Real-time monitoring and fault diagnosis method and system for sanding state of vehicle sanding device

Abstract

This invention relates to the field of intelligent operation and maintenance and train operation safety monitoring technology for rail transit, specifically a method and system for real-time monitoring and fault diagnosis of the sand discharge status of a vehicle sand spreading device. The method includes: collecting real-time vibration sensing data of the sand spreading pipeline and the current physical environment parameters of the vehicle; performing signal separation on the input data and outputting fluid impact feature vectors and background noise feature vectors; calculating the probability distribution of the fluid impact feature vectors and calculating the feature confidence entropy value representing the uncertainty of the feature distribution; and selecting the corresponding diagnostic logic to generate the sand discharge status result of the sand spreading device based on the comparison result between the feature confidence entropy value and a preset entropy threshold value. This invention effectively solves the technical problem that traditional methods are easily affected by environmental noise interference due to a lack of physical constraints.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and train operation safety monitoring technology for rail transit, specifically to a method and system for real-time monitoring and fault diagnosis of the sand discharge status of a vehicle sand spreading device. Background Technology

[0002] In the high-speed train operation control system, the sand spreading device is a key actuator for improving wheel-rail adhesion. Its operation is carried out in a complex dynamic physical environment, accompanied by the real-time generation of multi-source heterogeneous data.

[0003] To monitor the working status of sand-spreading devices, existing solutions generally employ threshold judgment logic based on the amplitude of a single vibration signal, or directly apply simple data-driven black-box deep learning models for feature recognition. While such solutions possess certain detection capabilities under static or low-speed conditions, in the extreme acoustic environment of high-speed vehicle operation, the sand-spreading pipeline is highly susceptible to strong interference from bogie structural vibration, aerodynamic noise, and structural acoustic coupling resonance. Because existing methods lack effective constraints based on physical mechanisms and fail to quantify the uncertainties in the signal, the system struggles to accurately separate the true fluid impact characteristics from a high-noise background. When faced with strong vibration signals other than sand-spreading, it is prone to feature confusion or model illusions, failing to effectively distinguish between valid sand-spreading signals and artifacts caused by structural resonance. This results in a high false alarm rate and low reliability of diagnostic results, making it difficult to meet the closed-loop control requirements of anti-skid control systems for high reliability and real-time performance of input signals. Therefore, how to effectively suppress environmental noise interference by incorporating physical mechanisms and improve the accuracy and robustness of status monitoring under complex dynamic conditions has become an urgent technical problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for real-time monitoring and fault diagnosis of the sand-spreading status of a vehicle's sand-spreading device. This addresses the problem of false alarms caused by structural-acoustic coupling resonance interference in the extreme noise environment of high-speed vehicle operation, as seen in existing technologies. By introducing a dual mechanism of physical constraint feature extraction and entropy monitoring, the invention effectively eliminates phantom sand-spreading false alarms caused by structural resonance, significantly improving the reliability of the input signal of the anti-skid control system. Specifically, the technical solution of this invention is as follows:

[0005] Real-time monitoring and fault diagnosis methods for the sand discharge status of vehicle sand spreading devices include:

[0006] Collect real-time vibration sensing data of the sand spreading pipeline and the current physical environment parameters of the vehicle;

[0007] The real-time vibration sensing data and the physical environment state parameters are input into a pre-constructed physical constraint feature extraction model. The model is configured to use fluid dynamics equations and structural modes as constraints to perform signal separation on the input data and output fluid impact feature vectors and background noise feature vectors.

[0008] The probability distribution of the fluid impact feature vector is calculated, and the feature confidence entropy value, which characterizes the uncertainty of the feature distribution, is calculated based on the probability distribution.

[0009] Based on the comparison result between the feature confidence entropy value and the preset entropy threshold, the corresponding diagnostic logic is selected to generate the sand output status result of the sand spreading device.

[0010] Preferably, the construction of the physical constraint feature extraction model includes using a deep neural network as the backbone network to extract time-frequency features, and using the inherent frequency modes of the sand-spreading pipeline and the gas-solid two-phase flow motion equation as regularization terms of the loss function.

[0011] The separation and output includes suppressing signal components that overlap with the structural resonance frequency under the physical environment state parameters, retaining the frequency band energy that conforms to the physical laws of fluid inelastic collisions, and generating the fluid impact feature vector.

[0012] Preferably, the calculation of the feature confidence entropy value includes normalizing the fluid impact feature vector to obtain a feature probability distribution, and calculating the information entropy of the feature probability distribution;

[0013] When the energy distribution of the fluid impact feature vector in the time-frequency domain is highly discrete, a high value of the feature confidence entropy is generated, indicating that the current signal has structural acoustic coupling resonance interference.

[0014] Preferably, the comparison result between the feature confidence entropy value and the preset entropy threshold includes:

[0015] When the feature confidence entropy value is lower than the preset entropy threshold, the fluid impact feature vector is input into a preset classification neural network, and the sand state category is directly output.

[0016] When the feature confidence entropy value is higher than or equal to the preset entropy threshold, the physical residual analysis logic is activated, the theoretical fluid energy is calculated based on the physical environment state parameters, the residual between the fluid impact feature vector and the theoretical fluid energy is calculated, and the sand discharge state result is determined based on the residual.

[0017] Preferably, the physical residual analysis logic is configured as follows: if the residual is within the preset physical consistency value range, it is determined to be a normal sand output state; if the residual exceeds the physical consistency value range, it is determined to be an artifact interference signal caused by structural resonance, and a no-sand-output or fault state is output.

[0018] A real-time monitoring and fault diagnosis system for the sand discharge status of a vehicle sand spreading device includes:

[0019] Multidimensional sensing device for acquiring raw signals from the vehicle bogie area;

[0020] The monitoring and diagnostic host establishes a communication connection with the multi-dimensional sensing device to transmit real-time data, and includes,

[0021] The data acquisition module is connected to the multidimensional sensing device;

[0022] The physical noise reduction module receives the data output by the data acquisition module;

[0023] The entropy monitoring module calculates the confidence entropy value based on the feature vector output by the physical denoising module.

[0024] The dynamic decision-making module selects the diagnostic path based on the feature confidence entropy value of the entropy monitoring module.

[0025] The data acquisition module synchronously acquires the vibration signal of the sand-spreading pipe wall and the main control physical parameters of the vehicle, and transmits them to the physical denoising module. The physical denoising module uses the embedded physical equation to constrain the deep learning model, removes structural resonance noise, and outputs the fluid impact feature vector to the entropy monitoring module.

[0026] The entropy monitoring module calculates the feature confidence entropy value in real time and sends it to the dynamic decision module. The dynamic decision module compares the feature confidence entropy value with a preset threshold. When the entropy value is low, it calls the classification network diagnosis; when the entropy value is high, it calls the physical residual analysis diagnosis. Finally, it outputs the sand state result.

[0027] Preferably, the multi-dimensional sensing device includes a vibration acceleration sensor installed at the end of the sand spreading pipeline, and a gateway interface for acquiring vehicle bus data;

[0028] The data acquisition module performs analog-to-digital conversion on the analog signals acquired by the vibration acceleration sensor and timestamps them with the vehicle speed and wind pressure data obtained from the gateway interface.

[0029] Preferably, the physical denoising module includes a physical constraint layer and a feature reconstruction layer;

[0030] The physical constraint layer generates a corresponding pipeline mode suppression mask based on the current physical environment state parameters. The feature reconstruction layer uses the pipeline mode suppression mask to perform weighted filtering on the input vibration signal and outputs a vector sequence containing only sand and fluid features.

[0031] Preferably, the dynamic decision-making module includes an adaptive switching unit; the adaptive switching unit is configured to dynamically adjust the diagnostic strategy within the subsequent time window in response to the duration for which the feature confidence entropy value continuously exceeds the threshold, and to lock the use of physical residual analysis logic when a continuous high entropy value structural resonance interference is detected until the entropy value falls back.

[0032] Preferably, the monitoring and diagnostic host also includes a result feedback interface;

[0033] The result feedback interface receives the status codes of sand discharge, blockage, idling, or sensor failure output by the dynamic decision module, and sends the status codes to the vehicle's anti-skid control system so that the anti-skid control system can adjust the braking strategy.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. This invention constructs a physical constraint feature extraction model, which introduces fluid dynamics equations and structural modes into the loss function of a deep neural network, realizing a deep coupling between data-driven and physical mechanisms. This method can force the model to learn real physical laws, thereby accurately removing background noise that conforms to the structural resonance law in the extreme acoustic environment of high-speed vehicle operation, and separating high-purity fluid impact features from multi-source heterogeneous data, effectively solving the technical problem that traditional methods are easily interfered with by environmental noise due to the lack of physical constraints.

[0036] 2. This invention introduces a feature confidence entropy evaluation mechanism, which quantitatively evaluates the dispersion of energy distribution of features in the frequency domain or time-frequency domain by calculating the probability distribution entropy of fluid impact feature vectors. This mechanism can keenly capture spectral anomalies caused by structural acoustic coupling resonance and monitor the uncertainty of model output in real time, thereby effectively preventing deep learning models from producing hallucinations or feature confusion under strong noise interference, and providing a quantitative basis for the credibility of the diagnostic system.

[0037] 3. This invention designs a dual-path diagnostic logic based on entropy monitoring and physical residual analysis. Under low-entropy steady state, a classification network is used to ensure response speed, while under high-entropy uncertain state, the law of conservation of physical energy is used for logical decision-making. This mechanism uses the first principles of physics to falsify high-energy artifact signals, eliminating false alarms caused by wheel-rail impact or bogie resonance, and significantly improving the reliability of the anti-skid control system input signal while ensuring real-time performance.

[0038] 4. This invention provides a refined state feedback interface and adaptive locking strategy, which can distinguish various fault modes such as physical blockage, pipeline venting, and sensor failure, and output detailed state codes to the anti-skid control system. Combined with a time window locking mechanism for continuous severe operating conditions, it effectively avoids frequent jumps in diagnostic results under complex dynamic environments, and provides rich and stable data support for the vehicle to achieve more accurate closed-loop braking control and subsequent maintenance and troubleshooting. Attached Figure Description

[0039] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0040] Figure 1 This is a flowchart of the method of the present invention;

[0041] Figure 2 This is a structural diagram of the system of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0043] Example 1:

[0044] Please see Figure 1 Real-time monitoring and fault diagnosis methods for the sand output status of vehicle sand spreading devices, including:

[0045] Collect real-time vibration sensing data of the sand spreading pipeline and the current physical environment parameters of the vehicle;

[0046] Real-time vibration sensing data and physical environment state parameters are input into a pre-constructed physical constraint feature extraction model. This model is configured to use fluid dynamics equations and structural modes as constraints to perform signal separation on the input data and output fluid impact feature vectors and background noise feature vectors.

[0047] The probability distribution of the fluid impact feature vector is calculated, and the feature confidence entropy value, which represents the uncertainty of the feature distribution, is calculated based on the probability distribution.

[0048] Based on the comparison results between the feature confidence entropy value and the preset entropy threshold, the corresponding diagnostic logic is selected to generate the sand output status result of the sand spreading device.

[0049] This embodiment details the specific execution process of the method, aiming to solve the technical problem that traditional threshold-based or simple black-box deep learning methods are easily affected by structural-acoustic coupling resonance interference and generate false alarms in high-speed trains under extreme noise environments.

[0050] The system performs a synchronous acquisition step of multi-source heterogeneous data, collecting real-time vibration sensing data of the sand spreading pipeline and the current physical environment status parameters of the vehicle;

[0051] Here, real-time vibration sensing data refers to the high-frequency vibration sequence collected by an accelerometer installed at the nozzle at the end of the sand-spreading pipeline. Its sampling frequency is set above 20kHz to cover the high-frequency acoustic emission characteristics of sand particles impacting the pipe wall. Physical environment state parameters refer to the key physical quantities obtained from the Train Network Control System (TCMS) that determine the hydrodynamic characteristics of the sand particles, mainly including train speed. Sand spreading air pressure and external wind speed ;

[0052] The system performs signal separation based on a physical constraint model. Real-time vibration sensing data and physical environment state parameters are input into a pre-constructed physical constraint feature extraction model. This model is configured to use fluid dynamics equations and structural modes as constraints. Inside the model, signal blind source separation is performed, and the output is a fluid impact feature vector characterizing the inelastic collision signal generated by sand particles impacting the pipe wall under this physical environment. And the background noise feature vector characterizing signals generated by wheel-rail impact, aerodynamic noise and pipeline structure resonance. ;

[0053] To quantify the determinism of the features extracted by the model and prevent the model from exhibiting illusions under strong noise, the system calculates the feature confidence entropy value, performs probability distribution calculation on the fluid impact feature vector, and calculates the feature confidence entropy value accordingly. The entropy value is a dimensionless scalar used to characterize the degree of dispersion of the energy distribution of the current fluid impact characteristics in the frequency domain or time-frequency domain.

[0054] The system executes a dual-path diagnostic logic decision, dynamically selecting the diagnostic path based on the comparison between the feature confidence entropy value and the preset entropy threshold: in response to a low-entropy state, the system is confident that the signal is pure and directly uses the classification network to generate the result; in response to a high-entropy state, that is, when the system detects high uncertainty, physical residual analysis is initiated, and the final decision is made using the first principles of physics.

[0055] This embodiment introduces a dual mechanism of entropy monitoring and physical residuals to construct a deeply coupled architecture of data-driven and physical principles. In the complex acoustic environment of high-speed vehicle operation, this method can use entropy to keenly capture the discretization characteristics of energy distribution caused by structural resonance, and use physical residual logic to falsify high-energy artifact signals in milliseconds. Thus, while ensuring the diagnostic response speed under normal operating conditions, it eliminates the false alarm phenomenon of ghost sand spreading caused by bogie structural resonance, and significantly improves the reliability of the input signal of the anti-skid control system.

[0056] Example 2:

[0057] The construction of the physical constraint feature extraction model includes using a variant of the U-Net architecture as the backbone network to extract time-frequency features. Its encoder part contains four one-dimensional convolutional layers for extracting multi-scale features, and the decoder part contains corresponding deconvolutional layers. Specifically, the kernel size of the four one-dimensional convolutional layers of the encoder is set to 3, the stride is set to 1, and the number of channels increases sequentially to 64, 128, 256, and 512. Each convolutional layer is followed by a ReLU activation function and a max pooling layer with a stride of 2 to match the feature extraction requirements of 20kHz high-frequency sampling signals.

[0058] The natural frequency modes of the sand-spreading pipeline and the motion equations of the gas-solid two-phase flow are used as regularization terms of the loss function.

[0059] The signal components that suppress the overlap with the structural resonance frequency under the physical environment state parameters are separated and output, while retaining the frequency band energy that conforms to the physical laws of fluid inelastic collision, thus generating a fluid impact feature vector.

[0060] This embodiment further illustrates the construction and operation mechanism of the physical constraint feature extraction model; in order to force the deep neural network to learn the real physical laws, rather than just fitting noise, this embodiment uses a composite loss function in the model training stage;

[0061] Define the physical regularization loss function:

[0062] ;

[0063] in, This represents the total loss value during model training.

[0064] This is a data-driven term that measures the degree of fit between the model output and the input data. In this embodiment, the input is the original signal. Pre-converted into a time-frequency spectrum using a short-time Fourier transform ;

[0065] Calculate the input time-frequency spectrum The sum of the time-frequency features reconstructed from the model The Frobenius norm error between them, i.e. ,in The Frobenius norm of the matrix is ​​used to ensure energy conservation and information integrity during signal separation.

[0066] and In this embodiment, the hyperparameter weights are used to balance the physical constraints and the data fitting. Preferred setting is Within the interval, Preferred setting is Within the range;

[0067] Based on this, we will focus on introducing fluid dynamics constraints. With structural modal constraints ;

[0068] Specifically, the formula for calculating the structural modal constraint term is:

[0069] ;

[0070] in, This represents the Hadamard product, which is the element-wise multiplication of matrices. The modal mask vector is dynamically generated based on the current vehicle speed, although the natural frequency of the sand spreading pipe... This is an inherent property of the structure, but only if the current vehicle speed... The corresponding aerodynamic excitation main frequency band covers this At that time, the mask vector is in the first three natural frequency ranges of the sand-spraying pipeline. ,in Let n be the nth natural frequency of the sand-spreading pipeline. For example, the preset resonance effect frequency band half bandwidth The value is 50Hz, when the frequency is at Within the specified interval, the time mask vector is set to 1, and in other frequency bands, it is set to 0, thus penalizing... The energy accumulation in the pipeline's inherent resonant frequency band forces the model to classify the energy in these frequency bands as ;

[0071] Meanwhile, fluid dynamics constraints The energy conservation form derived from the gas-solid two-phase flow equations of motion is used in the following calculation formula:

[0072] ;

[0073] in, The L2 norm of a vector is represented by the square of the Euclidean norm.

[0074] : Short-time energy integral for feature extraction;

[0075] : is the theoretical energy flux boundary derived from the gas-solid two-phase flow theory;

[0076] According to Bernoulli's equation in fluid mechanics, the fluid velocity at the nozzle... and It is directly proportional to the square root, that is mass flow With constant orifice diameter and flow velocity Proportional, that is Therefore, the total kinetic energy per unit time, i.e., the energy flux. ;

[0077] in The conversion coefficient for the pressure-energy flux mapping relationship depends on the sensor's sensitivity and sampling gain, and needs to be determined through a standard sand calibration experiment. The specific steps of this calibration experiment include:

[0078] With the vehicle stationary and no external wind interference, five different standard sand-spreading wind pressure values ​​were set. The range covers 200 kPa to 600 kPa;

[0079] For each wind pressure value, an accelerometer was used to collect a 10-second sand-spreading vibration signal, and its short-time energy integral average was calculated. ;

[0080] Using the least squares method to analyze data pairs Perform fitting and solve the equation The conversion coefficient in the pressure-energy flux mapping relationship ;

[0081] It constrains the extraction The energy amplitude must meet the wind pressure requirements. The sand grain kinetic energy theorem under the model forces the model to accept that only signals with specific energy levels and decay characteristics are sand-spreading signals.

[0082] The Physical Information Neural Network (PINN) architecture used in this embodiment essentially applies hard constraints of physical laws to the optimization surface of the neural network. This design enables the model to perform logical deduction based on the embedded fluid dynamics equations and structural modal information when facing extreme operating noise that has not appeared in the training set. It automatically suppresses those artifact signals with huge energy but frequency characteristics that conform to the pipeline resonance law, thereby achieving signal blind source separation based on physical essence without relying on massive labeled data.

[0083] Example 3:

[0084] The calculation of feature confidence entropy includes normalizing the fluid impact feature vector to obtain the feature probability distribution, and calculating the information entropy of the feature probability distribution;

[0085] When the energy distribution of the fluid impact feature vector in the time-frequency domain is highly discrete, a high numerical value of feature confidence entropy is generated, which indicates that the current signal has structural acoustic coupling resonance interference.

[0086] This embodiment details the calculation logic of feature confidence entropy value;

[0087] To assess whether the extracted fluid features are affected by structural resonances disguised as sand grain acoustic signatures, this embodiment defines a confidence entropy. The calculation formula is:

[0088] ;

[0089] in, For frequency index, This represents the total number of discrete frequency points of the fluid impact eigenvector in the frequency domain.

[0090] : is the fluid impact eigenvector at the th The normalized energy probability at each frequency component is calculated as follows: Divide by total energy;

[0091] Here, The fluid impact eigenvector derived from the physical constraint model at frequency The amplitude at that point;

[0092] In a real sand-spreading process, the acoustic emission signals generated by sand particles impacting the pipe wall are usually concentrated on a specific high-frequency broadband, with a relatively concentrated energy distribution. The entropy value calculated under these conditions... Lower;

[0093] Conversely, when structural acoustic coupling resonance occurs, although the total vibration energy is large, its energy is often discretely distributed across multiple structural modal frequency points, or exhibits a chaotic state across the entire frequency band. In this case, the normalized probability distribution... The surface tends to flatten, resulting in a calculated entropy value. Significantly increased;

[0094] This embodiment cleverly transforms the difficult-to-quantify signal purity into a calculable scalar index by leveraging the high sensitivity of information entropy to distribution dispersion. This index can keenly capture spectral whitening or multi-peak dispersion caused by structural resonance, providing a precise decision-making basis for the subsequent dynamic switching between low-cost statistical classification models and high-cost physical analysis models.

[0095] Example 4:

[0096] The comparison results between the feature confidence entropy value and the preset entropy threshold include:

[0097] When the feature confidence entropy value is lower than a preset entropy threshold, the fluid impact feature vector is input into a preset classification neural network, which directly outputs the sand state category. The preset classification neural network uses a lightweight fully connected network, containing an input layer, two hidden layers with 128 and 64 nodes respectively, and a Softmax output layer. The input is... The statistical characteristics, including mean, variance, skewness, and kurtosis, are output as the probability distribution of the corresponding sand production state.

[0098] When the feature confidence entropy value is higher than or equal to the preset entropy threshold, the physical residual analysis logic is activated. The theoretical fluid energy is calculated based on the physical environment state parameters, the residual between the fluid impact feature vector and the theoretical fluid energy is calculated, and the sand state result is determined based on the residual.

[0099] The physical residual analysis logic is configured as follows:

[0100] If the residual is within the preset physical consistency value range, it is determined to be a normal sand output state; if the residual exceeds the physical consistency value range, it is determined to be an artifact interference signal caused by structural resonance, and the output is a no-sand-output or fault state.

[0101] This embodiment details the physical residual analysis logic, which is the system's last line of defense during high entropy crises;

[0102] Response to feature confidence entropy value If the value is greater than or equal to a preset entropy threshold, the system determines that it is currently in a high-risk interference zone and initiates verification based on physical residuals.

[0103] System computational fluid energy physics residuals:

[0104] ;

[0105] in, To observe fluid energy, specifically based on the fluid impact feature vector output by the model. The total energy of the sum of squares is calculated within the current diagnostic time window, such as 100ms.

[0106] For theoretical fluid energy, the specific calculation follows a semi-empirical formula:

[0107]

[0108] in, Euler's number is a natural constant.

[0109] Among them, the maximum value function is introduced. The aim is to eliminate numerical calculation anomalies caused by negative baselines due to sensor zero drift or lack of wind pressure establishment.

[0110] : The minimum critical wind pressure constant for the sand spreading system is preset in the system memory, or the dynamic threshold is read in real time by the vehicle control unit;

[0111] This term is used to correct for the attenuation of effective impact energy caused by external flow field disturbances;

[0112] in, This is the longitudinal aerodynamic damping coefficient caused by the train's running speed. External wind speed The resulting transverse dissipation coefficient, in units of It is used to characterize the dissipation effect of complex flow fields on effective impact energy; For example, a value of 200 kPa, the unit must be consistent with... Maintain consistency;

[0113] in, and These are the fluid dynamic coefficients fitted based on bench experiments; The energy mapping reference coefficient contains the conversion factor for converting the pressure dimension to the energy dimension and the system energy conversion efficiency, which needs to be determined through standard sand volume calibration experiments. The pressure response index is theoretically 1.5 based on the fluid kinetic energy theorem. However, considering the adiabatic expansion effect of the gas, the actual engineering value is set between [1.4, 1.5].

[0114] The system executes the judgment logic:

[0115] In response to Within the preset physical consistency value range ,in The preset residual tolerance threshold is defined by its unit and energy integral. Consistent, unit joule, the value is set to twice the standard deviation of the residual distribution between the theoretical energy calculated under historical normal sand spreading conditions and the observed energy. This indicates that the observed vibration energy matches the energy that should be generated under the current wind pressure and vehicle speed conditions in terms of order of magnitude. At this time, even if the entropy value is high, the system still judges it as a normal sand discharge state.

[0116] In response to Exceeding the stated range indicates a significant deviation between the observed energy and the theoretical value: if Significantly higher than This means that the sensor detected a large amount of non-aerodynamic vibration energy, which the system determined to be an artifact interference signal caused by structural resonance; if Significantly lower than This means that the air passage is normal but there is no sand impact. If there is a blockage, both of these situations correspond to no sand coming out or a faulty state, thus achieving accurate diagnosis by utilizing physical inconsistencies.

[0117] This embodiment realizes a paradigm shift from voiceprint feature recognition to energy conservation verification. By introducing physical residual analysis, the system can effectively identify deceptive signals that are extremely realistic in terms of spectral features but cannot be consistent with the law of energy conservation, such as periodic strong impacts caused by wheel abrasions. This adjudication mechanism based on the first principles of physics ensures that the system has a zero false alarm rate under extreme conditions such as train passing through junctions and severe wheel-rail impacts.

[0118] Example 5:

[0119] Please see Figure 2 A real-time monitoring and fault diagnosis system for the sand discharge status of a vehicle sand spreading device, employing any of the aforementioned real-time monitoring and fault diagnosis methods for the sand discharge status of a vehicle sand spreading device; and a multi-dimensional sensing device for acquiring the original signal of the vehicle bogie area.

[0120] The monitoring and diagnostic host establishes a communication connection with the multi-dimensional sensing device to transmit real-time data, and includes: a data acquisition module connected to the multi-dimensional sensing device;

[0121] The physical noise reduction module receives data output from the data acquisition module;

[0122] The entropy monitoring module calculates the confidence entropy value based on the feature vector output by the physical denoising module.

[0123] The dynamic decision-making module selects the diagnostic path based on the feature confidence entropy value of the entropy monitoring module.

[0124] The data acquisition module synchronously collects vibration signals from the sand-spreading pipe wall and the vehicle's main control physical parameters;

[0125] The data is transmitted to the physical denoising module, which uses embedded physical equations to constrain the deep learning model, removes structural resonance noise, and outputs the fluid impact feature vector to the entropy monitoring module.

[0126] The entropy monitoring module calculates the feature confidence entropy value in real time and sends it to the dynamic decision module. The dynamic decision module compares the feature confidence entropy value with a preset threshold. When the entropy value is low, it calls the classification network diagnosis, and when the entropy value is high, it calls the physical residual analysis diagnosis, and finally outputs the sand state result.

[0127] The multi-dimensional sensing device includes a vibration acceleration sensor installed at the end of the sand spreading pipeline and a gateway interface for acquiring vehicle bus data.

[0128] The data acquisition module performs analog-to-digital conversion on the analog signals acquired by the vibration acceleration sensor and aligns them with the vehicle speed and wind pressure data obtained from the gateway interface using timestamps.

[0129] The physical denoising module includes a physical constraint layer and a feature reconstruction layer; the physical constraint layer generates the corresponding pipeline mode suppression mask based on the current physical environment state parameters;

[0130] The feature reconstruction layer uses a pipeline mode suppression mask to weight and filter the input vibration signal, and outputs a vector sequence containing only the features of sand and fluid.

[0131] This embodiment illustrates a real-time monitoring and fault diagnosis system for the sand discharge status of a vehicle sand spreading device that implements the above method;

[0132] The multi-dimensional sensing device includes a vibration acceleration sensor installed at the end of the sand-spreading pipeline, and a gateway interface for connecting to the MVB / WTB vehicle bus to obtain millisecond-level timestamps. and Data; for external environmental wind speed Meteorological data is received via a gateway interface, or defaults to 0 if missing. To ensure the effectiveness of physical constraints, the data acquisition module performs hard real-time synchronization to ensure the vibration signal... Timing and bus data Timing error less than 10ms;

[0133] The physical denoising module embeds a physical constraint model inference unit. The physical constraint layer stores pre-obtained data on the natural frequencies and mode shapes of the sand-spreading pipeline, acquired through finite element simulation or experimental modal analysis. It will look up tables or calculate the pipeline resonance sensitive frequency under the current working conditions and generate a pipeline mode suppression mask; the feature reconstruction layer will perform Hadamard product operation or weighted filtering on the input original vibration spectrum and the mask, and physically shield the structural resonance frequency band in the feature extraction stage, and output a high-purity sand-fluid feature vector sequence.

[0134] The entropy monitoring module has a built-in FPGA or DSP chip to perform entropy calculation in real time, while the dynamic decision-making module performs logical branch judgment.

[0135] This embodiment uses a collaborative design of the hardware architecture to embed physical prior knowledge into edge computing nodes. In particular, the generation and application of modal suppression masks in the physical denoising module is equivalent to applying a physical filter before the data enters the neural network. This not only significantly reduces the computational load of subsequent algorithms in processing high-noise signals, but also ensures the purity of feature vectors from the signal source, enabling the entire monitoring system to have industrial-grade anti-interference capabilities.

[0136] Example 6:

[0137] The dynamic decision-making module includes an adaptive switching unit;

[0138] The adaptive switching unit is configured to dynamically adjust the diagnostic strategy within the subsequent time window in response to the duration for which the feature confidence entropy value continuously exceeds the threshold. When a continuous high entropy value structural resonance interference is detected, the physical residual analysis logic is locked until the entropy value falls back.

[0139] This embodiment describes the implementation details of the adaptive switching unit in the dynamic decision-making module, mainly involving time window locking of the diagnostic strategy. During train operation, structural resonance, such as after passing through a section of corrugated track, often lasts for a period of time. If the system frequently jumps between low-entropy logic and high-entropy logic, it will cause the output results to jitter. Therefore, the adaptive switching unit introduces a locking mechanism:

[0140] Response to feature confidence entropy value The duration of time exceeding the threshold reaches the preset time. For example, at 500ms, the unit determines that the train has entered a region of continuous severe operating conditions. At this time, the system locks in the use of physical residual analysis logic and does not switch back to the classification network even if there is a brief drop in entropy value in the middle, until the entropy value is continuously below the threshold in the subsequent time window and returns to steady state.

[0141] This embodiment effectively solves the decision oscillation problem under long-term harsh operating conditions, such as long tunnels, windy areas, and specific corrugated sections, by introducing a hysteresis comparison strategy. This adaptive time window locking mechanism enhances the system's robustness to continuous environmental changes, ensures the smoothness and consistency of diagnostic results output, and avoids control system fluctuations that may be caused by frequent logic switching.

[0142] Example 7:

[0143] The monitoring and diagnostic host also includes a result feedback interface;

[0144] The result feedback interface receives the status codes of sand discharge, blockage, idling, or sensor failure output by the dynamic decision module, and sends the status codes to the vehicle's anti-skid control system so that the anti-skid control system can adjust the braking strategy.

[0145] This embodiment describes the implementation details of the result feedback interface;

[0146] The result feedback interface does not simply output simple 0 / 1 signals, but sends status codes containing rich diagnostic information to the vehicle anti-skid control system:

[0147] State 00 indicates normal sand production, at which point the residual is low and the classification confidence is high.

[0148] State 01 indicates physical blockage, at which point the wind pressure is... Normal, but the eigenvector energy is extremely low and there is no resonance interference;

[0149] State 02 indicates that the pipeline is venting. At this time, the wind pressure is normal and there are airflow noise characteristics, but there is a lack of high-frequency texture of sand particle collision.

[0150] State 11 indicates that the sensor is faulty or the confidence level is unavailable, in response to When the physical residual calculation is extremely high and divergent, the system self-diagnoses it as a sensor fault.

[0151] The fine-grained status coding output provided in this embodiment upgrades traditional sand-spreading monitoring to sand-spreading health management. By distinguishing between physical blockage, idling, and sensor failure, the vehicle anti-skid control system can implement a more refined closed-loop control strategy. For example, when physical blockage is detected, it can actively reduce the braking force demand of the axle instead of blindly waiting for the friction coefficient to rise, thereby maximizing braking efficiency and wheel-rail protection while ensuring driving safety.

[0152] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for real-time monitoring and fault diagnosis of sand discharge status of a vehicle sand spreading device, characterized in that: include: Collect real-time vibration sensing data of the sand spreading pipeline and the current physical environment parameters of the vehicle; The real-time vibration sensing data and the physical environment state parameters are input into a pre-constructed physical constraint feature extraction model. The model is configured to use fluid dynamics equations and structural modes as constraints to perform signal separation on the input data and output fluid impact feature vectors and background noise feature vectors. The probability distribution of the fluid impact feature vector is calculated, and the feature confidence entropy value, which characterizes the uncertainty of the feature distribution, is calculated based on the probability distribution. Based on the comparison result between the feature confidence entropy value and the preset entropy threshold, the corresponding diagnostic logic is selected to generate the sand output status result of the sand spreading device; The construction of the physical constraint feature extraction model includes using a deep neural network as the backbone network to extract time-frequency features, and using the inherent frequency modes of the sand-spreading pipeline and the gas-solid two-phase flow motion equation as the regularization term of the loss function. The separation and output includes suppressing signal components that overlap with the structural resonance frequency under the physical environment state parameters, retaining the frequency band energy that conforms to the physical laws of fluid inelastic collisions, and generating the fluid impact feature vector.

2. The method for real-time monitoring and fault diagnosis of the sand discharge status of a vehicle sand spreading device according to claim 1, characterized in that: The calculation of the feature confidence entropy value includes normalizing the fluid impact feature vector to obtain a feature probability distribution, and calculating the information entropy of the feature probability distribution. When the energy distribution of the fluid impact feature vector in the time-frequency domain is highly discrete, a high value of the feature confidence entropy is generated, indicating that the current signal has structural acoustic coupling resonance interference.

3. The method for real-time monitoring and fault diagnosis of the sand discharge status of a vehicle sand spreading device according to claim 1, characterized in that: The comparison result based on the feature confidence entropy value and the preset entropy threshold includes: When the feature confidence entropy value is lower than the preset entropy threshold, the fluid impact feature vector is input into a preset classification neural network, and the sand state category is directly output. When the feature confidence entropy value is higher than or equal to the preset entropy threshold, the physical residual analysis logic is activated, the theoretical fluid energy is calculated based on the physical environment state parameters, the residual between the fluid impact feature vector and the theoretical fluid energy is calculated, and the sand discharge state result is determined based on the residual.

4. The method for real-time monitoring and fault diagnosis of the sand discharge status of the vehicle sand spreading device according to claim 3, characterized in that: The physical residual analysis logic is configured as follows: if the residual is within the preset physical consistency value range, it is determined to be a normal sand output state; if the residual exceeds the physical consistency value range, it is determined to be an artifact interference signal caused by structural resonance, and outputs a no-sand-output or fault state.

5. A real-time monitoring and fault diagnosis system for the sand discharge status of a vehicle sand spreading device, characterized in that: A method for real-time monitoring and fault diagnosis of the sand discharge status of a vehicle sand spreading device as described in any one of claims 1 to 4; and, Multidimensional sensing device for acquiring raw signals from the vehicle bogie area; The monitoring and diagnostic host establishes a communication connection with the multi-dimensional sensing device to transmit real-time data, and includes, The data acquisition module is connected to the multidimensional sensing device; The physical noise reduction module receives the data output by the data acquisition module; The entropy monitoring module calculates the confidence entropy value based on the feature vector output by the physical denoising module. The dynamic decision-making module selects a diagnostic path based on the indicators from the entropy monitoring module. The data acquisition module synchronously acquires the vibration signal of the sand-spreading pipe wall and the main control physical parameters of the vehicle, and transmits them to the physical denoising module. The physical denoising module uses the embedded physical equation to constrain the deep learning model, removes structural resonance noise, and outputs the fluid impact feature vector to the entropy monitoring module. The entropy monitoring module calculates the feature confidence entropy value in real time and sends it to the dynamic decision module. The dynamic decision module compares the feature confidence entropy value with a preset threshold. When the entropy value is low, it calls the classification network diagnosis; when the entropy value is high, it calls the physical residual analysis diagnosis. Finally, it outputs the sand state result.

6. The real-time monitoring and fault diagnosis system for the sand discharge status of the vehicle sand spreading device according to claim 5, characterized in that: The multi-dimensional sensing device includes a vibration acceleration sensor installed at the end of the sand spreading pipeline, and a gateway interface for acquiring vehicle bus data. The data acquisition module performs analog-to-digital conversion on the analog signals acquired by the vibration acceleration sensor and timestamps them with the vehicle speed and wind pressure data obtained from the gateway interface.

7. The real-time monitoring and fault diagnosis system for the sand discharge status of the vehicle sand spreading device according to claim 5, characterized in that: The physical denoising module includes a physical constraint layer and a feature reconstruction layer; The physical constraint layer generates a corresponding pipeline mode suppression mask based on the current physical environment state parameters. The feature reconstruction layer uses the pipeline mode suppression mask to perform weighted filtering on the input vibration signal and outputs a vector sequence containing only sand and fluid features.

8. The real-time monitoring and fault diagnosis system for the sand discharge status of the vehicle sand spreading device according to claim 5, characterized in that: The dynamic decision-making module includes an adaptive switching unit; the adaptive switching unit is configured to dynamically adjust the diagnostic strategy within the subsequent time window in response to the duration for which the feature confidence entropy value continuously exceeds the threshold, and to lock the use of physical residual analysis logic when a continuous high entropy value structural resonance interference is detected until the entropy value falls back.

9. The real-time monitoring and fault diagnosis system for the sand discharge status of the vehicle sand spreading device according to claim 5, characterized in that: The monitoring and diagnostic host also includes a result feedback interface; The result feedback interface receives the status codes of sand discharge, blockage, idling, or sensor failure output by the dynamic decision module, and sends the status codes to the vehicle's anti-skid control system so that the anti-skid control system can adjust the braking strategy.

Images