Railway vehicle traction system fault early warning method based on multi-source sensor fusion

Abstract

The present application relates to the technical field of data processing, more particularly, the present application relates to a multi-source sensor fusion rail vehicle traction system fault early warning method, the method comprising: collecting traction motor stator voltage and current, using damping integral and cross multiplication operator to calculate stator flux linkage vector component and instantaneous electromagnetic drive potential; obtain the mechanical angular velocity, using the dynamic impedance model containing phase lag compensation to map the mechanical angular velocity and the instantaneous electromagnetic drive potential, obtain the theoretical background vibration energy density; calculate the measured vibration acceleration signal measured vibration energy density, using the difference between the measured vibration energy density and the theoretical background vibration energy density to obtain the normalized fault residual; based on the normalized fault residual, construct the residual time series and calculate the weighted permutation entropy value, combine the normalized fault residual and the weighted permutation entropy value to confirm the fault. The present application excludes non-fault information interference and realizes accurate early warning.

Description

Technical Field

[0001] This invention relates to the field of data processing technology. More specifically, this invention relates to a fault early warning method for rail vehicle traction systems based on multi-source sensor fusion. Background Technology

[0002] As the core power system of a train, the traction system of a rail vehicle operates in an extremely harsh environment. Existing fault monitoring technologies are typically based on energy threshold logic, which assumes a fault exists when the vibration amplitude exceeds a preset threshold. This logic works under steady-state, uniform speed conditions, but under complex transient conditions such as sudden traction, electric braking, and crossing phase-separation zones, the system faces severe challenges from non-fault information interference.

[0003] First, there is the challenge of misdiagnosing endogenous strong coupling interference. Traction motors are strongly coupled electromechanical systems. During sudden changes in operating conditions, a rapid change in current can trigger intense electromagnetic force waves, causing high-amplitude background vibrations in the motor casing. This vibration is a physical property of normal equipment operation, not a fault. Furthermore, due to the inherent inertial lag of the mechanical transmission chain, this normal vibration often lags behind the current excitation in time. Current technology cannot distinguish between fault impacts and normal electromagnetic mechanical responses, often misreporting the motor's normal high-energy response under changing operating conditions as a mechanical fault.

[0004] Secondly, there is the challenge of confusion caused by external random interference. Vehicles operate on complex track surfaces, and high-energy random impacts can occur at switches, joints, or wheel-rail hard points. These external impacts, in terms of time-domain waveforms and energy amplitudes, are easily confused with early damage impacts from mechanical components such as gears or bearings. Current technologies lack effective feature identification methods and cannot eliminate environmental noise from random road surface irregularities, leading to frequent false alarms when vehicles pass through switch areas.

[0005] Therefore, the core deficiency of existing technologies lies in the lack of a physical-based interference stripping mechanism under complex operating conditions with multi-source strong interference coupling. The system cannot distinguish the source of vibration energy, conflating intrinsic normal responses and extrinsic environmental noise with fault signals, thus causing the accuracy of fault warnings to drop sharply in dynamic environments. Summary of the Invention

[0006] To address the problem that existing technologies cannot effectively eliminate interference from non-fault information under complex dynamic operating conditions, resulting in a high false alarm rate, this invention proposes a fault early warning method for rail vehicle traction systems based on multi-source sensor fusion. This method includes the following steps:

[0007] The stator voltage and stator current of the traction motor are collected. The flux linkage and torque are calculated by damping integral calculation and cross product calculation to obtain the instantaneous electromagnetic driving potential.

[0008] The mechanical angular velocity of the traction motor is obtained, and the mechanical angular velocity and instantaneous electromagnetic driving potential are nonlinearly mapped using a preset dynamic impedance model to obtain the theoretical background vibration energy density. The dynamic impedance model includes phase hysteresis compensation parameters for the inertial effect of the mechanical transmission chain.

[0009] The measured vibration acceleration signal of the traction system is collected and the measured vibration energy density is calculated. The difference between the measured vibration energy density and the theoretical background vibration energy density is calculated using a dimensionless signal-to-noise ratio operator to obtain the normalized fault residual.

[0010] Based on continuous normalized fault residuals, a residual time series is constructed. The weighted permutation entropy value of the residual time series is calculated using the weighted permutation entropy algorithm. The normalized fault residuals and the weighted permutation entropy value are processed for fault confirmation using dual threshold logic to achieve fault early warning of rail vehicle traction system.

[0011] This invention first accurately describes the intrinsic physical excitation that triggers vibration by collecting electrical quantities and calculating the instantaneous electromagnetic driving potential. Second, it calculates the theoretical background vibration energy density using a dynamic impedance model that includes phase hysteresis compensation parameters, logically forcing the alignment of electromagnetic excitation and mechanical response on the time axis. This effectively isolates the intrinsic normal background vibration interference caused by mechanical inertia under transient conditions such as rapid traction or braking. Finally, by combining normalized fault residuals and a weighted permutation entropy algorithm, a dual verification logic of relative energy intensity and microscopic sequence order is constructed. This logic effectively distinguishes between exogenous interference caused by random road impacts and deterministic impacts caused by mechanical faults, thereby significantly improving the accuracy and robustness of fault warnings for rail vehicle traction systems in complex operating environments.

[0012] Preferably, obtaining the instantaneous electromagnetic driving potential includes:

[0013] Weighting coefficients are constructed using the cutoff time constant of the integrator. Based on the weighting coefficients, the sampling time interval, and the stator resistance, the stator voltage, stator current, and the stator flux vector component at the current moment are calculated using weighted iterative calculation to obtain the stator flux vector component at the current moment.

[0014] The difference between the cross product of the stator flux vector component and the stator current in the two-phase stationary coordinate system is calculated, and the difference is multiplied by a coefficient related to the number of motor pole pairs to obtain the instantaneous electromagnetic driving potential.

[0015] This invention introduces damped integral calculation and uses the cutoff time constant to weight the integration process, effectively suppressing the DC bias drift generated during the pure integration of low-frequency signals. This ensures the convergence and accuracy of the stator flux vector component calculation, thus providing a precise electromagnetic force reference for subsequent calculation of instantaneous electromagnetic driving potential.

[0016] Preferably, the stator resistance is taken from the average value of the three-phase resistance measured offline in a cold state; the cutoff time constant is determined based on the ratio of the fundamental frequency of the traction motor under rated operating conditions to the multiplier factor, so as to ensure that the cutoff frequency is lower than the fundamental frequency of the motor.

[0017] Preferably, obtaining the theoretical background vibration energy density includes:

[0018] The rotational component is obtained by calculating the square of the mechanical angular velocity and the product of the rotation and vibration transmission coefficients.

[0019] The historical instantaneous electromagnetic driving potential corresponding to the current moment is determined based on the number of hysteresis steps. The electromagnetic component is obtained by multiplying the nonlinear exponent of the historical instantaneous electromagnetic driving potential by the coupling coefficient between electromagnetic and vibration.

[0020] The theoretical background vibration energy density is obtained by calculating the sum of the rotational component, the electromagnetic component, and the system's fundamental noise floor constant.

[0021] This invention effectively simulates the physical transmission delay of a mechanical transmission chain by referencing the instantaneous electromagnetic driving potential at historical moments in the dynamic impedance model. This allows the calculated theoretical background vibration energy density to be precisely aligned with the measured mechanical response on the time axis, eliminating spurious residuals caused by time misalignment.

[0022] Preferably, the method for obtaining the rotation and vibration transmission coefficients and the system's fundamental noise floor constant includes:

[0023] The vehicle was controlled to coast with traction removed, and vibration data was collected at different speeds.

[0024] The least squares method is used to regress and fit the measured vibration energy density obtained from the vibration data, so as to minimize the sum of squared errors between the measured vibration energy density and the theoretical background vibration energy density calculated from the rotation angle, thereby determining the values ​​of the rotation and vibration transmission coefficient and the system's basic noise floor constant.

[0025] This invention utilizes the physical characteristic of no electromagnetic interference under coasting conditions to achieve independent calibration of the transmission coefficients of rotation and vibration related to mechanical rotation, thus avoiding parameter identification errors caused by multivariable coupling.

[0026] Preferably, the method for obtaining the coupling coefficient between electromagnetic fields and vibration includes:

[0027] Control the vehicle to run at a constant speed and apply a stepped traction torque;

[0028] Calculate the remaining vibration energy after subtracting the rotational component and the system's foundation noise floor constant from the measured vibration energy density;

[0029] The relationship between the residual vibration energy and the traction torque is fitted using the least squares method to minimize the fitting error, thereby determining the coupling coefficient between the electromagnetic field and the vibration.

[0030] Preferably, the method for obtaining the number of hysteresis steps includes:

[0031] Apply a torque step signal and record the moment of current step and the moment of vibration envelope peak.

[0032] Calculate the time difference between the peak moment of the vibration envelope and the moment of the current step.

[0033] The time difference is divided by the sampling time interval and rounded to obtain the number of hysteresis steps.

[0034] Preferably, obtaining the normalized fault residual includes:

[0035] Calculate the absolute value of the difference between the measured vibration energy density and the theoretical background vibration energy density;

[0036] The normalized fault residual is obtained by dividing the absolute value by the denominator, wherein the denominator is the sum of the theoretical background vibration energy density and the thermal noise protection constant.

[0037] This invention constructs a normalized fault residual, so that the fault judgment criterion no longer depends on the absolute energy amplitude, but on the degree of deviation of the measured value from the theoretical benchmark. This enables the system to automatically adapt to the energy base fluctuation under different power conditions, avoiding hypersensitive alarms under high power and false alarms under low power.

[0038] Preferably, the method for obtaining the weighted permutation entropy value includes:

[0039] Statistically determine the probability of occurrence of various permutation patterns in the residual time series;

[0040] Calculate the product of the probability of occurrence of each permutation pattern and the logarithm of that probability, and sum the products of all permutations;

[0041] The accumulated result is normalized to obtain the weighted permutation entropy value.

[0042] This invention describes the microscopic arrangement complexity of residual time series by calculating the weighted permutation entropy value, thereby effectively distinguishing between deterministic regular mutations caused by mechanical failures and random disorder fluctuations caused by external environmental interference, and improving the anti-interference capability of the early warning system.

[0043] Preferably, the fault confirmation process includes:

[0044] When the normalized fault residual is less than or equal to the first preset threshold, regardless of the weighted permutation entropy value, the current traction system is determined to be in normal operation and no alarm is triggered.

[0045] When the normalized fault residual is greater than a first preset threshold, it is further classified according to the entropy value:

[0046] If the weighted permutation entropy value is greater than the second preset threshold, the system determines it to be random road surface interference. At this time, only the event is recorded but no fault shutdown alarm is triggered.

[0047] If the weighted permutation entropy value is less than or equal to the second preset threshold, the system determines it to be a mechanical fault and immediately triggers a fault warning signal.

[0048] The present invention has the following beneficial effects:

[0049] This invention does not employ simple filtering methods, but instead constructs a dynamic impedance model that includes phase hysteresis compensation. This model utilizes electrical quantities, i.e., instantaneous electromagnetic driving potential, as prior knowledge, combined with mechanical inertia parameters, to accurately calculate the theoretical background vibration that the motor should produce under the current operating conditions. Subsequently, by subtracting this theoretical benchmark from the measured total energy, the interference of background vibration is eliminated. Thus, regardless of how drastically the operating conditions fluctuate, the system can completely isolate normal high-energy vibrations caused by sudden current changes and mechanical inertia, preventing them from being misjudged as faults.

[0050] Furthermore, after eliminating internal interference, this invention, considering the low-entropy characteristics of mechanical faults with deterministic periodicity and the high-entropy characteristics of road impacts with randomness, introduces weighted permutation entropy to distinguish between mechanical faults and road impacts. This ensures that an alarm is triggered only when the signal energy is sufficiently strong and possesses obvious structural features. This enables the system to intelligently identify genuine faults and ignore external environmental noise interference such as from switches and joints. Attached Figure Description

[0051] Figure 1 This is a flowchart of the steps of the fault early warning method for rail vehicle traction system based on multi-source sensor fusion provided in the embodiments of the present invention;

[0052] Figure 2 This is a schematic diagram comparing the multi-source sensing data provided in the embodiments of the present invention with the theoretical background vibration energy density as a function of working conditions;

[0053] Figure 3 This is a schematic diagram illustrating the effect of prior art fault monitoring provided in an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram illustrating the effect of the fault monitoring provided in an embodiment of the present invention. Detailed Implementation

[0055] Please see Figure 1 The diagram illustrates the steps of the fault early warning method for a rail vehicle traction system based on multi-source sensor fusion provided in Embodiment 1. The method includes the following steps:

[0056] S1: Collect the stator voltage and stator current of the traction motor, and use damping integral calculation and cross product calculation to observe the flux linkage and calculate the torque of the stator voltage and stator current to obtain the instantaneous electromagnetic driving potential.

[0057] It is important to note that, in order to accurately identify faults in subsequent steps, a dynamic reference benchmark capable of describing the normal background vibration under the current operating conditions must first be constructed. Electromagnetic extensor vibration caused by electromagnetic force is an important component of this normal background vibration, thus requiring accurate reconstruction of this vibration. Furthermore, the electromagnetic driving potential, as the excitation source of this extensor vibration, needs to be constructed first as an instantaneous electromagnetic driving potential. Since the calculation of the instantaneous electromagnetic driving potential is highly dependent on the accuracy of the stator flux linkage vector, pure integration calculations using a conventional voltage model are prone to flux linkage trajectory divergence due to accumulated sensor zero drift during the low-speed or startup phases of the traction motor. If the flux linkage is distorted, the calculated instantaneous electromagnetic driving potential will not accurately reflect the internal state of the motor, leading to the failure of the subsequent vibration benchmark. Therefore, the core objective of this step is to introduce a damped integral operator with physical dissipation characteristics to replace the pure integral stage, eliminating DC bias while ensuring the convergence of flux linkage observation, thereby achieving accurate reconstruction of the instantaneous electromagnetic driving potential.

[0058] Preferably, as an example, the stator voltage and stator current of the traction motor are collected, and the flux linkage and torque are calculated using damping integral calculation and cross product calculation to obtain the instantaneous electromagnetic driving potential, including:

[0059] First, the stator voltage and stator current of the traction motor at the current sampling moment are collected using a high-frequency current transformer.

[0060] Next, the stator voltage and stator current are subjected to flux linkage observation processing using a first-order low-pass damping integral operator to obtain the stator flux linkage vector components at the current moment.

[0061] The components of the stator flux linkage vector at the current moment, along the vertical axis of the two-phase stationary coordinate system, specifically satisfy the following relationship:

[0062]

[0063] In the formula, the subscript This represents the index of the current time step in a discrete time series. Indicates the first The components of the stator flux linkage vector at each sampling time on the vertical axis of the two-phase stationary coordinate system. This represents the cutoff time constant of the integrator. Indicates the sampling time interval. Indicates the first Stator flux vector components at each sampling time, Indicates the first The components of the stator voltage at each sampling moment on the horizontal axis of the two-phase stationary coordinate system. Indicates stator resistance. Indicates the first The components of the stator current on the horizontal axis of the two-phase stationary coordinate system at each sampling time. This represents the weighting coefficient.

[0064] Understandably, in relational expressions The term constitutes the historical state decay factor, whose value is slightly less than 1. In the formula... The term constitutes the current back EMF excitation, characterizing the actual rate of change of magnetic flux inside the motor. For the DC bias component caused by sensor zero drift or sampling error, since it remains constant on the time axis, its value will be forced to converge to zero through repeated iterative multiplication of the historical state attenuation factor, thus effectively blocking the accumulation path of integral drift; while for the fundamental component reflecting the normal operation of the motor, since its changing frequency is much higher than the cutoff frequency, its amplitude has been updated before attenuation occurs.

[0065] In this way, by dynamically adjusting the attenuation factor and the excitation term, the high-pass filtering characteristics of the physical system are successfully simulated, ensuring that the calculated stator flux linkage vector components always maintain an ideal circular trajectory on the phase plane without spiral divergence.

[0066] It should be noted that the methods for obtaining the stator resistance and the cutoff time constant include:

[0067] When the motor is offline and cold, the DC resistance of the stator winding of the traction motor is measured using a high-precision bridge, and the average value of the three-phase resistance is taken as the stator resistance.

[0068] The fundamental frequency of the traction motor under rated operating conditions is obtained, and the cutoff time constant is calculated by dividing the multiplication factor by the product of twice pi and the fundamental frequency. In this embodiment, the multiplication factor ranges from five to ten to ensure that the cutoff frequency is much lower than the motor's operating fundamental frequency.

[0069] Similarly, the stator flux linkage vector component in the vertical direction at the current moment is obtained by using the same operator.

[0070] Finally, the cross product operator is used to perform torque calculation on the stator flux vector components and the stator current to obtain the instantaneous electromagnetic driving potential at the current moment, which specifically satisfies the following relationship:

[0071]

[0072] In the formula, Indicates the first The instantaneous electromagnetic driving potential at each sampling moment, Indicates the number of pole pairs of the motor. Indicates the first The components of the stator current on the vertical axis of the two-phase stationary coordinate system at each sampling time. Indicates the first The components of the stator flux linkage vector at each sampling time on the vertical axis of the two-phase stationary coordinate system. Indicates the first The components of the stator current on the horizontal axis of the two-phase stationary coordinate system at each sampling time. Indicates the first The components of the stator flux linkage vector at each sampling time on the vertical axis of the two-phase stationary coordinate system.

[0073] Understandably, this relation, based on the principle of electromagnetic-mechanical decoupling, constructs a pure vibration source observer. The cross product term in the relation... The orthogonal component between the magnetic field vector and the current vector was extracted; this component is the sole source of the tangential electromagnetic force. The input variables in the equation include only electrical quantities such as voltage and current, completely independent of mechanical quantities such as gearbox vibration and bearing wear. This logical relationship of deriving mechanical output from electrical input ensures that the calculated instantaneous electromagnetic driving potential only reflects the theoretical excitation force that should have been generated, without including the actual mechanical response.

[0074] Thus, this indicator naturally shields the interference of friction, gaps, and resonance in the mechanical transmission chain, providing an absolutely pure reference benchmark for subsequent steps to align and compare the background vibration caused by electromagnetic force with the abnormal vibration caused by mechanical faults.

[0075] S2: Obtain the mechanical angular velocity of the traction motor, and use the preset dynamic impedance model to perform nonlinear mapping processing on the mechanical angular velocity and instantaneous electromagnetic driving potential to obtain the theoretical background vibration energy density. The dynamic impedance model includes phase hysteresis compensation parameters for the inertial effect of the mechanical transmission chain.

[0076] It should be noted that, in order to accurately identify faults in subsequent steps, a dynamic reference benchmark, namely the theoretical background vibration energy density, must first be constructed to describe the normal background vibration under the current operating conditions in real time. The vibration energy sources of rail vehicles mainly include the mechanical fundamental frequency vibration caused by rotor rotation and the electromagnetic stretching vibration caused by electromagnetic force. Therefore, the theoretical background vibration energy density can be constructed based on this.

[0077] Preferably, as an example, the mechanical angular velocity of the traction motor is obtained, and a preset dynamic impedance model is used to perform nonlinear mapping processing on the mechanical angular velocity and the instantaneous electromagnetic driving potential to obtain the theoretical background vibration energy density. The dynamic impedance model includes phase hysteresis compensation parameters for the inertial effect of the mechanical transmission chain, including:

[0078] By using a pre-defined dynamic impedance model to perform nonlinear mapping on the mechanical angular velocity and the instantaneous electromagnetic driving potential at historical moments, the theoretical background vibration energy density at the current moment is obtained, which specifically satisfies the following relationship:

[0079]

[0080] In the formula, Indicates the first The theoretical background vibration energy density at each sampling time. The transmission coefficient between rotation and vibration. Indicates the first Mechanical angular velocity at each sampling time, This represents the coupling coefficient between electromagnetic fields and vibration. Indicates the first The instantaneous electromagnetic driving potential at each sampling moment, This represents the magnetostrictive nonlinearity index. Represents the system's fundamental noise floor constant. Indicates the number of hysteresis steps. For dynamic impedance model, Represents the rotation component. Represents electromagnetic components.

[0081] Understandably, this relationship is a multiphysics coupling model that describes three orthogonal physical sources of vibration energy in rail vehicles. The term characterizes the mechanical rotation effect. Since the centrifugal force is proportional to the square of the rotational speed, this term accurately describes the rotor imbalance and the background vibration basis generated by the bearing balls when rotating at high speed. The term characterizes the electromagnetic stretching effect, where the subscript... It enables time rewinding, simulating the physical delay of force transmission from the air gap to the casing, ensuring that the calculation conforms to the causal law of excitation preceding response, thereby effectively eliminating spatiotemporal misalignment under variable operating conditions such as rapid traction or rapid braking; exponential This fits the nonlinear relationship between stress and magnetic field strength in the magnetic saturation region of ferromagnetic materials; The term characterizes the system's fundamental noise floor constant, encompassing steady-state components such as wheel-rail contact and aerodynamic noise.

[0082] In this way, by superimposing these three nonlinear terms, a model that changes in real time with the working conditions is constructed, enabling accurate prediction of normal physical vibrations.

[0083] It should be noted that the methods for obtaining the transmission coefficients of rotation and vibration, the system's fundamental noise floor constant, the coupling coefficients of electromagnetics and vibration, and the magnetostrictive nonlinear index in the dynamic impedance model include:

[0084] The vehicle is controlled to coast on a straight track with traction removed. An accelerometer mounted on the motor surface collects vibration data at different speeds. A preset sliding window is moved across the vibration data, and after each movement, the sum of the squares of the vibration data within the window is calculated to obtain the measured vibration energy density. Based on the measured vibration energy density, a first objective function is constructed. The least squares method is used to regress and fit the data to minimize the first objective function, thereby obtaining the rotational transfer coefficient and the system's fundamental noise floor constant. The first objective function satisfies the following relationship:

[0085]

[0086] In the formula, each vibration data point is recorded as a sample point. The first objective function represents the inertial gliding calibration process; This indicates the total number of sample points collected; Indicates the sample point index; Indicates the first Measured vibration energy density at each sample point; This represents the transmission coefficients of rotation and vibration to be solved; Indicates the first The mechanical angular velocity corresponding to each sample point; This represents the fundamental noise floor constant of the system to be solved. This represents the theoretical background vibration energy density under conditions without traction.

[0087] The vehicle is controlled to operate at a constant speed, and a stepped traction torque is applied. An accelerometer mounted on the motor surface collects vibration data at different speeds. A preset sliding window is moved across the vibration data, and after each movement, the sum of the squares of the vibration data within the preset sliding window is calculated to obtain the measured vibration energy density for each vibration data point. The known transmission coefficients of rotation and vibration, and the system's fundamental noise floor constant are substituted into a second objective function to fit the relationship between the remaining vibration energy and torque. Minimizing this second objective function yields the electromagnetic-vibration coupling coefficient and the magnetostrictive nonlinearity exponent. The second objective function satisfies the following relationship:

[0088]

[0089] In the formula, each vibration data point is recorded as a sample point. The second objective function represents the steady-state loading calibration process; This indicates the total number of sample points collected; Indicates the first The measured vibration energy density collected from each sample point; Represents the known transmission coefficients of rotation and vibration; Indicates the first The mechanical angular velocity corresponding to each sample point; This represents the known system fundamental noise floor constant; This represents the coupling coefficient between the electromagnetic field and the vibration to be solved; Indicates the first The traction torque corresponding to each sample point; The term represents the magnetostrictive nonlinear index to be solved. In this embodiment, the search range for solving the magnetostrictive nonlinear index is set to 1.5 to 2.5.

[0090] It should also be noted that the methods for obtaining the hysteresis steps include:

[0091] A torque step signal is applied during bench testing, and the time of current step and the time of vibration envelope peak are recorded. The time difference between the vibration envelope peak and the current step is calculated. The time difference is divided by the sampling time interval and rounded to obtain the hysteresis step number.

[0092] S3: Collect the measured vibration acceleration signal of the traction system and calculate the measured vibration energy density. Use the dimensionless signal-to-noise ratio operator to calculate the degree of difference between the measured vibration energy density and the theoretical background vibration energy density, and obtain the normalized fault residual.

[0093] It should be noted that in actual operation, the measured vibration energy density is the sum of normal background vibration and abnormal fault impact. Currently, the theoretical background vibration energy density under the current operating conditions has been accurately predicted. Therefore, by performing a difference operation between the measured and theoretical values, the normal vibration components caused by rotational speed and electromagnetic force can be eliminated, and the remaining residual is the pure fault energy. However, obtaining only the absolute fault energy is insufficient, because the same fault energy is extremely significant at low speeds but may be submerged at high speeds. Therefore, this step further compares the extracted fault energy with the current background energy baseline to construct a normalized index, thereby achieving a relative measurement of fault severity.

[0094] Preferably, as an example, the measured vibration acceleration signal of the traction system is acquired and the measured vibration energy density is calculated. The difference between the measured vibration energy density and the theoretical background vibration energy density is calculated using a dimensionless signal-to-noise ratio operator, resulting in the normalized fault residual, including:

[0095] First, during the operation of the rail vehicle, vibration data is collected using an accelerometer installed on the surface of the motor. The sliding window is aligned with the vibration data at the current moment, and the sum of the squares of the vibration data within the sliding window is calculated to obtain the measured vibration energy density at the current moment.

[0096] Next, the measured vibration energy density and the theoretical background vibration energy density are differentially normalized to obtain the normalized fault residual at the current moment, which specifically satisfies the following relationship:

[0097]

[0098] In the formula, Indicates the first Normalized fault residuals at each sampling time. Indicates the first The measured vibration energy density at each sampling time. Indicates the first The theoretical background vibration energy density at each sampling time. This represents the thermal noise protection constant.

[0099] Understandably, the numerator extracts the residual energy in the measured signal that cannot be explained by the theoretical model; this energy represents the potential fault component. The denominator acts as a dynamic normalization factor: under high-speed, heavy-load, high-energy conditions, the denominator is large, making it insensitive to normal fluctuations in large values ​​and preventing false alarms; under low-speed, no-load, low-energy conditions, the denominator is small, ensuring sensitivity to minor fault impacts and preventing missed alarms. (Constant) The introduction of this feature effectively avoids the problem of the denominator being zero when the vehicle is stationary, and also sets a minimum sensitivity limit.

[0100] It should be noted that the methods for obtaining the thermal noise protection constant include:

[0101] The vehicle is kept in a standby state with no high voltage input and the background noise signal output by the acceleration sensor is collected; the variance of the background noise signal is calculated and recorded as the system static noise floor variance; the thermal noise protection constant is set as a preset multiple of the system static noise floor variance, and in this embodiment, the preset multiple is 1.5.

[0102] S4: Construct a residual time series based on continuous normalized fault residuals, calculate the weighted permutation entropy value of the residual time series using the weighted permutation entropy algorithm, and perform fault confirmation processing on the normalized fault residuals and weighted permutation entropy value using double threshold logic to achieve fault early warning of rail vehicle traction system.

[0103] It should be noted that during the operation of rail vehicles, impacts from road switches, joints, or foreign objects often generate extremely high instantaneous energy, causing a sudden increase in fault residuals. Judging solely based on residual amplitude can easily lead to misreporting such external interference as mechanical faults. However, these two phenomena differ fundamentally in their signal topology: external road impacts are nondeterministic random processes, producing chaotic data sequences similar to white noise, exhibiting disorder; internal mechanical faults, on the other hand, are deterministic dynamic processes, with accompanying impacts often following specific physical periods, exhibiting a kind of implicit order. Therefore, this step introduces a novel observation dimension—microscopic order—to address the challenge of distinguishing between random external road impacts and internal mechanical structural faults by relying solely on energy amplitude.

[0104] Preferably, as an example, a residual time series is constructed based on continuous normalized fault residuals. The weighted permutation entropy value of the residual time series is calculated using a weighted permutation entropy algorithm. A dual-threshold logic is then used to perform fault confirmation processing on the normalized fault residuals and the weighted permutation entropy value to achieve fault early warning for the rail vehicle traction system, including:

[0105] First, normalized fault residuals from several historical moments prior to the current moment are obtained. These, along with the normalized fault residuals from the current moment, are then arranged chronologically to obtain a residual time series. The weighted permutation entropy algorithm is then used to perform microscopic disorder measurement on the residual time series, yielding a weighted permutation entropy value that satisfies the following relationship:

[0106]

[0107] In the formula, Indicates the first The weighted permutation entropy value at each sampling time. The embedding dimension represents the permutation entropy. The factorial represents the embedding dimension, which describes the total number of all possible permutations. Indicates the first The weighted probability of each permutation pattern occurring in the residual time series. This is the index variable for the arrangement pattern.

[0108] Understandably, this relation measures the inherent complexity and information density of the residual time series from an information theory perspective. When a vehicle encounters road irregularities such as switches or joint impacts, the resulting residual time series is a completely random white noise process, with the probability of various permutations occurring tending to be equal. In this case, the calculated entropy value approaches its maximum value of 1, representing a disordered and chaotic state of the system. Conversely, when a vehicle experiences mechanical failures such as broken gear teeth or detached bearings, periodic impact pulses appear in the residual time series, causing specific permutations to occur frequently, resulting in a significant skew in the probability distribution. In this case, the entropy value decreases significantly. By analyzing the entropy dimension, we can penetrate the surface appearance of energy amplitude and capture the underlying dynamic differences in the signal, thereby accurately eliminating environmental interference.

[0109] It should be noted that the method for obtaining the embedding dimension includes: selecting it based on the ratio between the sampling frequency and the fault characteristic frequency; in this embodiment, the embedding dimension is set to 4.

[0110] Next, a fault confirmation process is performed on the normalized fault residual and the weighted permutation entropy value using dual-threshold logic to obtain an early warning signal. The specific decision logic is as follows:

[0111] When the normalized fault residual is less than or equal to the first preset threshold, regardless of the weighted permutation entropy value, the current traction system is determined to be in normal operation and no alarm is triggered.

[0112] When the normalized fault residual is greater than a first preset threshold, it is further classified according to the entropy value:

[0113] If the weighted permutation entropy value is greater than the second preset threshold, the system determines it to be random road surface interference. At this time, only the event is recorded but no fault shutdown alarm is triggered.

[0114] If the weighted permutation entropy value is less than or equal to the second preset threshold, the system determines it to be a mechanical fault and immediately triggers a fault warning signal.

[0115] It should be noted that the methods for obtaining the first preset threshold and the second preset threshold include:

[0116] Collect sample data of vehicles running in a healthy state, and calculate the normalized fault residual and weighted permutation entropy value corresponding to each sample data; statistically analyze the probability distribution of the normalized fault residual of all sample data, and take its 99.7 percentile as the first preset threshold; statistically analyze the probability distribution of the weighted permutation entropy value of all sample data, and take its 5 percentile as the second preset threshold.

[0117] To demonstrate the effectiveness of the solution, relevant experiments were conducted. Below are the images obtained from the experiments:

[0118] Figure 2 This diagram illustrates the comparison between multi-source sensor data and theoretical background vibration energy density under varying operating conditions. The dashed curve represents the mechanical angular velocity variation of the traction motor of the rail vehicle, demonstrating the complete variable operating condition process from standstill to rapid acceleration, coasting, braking, and finally to steady speed. The solid curve represents the measured vibration energy density, and the dotted curve represents the theoretical background vibration energy density. The shaded area represents the time interval during which actual mechanical failures occurred.

[0119] pass Figure 2 As can be seen, during the rapid acceleration phase on the left, due to the large torque input, both the theoretical background vibration (shown by the dotted line) and the measured vibration (shown by the solid line) exhibit huge peak values. At this point, although a fault impact exists within the first shaded area, this fault energy is completely enveloped and submerged by the high-amplitude background energy, making it difficult to distinguish between normal high-current vibration and fault impact with the naked eye at the raw data level. This proves that relying solely on the raw vibration amplitude is insufficient to identify faults under severe operating conditions, establishing the necessity of introducing background decoupling.

[0120] Figure 3 This is a schematic diagram of the effect of existing technology for fault monitoring. The solid curve represents the measured vibration energy density, and the dotted and dashed lines represent the traditional fixed thresholds set in the existing technology.

[0121] pass Figure 3 It can be seen that during the acceleration phase on the left, the measured vibration energy significantly exceeded the fixed threshold due to normal electromagnetic excitation, causing the system to continuously issue false alarms even when there was no fault.

[0122] Figure 4 This diagram illustrates the effect of fault monitoring in this invention. The solid curve represents the normalized fault residual, characterizing the degree of pure fault after stripping away the background of normal operating conditions. The dashed line represents the first preset threshold set by this invention.

[0123] pass Figure 4As can be seen, during the acceleration phase on the left, the previously huge energy peak disappeared, and the normalized fault residual smoothly approached the zero axis. This indicates that the dynamic impedance model accurately predicted and counteracted normal operating interference, achieving suppression of background noise. Furthermore, regardless of whether the fault was located in the high-energy region on the left or the low-energy region on the right, the normalized fault residual exhibited clear and independent spikes, with highly consistent amplitudes. Both fault spikes significantly exceeded the first preset threshold. The fault warning system of this invention can achieve zero false alarms and zero false negatives.

[0124] This concludes the embodiment.

[0125] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fault early warning method for rail vehicle traction systems based on multi-source sensor fusion, characterized in that, include: The stator voltage and stator current of the traction motor are collected. The flux linkage and torque are calculated by damping integral calculation and cross product calculation to obtain the instantaneous electromagnetic driving potential. The mechanical angular velocity of the traction motor is obtained, and the mechanical angular velocity and instantaneous electromagnetic driving potential are nonlinearly mapped using a preset dynamic impedance model to obtain the theoretical background vibration energy density. The dynamic impedance model includes phase hysteresis compensation parameters for the inertial effect of the mechanical transmission chain. The measured vibration acceleration signal of the traction system is collected and the measured vibration energy density is calculated. The difference between the measured vibration energy density and the theoretical background vibration energy density is calculated using a dimensionless signal-to-noise ratio operator to obtain the normalized fault residual. Based on continuous normalized fault residuals, a residual time series is constructed. The weighted permutation entropy value of the residual time series is calculated using the weighted permutation entropy algorithm. The normalized fault residuals and the weighted permutation entropy value are processed for fault confirmation using dual threshold logic to achieve fault early warning of rail vehicle traction system.

2. The fault early warning method for rail vehicle traction system based on multi-source sensor fusion according to claim 1, characterized in that, The process of obtaining the instantaneous electromagnetic driving potential includes: Weighting coefficients are constructed using the cutoff time constant of the integrator. Based on the weighting coefficients, the sampling time interval, and the stator resistance, the stator voltage, stator current, and the stator flux vector component at the current moment are calculated using weighted iterative calculation to obtain the stator flux vector component at the current moment. The difference between the cross product of the stator flux vector component and the stator current in the two-phase stationary coordinate system is calculated, and the difference is multiplied by a coefficient related to the number of motor pole pairs to obtain the instantaneous electromagnetic driving potential.

3. The fault early warning method for rail vehicle traction system based on multi-source sensor fusion according to claim 2, characterized in that, The stator resistance is taken from the average value of the three-phase resistance measured offline in a cold state; the cutoff time constant is determined based on the ratio of the fundamental frequency of the traction motor under rated operating conditions to the multiplier factor, so as to ensure that the cutoff frequency is lower than the motor's operating fundamental frequency.

4. The fault early warning method for rail vehicle traction system based on multi-source sensor fusion according to claim 1, characterized in that, The obtained theoretical background vibration energy density includes: The rotational component is obtained by calculating the square of the mechanical angular velocity and the product of the rotation and vibration transmission coefficients. The historical instantaneous electromagnetic driving potential corresponding to the current moment is determined based on the number of hysteresis steps. The electromagnetic component is obtained by multiplying the nonlinear exponent of the historical instantaneous electromagnetic driving potential by the coupling coefficient between electromagnetic and vibration. The theoretical background vibration energy density is obtained by calculating the sum of the rotational component, the electromagnetic component, and the system's fundamental noise floor constant.

5. The fault early warning method for rail vehicle traction system based on multi-source sensor fusion according to claim 4, characterized in that, The methods for obtaining the rotation and vibration transmission coefficients and the system's fundamental noise floor constant include: The vehicle was controlled to coast with traction removed, and vibration data was collected at different speeds. The least squares method is used to regress and fit the measured vibration energy density obtained from the vibration data, so as to minimize the sum of squared errors between the measured vibration energy density and the theoretical background vibration energy density calculated from the rotation angle, thereby determining the values ​​of the rotation and vibration transmission coefficient and the system's basic noise floor constant.

6. The fault early warning method for rail vehicle traction system based on multi-source sensor fusion according to claim 4, characterized in that, The method for obtaining the coupling coefficient between electromagnetic fields and vibration includes: Control the vehicle to run at a constant speed and apply a stepped traction torque; Calculate the remaining vibration energy after subtracting the rotational component and the system's foundation noise floor constant from the measured vibration energy density; The relationship between the residual vibration energy and the traction torque is fitted using the least squares method to minimize the fitting error, thereby determining the coupling coefficient between the electromagnetic field and the vibration.

7. The fault early warning method for rail vehicle traction system based on multi-source sensor fusion according to claim 5, characterized in that, The method for obtaining the number of hysteresis steps includes: Apply a torque step signal and record the moment of current step and the moment of vibration envelope peak. Calculate the time difference between the peak moment of the vibration envelope and the moment of the current step. The time difference is divided by the sampling time interval and rounded to obtain the number of hysteresis steps.

8. The fault early warning method for rail vehicle traction system based on multi-source sensor fusion according to claim 1, characterized in that, The obtained normalized fault residual includes: Calculate the absolute value of the difference between the measured vibration energy density and the theoretical background vibration energy density; The normalized fault residual is obtained by dividing the absolute value by the denominator, wherein the denominator is the sum of the theoretical background vibration energy density and the thermal noise protection constant.

9. The fault early warning method for rail vehicle traction system based on multi-source sensor fusion according to claim 1, characterized in that, The method for obtaining the weighted permutation entropy value includes: Statistically determine the probability of occurrence of various permutation patterns in the residual time series; Calculate the product of the probability of occurrence of each permutation pattern and the logarithm of that probability, and sum the products of all permutations; The accumulated result is normalized to obtain the weighted permutation entropy value.

10. The fault early warning method for rail vehicle traction system based on multi-source sensor fusion according to claim 1, characterized in that, The fault confirmation process includes: When the normalized fault residual is less than or equal to the first preset threshold, regardless of the weighted permutation entropy value, the current traction system is determined to be in normal operation and no alarm is triggered. When the normalized fault residual is greater than a first preset threshold, it is further classified according to the entropy value: If the weighted permutation entropy value is greater than the second preset threshold, the system determines it to be random road surface interference. At this time, only the event is recorded but no fault shutdown alarm is triggered. If the weighted permutation entropy value is less than or equal to the second preset threshold, the system determines it to be a mechanical fault and immediately triggers a fault warning signal.

Images