Suspension controller fault diagnosis method and device, electronic equipment and storage medium

Abstract

The application discloses a fault diagnosis method and device of a suspension controller, electronic equipment and a storage medium, aiming to realize more rapid, efficient, accurate and effective fault diagnosis of the suspension controller. The method comprises the following steps: acquiring multi-channel time sequence data of the suspension controller, and determining a system step response parameter of the suspension controller; processing the multi-channel time sequence data and the system step response parameter by using a preset constraint model to obtain a time sequence window length; intercepting target multi-channel time sequence data corresponding to the time sequence window length from the multi-channel time sequence data; extracting multi-scale time sequence features of the target multi-channel time sequence data by using a feature extraction network of a preset diagnosis model, performing dimension reduction processing on the multi-scale time sequence features by using a fault classification network of the preset diagnosis model, and determining a fault diagnosis result of the suspension controller according to the dimension reduction features.

Description

Technical Field

[0001] This application relates to the field of rail transit technology, and in particular to a fault diagnosis method, device, electronic equipment, and computer-readable storage medium for a suspension controller. Background Technology

[0002] Currently, high-speed maglev transportation technology has entered a new stage of rapid development. With the continuous expansion of the scale of maglev transportation lines and the increasing operating mileage, the suspension system, as a key core component to ensure the stable operation of trains, directly determines the overall operational safety of trains and has become the top priority to support the efficient operation and maintenance of large-scale maglev transportation networks.

[0003] As the "brain" of the suspension guidance system, the suspension controller undertakes the core function of integrating information from multiple sensors and achieving stable control of the suspension electromagnets. However, during long-term train operation, the suspension controller is susceptible to failure due to multiple factors, posing a significant threat to operational safety. To address the risk of sensor failure, the suspension control system adopts a state observer design based on an overlapping structure. By collecting multi-dimensional signals such as gaps, current, and voltage to estimate faulty sensor data, it significantly reduces the system failure rate. However, the multi-parameter redundancy strategy also brings massive amounts of high-dimensional time-series data, posing a huge challenge to subsequent feature extraction and fault diagnosis: the surge in data dimensionality leads to a sharp drop in the efficiency of traditional diagnostic methods, making it difficult to accurately identify faults while retaining key feature information.

[0004] Therefore, how to achieve faster, more efficient, accurate and effective fault diagnosis of suspension controllers is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this application is to provide a fault diagnosis method for a suspension controller, which can achieve faster, more efficient, more accurate and effective fault diagnosis of the suspension controller; another purpose of this application is to provide a fault diagnosis device, electronic device and computer-readable storage medium for a suspension controller, all of which have the above-mentioned beneficial effects.

[0006] In a first aspect, this application discloses a fault diagnosis method for a suspension controller, comprising: Acquire multi-channel timing data of the suspension controller and determine the system step response parameters of the suspension controller; The multi-channel time-series data and the system step response parameters are processed using a preset constraint model to obtain the time-series window length; Extract the target multi-channel time series data corresponding to the time series window length from the multi-channel time series data; Multi-scale time-series features of the target multi-channel time-series data are extracted using the feature extraction network of the preset diagnostic model. The multi-scale time-series features are then dimensionality-reduced using the fault classification network of the preset diagnostic model. Based on the dimensionality-reduced features, the fault diagnosis result of the suspension controller is determined.

[0007] Optionally, the preset constraint model includes system response time constraints and autocorrelation function decay constraints; Accordingly, the multi-channel time-series data and the system step response parameters are processed using a preset constraint model to obtain the time-series window length, including: The system step response parameters are processed using the system response time constraint in the preset constraint model to obtain the minimum data length for the suspension controller to reach step response stability. The multi-channel time series data is processed using the autocorrelation function decay constraint in the preset constraint model to obtain the minimum hysteresis step size at which the autocorrelation coefficient of the multi-channel time series data decreases to within a preset threshold. The maximum value between the minimum data length and the minimum lag step size is used as the time series window length.

[0008] Optionally, the feature extraction network is specifically a residual neural network, and the residual neural network incorporates a skip connection structure; Accordingly, the feature extraction network of the pre-defined diagnostic model is used to extract multi-scale temporal features of the target multi-channel time-series data, including: The residual neural network is used to extract features from the target multi-channel time-series data to obtain shallow detail features and deep semantic features. The shallow detail features and the deep semantic features are fused using the skip connection structure to obtain the multi-scale temporal features.

[0009] Optionally, the fault classification network is specifically an Arnold neural network; Accordingly, the multi-scale temporal features are processed by dimensionality reduction using the fault classification network of the preset diagnostic model, and the fault diagnosis result of the suspension controller is determined based on the dimensionality reduction features, including: Each feature component in the multi-scale time series features is subjected to univariate nonlinear transformation processing to obtain each univariate nonlinear transformation feature; The weighted summation and aggregation of each of the univariate nonlinear transformation features are used to obtain intermediate feature variables; The intermediate feature variables are subjected to univariate nonlinear mapping processing to obtain nonlinear mapping features; The nonlinear mapping features are superimposed and output to obtain the fault diagnosis result.

[0010] Optionally, after acquiring the multi-channel timing data of the suspension controller, the process also includes: Preprocessing operations are performed on the multi-channel time-series data; the preprocessing operations include one or more combinations of time synchronization processing, noise filtering processing, and feature extraction processing.

[0011] Optionally, the fault diagnosis method for the suspension controller further includes: Build the simulation environment corresponding to the suspension controller in the simulation platform; Under the fault-free injection state of the simulation environment, the timing data of the suspension controller is collected to obtain the normal state dataset; When a gradual fault state is injected into the simulation environment, the timing data of the suspension controller is collected to obtain a gradual fault dataset. When intermittent noise fault conditions are injected into the simulation environment, the timing data of the suspension controller is collected to obtain the intermittent noise fault dataset; The preset diagnostic model is obtained by training the normal state dataset, the gradual fault dataset, and the intermittent noise fault dataset.

[0012] Optionally, the fault diagnosis method for the suspension controller further includes: The preset diagnostic model is evaluated using a preset three-dimensional evaluation system to obtain evaluation results; the preset three-dimensional evaluation system includes precision evaluation index, recall evaluation index, and F1 score evaluation index. When the evaluation result indicates that the preset diagnostic model fails the evaluation, the preset diagnostic model is updated.

[0013] Secondly, this application discloses a fault diagnosis device for a suspension controller, comprising: The acquisition module is used to acquire multi-channel timing data of the suspension controller and determine the system step response parameters of the suspension controller; The processing module is used to process the multi-channel time series data and the system step response parameters using a preset constraint model to obtain the time series window length; The extraction module is used to extract the target multi-channel time series data corresponding to the time series window length from the multi-channel time series data; The diagnostic module is used to extract multi-scale time-series features of the target multi-channel time-series data using the feature extraction network of the preset diagnostic model, to perform dimensionality reduction processing of the multi-scale time-series features using the fault classification network of the preset diagnostic model, and to determine the fault diagnosis result of the suspension controller based on the dimensionality reduction features.

[0014] Thirdly, this application discloses an electronic device, including: Memory, used to store computer programs; A processor, used to execute the computer program to implement the steps of any of the suspension controller fault diagnosis methods described above.

[0015] Fourthly, this application discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for diagnosing faults in a levitation controller.

[0016] This application provides a fault diagnosis method for a suspension controller, comprising: acquiring multi-channel time-series data of the suspension controller and determining the system step response parameters of the suspension controller; processing the multi-channel time-series data and the system step response parameters using a preset constraint model to obtain a time series window length; extracting target multi-channel time-series data corresponding to the time series window length from the multi-channel time-series data; extracting multi-scale time-series features of the target multi-channel time-series data using a feature extraction network of a preset diagnostic model; performing dimensionality reduction processing on the multi-scale time-series features using a fault classification network of the preset diagnostic model; and determining the fault diagnosis result of the suspension controller based on the dimensionality reduction features.

[0017] The technical solution provided in this application first collects multi-channel time-series data of the suspension controller and determines the system step response parameters of the suspension controller. Then, both are input into a preset constraint model to determine the most suitable time-series window length for the suspension controller. This allows for the extraction of target multi-channel time-series data within this window length from the multi-channel time-series data, which is then processed using a preset diagnostic model to achieve fault diagnosis of the suspension controller. Therefore, this technical solution utilizes a preset constraint model to determine the time-series window length, avoiding inaccurate fault diagnosis results caused by windows that are too short or too long. Furthermore, the preset diagnostic model includes a feature extraction network and a fault classification network. The former is used for multi-scale time-series feature extraction, while the latter is used for feature dimensionality reduction and fault classification, effectively reducing computational complexity and improving fault diagnosis efficiency. Therefore, this technical solution can achieve faster, more efficient, and more accurate fault diagnosis of the suspension controller.

[0018] The fault diagnosis device, electronic device, and computer-readable storage medium for the suspension controller provided in this application also have the above-mentioned technical effects, and will not be described in detail here. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the prior art and the embodiments of this application, the accompanying drawings used in the description of the prior art and the embodiments of this application will be briefly introduced below. Of course, the accompanying drawings described below with respect to the embodiments of this application are only a part of the embodiments in this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort, and such other drawings also fall within the protection scope of this application.

[0020] Figure 1 This is a flowchart illustrating a fault diagnosis method for a suspension controller provided in an embodiment of this application. Figure 2 This is a schematic diagram of the dynamic response curve of a suspension system provided in an embodiment of this application; Figure 3 This is a schematic diagram of the control structure of a suspension controller provided in an embodiment of this application; Figure 4 This is a flowchart illustrating another method for diagnosing faults in a suspension controller provided in an embodiment of this application. Figure 5 This is a schematic diagram illustrating a field application example of a suspension controller provided in this application embodiment; Figure 6 A schematic diagram of the structure of a fault diagnosis device for a suspension controller provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0021] The core of this application is to provide a fault diagnosis method for a suspension controller, which can achieve faster, more efficient, and more accurate fault diagnosis of the suspension controller; another core aspect of this application is to provide a fault diagnosis device, electronic device, and computer-readable storage medium for a suspension controller, all of which have the aforementioned beneficial effects.

[0022] To provide a clearer and more complete description of the technical solutions in the embodiments of this application, the technical solutions in the embodiments of this application will be described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0023] This application provides a method for diagnosing faults in a suspension controller.

[0024] Please refer to Figure 1 , Figure 1This is a flowchart illustrating a fault diagnosis method for a suspension controller provided in an embodiment of this application. The fault diagnosis method for the suspension controller may include, but is not limited to, the following S101~S104.

[0025] S101: Acquire multi-channel timing data of the suspension controller and determine the system step response parameters of the suspension controller.

[0026] This step aims to acquire multi-channel timing data of the suspension controller and determine the system step response parameters. Multi-channel timing data refers to timing data from multiple different physical dimensions. In one possible implementation, this multi-channel timing data may include suspension controller output voltage timing data, electromagnet current timing data, sensor gap value timing data, and acceleration response timing data. The system step response parameters refer to the parameters related to the suspension controller reaching a preset stability condition (e.g., 95% or 98% stability). In one possible implementation, this system step response parameter can specifically be the step response settling time, i.e., the duration required for the suspension controller to reach the preset stability condition.

[0027] In one embodiment of this application, after acquiring the multi-channel timing data of the suspension controller, the process may further include: performing preprocessing operations on the multi-channel timing data; the preprocessing operations include one or more combinations of time synchronization processing, noise filtering processing, and feature extraction processing. It is understood that data preprocessing operations can effectively ensure data format uniformity, making subsequent data calculations more convenient, thereby ensuring the accuracy and efficiency of suspension controller fault diagnosis.

[0028] S102: Use a preset constraint model to process multi-channel time series data and system step response parameters to obtain the time series window length.

[0029] This step aims to determine the time series window length. It is understood that in time series fault detection tasks, the selection of the window length plays a decisive role in the model's ability to capture temporal dependencies and improve prediction accuracy. If the window is set too short, it may fail to cover the system's key dynamic response processes, leading to insufficient modeling of lag dependencies. If the window is set too long, it may introduce redundant or irrelevant information, increasing computational complexity and increasing the risk of overfitting. Therefore, the window length setting needs to strike a balance between information sufficiency and model generalization ability. Based on this, this application proposes a method for determining the time series window length based on a preset constraint model. This involves inputting multi-channel time series data and system step response parameters into a preset constraint model for processing, and the model's output is the most suitable time series window length for the current suspension controller.

[0030] In one embodiment of this application, the preset constraint model may include system response time constraints and autocorrelation function decay constraints. Accordingly, processing multi-channel time-series data and system step response parameters using the preset constraint model to obtain the time series window length may include: processing the system step response parameters using the system response time constraints in the preset constraint model to obtain the minimum data length for the levitation controller to reach a stable step response; processing the multi-channel time-series data using the autocorrelation function decay constraints in the preset constraint model to obtain the minimum lag step size for the autocorrelation coefficient of the multi-channel time-series data to decrease to within a preset threshold; and using the maximum value between the minimum data length and the minimum lag step size as the time series window length.

[0031] Understandably, considering that most time series data in industrial environments originate from complex physical processes or closed-loop control systems, their data not only possess a clear physical response time scale but also often exhibit long-term lags in statistical correlation structures. Therefore, this application constructs mathematical modeling criteria for window length from two dimensions: system response time constraints and autocorrelation function decay constraints. A joint constraint model—a preset constraint model—is also constructed to provide a theoretical basis and interpretability support for selecting the time series window length.

[0032] In the implementation process, it is assumed that the suspension controller system has a total of m sensors (m data channels), and each sensor can collect N time points, X∈ m×N This is a complete multi-channel time series data, where the i-th row x (i) =[x (i) 1,x (i) 2, ..., x (i) N (i=1,2,...,m) represents the observation (time series data) of the i-th sensor. To ensure that the model can fully capture the key dynamic features and time dependencies in all sensor dimensions, the time series window length can be constrained and modeled from the following two aspects: (1) System response time constraint: Assume that the system response time of the i-th sensor is t. (i) s The uniform sampling interval is Δt d The minimum data length required to complete 95% of the response is: L (i) time =[t (i) s / Δt d ].

[0033] (2) Autocorrelation function attenuation constraint: for each sensor x (i) Calculate its first-order autocorrelation function ρ (i)(k), and define the autocorrelation attenuation length L. (i) acf The minimum lag step that satisfies the following conditions: L (i) acf =min{k | ρ (i) (k)<τ}, where τ is the autocorrelation decay threshold, used to determine how many hours the autocorrelation decays to before the sequence is considered to be essentially no longer correlated.

[0034] Therefore, the final time series window length L can be set as: L≥max{L (i) time L (i) acf The maximum value between the minimum data length and the minimum lag step size is used as the time series window length.

[0035] In practical applications, to determine the dynamic response characteristics of the suspension unit in a high-speed maglev train suspension system, a step input with an amplitude of 1mm is applied to the suspension controller. This activates the PID controller's internal PID regulator and current closed-loop mechanism, causing it to adjust the suspension voltage and apply it to the suspension electromagnet via a power chopper. This excites the generation of a suspension current, which in turn forms a levitation force to adjust the target gap. The initial system gap was 10mm, eventually stabilizing at 11mm. Its 98% response time was measured to be 2.0768s. Figure 2 As shown, Figure 2 This is a schematic diagram of the dynamic response curve of a suspension system provided in an embodiment of this application. Based on the sampling interval Δt d =0.02s, which translates to 104 time steps for the response time. Considering modeling redundancy and system uncertainty, an empirical compensation δ=16 steps is introduced. Therefore, the minimum input window length under the physical response constraint is 120 steps. Furthermore, the autocorrelation function of the time series from this channel is calculated, and the minimum lag step for decay to the threshold is extracted, resulting in 56 steps. Therefore, based on the joint constraint modeling criterion, the required time series window length for time series modeling is 120 steps.

[0036] S103: Extract the target multi-channel time series data corresponding to the time series window length from the multi-channel time series data.

[0037] This step aims to obtain the target multi-channel time series data corresponding to the time series window length. The multi-channel time series data corresponding to the time series window length can be directly extracted from the multi-channel time series data as the target multi-channel time series data.

[0038] S104: Use the feature extraction network of the preset diagnostic model to extract multi-scale time-series features of the target multi-channel time-series data, use the fault classification network of the preset diagnostic model to reduce the dimensionality of the multi-scale time-series features, and determine the fault diagnosis result of the suspension controller based on the dimensionality reduction features.

[0039] This step aims to perform multi-channel time-series data processing based on a pre-defined diagnostic model to achieve fault diagnosis of the suspension controller. The pre-defined diagnostic model includes a feature extraction network and a fault classification network; the former is used for feature extraction, and the latter for feature dimensionality reduction and fault identification.

[0040] Understandably, the suspension controller is one of the core components of the suspension guidance system. It comprehensively utilizes sensor information to achieve stable control of the suspension electromagnet. Installed in the underbody interlayer of the vehicle, two adjacent suspension controllers jointly control a single overlapping structure, such as... Figure 3 As shown, Figure 3 This is a schematic diagram of the control structure of a suspension controller provided in an embodiment of this application. The suspension controller is interchangeable, distinguishing different suspension points through physical address codes and function codes. During train operation, problems such as gap sensor failure, excessively high ambient temperature, or unstable power supply can cause the suspension controller to fail to properly adjust the gap between the train and the track, thereby affecting the train's suspension stability. Therefore, fault diagnosis of the suspension controller is particularly important.

[0041] Currently, to ensure the normal levitation of the suspension unit even when one or two sensors fail, the suspension control system constructs a state observer based on an overlapping structure. This observer utilizes the entire overlapping structure to obtain signals such as gap, current, voltage, acceleration, velocity, and resistance to estimate the measurement data of the faulty sensor. This redundancy optimization strategy reduces the failure rate of the suspension system and ensures train operation safety. However, while this multi-parameter redundant control method for a single suspension position improves the reliability of the suspension system, the multi-parameter redundancy generates a large amount of data, posing significant challenges to feature extraction and fault diagnosis, increasing the complexity and difficulty of fault diagnosis. Furthermore, traditional technologies mainly use convolutional neural networks for feature extraction and fault diagnosis of the suspension controller. These networks are designed for single-modal data, and when processing fault diagnosis of multiple input parameters of the suspension controller, they rely on stacking model layers to ensure diagnostic accuracy, leading to increased model complexity and a higher risk of overfitting. The complex model structure also affects model performance. Therefore, it is necessary to develop more efficient feature selection and dimensionality reduction algorithms to reduce data dimensionality, retain feature information, and complete the fault diagnosis process.

[0042] Based on this, this application proposes a ResNet (Residual Neural Network) model based on KAN (Kolmogorov-Arnold Networks), using the Kolmogorov-Arnold representation theorem module as the ResNet classifier. The specific implementation process can be referred to in the following two embodiments.

[0043] In one embodiment of this application, the feature extraction network can specifically be a residual neural network, and the residual neural network incorporates a skip connection structure; correspondingly, extracting multi-scale temporal features of target multi-channel time-series data using the feature extraction network of a preset diagnostic model can include: using a residual neural network to extract features from the target multi-channel time-series data to obtain shallow detail features and deep semantic features; and using a skip connection structure to fuse the shallow detail features and deep semantic features to obtain multi-scale temporal features.

[0044] It is understandable that when the depth of a convolutional neural network increases to a certain extent, the training error often increases and the testing performance decreases. This problem is called the degradation problem. The main reasons for degradation are as follows: (1) Gradient vanishing and gradient explosion: In traditional deep networks, the gradient decays or amplifies layer by layer during backpropagation, eventually causing the parameters of the previous network layers to fail to be updated effectively; (2) Overfitting and training difficulties: The increase in depth brings an exponential increase in the number of parameters, making the model prone to overfitting in small datasets or high noise scenarios. At the same time, the optimization space becomes complex and difficult to converge. Based on this, the introduction of residual learning mechanism and skip connections in residual neural networks can effectively solve the gradient vanishing and degradation problems in deep network training.

[0045] Specifically, the key concept behind ResNet is residual mapping. Assume the target mapping function is H(X), where H(X) is the ideal output and X is the input. Traditional convolutional neural networks directly learn H(X), while ResNet assumes the network can more easily learn its residual part: F(X) = H(X) - X, thus H(X) = F(X) + X. In this way, the network only needs to learn the difference F(X) between the input and the target output. If the optimal solution is close to the identity mapping, the residual function can easily converge to zero, significantly reducing the learning difficulty. The advantage of this approach is that even if the residual part cannot learn effective features, the network can still directly output the input through the identity mapping, achieving no performance degradation.

[0046] Furthermore, in actual convolution operations, kernel size, stride, pooling, and other operations may change the dimensionality of the feature map (including spatial resolution and number of channels). In this case, the dimensions of the input X and the residual branch output F(X) are inconsistent, making direct addition impossible. Therefore, ResNet can introduce a linear mapping W. sThis is used to adjust the input dimension. In practice, this mapping can be implemented using a 1×1 convolutional layer to adjust the feature dimension so that the feature dimension of the residual branch is consistent with that of the main branch: y=F(X)+W s (X), thus ensuring the consistency of the input and residual in terms of dimension, making the addition operation reasonable. If the dimensions are naturally the same, then W s This is the identity matrix. The core advantage of the residual structure lies in providing a non-decaying shortcut for the gradient during backpropagation. Assuming the loss function is L, the gradient with respect to the output is: L / y, the input gradient is: L / X= L / y·( F(X) / X+I), where I is the identity matrix, thus it can be seen that even if F(X) / Even when X is small (close to vanishing gradient), the gradient can still be directly passed to the previous layer through the identity I, meaning that the gradient does not completely vanish in deep networks. This derivation reveals the mathematical basis for ResNet's ability to support training networks with hundreds or even thousands of layers.

[0047] From an information theory perspective, in traditional convolutional neural networks, the input signal needs to undergo multiple nonlinear transformations before reaching the output, and these nonlinear operations can lead to gradual information loss. Residual structures, however, allow the input to "directly traverse" several convolutional layers to reach the output through shortcut connections. This preserves low-level features, such as edge and texture information, which directly impacts higher layers, facilitating detailed representation. Higher-level features are superimposed and corrected, with the residual part F(X) acting as a "correction term" based on the input features, providing additional discriminative information. Therefore, ResNet can be viewed as a "parallel transfer" of information between network layers, preserving the integrity of the input while layer-by-layer superposition of residual information. Consider a traditional deep convolutional neural network with the goal of learning a function H(X). Due to gradient decay, shallow layers are almost impossible to update effectively, leading to model degradation. In ResNet, if the residual part learns no information (i.e., F(X) = 0), the output is y = X, equivalent to the network automatically degenerating into an identity mapping, without degrading performance. If the residual part learns effective features, the network output is: y = F(X) + X, at which point the performance is improved. In other words, ResNet is at least no worse than shallow networks, and at most better.

[0048] In one embodiment of this application, the fault classification network can specifically be an Arnold neural network (KAN network). Accordingly, the fault classification network using a pre-defined diagnostic model is used to reduce the dimensionality of multi-scale temporal features, and the fault diagnosis result of the suspension controller is determined based on the reduced dimensionality features. This can include: performing univariate nonlinear transformation on each feature component in the multi-scale temporal features to obtain each univariate nonlinear transformation feature; performing weighted summation and aggregation on each univariate nonlinear transformation feature to obtain intermediate feature variables; performing univariate nonlinear mapping on the intermediate feature variables to obtain nonlinear mapping features; and superimposing the nonlinear mapping features to output the fault diagnosis result.

[0049] As can be understood, KAN's theorem states that any continuous multivariable function can be represented by a finite number of single-variable functions and bivariate addition operations. Specifically, for any continuous function f defined on an n-dimensional unit cube: f: [0,1] n → There exists a set of univariate continuous functions ψ i,j [0,1]→ , i : → , so that: ; The above equation demonstrates that a high-dimensional continuous function f can be accurately expressed by a weighted sum of a finite number of univariate functions without requiring complex modeling in the high-dimensional space. This result provides a theoretical guarantee for degrading high-dimensional problems into low-dimensional combinatorial problems. Based on this, within the deep learning framework, a new classifier module can be constructed using univariate mapping and addition operations to replace the traditional fully connected layer, reducing the parameter scale and improving interpretability.

[0050] Suppose that after extraction by the ResNet backbone network, a feature vector of dimension d is obtained: X=[x1,x2,…,x…]. d ]∈ d Traditional classification heads typically use y = WX + b, where W ∈ C×d Here, b is a learnable bias parameter used to shift the linear mapping result, and C is the number of classes. Its limitations are: the parameter size is C×d, which makes it prone to overfitting when the number of classes or feature dimensions are large; and linear mapping struggles to capture complex high-dimensional nonlinear relationships, requiring a multi-layer structure. However, drawing on the KAN representation theorem, high-dimensional mappings can be decomposed into a multi-level structure of "univariate function transformation + additive combination + univariate nonlinearity," thus designing a KAN classifier module. Based on this, its specific implementation process is as follows: (1) Define a set of functions ψ i (·), acting on each component of the eigenvector: ψ i (x j )=g i (x j ), where g i It can be selected as a linear function, a polynomial function, a radial basis function, or a univariate nonlinear mapping parameterized by a shallow neural network.

[0051] (2) Summing and aggregating the features across all dimensions yields intermediate variables: ; This step essentially involves performing a non-linear transformation on features of different dimensions, followed by a weighted summation, thereby compressing the feature dimensions while preserving information. In other words, if X is a high-dimensional feature, then z... i It can be viewed as its projection onto a certain single-variable function basis.

[0052] (3) Define another set of single-variable functions i (·), acts on intermediate variables: i (z i )=h i (z i ), where h i It can be approximated by simple nonlinear functions (such as ReLU, Sigmoid, Tanh) or small neural networks.

[0053] (4) Finally, the classifier output is: ; Where, y∈ C Each component represents a score or probability for a category.

[0054] Therefore, the original d-dimensional feature X is projected onto 2d+1 scalars z. i This step is similar to linear projection in Principal Component Analysis (PCA), but the difference lies in ψ. i It is a nonlinear univariate function, thus preserving a richer pattern. Mathematically, this is equivalent to expanding on a basis of nonlinear functions: X→[z1,z2,…,z...]. 2d+1 This allows for a high-fidelity representation of the original space even in lower-dimensional spaces. Although each ψ i (x j It is a univariate function, but through the addition operation, indirect coupling is achieved between different features. Then, via... i (z iThe nonlinear mapping of ) yields higher-order combined effects. This process is equivalent to constructing multivariate nonlinear interactions based on univariate variables, thus overcoming the linear limitations of single-layer fully connected systems.

[0055] Understandably, traditional fully connected (FC) layers directly use matrix W to model the linear relationship between features; the KAN classifier, on the other hand, first uses ψ... i Each dimension is transformed into a univariate feature, and then aggregated into z. i Finally, by ψ i It performs nonlinear combinations. The difference between the two is that KAN decomposes complex multidimensional mappings into interpretable combinations of simple functions, avoiding the computational redundancy and optimization difficulties caused by direct learning of large matrices.

[0056] Based on this, assuming the loss function is L, then according to the chain rule: ; ; This indicates that the gradient propagation path in KAN is decomposed into the product and summation of local univariate gradients, avoiding the instability caused by large-scale matrix multiplication, and making training more robust.

[0057] Furthermore, a significant advantage of the KAN classifier is its interpretability. Because each... i and ψ i They are all unary functions, therefore, a single g can be visualized individually. i (x j The curve of z illustrates how the function transforms the input features; the intermediate variable z i This can be viewed as a "feature channel," the magnitude of which reflects the importance of features under a certain set of univariate mappings; the output consists of a finite number of ψ... i (z i The summation of these elements makes it easier to track the contribution of each factor to the final prediction. This differs from traditional black-box fully connected layers, making it easier to provide decision-making support and meeting the interpretability requirements of industrial scenarios.

[0058] Therefore, it is evident that in the complete network, ResNet is responsible for extracting multi-scale features, but its output often still contains a large amount of redundancy. The KAN classifier, as an alternative classification head, can: compress feature dimensions and reduce the number of parameters; enhance feature interactions by capturing inter-dimensional dependencies through nonlinear combinations; maintain stable training and avoid overfitting caused by large matrices in fully connected (FC) layers; and provide interpretability, allowing for the analysis of prediction contributions item by item. This model follows the form of the KAN representation theorem, transforming the high-dimensional classification task into a structured modeling process involving a finite number of univariate function combinations, thus forming the theoretical foundation of the KAN classifier.

[0059] Furthermore, the fault diagnosis method for the suspension controller provided in this application embodiment may further include: building a simulation environment corresponding to the suspension controller in a simulation platform; collecting time-series data of the suspension controller in a fault-free state in the simulation environment to obtain a normal state dataset; collecting time-series data of the suspension controller in a gradually changing fault state in the simulation environment to obtain a gradually changing fault dataset; collecting time-series data of the suspension controller in an intermittent noise fault state in the simulation environment to obtain an intermittent noise fault dataset; and training a preset diagnostic model using the normal state dataset, the gradually changing fault dataset, and the intermittent noise fault dataset.

[0060] This application provides a method for constructing a preset diagnostic model, which enables gradual fault diagnosis and intermittent noise fault diagnosis of a suspension controller. In the implementation process, based on the suspension system dynamics model and control law, a complete closed-loop suspension system simulation environment can be built on the Simulink platform. The simulation uses a fixed-step integrator (step size 0.0002 s), a total duration of 3000 s, and a sampling frequency of 50 Hz. Each data sample contains four channels, corresponding to the controller output voltage, electromagnet current, sensor gap value, and acceleration response, respectively, constituting typical high-dimensional industrial time-series input characteristics. Based on this, the following fault injection operation is performed: (1) After 500s, a gradual fault is injected into the suspension controller, causing the output voltage in the suspension controller to drift by -0.04V per second, eventually drifting to -100V (+91%). That is, a drift occurs between the theoretical output voltage calculated by the suspension controller and the actual output voltage after the chopper, denoted as F. SCD Fault (Float controller gradual fault).

[0061] (2) An intermittent noise fault of the suspension controller is introduced after 500s, with a period of 2s, a noise time of 50%, a phase delay of 1s, and a standard deviation of 10, denoted as F. SCIHF Fault (intermittent noise fault in suspension controller).

[0062] Thus, normal state datasets, gradual fault datasets, and intermittent noise fault datasets can be collected to effectively train the pre-set diagnostic model.

[0063] Furthermore, the fault diagnosis method for the suspension controller provided in this application embodiment may further include: evaluating a preset diagnostic model using a preset three-dimensional evaluation system to obtain an evaluation result; the preset three-dimensional evaluation system includes a precision evaluation index, a recall evaluation index, and an F1 score evaluation index; when the evaluation result indicates that the preset diagnostic model fails the evaluation, the preset diagnostic model is updated.

[0064] Specifically, to systematically evaluate the model's diagnostic capability in gradually changing fault scenarios, this application introduces a comprehensive evaluation system (pre-defined three-dimensional evaluation system) to measure fault classification performance indicators. These performance indicators include precision, recall, and F1 score, which measure the model's overall ability to identify normal and faulty samples. Of course, if the evaluation of the pre-defined diagnostic model fails, it can be continuously updated and corrected until a pre-defined diagnostic model that meets the evaluation requirements is obtained.

[0065] Furthermore, through data analysis and validation of the aforementioned pre-defined diagnostic model, the following conclusions can be drawn: For the normal category, the model's precision is 86.42%, recall is 90.67%, and F1 score is 88.49%; for the gradual fault of the suspension controller (F... SCD For the () category, the model achieved a precision of 92.02%, a recall of 88.43%, and an F1 score of 90.19%; for the intermittent noise of the suspension controller (F SCIHF For the ) category, the model achieved a precision of 98.65%, a recall of 97.85%, and an F1 score of 98.25%.

[0066] Finally, please refer to the following example for verification: Figure 4 , Figure 4 This is a flowchart illustrating another fault diagnosis method for a suspension controller provided in this application embodiment. During the operation of the high-speed maglev train, sensor data such as current, gap, voltage, and acceleration of each suspension unit are collected in real time through the on-board diagnostic network and pre-processed by the on-board controller. The data is transmitted to the ground operation and maintenance center via a dedicated 5G communication link. After time synchronization, noise filtering, and feature extraction of the received multi-source data streams, the operation and maintenance system inputs the processed features into the ResNet-KAN fault diagnosis model of this application. This model can quickly determine the status of the suspension system and identify specific fault types under complex operating conditions. Figure 5 As shown, Figure 5 This is a schematic diagram of a field application example of a suspension controller provided in this application embodiment. The field suspension controller experiences a controller interruption fault at 3623.5s. At this time, the suspension controller cannot output current, and the current drops to 0A. This data is transmitted back to the operation and maintenance system through the vehicle diagnostic network and input into the model for fault diagnosis. The model can detect the suspension controller interruption fault, i.e., the inability to output current, after 0.49s.

[0067] As can be seen, the fault diagnosis method for the suspension controller provided in this application first collects multi-channel time-series data of the suspension controller and determines the system step response parameters of the suspension controller. Then, it inputs both into a preset constraint model to determine the most suitable time-series window length for the suspension controller. This allows for the extraction of target multi-channel time-series data within this window length from the multi-channel time-series data, which is then processed using the preset diagnostic model to achieve fault diagnosis of the suspension controller. Therefore, this technical solution utilizes a preset constraint model to determine the time-series window length, avoiding inaccurate fault diagnosis results caused by windows that are too short or too long. Furthermore, the preset diagnostic model includes a feature extraction network and a fault classification network. The former is used for multi-scale time-series feature extraction, and the latter is used for feature dimensionality reduction and fault classification, effectively reducing computational complexity and improving fault diagnosis efficiency. Therefore, this technical solution can achieve faster, more efficient, and more accurate fault diagnosis of the suspension controller.

[0068] This application provides a fault diagnosis device for a suspension controller.

[0069] Please refer to Figure 6 , Figure 6 This is a schematic diagram of a fault diagnosis device for a suspension controller provided in an embodiment of this application. The fault diagnosis device for the suspension controller may include: Module 1 is used to acquire multi-channel timing data of the suspension controller and determine the system step response parameters of the suspension controller; Processing module 2 is used to process multi-channel time series data and system step response parameters using a preset constraint model to obtain the time series window length; The extraction module 3 is used to extract the target multi-channel time series data corresponding to the time series window length from the multi-channel time series data; The diagnostic module 4 is used to extract multi-scale time-series features of the target multi-channel time-series data using the feature extraction network of the preset diagnostic model, reduce the multi-scale time-series features using the fault classification network of the preset diagnostic model, and determine the fault diagnosis result of the suspension controller based on the reduced features.

[0070] As can be seen, the fault diagnosis device for the suspension controller provided in this application first collects multi-channel time-series data of the suspension controller and determines the system step response parameters of the suspension controller. Then, it inputs both into a preset constraint model to determine the most suitable time-series window length for the suspension controller. This allows for the extraction of target multi-channel time-series data within this window length from the multi-channel time-series data, which is then processed using the preset diagnostic model to achieve fault diagnosis of the suspension controller. Therefore, this technical solution utilizes a preset constraint model to determine the time-series window length, avoiding inaccurate fault diagnosis results caused by windows that are too short or too long. Furthermore, the preset diagnostic model includes a feature extraction network and a fault classification network. The former is used for multi-scale time-series feature extraction, while the latter is used for feature dimensionality reduction and fault classification, effectively reducing computational complexity and improving fault diagnosis efficiency. Therefore, this technical solution can achieve faster, more efficient, and more accurate fault diagnosis of the suspension controller.

[0071] In one embodiment of this application, the preset constraint model includes a system response time constraint and an autocorrelation function decay constraint. Accordingly, the processing module 2 can be specifically used to process the system step response parameters using the system response time constraint in the preset constraint model to obtain the minimum data length for the suspension controller to reach a stable step response time; to process the multi-channel time series data using the autocorrelation function decay constraint in the preset constraint model to obtain the minimum lag step length for the autocorrelation coefficient of the multi-channel time series data to decrease to within a preset threshold; and to use the maximum value between the minimum data length and the minimum lag step length as the time series window length.

[0072] In one embodiment of this application, the feature extraction network is specifically a residual neural network, and the residual neural network incorporates a skip connection structure; correspondingly, the diagnostic module 4 can be specifically used to extract features from the target multi-channel time series data using the residual neural network to obtain shallow detail features and deep semantic features; and to fuse the shallow detail features and deep semantic features using the skip connection structure to obtain multi-scale time series features.

[0073] In one embodiment of this application, the fault classification network is specifically an Arnold neural network; correspondingly, the diagnostic module 4 can be specifically used to perform univariate nonlinear transformation processing on each feature component in the multi-scale time series features to obtain each univariate nonlinear transformation feature; perform weighted summation and aggregation processing on each univariate nonlinear transformation feature to obtain intermediate feature variables; perform univariate nonlinear mapping processing on the intermediate feature variables to obtain nonlinear mapping features; and superimpose the nonlinear mapping features to output the fault diagnosis result.

[0074] In one embodiment of this application, the fault diagnosis device for the suspension controller may further include a preprocessing module for performing preprocessing operations on the multi-channel time-series data after acquiring the multi-channel time-series data of the suspension controller; the preprocessing operations include one or more combinations of time synchronization processing, noise filtering processing, and feature extraction processing.

[0075] In one embodiment of this application, the fault diagnosis device for the suspension controller may further include a training module, used to build a simulation environment corresponding to the suspension controller in a simulation platform; to collect time-series data of the suspension controller in a fault-free state in the simulation environment to obtain a normal state dataset; to collect time-series data of the suspension controller in a gradually changing fault state in the simulation environment to obtain a gradually changing fault dataset; to collect time-series data of the suspension controller in an intermittent noise fault state in the simulation environment to obtain an intermittent noise fault dataset; and to train a preset diagnostic model using the normal state dataset, the gradually changing fault dataset, and the intermittent noise fault dataset.

[0076] In one embodiment of this application, the fault diagnosis device of the suspension controller may further include an update module, which is used to evaluate the preset diagnostic model using a preset three-dimensional evaluation system to obtain an evaluation result; the preset three-dimensional evaluation system includes precision evaluation index, recall evaluation index, and F1 score evaluation index; when the evaluation result is that the preset diagnostic model fails the evaluation, the preset diagnostic model is updated.

[0077] For a description of the apparatus provided in the embodiments of this application, please refer to the above method embodiments; further details will not be repeated here.

[0078] This application provides an electronic device.

[0079] Please refer to Figure 7 , Figure 7 This application provides a schematic diagram of the structure of an electronic device, which may include: Memory 11 is used to store computer programs; The processor 10 is configured to execute a computer program that implements the steps of any of the above-described methods for diagnosing faults in a suspension controller.

[0080] like Figure 7 The diagram shows the structural composition of an electronic device, which may include a processor 10, a memory 11, a communication interface 12, and a communication bus 13. The processor 10, memory 11, and communication interface 12 all communicate with each other through the communication bus 13.

[0081] In this embodiment, the processor 10 may be a central processing unit (CPU), an application-specific integrated circuit, a digital signal processor, a field-programmable gate array, or other programmable logic devices.

[0082] The processor 10 can call the program stored in the memory 11. Specifically, the processor 10 can execute the operations in the embodiment of the fault diagnosis method for the suspension controller.

[0083] The memory 11 is used to store one or more programs. The programs may include program code, which includes computer operation instructions. In this embodiment, the memory 11 stores at least a program for implementing the following functions: Acquire multi-channel time-series data of the suspension controller and determine the system step response parameters of the suspension controller; process the multi-channel time-series data and system step response parameters using a preset constraint model to obtain the time series window length; extract the target multi-channel time-series data corresponding to the time series window length from the multi-channel time-series data; extract multi-scale time-series features of the target multi-channel time-series data using the feature extraction network of the preset diagnostic model; perform dimensionality reduction processing of the multi-scale time-series features using the fault classification network of the preset diagnostic model; and determine the fault diagnosis result of the suspension controller based on the dimensionality reduction features.

[0084] In one possible implementation, the memory 11 may include a program storage area and a data storage area, wherein the program storage area may store the operating system and applications required for at least one function; and the data storage area may store data created during use.

[0085] In addition, memory 11 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device or other volatile solid-state storage device.

[0086] Communication interface 12 can be an interface for the communication module, used to connect with other devices or systems.

[0087] Of course, it should be noted that, Figure 7 The structure shown does not constitute a limitation on the electronic device in the embodiments of this application. In practical applications, the electronic device may include more than Figure 7 More or fewer components as shown, or combinations of certain components.

[0088] This application provides a computer-readable storage medium.

[0089] The computer-readable storage medium provided in this application embodiment stores a computer program, which, when executed by a processor, can implement the steps of any of the above-described fault diagnosis methods for a suspension controller.

[0090] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0091] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.

[0092] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0093] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0094] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0095] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and core ideas of this application. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A fault diagnosis method for a suspension controller, characterized in that, include: Acquire multi-channel timing data of the suspension controller and determine the system step response parameters of the suspension controller; The multi-channel time-series data and the system step response parameters are processed using a preset constraint model to obtain the time-series window length; Extract the target multi-channel time series data corresponding to the time series window length from the multi-channel time series data; Multi-scale time-series features of the target multi-channel time-series data are extracted using the feature extraction network of the preset diagnostic model. The multi-scale time-series features are then dimensionality-reduced using the fault classification network of the preset diagnostic model. Based on the dimensionality-reduced features, the fault diagnosis result of the suspension controller is determined.

2. The fault diagnosis method for the suspension controller according to claim 1, characterized in that, The preset constraint model includes system response time constraints and autocorrelation function decay constraints; Accordingly, the multi-channel time-series data and the system step response parameters are processed using a preset constraint model to obtain the time-series window length, including: The system step response parameters are processed using the system response time constraint in the preset constraint model to obtain the minimum data length for the suspension controller to reach step response stability. The multi-channel time series data is processed using the autocorrelation function decay constraint in the preset constraint model to obtain the minimum hysteresis step size at which the autocorrelation coefficient of the multi-channel time series data decreases to within a preset threshold. The maximum value between the minimum data length and the minimum lag step size is used as the time series window length.

3. The fault diagnosis method for the suspension controller according to claim 1, characterized in that, The feature extraction network is specifically a residual neural network, and the residual neural network incorporates a skip connection structure; Accordingly, the feature extraction network of the pre-defined diagnostic model is used to extract multi-scale temporal features of the target multi-channel time-series data, including: The residual neural network is used to extract features from the target multi-channel time-series data to obtain shallow detail features and deep semantic features. The shallow detail features and the deep semantic features are fused using the skip connection structure to obtain the multi-scale temporal features.

4. The fault diagnosis method for the suspension controller according to claim 1, characterized in that, The fault classification network is specifically an Arnold neural network; Accordingly, the multi-scale temporal features are processed by dimensionality reduction using the fault classification network of the preset diagnostic model, and the fault diagnosis result of the suspension controller is determined based on the dimensionality reduction features, including: Each feature component in the multi-scale time series features is subjected to univariate nonlinear transformation processing to obtain each univariate nonlinear transformation feature; The weighted summation and aggregation of each of the univariate nonlinear transformation features are used to obtain intermediate feature variables; The intermediate feature variables are subjected to univariate nonlinear mapping processing to obtain nonlinear mapping features; The nonlinear mapping features are superimposed and output to obtain the fault diagnosis result.

5. The fault diagnosis method for the suspension controller according to claim 1, characterized in that, After acquiring the multi-channel timing data of the suspension controller, the following steps are also included: Preprocessing operations are performed on the multi-channel time-series data; the preprocessing operations include one or more combinations of time synchronization processing, noise filtering processing, and feature extraction processing.

6. The fault diagnosis method for the suspension controller according to any one of claims 1 to 5, characterized in that, Also includes: Build the simulation environment corresponding to the suspension controller in the simulation platform; Under the fault-free injection state of the simulation environment, the timing data of the suspension controller is collected to obtain the normal state dataset; When a gradual fault state is injected into the simulation environment, the timing data of the suspension controller is collected to obtain a gradual fault dataset. When intermittent noise fault conditions are injected into the simulation environment, the timing data of the suspension controller is collected to obtain the intermittent noise fault dataset; The preset diagnostic model is obtained by training the normal state dataset, the gradual fault dataset, and the intermittent noise fault dataset.

7. The fault diagnosis method for the suspension controller according to claim 6, characterized in that, Also includes: The preset diagnostic model is evaluated using a preset three-dimensional evaluation system to obtain the evaluation results; The preset three-dimensional evaluation system includes precision evaluation index, recall evaluation index, and F1 score evaluation index; When the evaluation result indicates that the preset diagnostic model fails the evaluation, the preset diagnostic model is updated.

8. A fault diagnosis device for a suspension controller, characterized in that, include: The acquisition module is used to acquire multi-channel timing data of the suspension controller and determine the system step response parameters of the suspension controller; The processing module is used to process the multi-channel time series data and the system step response parameters using a preset constraint model to obtain the time series window length; The extraction module is used to extract target multi-channel time series data corresponding to the time series window length from the multi-channel time series data; The diagnostic module is used to extract multi-scale time-series features of the target multi-channel time-series data using the feature extraction network of the preset diagnostic model, to perform dimensionality reduction processing of the multi-scale time-series features using the fault classification network of the preset diagnostic model, and to determine the fault diagnosis result of the suspension controller based on the dimensionality reduction features.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the fault diagnosis method for the suspension controller as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the fault diagnosis method for the suspension controller as described in any one of claims 1 to 7.

Images