Railway vehicle running gear fault diagnosis method and device

Abstract

A method and apparatus for diagnosing faults in the running gear of a rail vehicle, belonging to the field of rail transit vehicle fault diagnosis technology. The diagnostic method includes: acquiring monitoring signals from the running gear of the rail vehicle under test and determining preprocessed signals; constructing a reinforcement learning environment and building an agent; constructing a comprehensive reward function and training the agent using a dual-delay deep deterministic policy gradient algorithm to obtain the optimal policy output by the trained agent and the main network of the trained agent; updating the coefficients of the finite impulse response filter using the optimal policy, and performing Hilbert transform and envelope spectrum analysis on the preprocessed signals based on the updated finite impulse response filter to identify faults in the running gear of the rail vehicle under test. This invention can accurately identify the fault type and location of the running gear of a rail vehicle, improving the accuracy and reliability of fault diagnosis and solving the problem of low accuracy in identifying faults in the running gear of rail vehicles.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology for rail transit vehicles, and specifically relates to a method for diagnosing faults in the running gear of rail vehicles. This invention also relates to a device for diagnosing faults in the running gear of rail vehicles. Background Technology

[0002] The running gear is the core subsystem of a rail vehicle. Key components such as wheelsets, axle box bearings, and gear transmission systems in the running gear are subjected to harsh conditions of high speed, heavy load, and strong vibration for a long time, which makes them prone to failure and threatens driving safety. Fault diagnosis and health management are the focus of research in the field of rail transit.

[0003] Traditional fault diagnosis relies on signal processing methods such as spectral kurtosis, adaptive decomposition, wavelet transform, and sparse representation. These methods require manual setting of filtering parameters based on experience, resulting in poor robustness and insufficient generalization ability under strong noise and complex operating conditions, making it difficult to accurately extract fault pulse features.

[0004] In recent years, reinforcement learning has been gradually applied to signal enhancement and parameter optimization tasks. Among them, the Deep Q-Network (DQN) algorithm is only suitable for discrete action spaces and cannot handle high-dimensional continuous optimization problems such as finite impulse response (FIR) filter coefficients. It also suffers from Q-value overestimation and training oscillation defects. Although the Deep Deterministic Policy Gradient (DDPG) algorithm is suitable for continuous action spaces, it still suffers from Q-value overestimation and training instability caused by synchronous updates of the Actor network and Critic network.

[0005] Existing reinforcement learning research related to fault diagnosis of running gears often focuses on fault classification, neglecting the core link of envelope demodulation. Furthermore, it fails to integrate multiple indicators such as time domain, envelope spectrum sparsity, and peak signal-to-noise ratio for joint optimization, making it impossible to comprehensively and objectively evaluate the filtering enhancement effect and meet the requirements for high robustness fault feature extraction in complex environments.

[0006] Relevant patent document: CN117370837A discloses a train running gear fault diagnosis method based on scale-space guided fast spectrum correlation, including: collecting high-speed train axle box vibration data of wheelset damage and bearing failure at different speeds, thereby obtaining an experimental dataset containing different fault modes; processing the collected data through a fast spectrum correlation algorithm to obtain a fast spectrum coherence function; obtaining a guide curve by standardizing the spectrum coherence function using the L2 / L1 norm; dividing the guide curve into frequency bands using a scale-space method; applying fault quantization indicators to filter each divided frequency band, selecting the frequency band containing the most wheelset and axle box bearing fault information for enhanced envelope, obtaining an enhanced envelope spectrum, and realizing accurate identification of wheelset and axle box bearing faults.

[0007] The above technologies cannot accurately identify the type and location of faults in the running gear of rail vehicles. Summary of the Invention

[0008] The purpose of this invention is to provide a fault diagnosis method for the running gear of rail vehicles, which can accurately identify the fault type and location of the running gear of rail vehicles, improve the accuracy and reliability of fault diagnosis, and solve the problem of low accuracy in fault identification of the running gear of rail vehicles in the prior art.

[0009] Therefore, another objective of the present invention is to provide a fault diagnosis device for the running gear of a rail vehicle.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for diagnosing faults in the running gear of a rail vehicle, comprising the following steps: S1: Acquire the monitoring signals of the running gear of the track vehicle under test, and determine the preprocessing signals; S2: Construct a reinforcement learning environment and build an intelligent agent. The reinforcement learning environment uses the coefficients of a finite impulse response filter as the state space and the increments of the coefficients of the finite impulse response filter as the action space. The intelligent agent includes a main network and a target network. S3: Construct a comprehensive reward function and train the agent using a dual-delay deep deterministic policy gradient algorithm to obtain the optimal policy output by the trained agent and the main network of the trained agent. S4: Update the coefficients of the finite impulse response filter using the optimal strategy, and perform Hilbert transform and envelope spectrum analysis on the preprocessed signal based on the updated finite impulse response filter to identify the faults in the running gear of the track vehicle under test.

[0011] In the above technical solution, the preferred technical solution may be that step S1 specifically includes: S1.1: Acquire monitoring signals collected by the accelerometer during the operation of the running gear of the track vehicle under test, wherein the monitoring signals include at least one of vibration signals, noise signals, and acoustic emission signals.

[0012] S1.2: Perform DC component removal preprocessing on the monitoring signal to obtain the preprocessed signal.

[0013] In the above technical solution, a preferred technical solution may also be that, in step S2, the method for constructing the reinforcement learning environment specifically includes: S2.1: Define the basic environmental parameters, which include the finite impulse response filter length, action scaling factor, and maximum number of interaction steps in the environment; S2.2: Define the state space using the finite impulse response filter coefficients, and generate initial filter coefficients; S2.3: Define the action space by the increment of the coefficients of the finite impulse response filter; S2.4: Calculate the preprocessed signal, wherein the characteristic indicators of the preprocessed signal include kurtosis, spectral negative entropy, squared envelope Gini index, and peak signal-to-noise ratio.

[0014] In step 2.4, the method for calculating the preprocessed signal includes: S2.41: Calculate the kurtosis of the preprocessed signal using a first formula, wherein the first formula is: (1), in, The kurtosis of the preprocessed signal, The preprocessed signal; S2.42: Calculate the spectral negative entropy of the preprocessed signal using the second formula, which is: (2), in, The spectral negative entropy of the preprocessed signal, The energy entropy of the preprocessed signal. for The envelope spectrum amplitude of the preprocessed signal at the specified scale. It is the natural logarithm function. This is the full-scale statistical average; S2.43: Calculate the squared envelope Gini index of the preprocessed signal using the third formula, which is: (3), in, The squared envelope Gini index of the preprocessed signal. The total number of sampling points of the preprocessed signal. It is a sliding window. For discrete sampling point index, For the first Spectral energy at each sampling point for sequential Norm; S2.44: Calculate the peak signal-to-noise ratio of the preprocessed signal using the fourth formula, which is: (4), in, The peak signal-to-noise ratio of the preprocessed signal. The energy of the fault characteristic signal, This represents the energy of the noise signal.

[0015] In the above technical solution, a preferred embodiment may also be that, in step S2, the main network and the target network have the same structure, and both the main network and the target network consist of one Actor main network and two independent Critic networks; in step S3, the agent's main network is the agent's Actor main network. In step S3, the method for constructing the comprehensive reward function includes: S3.11: Input the current finite impulse response filter coefficients as the current state into the Actor main network, and output the current action; S3.12: Input the current action into the reinforcement learning environment, and update the current finite impulse response filter coefficients in combination with the action scaling coefficients; S3.13: Perform linear convolution filtering on the preprocessed signal using the updated finite impulse response filter coefficients to obtain the filtered signal; S3.14: Calculate the characteristic parameters of the filtered signal; S3.15: Calculate the relative improvement rate of the feature index of each filtered signal and the feature index of the corresponding preprocessed signal, and perform exponential moving average normalization on each relative improvement rate to obtain four normalized relative improvement rates. S3.16: The four relative improvement rates after normalization are weighted and summed to obtain the comprehensive reward function.

[0016] In the above technical solution, a preferred technical solution may also be that, in step S4, the method of updating the coefficients of the finite impulse response filter using the optimal strategy includes: The initial finite impulse response filter coefficients corresponding to the preprocessed signal are input into the Actor main network of the agent, and the increments of the corresponding finite impulse response filter coefficients are output. The initial finite impulse response (FIR) filter coefficients are iteratively updated using the increments of the FIR filter coefficients to obtain the optimal FIR filter coefficients.

[0017] In step S4, the method for identifying faults in the running gear of the track vehicle under test by performing Hilbert transform and envelope spectrum analysis on the preprocessed signal based on the updated finite impulse response filter includes: S4.1: Filter the preprocessed signal using the updated optimal finite impulse response filter to obtain the filtered output signal; S4.2: Perform a Hilbert transform on the filtered output signal to construct an analytic signal; S4.3: Perform a modulo operation on the analytical signal to extract the instantaneous envelope of the filtered output signal; S4.4: Perform a fast Fourier transform on the instantaneous envelope of the filtered output signal to obtain the envelope spectrum. Based on the comparison between the envelope spectrum and the theoretical fault characteristic frequencies of each key component of the running gear of the track vehicle under test, the faults of the running gear of the track vehicle under test can be identified.

[0018] A fault diagnosis device for the running gear of a rail vehicle, which uses the aforementioned fault diagnosis method for the running gear of a rail vehicle, the fault diagnosis device for the running gear of a rail vehicle includes: The acquisition module is used to acquire monitoring signals from the running gear of the track vehicle under test and determine the preprocessing signals.

[0019] A construction module is used to build a reinforcement learning environment and construct an agent. The reinforcement learning environment uses the coefficients of a finite impulse response filter as the state space and the increments of the coefficients of the finite impulse response filter as the action space. The agent contains a main network and a target network with the same structure, and both the main network and the target network are composed of one Actor main network and two independent Critic networks.

[0020] The training module is used to construct a comprehensive reward function and train the agent using a dual-delay deep deterministic policy gradient algorithm to obtain the optimal policy output by the trained agent and the Actor main network of the trained agent.

[0021] The identification module is used to update the coefficients of the finite impulse response filter using the optimal strategy, and based on the updated finite impulse response filter, to perform Hilbert transform and envelope spectrum analysis on the preprocessed signal, thereby identifying the faults in the running gear of the track vehicle under test.

[0022] The present invention provides a terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the fault diagnosis method for the running gear of a rail vehicle.

[0023] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the fault diagnosis method for the running gear of a rail vehicle.

[0024] This invention provides a method and apparatus for diagnosing faults in the running gear of a rail vehicle. The method involves acquiring monitoring signals from the running gear of the rail vehicle under test and determining preprocessed signals; constructing a reinforcement learning environment and building an agent; using finite impulse response (FIR) filter coefficients as the state space and the increments of the FIR filter coefficients as the action space; and including a main network and a target network (preferably with identical structures, each consisting of one Actor main network and two independent Critic networks); constructing a comprehensive reward function and training the agent using a dual-delay deep deterministic policy gradient algorithm; updating the FIR filter coefficients using the optimal policy output by the trained agent; and performing Hilbert transform and envelope spectrum analysis on the preprocessed signals based on the updated FIR filter to identify faults in the running gear of the rail vehicle under test. This invention updates the coefficients of a finite impulse response (FIR) filter using the optimal strategy output by a trained agent. Compared to traditional fixed filter coefficient settings, this invention dynamically adjusts filter parameters according to actual conditions, allowing the filter to better adapt to signal characteristics under different operating conditions. This more effectively filters out noise, highlights fault characteristic signals, and improves the accuracy and relevance of signal processing. Furthermore, by designing a reinforcement learning environment with the FIR filter coefficients and their increments as the state and action spaces respectively, employing an agent structure containing a master network and a target network, and using a dual-delay deep deterministic policy gradient algorithm to train the agent, the algorithm's stability, convergence, sample efficiency, and agent learning ability are improved. Simultaneously, by performing Hilbert transform and envelope spectrum analysis on the preprocessed signal based on the updated FIR filter, fault characteristic frequency components in the signal can be extracted more effectively, thereby more accurately identifying the fault type and location of the track vehicle's running gear and improving the accuracy and reliability of fault diagnosis.

[0025] In summary, the present invention provides a method and apparatus for diagnosing faults in the running gear of rail vehicles. It can accurately identify the fault type and location of the running gear of rail vehicles, improve the accuracy and reliability of fault diagnosis, and solve the problem of low accuracy in fault identification of the running gear of rail vehicles in the prior art. Attached Figure Description

[0026] Figure 1 This is a flowchart (block diagram, Example 1) of the fault diagnosis method for the running gear of a rail vehicle according to the present invention.

[0027] Figure 2 This is a schematic diagram of the interaction between the reinforcement learning environment and the intelligent agent provided in Embodiment 1 of the present invention.

[0028] Figure 3 This is a schematic diagram of the intelligent agent structure provided in Embodiment 1 of the present invention.

[0029] Figure 4 The flowchart of agent training and interaction provided for Embodiment 1 of the present invention.

[0030] Figure 5 This is a schematic diagram illustrating the iterative changes of the reward function provided in Embodiment 1 of the present invention.

[0031] Figure 6 This is a schematic diagram of the time-domain waveform of the simulated signal under the conventional method provided in Embodiment 1 of the present invention.

[0032] Figure 7 This is a schematic diagram of the envelope spectrum of the simulated signal under the conventional method provided in Embodiment 1 of the present invention.

[0033] Figure 8 This is a schematic diagram of the time-domain waveform of the simulation signal provided in Embodiment 1 of the present invention.

[0034] Figure 9 This is a schematic diagram of the envelope spectrum of the simulated signal provided in Embodiment 1 of the present invention.

[0035] Figure 10 This is a schematic diagram of the structure of the fault diagnosis device for the running gear of a rail vehicle provided in Embodiment 2 of the present invention.

[0036] Figure 11 This is a schematic diagram of a terminal provided in Embodiment 3 of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of this invention.

[0038] Example 1: As Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8 , Figure 9 As shown, the fault diagnosis method for the running gear of a rail vehicle of the present invention includes the following steps: S1: Acquire monitoring signals from the running gear of the track vehicle under test and determine the preprocessing signal. In this embodiment, the monitoring signals of the running gear of the track vehicle under test under actual operating conditions are acquired, and then the monitoring signals are preprocessed to obtain preprocessed signals for subsequent filtering optimization and fault feature extraction, providing standard input data for the reinforcement learning agent. Step S1 specifically includes: S1.1: Acquire monitoring signals collected by the accelerometer during the operation of the running gear of the track vehicle under test, wherein the monitoring signals include at least one of vibration signals, noise signals, and acoustic emission signals; in this embodiment, the monitoring signals include vibration signals and noise signals.

[0039] S1.2: Perform DC component removal preprocessing on the monitoring signal to obtain the preprocessed signal.

[0040] Specifically, this can be achieved by deploying accelerometers at key load-bearing and transmission components of the track vehicle under test, real-time monitoring signals of the vehicle operating at different speeds, loads, and track conditions. These monitoring signals include, but are not limited to, vibration, noise, and acoustic emission signals, comprehensively covering the dynamic response information of core components such as wheelsets, bearings, and gears. Then, the acquired monitoring signals undergo DC component removal preprocessing to eliminate static DC offsets introduced by sensor zero-point drift, power supply bias, etc., and to eliminate interference from invalid constant components on subsequent filtering optimization and fault feature extraction, resulting in a preprocessed signal with a stable baseline, no static offset, and suitable for direct input into a reinforcement learning environment.

[0041] For example, the monitoring signal is a vibration signal. The monitoring signal generated by the running gear of the track vehicle under test during operation is collected by an accelerometer and input into the system. Taking a simulated signal as an example, the sampling frequency is 20kHz and the signal length is 20,000 points. The DC component is removed from the monitoring signal to obtain the preprocessed signal corresponding to the vibration signal. This serves as the input to the reinforcement learning environment. Among them, the preprocessed signal corresponds to the vibration signal. It consists of four parts, namely: (5), in, The number of fault pulses. The fault pulse number, , , Both represent pulse amplitude. The resonant frequency, For attenuation parameters, The fault characteristic frequency, To simulate the random slippage effect of the rolling element, its value is randomly selected within the range of 1%-2% of the fault pulse period; The rotation frequency of the harmonic components. The phase of the harmonic component. For a moment, The noise follows a Gaussian distribution with a mean of 0 and a standard deviation of 0.3. The parameters of the simulated signal are shown in Table 1.

[0042] Table 1 Parameters of the simulated signal

[0043] This embodiment acquires monitoring signals from the running gear of the track vehicle under test and determines the pre-processed signals, effectively removing noise, interference, and other adverse factors from the original signals, thus improving signal purity and usability. This lays the foundation for accurate subsequent fault diagnosis and avoids misdiagnosis or missed diagnosis due to signal quality issues.

[0044] S2: Construct a reinforcement learning environment and build an agent. The reinforcement learning environment uses the coefficients of a finite impulse response filter as the state space and the increments of the coefficients as the action space. The agent includes a main network and a target network. The main network and the target network have the same structure, and both the main network and the target network consist of one Actor main network and two independent Critic networks.

[0045] In this embodiment, refer to Figure 2 and Figure 3 A reinforcement learning environment for enhancing the fault characteristics of the running gear of the track vehicle under test is constructed, and the coefficients of the finite impulse response filter are used. Defined as the state space of a reinforcement learning environment The increment of the coefficients of the finite impulse response filter Defined as the action space of a reinforcement learning environment This forms an interactive system centered on the optimization of finite impulse response (FIR) filter coefficients. Simultaneously, an agent based on a dual-delay deep deterministic policy gradient algorithm is constructed. This agent comprises a main network and a target network with identical structures, both employing a combination of one Actor main network and two independent Critic networks. The Actor main network generates continuous actions to adjust the FIR filter coefficients, the two independent Critic networks evaluate action value and suppress Q-value overestimation, and the target network stabilizes the training process and avoids network update oscillations.

[0046] In step S2, the method for constructing the reinforcement learning environment specifically includes: S2.1: Define the basic environmental parameters, which include the finite impulse response filter length, action scaling factor, and maximum number of interaction steps in the environment.

[0047] In step S2.1, the basic environmental parameters (of the reinforcement learning environment) are defined, namely, the length of the finite impulse response filter is set. Action scaling factor Maximum number of interaction steps in the environment .

[0048] S2.2: Define the state space using the finite impulse response filter coefficients and generate the initial filter coefficients.

[0049] In step S2.2, the coefficients of the finite impulse response filter are used. Define state space Generate initial filter coefficients and conduct Normalization, that is: (6), in, for Norm.

[0050] In this embodiment, the output of the state space is used to form the initial state s0 of the agent, which provides the initial basis for subsequent adaptive filtering optimization.

[0051] S2.3: Define the action space by the increment of the coefficients of the finite impulse response filter.

[0052] In step S2.3, the increment of the finite impulse response filter coefficients is used. Define action space This allows the intelligent agent to achieve refined and adaptive adjustment of the filter coefficients through continuous output actions.

[0053] S2.4: Calculate the preprocessed signal, wherein the characteristic indicators of the preprocessed signal include kurtosis, spectral negative entropy, squared envelope Gini index, and peak signal-to-noise ratio.

[0054] In step S2.4, the preprocessed signal is... Multi-dimensional feature index calculations are performed to comprehensively quantify the fault characteristic intensity, noise suppression level, and signal quality of the preprocessed signal, providing a quantitative basis for the subsequent comprehensive reward function construction. Step 2.4 involves calculating the preprocessed signal using the following methods: S2.41: Calculate the kurtosis of the preprocessed signal using a first formula, wherein the first formula is: (1), in, The kurtosis of the preprocessed signal, This refers to the preprocessed signal.

[0055] S2.42: Calculate the spectral negative entropy of the preprocessed signal using the second formula, which is: (2), in, For the preprocessed signal Spectral negative entropy, The energy entropy of the preprocessed signal. for The envelope spectrum amplitude of the preprocessed signal at the specified scale. It is the natural logarithm function. This is the full-scale statistical average.

[0056] S2.43: Calculate the squared envelope Gini index of the preprocessed signal using the third formula, which is: (3), in, For the preprocessed signal The squared envelope Gini index, The total number of sampling points of the preprocessed signal. It is a sliding window. For discrete sampling point index, For the first Spectral energy at each sampling point for sequential Norm.

[0057] S2.44: Calculate the peak signal-to-noise ratio of the preprocessed signal using the fourth formula, which is: (4), in, For the preprocessed signal Peak signal-to-noise ratio, The energy of the fault characteristic signal, This represents the energy of the noise signal.

[0058] S2.5: In one implementation, building an intelligent agent may include: Construct the Actor main network and output the increments of the finite impulse response filter coefficients; Construct a dual-crit master network to evaluate the value of actions, wherein the dual-crit master network includes a first crit master network and a second crit master network; Construct a target network with the same structure as the main network. The target network includes an Actor target network, a first Critic target network, and a second Critic target network. Store the interaction data through an experience replay pool. The main network parameters and target network parameters are synchronized by soft update. Truncation Gaussian noise is added to the target action. The main Actor network and the target Actor network are updated with a delay every two steps. Ornstein–Uhlenbeck noise is used to add exploratory perturbations to the actions, thus completing the construction of the agent.

[0059] Furthermore, the process of constructing an intelligent agent is as follows: S2.51: Constructing the Actor main network: The current finite impulse response filter coefficient vector As input, the output corresponds to a deterministic continuous action of the filter length. (i.e., the increment of the coefficients of the finite impulse response filter).

[0060] The network structure is a four-layer fully connected layer. The activation function used is Leaky ReLU, and the output layer is multiplied by an action scaling factor after being Tanh-limited. The optimization objective is to maximize the Q-value evaluated by the Critic, i.e., minimize the expected negative Q-value. The objective function is: (7), in, Let be the loss function of the Actor main network. As expected, This is the current state. The Q value, for The strategy of the Actor main network under the given state.

[0061] S2.52: Constructing a dual-Critic master network: using state vectors With action vectors The concatenation result is the input, and the output is the action evaluation Q-value assessment. .

[0062] The network structure is a four-layer fully connected network. Two independent Critic main networks are used, and the smaller output value is selected to suppress overestimation. The optimization objective is to minimize the mean square error of the time difference. The objective function is: (8), in, Let be the loss function of the two Critic main networks. The target value is the target Q value.

[0063] Among them, target value The calculation formula is: (9), in, , These are the predicted Q-values ​​of the two Critic main networks, respectively. For immediate return, As a discount factor, For the next action, This is the next state.

[0064] S2.53: Construct the target network: To improve training stability, this embodiment builds an Actor target network and a dual Critic target network, collectively referred to as the target network.

[0065] The target network has the same structure, input, output, and loss function as the main network, and the initial parameters are the same as those of the main network.

[0066] In this embodiment, the target network is only used to calculate the stable target Q value and does not participate in gradient backpropagation.

[0067] S2.54: Experience Replay Mechanism Create an experience replay buffer with a capacity of 100,000 to store experience tuples generated by the agent's interactions with the environment. Randomly sampled batches of data during training breaks down sample correlation and improves network convergence stability.

[0068] S2.55: Target network parameter soft update: A soft update strategy is adopted to slowly synchronize the parameters of the master network to the target network, that is: (10) in, For the target network parameters, Main network parameters, To update the step size, this embodiment takes... Updates are performed on the Actor target network, the first Critic target network, and the second Critic target network, respectively.

[0069] S2.56: Policy Noise and Delayed Updates Truncated Gaussian noise is added during the target policy generation phase, and delayed updates are performed. The noise distribution is as follows: The cutoff range is The update rule is to update the Actor main network and the Actor target network every two steps to reduce training oscillations.

[0070] S2.57: Exploratory noise (i.e., the Ornstein-Uhlenbeck process): The Ornstein-Uhlenbeck process is used to add continuous exploration noise to the action to balance exploration and convergence, specifically as follows: Motion noise is generated using the Ornstein-Uhlenbeck process: (11), in, For the first Step motion noise value, For the first Step motion noise value, This is the regression coefficient (i.e., the attenuation coefficient). This represents the long-term mean of the noise. The standard deviation of the noise. It is a standard normally distributed random variable.

[0071] Among them, the initial standard deviation The initial value is 0.1, gradually decreasing to 0.05 to balance exploration and convergence, i.e.: (12) in, For the first The noise standard deviation of the step, To train the number of steps.

[0072] S3: Construct a comprehensive reward function and train the agent using a dual-delay deep deterministic policy gradient algorithm to obtain the optimal policy output by the trained agent and the main network of the trained agent.

[0073] In this embodiment, based on the impulsiveness, cyclic stationarity, sparsity, and signal-to-noise ratio improvement effect of the preprocessed signal after passing through a finite impulse response filter, four feature indicators—kurtosis, spectral negative entropy, squared envelope Gini index, and peak signal-to-noise ratio—are fused to construct a weighted fusion comprehensive reward function, fully quantifying the fault feature enhancement effect of the filtered signal. Then, a dual-delay deep deterministic policy gradient algorithm is used to drive the interaction between the agent and the reinforcement learning environment. Through mechanisms such as experience replay, action noise exploration, policy delayed update, and target network soft update, the agent is iteratively trained with the goal of maximizing the comprehensive reward, enabling the agent to autonomously learn the optimal filtering strategy adapted to complex working conditions and high-noise environments.

[0074] In step S3, the method for constructing the comprehensive reward function includes: S3.11: Input the current finite impulse response filter coefficients as the current state into the Actor main network, and output the current action; S3.12: Input the current action into the reinforcement learning environment, and update the current finite impulse response filter coefficients in combination with the action scaling coefficients; S3.13: Perform linear convolution filtering on the preprocessed signal using the updated finite impulse response filter coefficients to obtain the filtered signal; S3.14: Calculate the characteristic parameters of the filtered signal; S3.15: Calculate the relative improvement rate of the feature index of each filtered signal and the feature index of the corresponding preprocessed signal, and perform exponential moving average normalization on each relative improvement rate to obtain four normalized relative improvement rates. S3.16: The four relative improvement rates after normalization are weighted and summed to obtain the comprehensive reward function.

[0075] Furthermore, the current finite impulse response filter coefficients obtained through iterative updates in the reinforcement learning environment are... As the current state (i.e., environmental input), it is used for subsequent action execution and coefficient updates, providing the current filtering parameters for signal filtering. These coefficients are then input into the Actor main network, where the current action output by the agent is weighted using action scaling factors and superimposed onto the current finite impulse response filter coefficients to complete the iterative update. The calculation formula is as follows: (13) in, These are the updated finite impulse response filter coefficients. These are the coefficients of the current finite impulse response filter. The current action output by the agent.

[0076] Among them, the current action The calculation formula is: (14) in, The scaling factor is taken as the motion scaling factor in this embodiment. ; This is a truncation function.

[0077] After the update, the coefficients are subjected to amplitude constraints to ensure the stability of the filter.

[0078] Then, the updated finite impulse response filter coefficients are used. For preprocessed signals Performing linear convolution filtering yields the filtered signal, i.e.: (15) in, This is the filtered signal.

[0079] The filtering process in this embodiment achieves noise suppression and fault impact feature enhancement, providing a high-quality signal for subsequent feature calculation.

[0080] For filtered signals Calculate the corresponding characteristic indices, i.e., the kurtosis of the filtered signal. Spectral negative entropy of the filtered signal The square envelope Gini index of the filtered signal Peak signal-to-noise ratio of the filtered signal .

[0081] Among them, the kurtosis of the filtered signal The spectral negative entropy of the filtered signal reflects the intensity of the fault impact. The squared envelope Gini index of the filtered signal reflects the degree of spectral concentration. Reflecting impulse sparsity, the peak signal-to-noise ratio of the filtered signal It reflects the noise suppression effect.

[0082] Then, using the characteristic indices of the preprocessed signal as a benchmark, the relative enhancement rate of the filtered signal is calculated to quantify the feature enhancement effect brought about by filtering. Take kurtosis as an example: (16) in, The relative increase rate of kurtosis.

[0083] The four relative improvement rates were normalized using an exponential moving average, specifically as follows: relative improvement rate for each characteristic indicator Their mean and standard deviation are respectively and The calculation formulas are as follows: (17) (18) in, The learning rate can be set to 0.1.

[0084] The formula for normalization is: (19) in, This is the standardized value of the relative improvement rate for each feature indicator.

[0085] Then, a weighted summation operation is performed on the four normalized relative improvement rates to obtain the comprehensive reward function. The calculation formula is: (20), in, The relative increase rate of kurtosis after normalization. The relative increase rate of the normalized spectral negative entropy. The relative increase rate of the normalized squared envelope Gini index. This represents the relative improvement rate of the peak signal-to-noise ratio after normalization. , , , These are the weighting coefficients, and the sum of the weighting coefficients is 1.

[0086] Reference Figure 4 Training an agent using a dual-delay deep deterministic policy gradient algorithm can include: Initialize the agent's main network and target network; The current state is input into the Actor main network to generate actions, the environment executes the actions, and the overall reward is calculated. The experience tuple is then stored in the experience replay pool. When the number of samples in the experience replay pool exceeds 256, batch data is sampled, the target Q value is calculated using the target network, and the first Critic main network and the second Critic main network are updated. The Actor main network is updated every two steps, and the target network is synchronized via soft updates. Repeat the iteration until the reward converges, thus completing the agent training.

[0087] Furthermore, the agent training process is as follows: S3.21: Initialize the main network and the target network: Initialize the weight parameters of the Actor main network, the first Critic main network, and the second Critic main network, and synchronously initialize the Actor target network, the first Critic target network, and the second Critic target network to make their parameters completely consistent with the main network parameters.

[0088] Initialize the experience replay pool and set its capacity to 100,000.

[0089] Configure the Adam optimizer: Actor learning rate is The Critic learning rate is .

[0090] Setting hyperparameters: Discount factor Soft update coefficient Strategy update interval step size .

[0091] S3.22: Input signal status The Actor's main network outputs actions. : Current state (i.e., the coefficients of the current finite impulse response filter) are input to the Actor main network. The Actor main network outputs a continuous action after forward propagation, which is then limited by the Tanh activation function and multiplied by the action scaling factor. And add Ornsteln-Uhlenbeck exploration noise to the actions, and restrict the actions to Within the scope, the final action is obtained. .

[0092] S3.23: The environment performs an action and returns a comprehensive reward. With the new state : Reinforcement learning environment receives actions Combined with motion scaling factor Update the coefficients of the current finite impulse response filter. Apply the updated filter to the preprocessed signal. Perform linear convolution filtering to obtain the filtered signal. The kurtosis, spectral negative entropy, squared envelope Gini index, and peak signal-to-noise ratio of the filtered signal are calculated. The corresponding relative boost rates are then calculated based on the characteristic indicators of the preprocessed signal. Each relative boost rate is normalized by exponential moving average and then weighted and summed to obtain the comprehensive reward. The updated finite impulse response filter coefficients are used as the next state (i.e., the new state). ).

[0093] S3.24: Empirical tuples Store in the experience replay pool: The state obtained from this interaction ,action ,award Next state The experience tuples are formed and stored in the experience replay pool for subsequent batch sampling training.

[0094] S3.25: When the number of samples exceeds 256, random sampling is performed from the empirical replay pool: The number of samples in the experience replay pool is determined in real time. If the number of samples is less than 256, data collection continues interactively. If the number of samples is greater than or equal to 256, a batch (256 groups) of samples is randomly and uniformly sampled from the experience replay pool for network updates.

[0095] S3.26: Update the Critic network, specifically including: S3.261: Change the next state The input is fed into the Actor target network to obtain the target action, and Gaussian noise is then added. Cut off to The formula for calculating the target action is: (twenty one), in, For the target action, This is a noise strategy.

[0096] S3.262: Change the next state With target action Input the first Critic target network and the second Critic target network respectively, calculate the Q value using the following formula, and take the smaller value to calculate the target Q value.

[0097] (9).

[0098] S3.263: Set the current state With action Input the first Critic master network and the second Critic master network to obtain the predicted Q value.

[0099] S3.264: Calculate the mean squared error loss and backpropagate to update the parameters of the first and second Critic master networks, i.e.: (twenty two).

[0100] S3.27: Update the Actor network every two steps: A delayed update strategy is adopted, where an Actor network update is performed after every two Critic network updates are completed. Current state Input the Actor main network to generate policy actions. Input the state and actions into the first Critic main network to obtain Q-value estimates. Calculate the policy gradient and update the Actor main network with the objective of maximizing the Q-value. The calculation formula is as follows: (7).

[0101] S3.28: Perform a soft update of the target network parameters: After the Actor main network is updated, a soft update is performed on the Actor target network, the first Critic target network, and the second Critic target network, using the following formula: (10).

[0102] S3.29: Repeat until training converges: Repeat steps S3.21 to S3.28, continuously interacting with the environment, sampling, and updating the network. (Refer to...) Figure 5 When the comprehensive reward function tends to be stable, the fluctuation is less than the threshold, and the filter performance no longer improves, the training is considered to have converged, and the optimal agent network parameters are saved.

[0103] S4: Update the coefficients of the finite impulse response filter using the optimal strategy, and perform Hilbert transform and envelope spectrum analysis on the preprocessed signal based on the updated finite impulse response filter to identify the faults in the running gear of the track vehicle under test.

[0104] In this embodiment, the optimal strategy output by the trained agent is invoked to automatically update the finite impulse response (FIR) filter and its coefficients, resulting in an optimal FIR filter adapted to the preprocessed signal. This optimal FIR filter is then used to perform adaptive noise reduction and fault pulse enhancement on the preprocessed signal. Hilbert transform and envelope spectrum analysis are then performed on the filtered enhanced signal to extract fault characteristic frequencies and their harmonics. Based on the amplitude enhancement of the fault characteristic frequencies in the envelope spectrum, the system accurately determines whether and what type of fault exists in key components such as wheels, bearings, and gears in the running gear of the test track vehicle, thus completing the intelligent diagnosis of faults in the running gear of the test track vehicle.

[0105] In step S4, the method of updating the finite impulse response filter coefficients using the optimal strategy includes: inputting the initial finite impulse response filter coefficients corresponding to the preprocessed signal into the Actor main network of the agent, and outputting the increment of the corresponding finite impulse response filter coefficients; iteratively updating the initial finite impulse response filter coefficients using the increment of the finite impulse response filter coefficients to obtain the optimal finite impulse response filter coefficients.

[0106] Furthermore, the initial finite impulse response (FIR) filter coefficients corresponding to the preprocessed signal are used as state inputs to the Actor main network of the trained agent. The Actor main network performs forward inference and outputs the FIR filter coefficient increments, which are consistent with the length dimension of the FIR filter. Then, the exploration noise and action perturbations used in the training phase are removed, retaining only the deterministic optimal increments. Using the FIR filter coefficient increments and preset action scaling factors, the initial FIR filter coefficients are weighted, superimposed, and iteratively updated. The updated coefficients that satisfy the stability constraints are taken as the optimal FIR filter coefficients, enabling the filter to possess optimal fault impact feature enhancement and noise suppression capabilities.

[0107] In step S4, the method for identifying faults in the running gear of the track vehicle under test by performing Hilbert transform and envelope spectrum analysis on the preprocessed signal based on the updated finite impulse response filter includes: S4.1: Filter the preprocessed signal using the updated optimal finite impulse response filter to obtain the filtered output signal; S4.2: Perform a Hilbert transform on the filtered output signal to construct an analytic signal; S4.3: Perform a modulo operation on the analytical signal to extract the instantaneous envelope of the filtered output signal; S4.4: Perform a fast Fourier transform on the instantaneous envelope of the filtered output signal to obtain the envelope spectrum. Based on the comparison between the envelope spectrum and the theoretical fault characteristic frequencies of each key component of the running gear of the track vehicle under test, the faults of the running gear of the track vehicle under test can be identified.

[0108] Furthermore, the preprocessed signal is convolved and filtered using an updated optimal finite impulse response filter to obtain a filtered output signal highlighting the periodic impact component of the fault. Then, a Hilbert transform is performed on the filtered output signal to construct an analytic signal composed of real and imaginary parts, achieving phase orthogonality. Modulus operation is performed on the analytic signal to extract the instantaneous envelope of the filtered output signal, highlighting the time-domain periodicity of the fault impact. A fast Fourier transform is then performed on the instantaneous envelope to transform the time-domain envelope features to the frequency domain, obtaining the envelope spectrum. Finally, the actual frequencies corresponding to significant peaks in the envelope spectrum are compared one by one with the theoretical fault characteristic frequencies of the bearings, gears, and wheelsets of the track vehicle under test. If the actual frequency matches the theoretical fault characteristic frequency or overlaps within the allowable error range, the corresponding key component is determined to have a fault, achieving accurate identification, location, and severity determination of faults in the track vehicle's running gear.

[0109] For example, refer to Figure 6 , Figure 7 The periodic pulses at the fault threshold are severely masked by interference signals unrelated to the fault, making the envelope spectral line corresponding to the fault frequency almost indistinguishable. Then, the original simulation signal is processed according to the embodiments of this application, and the processing result is referenced. Figure 8 and Figure 9 As can be seen, the present invention can effectively recover clear periodic fault pulses from filtered signals and achieve obvious fault frequency detection in the corresponding envelope spectrum. This provides a direct verification of the excellent performance of this embodiment in fault feature extraction.

[0110] This invention updates the coefficients of a finite impulse response (FIR) filter using the optimal strategy output by a trained agent. Compared to traditional methods with fixed filter coefficient settings, this invention dynamically adjusts filter parameters based on actual conditions, allowing the filter to better adapt to signal characteristics under different operating conditions. This results in more effective noise filtering, highlighting fault characteristic signals, and improving the accuracy and relevance of signal processing. Furthermore, by designing a reinforcement learning environment with the FIR filter coefficients and their increments as the state and action spaces, respectively, and employing an agent structure containing a master network and a target network, the invention utilizes a dual-delay deep deterministic policy gradient algorithm to train the agent. This improves the algorithm's stability, convergence, sample efficiency, and agent learning ability. Simultaneously, by performing Hilbert transform and envelope spectrum analysis on the preprocessed signal based on the updated FIR filter, fault characteristic frequency components in the signal can be extracted more effectively. This allows for more accurate identification of the fault type and location in the running gear of the rail vehicle, improving the accuracy and reliability of fault diagnosis.

[0111] The order of the steps in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0112] Example 2: Figure 10 As shown, the rail vehicle running gear fault diagnosis device of the present invention uses the above-described rail vehicle running gear fault diagnosis method, as described in Example 1. The rail vehicle running gear fault diagnosis device includes: Acquisition module 1 is used to acquire monitoring signals from the running gear of the track vehicle under test and determine a preprocessed signal. Specifically, acquisition module 1 can be used to: acquire monitoring signals collected by an accelerometer during the operation of the running gear of the track vehicle under test, wherein the monitoring signals include at least one of vibration signals, noise signals, and acoustic emission signals; and perform DC component removal preprocessing on the monitoring signals to obtain the preprocessed signal.

[0113] Module 2 is used to construct a reinforcement learning environment and build an agent. The reinforcement learning environment uses the coefficients of a finite impulse response filter as the state space and the increments of the coefficients of the finite impulse response filter as the action space. The agent includes a main network and a target network with the same structure. Both the main network and the target network are composed of one Actor main network and two independent Critic networks.

[0114] Module 2 can be specifically used to: define basic environmental parameters, including the finite impulse response filter length, action scaling factor, and maximum number of interaction steps in the environment; define the state space using the finite impulse response filter coefficients to generate initial filter coefficients; define the action space using the increments of the finite impulse response filter coefficients; and calculate the preprocessed signal, whose characteristic indices include kurtosis, spectral negative entropy, squared envelope Gini index, and peak signal-to-noise ratio. Module 2 is also specifically used to calculate the kurtosis, spectral negative entropy, squared envelope Gini index, and peak signal-to-noise ratio of the preprocessed signal using the first, second, third, and fourth formulas described above.

[0115] Training module 3 is used to construct a comprehensive reward function and train the agent using a dual-delay deep deterministic policy gradient algorithm to obtain the optimal policy output by the trained agent and the Actor main network of the trained agent.

[0116] In step S3, the agent's main network is the agent's Actor main network, meaning the main network includes the Actor main network. Module 3 can specifically be used to: input the current finite impulse response (FIR) filter coefficients as the current state into the Actor main network, and output the current action; input the current action into the reinforcement learning environment, and update the current FIR filter coefficients by combining the action scaling coefficients; perform linear convolution filtering on the preprocessed signal using the updated FIR filter coefficients to obtain a filtered signal; calculate the feature index of the filtered signal; calculate the relative improvement rate of each filtered signal's feature index with the corresponding feature index of the preprocessed signal, and perform exponential moving average normalization on each relative improvement rate to obtain four normalized relative improvement rates; and perform a weighted summation of the four normalized relative improvement rates to obtain the comprehensive reward function.

[0117] The identification module 4 is used to update the coefficients of the finite impulse response filter using the optimal strategy, and to perform Hilbert transform and envelope spectrum analysis on the preprocessed signal based on the updated finite impulse response filter, so as to identify the faults in the running gear of the track vehicle under test.

[0118] The identification module 4 can be specifically used to: input the initial finite impulse response (FIR) filter coefficients corresponding to the preprocessed signal into the Actor main network of the agent, and output the increment of the corresponding FIR filter coefficients; iteratively update the initial FIR filter coefficients using the increment of the FIR filter coefficients to obtain the optimal FIR filter coefficients. The identification module 4 can also be used to: filter the preprocessed signal using the updated optimal FIR filter to obtain a filtered output signal; perform a Hilbert transform on the filtered output signal to construct an analytic signal; perform a modulo operation on the analytic signal to extract the instantaneous envelope of the filtered output signal; perform a fast Fourier transform on the instantaneous envelope of the filtered output signal to obtain the envelope spectrum; and, based on the comparison of the envelope spectrum with the theoretical fault characteristic frequencies of each key component of the running gear of the track vehicle under test, identify the faults in the running gear of the track vehicle under test.

[0119] Example 3: As Figure 11 As shown, the present invention provides a terminal 5, including a processor 51, a memory 52, and a computer program 53 stored in the memory 52 and executable on the processor 51. When the processor 51 executes the computer program 53, it implements the steps of the fault diagnosis method for the running gear of the rail vehicle, for example... Figure 1 Steps S1 to S4 are shown. Alternatively, when the processor 51 executes the computer program 53, it implements the functions of each module / unit in the above-described device embodiments, for example... Figure 10 The functions of each module are shown.

[0120] For example, the computer program 53 can be divided into one or more modules / units, which are stored in the memory 5 and executed by the processor 51 to complete this embodiment. The one or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program 53 in the terminal. For example, the computer program 53 can be divided into... Figure 10 The modules shown.

[0121] The terminal 5 can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. The terminal 5 may include, but is not limited to, a processor 51 and a memory 52. ​​Those skilled in the art will understand that... Figure 11 This is merely an example of terminal 5 and does not constitute a limitation on terminal 5. It may include more or fewer components than shown, or combine certain components, or different components. For example, terminal 5 may also include input / output devices, network access devices, buses, etc.

[0122] The processor 51 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0123] The memory 52 can be an internal storage unit of the terminal 5, such as a hard disk or memory of the terminal 5. The memory 52 can also be an external storage device of the terminal 5, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the terminal 5. Furthermore, the memory 52 can include both internal storage units and external storage devices of the terminal 5. The memory 52 is used to store the computer program 53 and other programs and data required by the terminal 5. The memory 52 can also be used to temporarily store data that has been output or will be output.

[0124] The functional units and modules in this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Those skilled in the art will recognize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. 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 embodiment.

[0125] In this invention, the embodiments of terminal 5 and fault diagnosis device described are merely illustrative. For example, the division of modules or units is only a logical functional division, and there may be other division methods in actual implementation. The functional units in each embodiment can be integrated into one processing unit, or each unit can exist physically separately. The implementation of all or part of the processes in the above embodiments of this invention can also be accomplished by a computer program instructing related hardware.

[0126] The present invention also provides a computer-readable storage medium storing a computer program 53, which, when executed by a processor, implements the steps of the fault diagnosis method for the running gear of a rail vehicle. The computer program 53 includes computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0127] In summary, the above embodiments of the present invention provide a method and apparatus for diagnosing faults in the running gear of rail vehicles. It can accurately identify the fault type and location of the running gear of rail vehicles, improve the accuracy and reliability of fault diagnosis, and solve the problem of low accuracy in fault identification of the running gear of rail vehicles in the prior art.

Claims

1. A rail vehicle running gear fault diagnosis method, characterized by, It includes the following steps: S1: Acquire the monitoring signals of the running gear of the track vehicle under test, and determine the preprocessing signals; S2: Construct a reinforcement learning environment and build an intelligent agent. The reinforcement learning environment uses the coefficients of a finite impulse response filter as the state space and the increments of the coefficients of the finite impulse response filter as the action space. The intelligent agent includes a main network and a target network. S3: Construct a comprehensive reward function and train the agent using a dual-delay deep deterministic policy gradient algorithm to obtain the optimal policy output by the trained agent and the main network of the trained agent. S4: Update the coefficients of the finite impulse response filter using the optimal strategy, and perform Hilbert transform and envelope spectrum analysis on the preprocessed signal based on the updated finite impulse response filter to identify the faults in the running gear of the track vehicle under test.

2. The rail vehicle running-gear fault diagnostic method according to claim 1, characterized in that, Step S1 includes: S1.1: Acquire monitoring signals collected by the accelerometer during the operation of the running gear of the track vehicle under test, wherein the monitoring signals include at least one of vibration signals, noise signals, and acoustic emission signals; S1.2: Perform DC component removal preprocessing on the monitoring signal to obtain the preprocessed signal.

3. The rail vehicle running-gear fault diagnostic method according to claim 1, characterized in that, In step S2, the method for constructing the reinforcement learning environment includes: S2.1: Define the basic environmental parameters, which include the finite impulse response filter length, action scaling factor, and maximum number of interaction steps in the environment; S2.2: Define the state space using the finite impulse response filter coefficients, and generate initial filter coefficients; S2.3: Define the action space by the increment of the coefficients of the finite impulse response filter; S2.4: Calculate the preprocessed signal, wherein the characteristic indicators of the preprocessed signal include kurtosis, spectral negative entropy, squared envelope Gini index, and peak signal-to-noise ratio.

4. The rail vehicle running-gear fault diagnostic method according to claim 3, characterized in that, In step 2.4, the method for calculating the preprocessed signal includes: S2.41: Calculate the kurtosis of the preprocessed signal using a first formula, wherein the first formula is: （1）， in, The kurtosis of the preprocessed signal, The preprocessed signal; S2.42: Calculate the spectral negative entropy of the preprocessed signal using the second formula, which is: （2）， in, The spectral negative entropy of the preprocessed signal, The energy entropy of the preprocessed signal. for The envelope spectrum amplitude of the preprocessed signal at the specified scale. It is the natural logarithm function. This is the full-scale statistical average; S2.43: Calculate the squared envelope Gini index of the preprocessed signal using the third formula, which is: （3）， in, The squared envelope Gini index of the preprocessed signal. The total number of sampling points of the preprocessed signal. It is a sliding window. For discrete sampling point index, For the first Spectral energy at each sampling point for sequential Norm; S2.44: Calculate the peak signal-to-noise ratio of the preprocessed signal using the fourth formula, which is: （4）， in, The peak signal-to-noise ratio of the preprocessed signal. The energy of the fault characteristic signal, This represents the energy of the noise signal.

5. The method for diagnosing faults in the running gear of a rail vehicle according to claim 3, characterized in that, In step S2, the main network and the target network have the same structure, and both the main network and the target network consist of one Actor main network and two independent Critic networks; in step S3, the agent's main network is the agent's Actor main network.

6. The method for diagnosing faults in the running gear of a rail vehicle according to claim 5, characterized in that, In step S3, the method for constructing the comprehensive reward function includes: S3.11: Input the current finite impulse response filter coefficients as the current state into the Actor main network, and output the current action; S3.12: Input the current action into the reinforcement learning environment, and update the current finite impulse response filter coefficients in combination with the action scaling coefficients; S3.13: Perform linear convolution filtering on the preprocessed signal using the updated finite impulse response filter coefficients to obtain the filtered signal; S3.14: Calculate the characteristic parameters of the filtered signal; S3.15: Calculate the relative improvement rate of the feature index of each filtered signal and the feature index of the corresponding preprocessed signal, and perform exponential moving average normalization on each relative improvement rate to obtain four normalized relative improvement rates. S3.16: The four relative improvement rates after normalization are weighted and summed to obtain the comprehensive reward function.

7. The method for diagnosing faults in the running gear of a rail vehicle according to claim 5, characterized in that, In step S4, the method for updating the coefficients of the finite impulse response filter using the optimal strategy includes: The initial finite impulse response filter coefficients corresponding to the preprocessed signal are input into the Actor main network of the agent, and the increments of the corresponding finite impulse response filter coefficients are output. The initial finite impulse response (FIR) filter coefficients are iteratively updated using the increments of the FIR filter coefficients to obtain the optimal FIR filter coefficients.

8. The method for diagnosing faults in the running gear of a rail vehicle according to claim 1, characterized in that, In step S4, the method for identifying faults in the running gear of the track vehicle under test by performing Hilbert transform and envelope spectrum analysis on the preprocessed signal based on the updated finite impulse response filter includes: S4.1: Filter the preprocessed signal using the updated optimal finite impulse response filter to obtain the filtered output signal; S4.2: Perform a Hilbert transform on the filtered output signal to construct an analytic signal; S4.3: Perform a modulo operation on the analytical signal to extract the instantaneous envelope of the filtered output signal; S4.4: Perform a fast Fourier transform on the instantaneous envelope of the filtered output signal to obtain the envelope spectrum. Based on the comparison between the envelope spectrum and the theoretical fault characteristic frequencies of each key component of the running gear of the track vehicle under test, the faults of the running gear of the track vehicle under test can be identified.

9. A fault diagnosis device for the running gear of a rail vehicle, which uses the fault diagnosis method for the running gear of a rail vehicle according to any one of claims 1-8, characterized in that... The fault diagnosis device for the running gear of rail vehicles includes: The acquisition module is used to acquire monitoring signals from the running gear of the track vehicle under test and determine the preprocessed signals. The building module is used to build a reinforcement learning environment and construct an agent. The reinforcement learning environment uses the coefficients of a finite impulse response filter as the state space and the increment of the coefficients of the finite impulse response filter as the action space. The agent contains a main network and a target network with the same structure. Both the main network and the target network are composed of one Actor main network and two independent Critic networks. The training module is used to construct a comprehensive reward function and train the agent using a dual-delay deep deterministic policy gradient algorithm to obtain the optimal policy output by the trained agent and the Actor main network of the trained agent. The identification module is used to update the coefficients of the finite impulse response filter using the optimal strategy, and based on the updated finite impulse response filter, to perform Hilbert transform and envelope spectrum analysis on the preprocessed signal, thereby identifying the faults in the running gear of the track vehicle under test.

Images