system over time.

[0029] In a feasible embodiment, in step 120, for the acquired monitoring signal sequence, features can be extracted from both the time domain and frequency domain to construct corresponding prompts. Alternatively, the two types of prompts can be weighted and fused to form a time-frequency hybrid prompt set that simultaneously possesses the signal's time-domain evolution law and frequency-domain energy distribution characteristics.

[0030] In a feasible embodiment, in step 130, the gating weights used to adjust the weight ratios of different components in the time-frequency hybrid cue set can be calculated based on a pre-set basic cue set and the characteristic distribution of the currently input monitoring signal sequence.

[0031] In a feasible embodiment, in step 140, based on the monitoring signal sequence, the constructed time-frequency hybrid cue set, and the calculated gating weights, the complex monitoring signal can be decomposed into several shape primitives with clear physical meaning (i.e., interpretable shape primitives) through encoding, decoding, and vector quantization. At the same time, the shape parameters such as amplitude center, scale, starting position, and duration corresponding to each shape primitive can also be obtained.

[0032] In a feasible embodiment, in step 150, the extracted shape primitives can be used as basic units, and the original monitoring signal sequence can be reconstructed according to the parameters corresponding to each shape primitive to generate a reconstructed signal sequence corresponding to the original signal sequence.

[0033] In a feasible embodiment, in step 160, the residual between the original monitoring signal sequence and the reconstructed signal sequence can be calculated, the residual statistic can be compared with a preset fault threshold, and the current operating status of the traction system can be determined based on the comparison result, thereby obtaining the fault diagnosis result.

[0034] Overall, in steps 110 to 160, for the collected traction system monitoring signal sequence, features can be extracted from both the time and frequency domains to construct corresponding prompts. Then, the two types of prompts are weighted and fused to form a time-frequency hybrid prompt set. This approach considers both the time-domain evolution of the signal and the energy distribution characteristics in the frequency domain, improving feature adaptability in cross-domain scenarios. Next, based on a pre-set basic prompt set and the feature distribution of the current monitoring signal sequence, a gating weight is determined to adjust the weight ratio of different components in the time-frequency hybrid prompt set, thereby enhancing the specificity of the features. Subsequently, based on the original monitoring signal sequence, the time-frequency hybrid prompt set, and the gating weight, several [specific features] can be obtained through encoding / decoding and vector quantization processing. By clearly defining the physical meaning of shape primitives and their corresponding key parameters such as amplitude center, scale, starting position, and duration, a precise characterization of fine-grained features of fault signals can be achieved. Based on this, the extracted shape primitives can be used as basic units to reconstruct the original monitoring signal sequence according to the parameters corresponding to each shape primitive, generating a corresponding reconstructed signal sequence. Finally, the residual between the original monitoring signal sequence and the reconstructed signal sequence can be calculated. The residual statistics are compared with a pre-set fault threshold, and the comparison results determine whether the operating status of the traction system is abnormal, thus obtaining the final fault diagnosis result. This effectively reduces the false alarm rate and missed alarm rate under complex operating conditions, while also clarifying the specific time period and shape type of the anomaly.

[0035] In one feasible embodiment, such as Figure 2 As shown, the process of obtaining the monitoring signal sequence of the traction system in step 110 may include, but is not limited to, steps 210 to 230.

[0036] Step 210: Collect monitoring signal dataset; Step 220: Clean, denoise, and standardize the monitoring signal dataset to obtain a standardized signal set; Step 230: According to the preset time window, the standardized signal set is divided into sample sequences of equal length to obtain the monitoring signal sequence.

[0037] In one feasible embodiment, the monitoring signal dataset includes a source domain dataset. and target domain dataset Among them, the source domain dataset It consists of multiple sets of raw monitoring signals continuously collected from the traction system operating in a simulated scenario. Each set of signals is accompanied by a corresponding fault label. The label information can accurately reflect the key states of the traction system, such as the type and severity of the fault, that occur in the simulated scenario. Target domain dataset These are multiple sets of raw monitoring signals continuously collected from the traction system operating under real-world conditions. These signals directly reflect the actual operating status of the traction system in real-world application scenarios, but without additional fault labels.

[0038] In a feasible embodiment, in step 220, the cleaning process can remove invalid data, duplicate data, and outliers caused by equipment malfunctions, transmission interference, etc. during the data acquisition process; the noise reduction process can use filtering algorithms to filter out irrelevant noise such as environmental electromagnetic interference and equipment vibration interference mixed in the signal, and retain the effective features in the signal related to the operating status of the traction system; the standardization process can eliminate the scale differences between data from different sources by unifying the dimensions and numerical distribution range of the data, and finally obtain the source domain standardized signal set and the target domain standardized signal set respectively. The two are merged to form the standardized signal set required for subsequent processing, providing a consistent data foundation for feature extraction and modeling in cross-domain scenarios.

[0039] In a feasible embodiment, in step 230, a fixed preset time window length and window sliding step size can be set according to the duration of the traction system fault characteristics. According to this preset time window, all signal data in the standardized signal set are segmented into several sample sequences of equal length, completing the signal block processing. Each sample sequence contains signal information within a continuous time period, completely preserving the operating status characteristics of the traction system within that time period. These sample sequences together constitute the monitoring signal sequence.

[0040] In one feasible embodiment, such as Figure 3 As shown, the execution process of constructing a time-frequency hybrid prompt set based on the monitoring signal sequence in step 120 may include, but is not limited to, steps 310 to 350.

[0041] Step 310: Perform Fourier transform and standardization on the monitoring signal sequence to obtain the spectrum matrix; Step 320: Perform amplitude normalization on the spectrum matrix to obtain spectrum data; Step 330: Perform frequency band division and similarity pooling on the spectrum data to obtain frequency domain prompts; Step 340: Obtain the time-domain cue based on the time characteristics of the monitored signal sequence; Step 350: Weight the frequency domain prompt and the time domain prompt according to the preset time-frequency weight to obtain a time-frequency hybrid prompt set.

[0042] In a feasible embodiment, the monitoring signal sequence can be denoted as: , ,in For time indexing, This represents the total length of the sequence. For the input monitoring signal sequence... Apply a Fourier transform; if it is a short-time Fourier transform, it is denoted as... If it is a discrete Fourier transform, then it is denoted as... Where f is the frequency index; taking the amplitude of the transformed result yields the corresponding spectrum matrix or spectrum vector Y, expressed as Y = Or Y= .

[0043] In a feasible embodiment, step 320, based on step 310, may adjust the spectral amplitude in the frequency dimension or a pre-defined frequency band dimension. Amplitude normalization is performed to eliminate the differences in frequency scale between signals from different sources, ultimately obtaining cross-domain scale-aligned spectral data Y.

[0044] In one feasible embodiment, in step 330, a frequency band division operation can be performed first, and a set of continuous and non-overlapping frequency bands can be pre-defined. Each of them A preset frequency band interval is defined; then, for the spectral data within each frequency band after division, its power spectral density is calculated, and the calculation process is shown in equation (1):

[0045] For each frequency band Based on the cross-domain similarity of the source and target domain spectral data within the frequency band, a non-negative weight that satisfies the normalization condition is assigned to each frequency point f within the band. ( Subsequently, the power spectral density within the frequency band is weighted and aggregated within the band to obtain a single frequency domain cue component. The components corresponding to all frequency bands together form a set of frequency domain cue symbols. The calculation process is shown in equation (2):

[0046] The set of frequency domain prompts can be denoted as It is used to characterize the energy distribution of multiple frequency bands and the cross-domain spectral similarity relationship.

[0047] In one feasible embodiment, in step 340, the monitoring signal sequence is... Feature extraction is performed to uncover the temporal attributes inherent in the sequence. These temporal attributes include at least one or more of the following: trend, seasonality, and transient nature, which can fully characterize the temporal dependency structure of the signal sequence. Furthermore, feature fusion and dimensionality mapping are performed based on the extracted temporal attributes to obtain a set of temporal cueings. .

[0048] In one feasible embodiment, in step 350, learnable time-frequency weights may be used. right and Dimensionally weighted, a time-frequency hybrid cue set is obtained. The calculation process is shown in equation (3):

[0049] in This represents a stepwise multiplication and addition process, which enables an adaptive balance between time and frequency features within a unified cue space. The time-frequency hybrid cue set can be denoted as... .

[0050] In one feasible embodiment, the preset prompt set refers to a set of prompt prototypes, which can be pre-set by the system (based on prior knowledge of traction system faults) or trained through historical monitoring data (learning signal characteristics of typical faults), denoted as... Where K is the number of prototypes. This refers to a single cue prototype vector. In determining the gating weights based on a preset cue set and monitoring signal sequence, statistical features (such as the signal's mean, variance, and trend characteristics) can be extracted from the current monitoring signal sequence to obtain a sample statistical vector s. This vector directly reflects the core state characteristics of the current monitoring signal. For each prototype vector in the cue prototype set... The similarity between a vector and a sample statistical vector s can be calculated using similarity functions (such as cosine similarity, Euclidean distance, etc.). This is used to measure the degree of matching between the current monitoring signal and each prompt prototype. The similarity results for all prototypes are then analyzed. Applying the softmax function for normalization yields the gate weights. The calculation process is shown in equation (4):

[0051] The larger the weight value, the higher the matching degree between the corresponding prompt prototype and the current monitoring signal.

[0052] In one feasible embodiment, it is possible to Apply threshold filtering or Top-K sparsification (retaining only the top-ranked prototypes) to suppress cue noise irrelevant to the current sample and highlight key cue that is sensitive to fault morphology.

[0053] In one feasible embodiment, gating weights can perform sparse selection on a small number of shape primitives and frequency bands. Essentially, this involves differentially weighting and filtering shape primitives and frequency bands based on the characteristics of the current monitored signal, retaining only a few key elements that highly match the core features of the signal, while suppressing irrelevant or redundant parts. Specifically, each weight value of the gating weight corresponds to a type of shape primitive or a frequency band, and the weight directly reflects the degree of matching between the shape primitive / frequency band and the current signal characteristics. When implementing sparse selection, filtering can be based on weight values, that is, retaining only a small number of shape primitives and frequency bands with high weight rankings. These elements often carry the core characteristics of the traction system's operating state (especially fault states). At the same time, shape primitives and frequency bands with extremely low weight values ​​are reduced or discarded, thereby filtering out noise interference and irrelevant features.

[0054] In one feasible embodiment, such as Figure 4 As shown, the execution process of step 140 may include, but is not limited to, steps 410 to 430.

[0055] Step 410: Based on the sequence characteristics of the monitoring signal sequence, combined with the time-frequency hybrid cue set and gating weights, the attribute label is calculated; Step 420: Perform vector quantization on the attribute tags to obtain discrete codes; Step 430: Perform shape decoding on the discrete code to obtain shape primitives.

[0056] In one feasible embodiment, combined with Figure 5 As shown, step 410 is implemented using the Transformer encoder-decoder module. The core is to combine time-frequency cues and gating constraints to complete feature encoding. The specific process is as follows: First, the original monitoring signal sequence is normalized and segmented. Then, position encoding is superimposed to convert the segmented signal into a fixed-dimensional sequence embedding. ,in The time length is the block level. The feature dimension is defined by this sequence embedding, which represents the core sequence features of the monitored signal and is input to the Transformer encoder. The time-frequency hybrid cue, after gating and sparse filtering, is transformed into a fixed-dimensional conditional cue representation through linear mapping, serving as the key vector K and value vector V of the Transformer cross-attention decoder, respectively. Simultaneously, the sequence embedding output by the encoder undergoes further linear transformation to obtain the query vector Q. To control the impact of the cue on the decoding process, a gating factor is injected into the attention alignment (logits), and feature matching calculation is then performed using a scaled dot product attention formula.

[0057] In equation (5), g is the gating factor related to equation (4); This indicates element-wise multiplication. This is the dimension of the key vector. This step amplifies cue features that highly match the current signal and suppresses irrelevant interference.

[0058] After multi-layer cross-attention computation, the decoder outputs the intermediate reconstruction sequence and continuous attribute labels. Attribute markers All the core information required for subsequent shape primitives is encoded in the attribute space, including parameters such as amplitude center, scale, starting position, and duration.

[0059] In a feasible embodiment, step 420 is performed by the Vector Quantization (VQ-layer) module. The core of this module is to discretize continuous attribute tags to enhance the distinguishability and interpretability of shape primitives. The specific process is as follows: A fixed set of code vectors is pre-constructed as a codebook. Each code vector in the codebook corresponds to a typical signal feature pattern. The continuous attribute marker token obtained in step 410 is input into the VQ-layer, and the token is compared with all code vectors in the codebook. Distance, select the code vector with the closest distance. A discrete representation of a token.

[0060] Output discrete code after matching Meanwhile, by constraining the difference between the discrete code and the original attribute label through the quantization loss function (Equation (6)), it is ensured that the discretization process does not lose key features:

[0061] in This indicates that the gradient operator is stopped. To balance the coefficients, we need to ensure the discriminative power of the discrete code while making the codebook learning process more stable.

[0062] In a feasible embodiment, step 430 is completed by the shape decoder module, the core of which is to restore the discrete code into shape primitives with definite physical parameters. The specific process is as follows: The discrete code obtained in step 420 The input shape decoder will parse out the amplitude center of the code based on the code vector features corresponding to the discrete code. ,scale Starting position Duration Parameters, etc. Based on the parsed parameters, the decoder generates a set of shape primitives with distinct local features. Each shape element corresponds to a typical local feature pattern of the monitoring signal.

[0063] To enhance the interpretability and reconstruction consistency of shape primitives, a shape constraint loss is introduced during training. and reconstruction consistency loss ,in The calculation process is shown in equation (7). The calculation process is shown in equation (8):

[0064] in, The preset prior / teacher shape template is used to make the generated shape primitives conform to the known typical fault / normal state characteristics.

[0065]

[0066] in, Its function is to keep the reconstructed sequence numerically consistent with the original signal.

[0067] Through the and The joint optimization enables shape primitives to accurately reconstruct actual monitoring signals while maintaining good interpretability in the parameter space.

[0068] In a feasible embodiment, after calculating the shape primitives, the original monitoring signal sequence can be further reconstructed based on the shape primitives and their corresponding shape parameters to obtain a reconstructed signal sequence. The execution process is as follows: first, attribute encoding is performed based on the shape parameters and their corresponding discrete codes to obtain attribute tags; then, the original monitoring signal sequence is reconstructed based on the attribute tags to obtain the reconstructed signal sequence. Specifically, after obtaining the shape primitives and their corresponding shape parameters (amplitude center)... ,scale Starting position Duration After that, these shape parameters are combined with the discrete code. The common input attribute encoder fuses and encodes information from discrete codes and shape parameters to generate attribute tags. ,,mark It integrates all feature information of the shape primitives. It generates an input signal sequence decoder with attribute labels. The decoder uses these attribute labels as constraints, combined with the sequence features of the original monitored signal, to progressively generate the final reconstructed signal sequence. .

[0069] In a feasible embodiment, intermediate reconstruction sequences and final reconstruction sequences can be used. The original signal sequence x(t) is used as the constraint target, and consistency and alignment loss are employed. By applying consistency and alignment losses to the temporal statistical properties and shape space properties, the generated reconstructed sequence is made consistent with the target sequence in both statistical and shape sense, thereby improving the cross-domain robustness and engineering usability of the sequence.

[0070] In one feasible embodiment, such as Figure 6 As shown, the execution process of step 160 may include, but is not limited to, steps 610 and 620.

[0071] Step 610: Calculate the signal sequence residual value based on the monitored signal sequence and the reconstructed signal sequence; Step 620: Obtain the fault diagnosis result based on the signal sequence residual value and the preset fault threshold.

[0072] In a feasible embodiment, step 610 involves quantifying the difference between the original signal and the reconstructed signal to obtain core indicators for fault diagnosis. For example... Figure 7 As shown, the execution process may include, but is not limited to, steps 710 to 740.

[0073] Step 710: Obtain the residual sequence based on the monitoring signal sequence and the reconstructed signal sequence; Step 720: Calculate the mean residual of the signal sequence based on the residual sequence; Step 730: Calculate the standard deviation of the signal sequence residuals based on the mean of the signal sequence residuals; Step 740: Calculate the residual value of the signal sequence based on the mean residual and the standard deviation of the residual.

[0074] In one feasible embodiment, in step 710, the original monitoring signal sequence is... and reconstructed signal sequence Subtracting the elements at corresponding positions yields the residual at each time step. All residuals form a residual sequence.

[0075] In a feasible embodiment, in step 720, in order to uniformly measure the overall size of the residuals, the absolute value of each element in the residual sequence is taken (to avoid the cancellation of positive and negative deviations), and then the average value is calculated to obtain the residual mean value reflecting the average deviation. The calculation process is shown in equation (9):

[0076] in, This is the window length.

[0077] In a feasible embodiment, since the mean alone cannot reflect the fluctuation of the residuals, step 730 calculates the dispersion of the absolute value of the residuals based on the residual mean, thereby obtaining the residual standard deviation that reflects the degree of fluctuation. The calculation process is shown in equation (10):

[0078] In a feasible embodiment, in step 740, the mean and standard deviation of the residuals are combined, and an adjustment hyperparameter λ is introduced to weight and combine the two to obtain the final signal sequence residual value. λ can be adjusted according to the abnormal sensitivity requirements of different scenarios.

[0079] In a feasible embodiment, step 620 obtains a normal or abnormal diagnostic conclusion by comparing the residual value with a threshold. The threshold θ is determined as follows: during the training phase, the distribution of the residual values ​​j of the abnormal subset or normal samples in the source domain is adaptively estimated (e.g., by selecting quantiles or using a mean-weighted deviation method) to ensure that the threshold matches the normal fluctuation range of the signal.

[0080] In a feasible embodiment, during the process of obtaining a fault diagnosis result based on the signal sequence residual value and a preset fault threshold, if the number of times the signal sequence residual value is continuously greater than the preset fault threshold exceeds a first threshold, a fault diagnosis result indicating that the monitored signal sequence has an abnormal fault is obtained; if the number of times the signal sequence residual value is continuously less than or equal to the preset fault threshold exceeds a second threshold, a fault diagnosis result indicating that the monitored signal sequence does not have an abnormal fault is obtained. In short, if continuously... If the residual value j obtained from each calculation is greater than the threshold θ, it indicates that the signal deviation continues to exceed the normal range, triggering a fault alarm; if it continues... When the residual value j obtained from the second calculation falls back to θ or below, it indicates that the signal has returned to normal, and the fault alarm is then cleared.

[0081] In a feasible embodiment, after obtaining the fault diagnosis result, the fault frequency band in the monitoring signal sequence can be determined based on the fault diagnosis result, and the fault frequency band can be identified by using the time-frequency mixed cue set and shape primitives to obtain the fault type. Specifically, after obtaining the abnormal fault diagnosis result, the frequency band that deviates from the normal range in the monitoring signal (i.e., the fault frequency band) can be located first by combining information such as residual sequence and frequency band energy; then, the characteristics of the fault frequency band can be analyzed using the previously generated time-frequency mixed cue set (containing the feature information of each frequency band) and shape primitives (corresponding to the local feature patterns of the signal), and finally the specific fault type can be identified, including single faults and compound faults.

[0082] In one feasible embodiment, such as Figure 8 As shown, the fault identification process may include, but is not limited to, steps 810 to 850.

[0083] Step 810: Based on the time-frequency hybrid cue set, classify the fault modes of the fault frequency band to obtain the target fault mode; Step 820: Determine the fault message corresponding to the target fault mode; Step 830: Combine and merge the fault messages to obtain a composite fault message; Step 840: Obtain the probability distribution of fault categories based on the composite fault indications; Step 850: Interpret the probability distribution of fault categories based on shape primitives to obtain the fault type.

[0084] In one feasible embodiment, a time-frequency hybrid prompt set It includes signal characteristics of each frequency band, and based on this, preliminary pattern recognition is performed on the located faulty frequency band, such as determining whether the fault is a single mode (e.g., fault A, fault B) or a combined mode (e.g., fault A, fault B). , , , This allows us to determine the target failure mode.

[0085] In a feasible embodiment, for the target fault mode identified in step 810, the corresponding single fault prompt is extracted: if the target is a single fault (such as fault A), the prompt corresponding to fault A is extracted; if the target is a combined mode (such as A+B), the prompts corresponding to fault A and fault B are extracted respectively.

[0086] In a feasible embodiment, if the target fault mode is a combined mode, the corresponding single fault prompts need to be sequentially concatenated in the time and frequency band dimensions and aligned with the feature scales; then, these single fault prompts are integrated into a composite fault prompt through learnable linear fusion or gated fusion. This explicitly expresses the interaction relationships when multiple faults coexist.

[0087] In one feasible embodiment, a single fault indication will be provided. Or a combination of fault prompts The input classifier trains the model through multi-class / multi-label learning (using cross-entropy loss to optimize class probabilities during training), and finally outputs the probability distribution of the fault class corresponding to the current fault (e.g., "fault A probability 0.9, compound fault A+B probability 0.85", etc.). In a feasible embodiment, in step 850, the category corresponding to the maximum posterior in the probability distribution can be selected as the preliminary fault result; simultaneously, combining the shape primitives and the corresponding interpretability indicators, it can be explained which type of shape primitive plays a dominant role in the fault frequency band, ultimately obtaining a fault type that combines result and explanation. The interpretability indicators include: the shape primitive contribution vector obtained by multi-head attention weighting. This represents the relative contribution of primitives of different shapes at various time positions; the frequency band energy contribution vector is calculated based on the frequency band energy. Used to indicate the energy support status of each frequency band; based on a time-frequency hybrid cue set The cue-feature attribution heatmap, constructed with sequence features, is used to reveal the causal correspondence between cue and feature.

[0088] The following is combined with Figure 5 The fault diagnosis method of this system will be described in detail with a specific embodiment.

[0089] Based on the aforementioned fault diagnosis method processing logic and execution flow, this embodiment further constructs a system architecture for implementing the method. This system includes an input module, a hybrid prompt generation module, a separable shape element module, and a fault identification module. The functions and collaborative relationships of each module are as follows: The input module is responsible for acquiring the monitoring signal sequence of the traction system, specifically the source domain dataset. With the target domain dataset Cleaning, noise reduction, standardization, and segmentation are performed to form a standardized monitoring signal sequence. This sequence is fed to the hybrid cue generation module for the construction of the time-frequency hybrid cue set, and directly serves as the basic input for the separable shape element module to support the joint modeling of the signal's internal shape structure and energy distribution.

[0090] The core task of the hybrid prompt generation module is to construct a time-frequency hybrid prompt set based on the monitored signal sequence. The gating weight g is determined by combining a preset cue set with the monitoring signal sequence. The specific workflow is as follows: First, the monitoring signal sequence undergoes Fourier transform and standardization to generate spectral data, converting the time-domain signal into a stable frequency-domain representation to achieve band aggregation; then, the spectral data is divided into frequency bands and similarity pooled to obtain a set of frequency-domain cue symbols. Simultaneously, the temporal features of the monitoring signal sequence are extracted to form a set of time-domain cue symbols. Subsequently, the two types of cue sets are weighted dimension-wise using learnable time-frequency weights, ultimately generating a time-frequency hybrid cue set. Furthermore, the module's built-in prototype allocation unit can calculate the gating weight *g* based on a preset basic cue set and the characteristic distribution of the current monitored signal sequence, thereby adjusting the weight ratio of different components in the time-frequency hybrid cue set. The generated time-frequency hybrid cue set... Along with the gating weight g, it will be synchronously transmitted to the separable shape element module for subsequent processing.

[0091] Separable shape element modules are based on monitoring signal sequences and time-frequency mixed cue sets. The gating weights g are used to perform the work, and finally the shape primitives s and the reconstructed signal sequence are output. The specific steps are as follows: First, the input signal sequence is normalized and divided into blocks, and then positional encoding is superimposed before being input into a multi-layer Transformer encoding / decoding structure; in the Transformer cross-attention mechanism, the linear transformation result of the sequence feature Z is used as the query vector Q by the attribute decoder, and the gated time-frequency hybrid cue set is then used as the query vector Q. The linear transformation results are used as key vector K and value vector V, and gating weights g are injected into the attention alignment stage to complete conditional decoding, outputting intermediate reconstruction results and continuous attribute labels z; the labels z are quantized into discrete codes through the VQ-layer. Shape decoding yields the shape primitive 's' and its corresponding amplitude center, scale, starting position, duration, and other parameters; simultaneously, a generation branch is set up to generate the discrete code. The corresponding shape parameters are input into the attribute encoder to generate attribute tags. Then, the signal sequence decoder generates the reconstructed signal sequence. Finally, consistency and alignment losses are applied. Constraining the temporal statistical properties and shape space properties enhances the cross-domain robustness of the sequence. It should be noted that the specific workflow of the separable shape element module is consistent with the processing procedures in steps 410 to 430 above; for more details, please refer to the corresponding descriptions above, which will not be elaborated here.

[0092] The fault identification module diagnoses faults based on a preset fault threshold and the reconstructed signal sequence. Its working logic is as follows: First, it calculates the residual value j between the monitored signal sequence and the reconstructed signal sequence, i.e., the reconstruction error. The error can be calculated using the mean square error (MSE) or Euclidean distance algorithm. If the residual value j obtained from L1 consecutive calculations is greater than the threshold θ, it indicates that the signal deviation continues to exceed the normal range, and the system immediately triggers a fault alarm. If the residual value j obtained from L2 consecutive calculations falls back to or below the threshold θ, the signal is determined to have returned to normal, and the fault alarm is deactivated. In the event of a triggered fault alarm, this module further incorporates a time-frequency hybrid cue set. With shape primitive s, single or compound fault identification is performed on the fault frequency band, and the final output is an interpretable diagnostic result, including fault category, shape primitive contribution vector, frequency band energy contribution vector and corresponding visualization chart (such as attribution heatmap).

[0093] It is worth noting that by leveraging the correspondence between multi-head attention weights and time-frequency hybrid cues, we can further construct shape primitive contribution vectors, frequency band energy contribution vectors, and cue-feature attribution thermodynamic mappings. Through this series of visualization methods, we can intuitively explain which type of shape, where, and in which frequency band it causes signal anomalies. This enables highly sensitive detection of weak and complex faults in the traction system while outputting interpretable diagnostic results with practical engineering value.

[0094] See Figure 9 The diagram illustrates the complete workflow of time-frequency hybrid cue learning and shape element processing in traction system fault diagnosis. The collaborative logic of each module in the system is as follows: The hybrid cue generation module's workflow includes: inputting single fault signals and composite fault signals, first splitting the signals to obtain features in three frequency bands: low frequency, mid-high frequency, and high frequency. Then, calculating the correlation degree of each frequency band feature through cosine similarity and cross-attention, and then pooling to obtain a frequency domain cue set. Simultaneously, the composite fault signal is input into the semantic prototype unit, combined with a preset basic cue set to generate prototype features, and then fused with the prototype to obtain another cue information. Finally, the two types of cue information are superimposed to form the final time-frequency hybrid cue set. The formula shown in the diagram is the gating weight constraint condition, which essentially limits the weight ratio during prototype fusion to avoid a certain type of feature having an excessively high proportion in the hybrid cue set, ensuring a balanced fusion of different features, thereby making the generated time-frequency hybrid cue set more stable. The separable shape element module's workflow includes: input signals are first normalized and segmented, then superimposed with position encoding and input into the Transformer encoder. The encoded features are output as continuous attribute labels by the attribute decoder, then quantized into discrete codes by the codebook, and finally processed by the shape decoder to obtain shape primitives and their parameters. Subsequently, the shape primitives and discrete codes are input into the attribute encoder, processed by the Transformer decoder, and then enter the MLP frequency learning stage. Features are first processed by frequency band, then cross-domain adaptation is achieved through a frequency multilayer perceptron and domain inversion, and finally, the output features are used by the fault identification module. After the above processing, the features are used by the MLP frequency learning classification unit to determine the fault category, and simultaneously, combined with a time-frequency hybrid cue set, interpretable results are output, such as the fault frequency band and the contribution of the shape primitives.

[0095] In a feasible embodiment, after acquiring the monitoring signal sequence of the traction system, the organization and division of domain data can be carried out simultaneously. The source domain data consists of historical operation and fault signals of the traction system, while the target domain data consists of the real-time signals to be monitored. Specifically, the signals from the source and target domains are divided into training and testing sets according to a preset ratio. The training set is used for training and fitting subsequent modules, while the testing set is used to verify the diagnostic accuracy and generalization ability of the modules. Simultaneously, a small number of real-time signal samples from the target domain are retained as basic data for constructing target domain prompts. Based on this, a unique domain identifier needs to be added to the signals from the source and target domains. This identifier serves as conditional metadata and participates throughout the subsequent construction of the time-frequency hybrid prompt set, the alignment and matching of signal features from different domains, and the quantitative evaluation of the module's final diagnostic effect.

[0096] It is understood that this application can be used for weak fault detection and complex fault identification in multi-condition and high-noise environments such as rail transit traction systems. The overall process is as follows: First, the source domain historical signal and the target domain real-time signal of the traction system are divided into data, reserving a small number of target domain samples for subsequent prompt construction; then, time-frequency hybrid prompts are constructed for the signal. Specifically, the signal is first subjected to spectrum transformation, frequency band division, and power spectral density calculation. Combined with cross-domain similarity weighted aggregation, frequency domain prompts are obtained, and then the signal time attributes are extracted to obtain time domain prompts. The two are fused through learnable weights to form a time-frequency hybrid prompt set; then, based on the prompt set and signal features, shape primitives are obtained through attribute labeling, vector quantization, shape primitive decoding, and other steps; then, the presence of a fault is determined by calculating the residual between the original signal and the reconstructed signal; if an anomaly is determined, single fault identification and complex fault identification are performed on the fault frequency band by combining the time-frequency hybrid prompt set and the shape primitives, and the cause of the fault is explained by indicators such as shape primitive contribution and frequency band energy contribution.

[0097] In a feasible embodiment, this application achieves high recognition accuracy in both single-fault and compound-fault diagnosis. Figure 10(a) is the confusion matrix for a single-fault scenario, with the diagnostic objects including five types of faults: apq, drift, gain, periodic, and shock. From the diagonal of the matrix, it can be seen that the classification accuracy for each type of fault is apq 98.50%, drift 97.00%, gain 97.50%, periodic 97.00%, and shock 97.50%, respectively. The values ​​in the off-diagonal areas are generally no more than 1.00%, with only a small amount of confusion (0.20% to 1.00%) between adjacent categories. This indicates that the method can stably distinguish different faults in single-fault mode, with very low probabilities of missed detection and false detection. Figure 10(b) is the confusion matrix for a compound-fault scenario, with the diagnostic objects being three types of compound faults: apqperiodic, apqshock, and driftapq. The classification accuracy for diagonal faults reached 90.50%, 92.00%, and 94.00%, respectively, while the accuracy for off-diagonal faults fluctuated only between 2.00% and 6.50%. Even with overlapping faults, the method can effectively distinguish between different compound fault modes with very few misclassifications. These two results demonstrate that the method achieves high diagnostic accuracy and low confusion in both single and compound fault scenarios, validating its fault identification capabilities and practical value under complex operating conditions in traction systems.

[0098] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A traction system fault diagnosis method based on shape primitives and time-frequency hybrid prompts, characterized in that, include: Acquire the monitoring signal sequence of the traction system; Based on the monitoring signal sequence, a time-frequency hybrid cue set is constructed; The gating weight is determined based on the preset prompt set and the monitoring signal sequence; Based on the monitoring signal sequence, the time-frequency hybrid cue set, and the gating weights, shape primitives are obtained; The monitoring signal sequence is reconstructed based on the shape primitives to obtain a reconstructed signal sequence. Based on the preset fault threshold and the reconstructed signal sequence, the fault diagnosis result is obtained.

2. The traction system fault diagnosis method according to claim 1, characterized in that, The method of obtaining shape primitives based on the monitoring signal sequence, the time-frequency hybrid cue set, and the gating weights includes: Based on the sequence characteristics of the monitored signal sequence, combined with the time-frequency hybrid cue set and the gating weight, the attribute label is calculated; The attribute markers are vector-quantized to obtain discrete codes; The discrete code is subjected to shape decoding to obtain shape primitives.

3. The traction system fault diagnosis method according to claim 2, characterized in that, The step of reconstructing the monitoring signal sequence based on the shape primitives to obtain the reconstructed signal sequence includes: Attributes are encoded based on the shape parameters of the shape primitives and the corresponding discrete codes to obtain attribute tags. The monitoring signal sequence is reconstructed based on the attribute tags to obtain the reconstructed signal sequence.

4. The traction system fault diagnosis method according to claim 1, characterized in that, The acquisition of the monitoring signal sequence of the traction system includes: Collect monitoring signal datasets; The monitoring signal dataset is cleaned, denoised, and standardized to obtain a standardized signal set; According to a preset time window, the standardized signal set is divided into sample sequences of equal length to obtain the monitoring signal sequence.

5. The traction system fault diagnosis method according to claim 1, characterized in that, The step of constructing a time-frequency hybrid cue set based on the monitored signal sequence includes: The monitoring signal sequence is subjected to Fourier transform and standardization to obtain the spectrum matrix; The spectrum matrix is ​​subjected to amplitude normalization to obtain spectrum data; The frequency spectrum data is divided into frequency bands and subjected to similarity pooling to obtain frequency domain prompts; Based on the temporal characteristics of the monitored signal sequence, a time-domain cue is obtained; The frequency domain prompt and the time domain prompt are weighted according to the preset time-frequency weights to obtain a time-frequency hybrid prompt set.

6. The traction system fault diagnosis method according to claim 1, characterized in that, The step of obtaining the fault diagnosis result based on the preset fault threshold and the reconstructed signal sequence includes: The residual value of the signal sequence is calculated based on the monitoring signal sequence and the reconstructed signal sequence; The fault diagnosis result is obtained based on the residual value of the signal sequence and the preset fault threshold.

7. The traction system fault diagnosis method according to claim 6, characterized in that, The step of calculating the signal sequence residual value based on the monitored signal sequence and the reconstructed signal sequence includes: Based on the monitoring signal sequence and the reconstructed signal sequence, the residual sequence is obtained; The mean residual of the signal sequence is calculated based on the residual sequence. The standard deviation of the signal sequence residuals is calculated based on the mean of the signal sequence residuals. The residual value of the signal sequence is calculated based on the mean residual value and the standard deviation of the residual value of the signal sequence.

8. The traction system fault diagnosis method according to claim 6, characterized in that, The step of obtaining the fault diagnosis result based on the signal sequence residual value and a preset fault threshold includes: When the number of times the residual value of the signal sequence is continuously greater than the preset fault threshold exceeds the first threshold, a fault diagnosis result indicating that there is an abnormal fault in the monitored signal sequence is obtained. When the number of times the residual value of the signal sequence is continuously less than or equal to the preset fault threshold exceeds the second threshold, a fault diagnosis result is obtained that there is no abnormal fault in the monitored signal sequence.

9. The traction system fault diagnosis method according to claim 1, characterized in that, After obtaining the fault diagnosis result, the method further includes: Based on the fault diagnosis results, the fault frequency band in the monitoring signal sequence is determined; The fault type is obtained by identifying the fault frequency band based on the time-frequency hybrid cue set and the shape primitive.

10. The traction system fault diagnosis method according to claim 9, characterized in that, The step of identifying the fault frequency band based on the time-frequency hybrid cue set and the shape primitive to obtain the fault type includes: Based on the time-frequency hybrid cue set, the fault mode of the fault frequency band is classified to obtain the target fault mode; Determine the fault indication corresponding to the target fault mode; The fault prompts are spliced ​​and merged to obtain composite fault prompts; The probability distribution of fault categories is obtained based on the composite fault indication; The fault type is obtained by interpreting the probability distribution of the fault category based on the shape primitive.

Images