A train derailment risk diagnosis system based on bogie fault feature analysis

Abstract

The present application belongs to the technical field of rail transit train operation safety monitoring, and specifically discloses a train derailment risk diagnosis system based on bogie fault feature analysis, which comprises a fault feature extraction module, a fault feature calculation module, a derailment risk diagnosis module and a diagnosis output module. The fault feature extraction module maps the collected vibration acceleration signals and bearing temperature signals into equivalent derailment coefficients and equivalent wheel load reduction rates. The fault feature calculation module calculates the change rates of the equivalent derailment coefficients and equivalent wheel load reduction rates and combines them into a bogie fault feature vector. The derailment risk diagnosis module obtains the safety residual time in two dimensions according to the difference between the safety boundary and the current value divided by the change rate, and takes the smaller value as the derailment risk diagnosis result. The diagnosis output module outputs a derailment early warning signal. The present application realizes quantitative prediction of derailment risk and direct output of safety residual time, facilitates advance maintenance arrangement and reduces the risk of derailment accidents.

Description

Technical Field

[0001] This invention belongs to the field of rail transit train operation safety monitoring technology, specifically relating to a train derailment risk diagnosis system based on bogie fault characteristic analysis. Background Technology

[0002] The bogie of a rail transit train is a critical running gear that bears the weight of the car body and transmits wheel-rail forces; its health directly determines the safety of train operation. Derailment accidents are mainly caused by faults such as bogie bearing wear, gear damage, and suspension system degradation. These faults lead to the deterioration of wheel-rail dynamic parameters, and derailment occurs when the derailment coefficient or wheel load reduction rate exceeds safety limits. Existing technologies typically monitor the running gear condition by installing vibration and temperature sensors on key parts of the bogie, outputting alarm signals when monitored parameters exceed preset thresholds, or by directly measuring the derailment coefficient and wheel load reduction rate using wheel-rail force sensors installed on the wheelsets for real-time judgment.

[0003] However, existing technologies suffer from the following fundamental flaws: traditional methods can only determine whether the current running gear status exceeds a safety threshold, failing to address the core issue of how much safe operating time remains before derailment occurs. Since derailment is the result of cumulative fault evolution, simply monitoring current status parameters cannot predict the development trend of derailment risk, leaving maintenance personnel only able to passively respond to alarm signals and unable to proactively arrange preventative maintenance. Furthermore, wheel-rail force sensors are expensive and require regular calibration, making large-scale deployment on operating trains difficult; while diagnostic methods based on indirect signals such as vibration and temperature lack the ability to accurately map sensor data to derailment risk. Therefore, how to achieve early quantitative prediction of derailment risk and accurate output of remaining safe operating time without adding expensive wheel-rail force sensors remains a long-standing and unresolved technical challenge in this field.

[0004] Therefore, this invention proposes a train derailment risk diagnosis system based on bogie fault characteristic analysis. Summary of the Invention

[0005] This invention provides a train derailment risk diagnosis system based on bogie fault characteristic analysis to solve at least one of the technical problems mentioned above.

[0006] To address the aforementioned technical problems, this invention discloses a train derailment risk diagnosis system based on bogie fault characteristic analysis, comprising:

[0007] The fault feature extraction module is used to collect vibration acceleration signals and bearing temperature signals of the bogie axle box. The vibration acceleration signals and bearing temperature signals are input into a pre-established state mapping model. The state mapping model outputs the corresponding equivalent derailment coefficient and equivalent wheel load reduction rate based on the current vibration acceleration signal and the current bearing temperature signal.

[0008] The fault feature calculation module is used to fit the equivalent derailment coefficient obtained from multiple consecutive sampling times into a first time evolution curve, calculate the first derivative of the first time evolution curve at the current time as the derailment coefficient margin consumption rate, fit the equivalent wheel load reduction rate obtained from multiple consecutive sampling times into a second time evolution curve, calculate the first derivative of the second time evolution curve at the current time as the wheel load reduction rate margin consumption rate, and combine the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate into a bogie fault feature vector.

[0009] The derailment risk diagnosis module is used to calculate the remaining safe time in the derailment coefficient dimension when the derailment coefficient margin consumption rate is greater than zero. This is done by dividing the difference between the derailment coefficient safety boundary and the current equivalent derailment coefficient by the derailment coefficient margin consumption rate in the bogie fault feature vector. When the derailment coefficient margin consumption rate is not greater than zero, the remaining safe time in the derailment coefficient dimension is recorded as the preset maximum safe time. Similarly, the module calculates the remaining safe time in the wheel load reduction rate dimension by dividing the difference between the wheel load reduction rate safety boundary and the current equivalent wheel load reduction rate by the wheel load reduction rate margin consumption rate in the bogie fault feature vector. When the wheel load reduction rate margin consumption rate is not greater than zero, the remaining safe time in the wheel load reduction rate dimension is recorded as the preset maximum safe time. The smaller of the remaining safe time in the derailment coefficient dimension and the remaining safe time in the wheel load reduction rate dimension is taken as the derailment risk diagnosis result.

[0010] The diagnostic output module is used to output a derailment warning signal when the derailment risk diagnosis result is less than a preset time threshold, and to display the bogie fault feature vector and the derailment risk diagnosis result together on the driver's cab display screen.

[0011] Preferably, the state mapping model in the fault feature extraction module adopts a dual-channel neural network architecture, which includes a vibration feature processing channel and a temperature feature processing channel.

[0012] The vibration feature processing channel is composed of multiple convolutional layers and multiple pooling layers stacked together. The input of the vibration feature processing channel is the time-frequency plot of the vibration acceleration signal, and the output of the vibration feature processing channel is the vibration deep feature vector.

[0013] The temperature feature processing channel is composed of multiple fully connected layers stacked together. The input of the temperature feature processing channel is the time series of the bearing temperature signal, and the output of the temperature feature processing channel is the deep temperature feature vector.

[0014] The dual-channel neural network architecture also includes a feature fusion layer and a regression output layer. The feature fusion layer concatenates the deep vibration feature vector and the deep temperature feature vector and inputs them into the regression output layer. The regression output layer outputs the equivalent derailment coefficient and the equivalent wheel load reduction rate. During the training process of the dual-channel neural network architecture, the measured derailment coefficient and the measured wheel load reduction rate used to calculate the output error are obtained by a wheel-rail force measuring device installed on the bogie wheelset.

[0015] Preferably, the training process of the dual-channel neural network architecture introduces an adversarial domain adaptation mechanism, which is used to eliminate the distribution differences of vibration acceleration signals under different line conditions and the distribution differences of bearing temperature signals under different ambient temperatures.

[0016] The adversarial domain adaptation mechanism includes a feature extractor, a domain discriminator, and a regressor. The feature extractor maps the input signal to a common feature space. The domain discriminator is used to determine which line conditions or which ambient temperature range the features in the common feature space originate from. The regressor is used to regress the equivalent derailment coefficient and the equivalent wheel load reduction rate from the features in the common feature space.

[0017] During the training process of the dual-channel neural network architecture, the optimization objective of the feature extractor is to minimize the prediction error of the regressor and maximize the discrimination error of the domain discriminator. The optimization objective of the domain discriminator is to minimize the domain discrimination error. Through adversarial training, the feature distribution in the common feature space is decoupled from the line conditions and ambient temperature.

[0018] Preferably, the fault feature calculation module fits the equivalent derailment coefficients obtained at multiple consecutive sampling times into a first-time evolution curve by modeling the sequence of equivalent derailment coefficients changing with time as a linear time-varying system, and using a strong tracking Kalman filter to estimate the state of the linear time-varying system online.

[0019] The equivalent derailment coefficient is processed using a first strong tracking Kalman filter. The state vector of the first strong tracking Kalman filter includes the equivalent derailment coefficient, the first derivative of the equivalent derailment coefficient, and the second derivative of the equivalent derailment coefficient. The equivalent wheel load reduction rate is processed using a second strong tracking Kalman filter. The state vector of the second strong tracking Kalman filter includes the equivalent wheel load reduction rate, the first derivative of the equivalent wheel load reduction rate, and the second derivative of the equivalent wheel load reduction rate.

[0020] The fading factor of the strong tracking Kalman filter is adjusted in real time according to the covariance matrix of the state estimation error, so that the strong tracking Kalman filter can maintain its adaptive tracking capability to the real state when the equivalent derailment coefficient changes abruptly.

[0021] The first derivative component in the state estimate output by the strong tracking Kalman filter is used as the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate, while the second derivative component is used as the derailment coefficient margin consumption acceleration and the wheel load reduction rate margin consumption acceleration.

[0022] Preferably, the real-time adjustment method for the fading factor in the strong tracking Kalman filter is as follows:

[0023] Calculate the autocorrelation matrix of the state estimation residual sequence, and use the ratio of the trace of the autocorrelation matrix to the preset residual threshold as the basic fading factor;

[0024] When the absolute value of the second derivative of the equivalent derailment coefficient exceeds the preset threshold for the second derivative of the derailment coefficient or the absolute value of the second derivative of the equivalent wheel load reduction rate exceeds the preset threshold for the second derivative of the wheel load reduction rate, the basic fading factor is multiplied by the preset amplification factor to obtain the final fading factor. At the same time, the derailment coefficient margin consumption rate is added to the derailment coefficient margin consumption acceleration and multiplied by the preset prediction time step to obtain the predicted derailment coefficient margin consumption rate. The wheel load reduction rate margin consumption rate is added to the wheel load reduction rate margin consumption acceleration and multiplied by the preset prediction time step to obtain the predicted wheel load reduction rate margin consumption rate. The predicted derailment coefficient margin consumption rate and the predicted wheel load reduction rate margin consumption rate are used to replace the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate in the bogie fault feature vector.

[0025] When the absolute value of the second derivative of the equivalent derailment coefficient does not exceed the preset threshold for the second derivative of the derailment coefficient and the absolute value of the second derivative of the equivalent wheel load reduction rate does not exceed the preset threshold for the second derivative of the wheel load reduction rate, the basic fading factor is directly used as the final fading factor, keeping the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate in the bogie fault feature vector unchanged.

[0026] Preferably, after the derailment risk diagnosis module takes the smaller of the safe remaining time in the derailment coefficient dimension and the safe remaining time in the wheel load reduction rate dimension as the derailment risk diagnosis result, it further corrects the derailment risk diagnosis result based on the ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate margin consumption rate in the bogie fault feature vector.

[0027] The correction method for the derailment risk diagnosis result is as follows: when the ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate margin consumption rate exceeds the preset normal ratio range, the derailment risk diagnosis result is divided by a penalty coefficient greater than 1. The magnitude of the penalty coefficient is positively correlated with the degree to which the ratio deviates from the preset normal ratio range.

[0028] The preset normal ratio range is determined based on the statistical distribution of the ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate consumption rate under the historical normal operating conditions of the bogie.

[0029] Preferably, the penalty coefficient is calculated as follows:

[0030] Calculate the current ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate margin consumption rate, and calculate the deviation ratio of the current ratio from the center value of the preset normal ratio range.

[0031] The square of the deviation multiplier is used as the base penalty value, and the smaller of the base penalty value and the preset penalty upper limit value is used as the penalty coefficient.

[0032] Preferably, it also includes a fault type identification module, which is used to input the bogie fault feature vector into a pre-trained fault classifier when the derailment risk diagnosis result is less than a preset time threshold, and the fault classifier outputs the fault type identification result of the current bogie.

[0033] The fault classifier uses a support vector machine, and the kernel function of the support vector machine is a radial basis function. The width parameter of the radial basis function is determined based on the average Euclidean distance between the bogie fault feature vector samples.

[0034] The fault type identification results include bearing fault categories, gear fault categories, suspension fault categories, and no fault categories.

[0035] Preferably, the training samples for the fault classifier are amplified using synthetic minority class oversampling technology to balance the number of samples for each fault type.

[0036] The amplification method of synthetic minority class oversampling technology is as follows: for fault types with fewer than a preset threshold number of samples, two adjacent samples of the fault type are selected in the bogie fault feature vector space, and new samples are generated by random interpolation on the line connecting the two adjacent samples. The interpolation operation is repeated until the number of samples of the fault type reaches the preset threshold number of samples.

[0037] Preferably, it also includes a prediction uncertainty quantification module, which is used to output a first prediction interval of the equivalent derailment coefficient and a second prediction interval of the equivalent wheel load reduction rate while the state mapping model outputs the equivalent derailment coefficient and the equivalent wheel load reduction rate.

[0038] The first and second prediction intervals were obtained through multiple forward inferences after randomly discarding some neurons in the state mapping model.

[0039] When calculating the derailment risk diagnosis results, the derailment risk diagnosis module uses the upper boundary of the first prediction interval to replace the equivalent derailment coefficient and the upper boundary of the second prediction interval to replace the equivalent wheel load reduction rate, thus obtaining a conservative estimate of the derailment risk diagnosis results.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] Existing train derailment risk diagnosis technologies can only monitor whether the current state of the bogie exceeds a preset threshold using sensors such as vibration and temperature sensors, or directly measure the derailment coefficient and wheel load reduction rate using expensive wheel-rail force sensors for real-time judgment. This has two fundamental drawbacks: first, it cannot answer the core question of "how much safe operating time is left from the current state until derailment occurs," leading maintenance personnel to only passively respond to alarms and unable to proactively arrange preventative maintenance; second, it lacks the ability to accurately map low-cost vibration and temperature signals to derailment risk. This invention constructs a state mapping model to convert vibration acceleration signals and bearing temperature signals into equivalent derailment coefficients and equivalent wheel load reduction rates. It then calculates the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate, ultimately outputting the smaller of the safe remaining time in the derailment coefficient dimension and the safe remaining time in the wheel load reduction rate dimension as the derailment risk diagnosis result. This achieves a leap from "current state monitoring" to "safe remaining time prediction," enabling maintenance personnel to know in advance the remaining operating time before a derailment occurs, thereby proactively arranging maintenance plans and significantly reducing the incidence of derailment accidents and the risk of line operation interruptions. Attached Figure Description

[0042] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0043] Figure 1 This is a module architecture diagram of the train derailment risk diagnosis system in an embodiment of the present invention;

[0044] Figure 2 This is an internal logic diagram of the derailment risk diagnosis module in an embodiment of the present invention;

[0045] Figure 3 This is a flowchart of signal processing and calculation in an embodiment of the present invention. Detailed Implementation

[0046] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0047] Furthermore, in this invention, the use of terms such as "first" and "second" is for descriptive purposes only and does not specifically refer to any order or sequence, nor is it intended to limit the invention. They are merely used to distinguish components or operations described using the same technical terms and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions and features of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If a combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0048] The present invention provides the following embodiments.

[0049] Example 1

[0050] This invention provides a train derailment risk diagnosis system based on bogie fault feature analysis, such as... Figure 1-3 As shown, it includes:

[0051] The fault feature extraction module is used to collect vibration acceleration signals and bearing temperature signals of the bogie axle box. The vibration acceleration signals and bearing temperature signals are input into a pre-established state mapping model. The state mapping model outputs the corresponding equivalent derailment coefficient and equivalent wheel load reduction rate based on the current vibration acceleration signal and the current bearing temperature signal.

[0052] The fault feature calculation module is used to fit the equivalent derailment coefficient obtained from multiple consecutive sampling times into a first time evolution curve, calculate the first derivative of the first time evolution curve at the current time as the derailment coefficient margin consumption rate, fit the equivalent wheel load reduction rate obtained from multiple consecutive sampling times into a second time evolution curve, calculate the first derivative of the second time evolution curve at the current time as the wheel load reduction rate margin consumption rate, and combine the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate into a bogie fault feature vector.

[0053] The derailment risk diagnosis module is used to calculate the remaining safe time in the derailment coefficient dimension when the derailment coefficient margin consumption rate is greater than zero. This is done by dividing the difference between the derailment coefficient safety boundary and the current equivalent derailment coefficient by the derailment coefficient margin consumption rate in the bogie fault feature vector. When the derailment coefficient margin consumption rate is not greater than zero, the remaining safe time in the derailment coefficient dimension is recorded as the preset maximum safe time. Similarly, the module calculates the remaining safe time in the wheel load reduction rate dimension by dividing the difference between the wheel load reduction rate safety boundary and the current equivalent wheel load reduction rate by the wheel load reduction rate margin consumption rate in the bogie fault feature vector. When the wheel load reduction rate margin consumption rate is not greater than zero, the remaining safe time in the wheel load reduction rate dimension is recorded as the preset maximum safe time. The smaller of the remaining safe time in the derailment coefficient dimension and the remaining safe time in the wheel load reduction rate dimension is taken as the derailment risk diagnosis result.

[0054] The diagnostic output module is used to output a derailment warning signal when the derailment risk diagnosis result is less than a preset time threshold, and to display the bogie fault feature vector and the derailment risk diagnosis result together on the driver's cab display screen.

[0055] The working principle and beneficial effects of the above technical solution are as follows:

[0056] The state mapping model employs a dual-channel neural network architecture. The vibration feature processing channel consists of three convolutional layers and three max-pooling layers stacked alternately, with each convolutional layer using a modified linear unit (MRU) as the activation function. The input to this channel is the short-time Fourier transform (SFT) time-frequency plot of the vibration acceleration signal, with a size of 128×128 pixels, and the output is a 512-dimensional vibration deep feature vector. The temperature feature processing channel consists of four fully connected layers stacked, with 64, 32, 16, and 8 neurons per layer, respectively, and each layer using a modified linear unit (MRU) as the activation function. The input to this channel is a time series of bearing temperature signals sampled at 1-second intervals over the past 60 seconds, totaling 60 values, and the output is an 8-dimensional temperature deep feature vector. The outputs of the two channels are concatenated in a feature fusion layer to obtain a 520-dimensional fused feature vector. This fused feature vector is then input to a regression output layer consisting of two fully connected layers, which uses a linear activation function to output the equivalent derailment coefficient and the equivalent wheel load reduction rate, ranging from 0 to 2. During training, the dual-channel neural network uses mean squared error as the loss function, and the measured derailment coefficient and measured wheel load reduction rate obtained through the wheel-rail force measurement device are used as supervision signals. The Adam optimizer is used to adjust the weights of the convolutional kernels and the weights of the fully connected layers in the network through the backpropagation algorithm.

[0057] The fault characteristic calculation module employs a strong-tracking Kalman filter to fit the equivalent derailment coefficient and equivalent wheel load reduction rate obtained at discrete sampling times into continuous first-time evolution curves and second-time evolution curves. The state vector of the strong-tracking Kalman filter comprises the equivalent derailment coefficient, its first derivative, its second derivative, the equivalent wheel load reduction rate, its first derivative, and its second derivative, totaling six dimensions. The state transition matrix is ​​set as a 6x6 matrix based on the constant acceleration motion model, and the observation matrix is ​​set as a 2x6 matrix. The first two columns of the observation matrix correspond to the direct observations of the equivalent derailment coefficient and the equivalent wheel load reduction rate. This filter estimates continuous state values ​​online from discrete observations containing measurement noise through two stages: a prediction step and an update step. The core difference of this filter lies in the introduction of a fading factor. This fading factor is calculated in real-time based on the autocorrelation matrix of the state estimation residual sequence. When the equivalent derailment coefficient or equivalent wheel load reduction rate undergoes a sudden change, the fading factor automatically increases, causing the filter gain to increase accordingly, thus rapidly tracking changes in the actual state. In the output state estimate, the first dimension serves as the equivalent derailment coefficient filter value, the second dimension as the derailment coefficient margin consumption rate, the third dimension as the derailment coefficient margin consumption acceleration, the fourth dimension as the equivalent wheel load reduction rate filter value, the fifth dimension as the wheel load reduction rate margin consumption rate, and the sixth dimension as the wheel load reduction rate margin consumption acceleration.

[0058] The derailment factor margin consumption rate and the wheel load reduction rate margin consumption rate are combined sequentially to form a two-dimensional bogie fault characteristic vector. The first dimension of this vector is the derailment factor margin consumption rate, and the second dimension is the wheel load reduction rate margin consumption rate.

[0059] The derailment risk diagnosis module first determines whether the derailment coefficient margin consumption rate is greater than 0. If it is greater than 0, it calculates the difference between the derailment coefficient safety boundary and the current equivalent derailment coefficient, divides it by the derailment coefficient margin consumption rate, and obtains the remaining safe time in the derailment coefficient dimension. If the derailment coefficient margin consumption rate is not greater than 0, the remaining safe time in the derailment coefficient dimension is recorded as the preset maximum safe duration. Similarly, it determines whether the wheel load reduction rate margin consumption rate is greater than 0. If it is greater than 0, it calculates the difference between the wheel load reduction rate safety boundary and the current equivalent wheel load reduction rate, divides it by the wheel load reduction rate margin consumption rate, and obtains the remaining safe time in the wheel load reduction rate dimension. If it is not greater than 0, the remaining safe time in the wheel load reduction rate dimension is recorded as the preset maximum safe duration. According to safety standards in the rail transit field, the derailment coefficient safety boundary is set to 1.0, and the wheel load reduction rate safety boundary is set to 0.8. Finally, it compares the remaining safe time in the derailment coefficient dimension with the remaining safe time in the wheel load reduction rate dimension, and takes the smaller value as the derailment risk diagnosis result.

[0060] The preset maximum safe duration is a fixed value that is much larger than the duration of a single train operation, such as 10,000 hours. This is used to indicate that there is no time constraint in the current dimension when the margin consumption rate is not greater than 0.

[0061] The preset time threshold is set in advance according to the actual situation of the train operation line. For example, it is preset to 30 minutes in high-risk sections that require emergency braking and 5 minutes in normal operation sections.

[0062] When the derailment risk diagnosis result is less than the preset time threshold, the diagnostic output module sends a derailment warning signal to the train control system, triggering an audible and visual alarm in the driver's cab. At the same time, the diagnostic output module displays the bogie fault feature vector and the derailment risk diagnosis result in numerical form on the driver's cab display screen, for example, "Derailment coefficient margin consumption rate: 0.1 per hour, wheel load reduction rate margin consumption rate: 0.05 per hour, estimated safe remaining time: 15 minutes".

[0063] The fault feature extraction module, fault feature calculation module, derailment risk diagnosis module, and diagnosis output module provided in this embodiment work together to achieve a complete closed loop from vibration acceleration signals and bearing temperature signals to derailment risk diagnosis results. The fault feature extraction module converts low-cost sensor signals into equivalent derailment coefficients and equivalent wheel load reduction rates through a dual-channel neural network, eliminating the need for expensive wheel-rail force measurement devices. The fault feature calculation module calculates the derailment coefficient margin consumption rate and wheel load reduction rate margin consumption rate through a strong-tracking Kalman filter and combines them into a bogie fault feature vector, upgrading static threshold monitoring to dynamic trend monitoring. When the derailment coefficient margin consumption rate is greater than zero, the derailment risk diagnosis module calculates the remaining safe time in the derailment coefficient dimension by dividing the difference between the derailment coefficient safety boundary and the current equivalent derailment coefficient by the derailment coefficient margin consumption rate. When the derailment coefficient margin consumption rate is not greater than zero, the remaining safe time in the derailment coefficient dimension is recorded as the preset maximum safe duration. When the wheel load reduction rate margin consumption rate is greater than zero, the module calculates the remaining safe time in the wheel load reduction rate dimension by dividing the difference between the wheel load reduction rate safety boundary and the current equivalent wheel load reduction rate by the wheel load reduction rate margin consumption rate. When the wheel load reduction rate margin consumption rate is not greater than zero, the remaining safe time in the wheel load reduction rate dimension is recorded as the preset maximum safe duration. The smaller value of the remaining safe time in the two dimensions is used as the derailment risk diagnosis result, achieving a fundamental leap from "whether the current state exceeds the standard" to "how much safe operating time is left before derailment". When the derailment risk diagnosis result is less than a preset time threshold, the diagnostic output module outputs a derailment warning signal and displays the bogie fault feature vector and derailment risk diagnosis result on the driver's cab display screen, enabling maintenance personnel to know the derailment risk evolution trend in advance and proactively arrange preventive maintenance. The four modules together form a complete technical chain from data acquisition, feature extraction, risk quantification to warning output, achieving a synergistic technical effect that goes beyond the simple superposition of the functions of each module.

[0064] Example 2

[0065] Based on Example 1, the state mapping model in the fault feature extraction module adopts a dual-channel neural network architecture, which includes a vibration feature processing channel and a temperature feature processing channel.

[0066] The vibration feature processing channel is composed of multiple convolutional layers and multiple pooling layers stacked together. The input of the vibration feature processing channel is the time-frequency plot of the vibration acceleration signal, and the output of the vibration feature processing channel is the vibration deep feature vector.

[0067] The temperature feature processing channel is composed of multiple fully connected layers stacked together. The input of the temperature feature processing channel is the time series of the bearing temperature signal, and the output of the temperature feature processing channel is the deep temperature feature vector.

[0068] The dual-channel neural network architecture also includes a feature fusion layer and a regression output layer. The feature fusion layer concatenates the deep vibration feature vector and the deep temperature feature vector and inputs them into the regression output layer. The regression output layer outputs the equivalent derailment coefficient and the equivalent wheel load reduction rate. During the training process of the dual-channel neural network architecture, the measured derailment coefficient and the measured wheel load reduction rate used to calculate the output error are obtained by a wheel-rail force measuring device installed on the bogie wheelset.

[0069] The working principle and beneficial effects of the above technical solution are as follows:

[0070] The vibration feature processing channel of the dual-channel neural network architecture consists of five convolutional layers and five max-pooling layers stacked alternately. Each convolutional layer has a 3x3 kernel size and a stride of 1, followed by a modified linear unit activation function. The input to this channel is a time-frequency image generated by short-time Fourier transform of the vibration acceleration signal. The time-frequency image has a size of 128×128 pixels and a grayscale value range of 0 to 255. This channel extracts local texture features from the time-frequency image through layer-by-layer convolution operations and reduces the feature dimensionality through pooling operations, ultimately outputting a 512-dimensional vibration deep feature vector.

[0071] The temperature feature processing channel of the dual-channel neural network architecture consists of four stacked fully connected layers: the first fully connected layer has 64 neurons, the second has 32 neurons, the third has 16 neurons, and the fourth has 8 neurons. Each fully connected layer is followed by a modified linear unit activation function. The input to this channel is a time series of bearing temperature signals collected at 1-second intervals over the past 60 seconds, totaling 60 temperature values, each in degrees Celsius. This channel performs a nonlinear transformation on the temperature time series through layer-by-layer fully connected operations, ultimately outputting an 8-dimensional deep temperature feature vector.

[0072] The feature fusion layer concatenates the 512-dimensional vibration deep feature vector with the 8-dimensional temperature deep feature vector, resulting in a 520-dimensional fused feature vector. The regression output layer consists of two stacked fully connected layers: the first fully connected layer has 32 neurons and uses a modified linear unit activation function, while the second fully connected layer has 2 neurons and uses a linear activation function. The regression output layer receives the 520-dimensional fused feature vector as input and outputs the equivalent derailment coefficient and the equivalent wheel load reduction rate. The equivalent derailment coefficient ranges from 0 to 2, and the equivalent wheel load reduction rate ranges from 0 to 1.

[0073] During the training of the dual-channel neural network architecture, the mean squared error loss function is used. The supervisory signals used to calculate the output error are the measured derailment coefficient and the measured wheel load reduction rate, which are obtained through a wheel-rail force measurement device installed on the bogie wheelset. The wheel-rail force measurement device has built-in strain gauges, which are attached to the surface of the wheelset spokes. When the wheel is subjected to vertical and lateral forces from the wheel and rail, the spokes produce strain, which the strain gauges convert into voltage signals. The measured derailment coefficient and the measured wheel load reduction rate are then calculated using calibration formulas. During training, vibration acceleration signals, bearing temperature signals, measured derailment coefficients, and measured wheel load reduction rates collected at the same time are used as a training sample. A training set of 10,000 training samples is constructed. The Adam optimizer is used to iteratively adjust the weights of the convolutional kernels and fully connected layers in the network through backpropagation, so that the equivalent derailment coefficient and the equivalent wheel load reduction rate output by the network gradually approach the measured derailment coefficient and the measured wheel load reduction rate.

[0074] This embodiment, based on Embodiment 1, further defines the specific structure of the dual-channel neural network architecture and the method of acquiring training data. The vibration feature processing channel extracts local texture features from the time-frequency map through stacked convolutional and pooling layers, while the temperature feature processing channel extracts nonlinear variation features of the temperature time series through stacked fully connected layers. The heterogeneous features of the two channels are concatenated in the feature fusion layer and then jointly input into the regression output layer, enabling the network to simultaneously learn the joint influence of vibration and temperature signals on the derailment coefficient. The measured derailment coefficient and measured wheel load reduction rate obtained through the wheel-rail force measurement device are used as supervisory signals, allowing the dual-channel neural network to establish an accurate mapping from vibration acceleration signals and bearing temperature signals to the equivalent derailment coefficient and equivalent wheel load reduction rate in a supervised learning manner. This provides reliable basic parameters for subsequent derailment risk diagnosis, further improving the accuracy and engineering practicality of the diagnostic system.

[0075] Example 3

[0076] Based on Example 2, the training process of the dual-channel neural network architecture introduces an adversarial domain adaptation mechanism, which is used to eliminate the distribution differences of vibration acceleration signals under different line conditions and the distribution differences of bearing temperature signals under different ambient temperatures.

[0077] The adversarial domain adaptation mechanism includes a feature extractor, a domain discriminator, and a regressor. The feature extractor maps the input signal to a common feature space. The domain discriminator is used to determine which line conditions or which ambient temperature range the features in the common feature space originate from. The regressor is used to regress the equivalent derailment coefficient and the equivalent wheel load reduction rate from the features in the common feature space.

[0078] During the training process of the dual-channel neural network architecture, the optimization objective of the feature extractor is to minimize the prediction error of the regressor and maximize the discrimination error of the domain discriminator. The optimization objective of the domain discriminator is to minimize the domain discrimination error. Through adversarial training, the feature distribution in the common feature space is decoupled from the line conditions and ambient temperature.

[0079] The working principle and beneficial effects of the above technical solution are as follows:

[0080] The feature extractor in the adversarial domain adaptation mechanism consists of vibration feature processing channels and temperature feature processing channels from a dual-channel neural network. The input to the feature extractor is the time-frequency plot of the vibration acceleration signal and the time series of the bearing temperature signal; the output is a 520-dimensional fused feature vector in a common feature space. This common feature space is a high-dimensional vector space in which samples collected under different line conditions or ambient temperatures exhibit their own distribution characteristics.

[0081] The domain discriminator in the adversarial domain adaptation mechanism consists of three stacked fully connected layers: the first layer has 256 neurons, the second layer has 128 neurons, and the third layer has K neurons, where K is the total number of domain categories. These domain categories include different line condition categories and different ambient temperature range categories. For example, line condition categories include straight segments, curve segments with small radii of curvature, and curve segments with large radii of curvature; ambient temperature range categories include low temperature range, normal temperature range, and high temperature range. The input to the domain discriminator is a 520-dimensional fused feature vector output from the feature extractor, and the output is a K-dimensional probability distribution vector, where each dimension represents the probability that the input feature belongs to the corresponding domain category. The domain discriminator uses the cross-entropy loss function to calculate the domain discrimination error.

[0082] The regressor in the adversarial domain adaptation mechanism consists of a regression output layer from a dual-channel neural network. The input to the regressor is a 520-dimensional fused feature vector from the feature extractor, and the output is the equivalent derailment coefficient and the equivalent wheel load reduction rate. The regressor uses the mean squared error loss function to calculate the regression prediction error.

[0083] The adversarial domain adaptation mechanism connects the feature extractor and the domain discriminator through a gradient inversion layer. During the forward propagation phase, the gradient inversion layer directly passes the output of the feature extractor to the domain discriminator without modification. During the backpropagation phase, the gradient inversion layer multiplies the gradient from the domain discriminator by a negative coefficient and then feeds it back to the feature extractor. The negative coefficient is set to 0 in the early stages of training and gradually increases to 1 with each training iteration. The feature extractor's optimization objective is to minimize the regressor's prediction error while maximizing the domain discriminator's discrimination error, even if the domain discriminator cannot distinguish which line condition or ambient temperature range the features in the common feature space originate from. The domain discriminator's optimization objective is to minimize its own domain discrimination error. Through this adversarial training method, the feature extractor and the domain discriminator compete against each other, ultimately decoupling the feature distribution in the common feature space from line conditions and ambient temperature. This means that samples collected under different line conditions or ambient temperatures exhibit similar or identical distributions in the common feature space.

[0084] This embodiment, building upon Embodiment 2, further introduces an adversarial domain adaptation mechanism. Through adversarial training between the feature extractor and the domain discriminator, the feature distribution in the common feature space is decoupled from track conditions and ambient temperature. Differences in the distribution of vibration acceleration signals under different track conditions and bearing temperature signals under different ambient temperatures are effectively eliminated, ensuring consistent mapping accuracy of the dual-channel neural network across various operating environments. When a train travels from one track to another, or from a low-temperature environment to a high-temperature environment, the equivalent derailment coefficient and equivalent wheel load reduction rate output by the state mapping model do not shift due to changes in track conditions or ambient temperature, significantly improving the generalization ability and robustness of the derailment risk diagnosis system under different operating environments.

[0085] Example 4

[0086] Based on Example 1, the fault feature calculation module fits the equivalent derailment coefficient obtained at multiple consecutive sampling times into a first time evolution curve by modeling the sequence of equivalent derailment coefficient changes over time as a linear time-varying system and using a strong tracking Kalman filter to estimate the state of the linear time-varying system online.

[0087] The equivalent derailment coefficient is processed using a first strong tracking Kalman filter. The state vector of the first strong tracking Kalman filter includes the equivalent derailment coefficient, the first derivative of the equivalent derailment coefficient, and the second derivative of the equivalent derailment coefficient. The equivalent wheel load reduction rate is processed using a second strong tracking Kalman filter. The state vector of the second strong tracking Kalman filter includes the equivalent wheel load reduction rate, the first derivative of the equivalent wheel load reduction rate, and the second derivative of the equivalent wheel load reduction rate.

[0088] The fading factor of the strong tracking Kalman filter is adjusted in real time according to the covariance matrix of the state estimation error, so that the strong tracking Kalman filter can maintain its adaptive tracking capability to the real state when the equivalent derailment coefficient changes abruptly.

[0089] The first derivative component in the state estimate output by the strong tracking Kalman filter is used as the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate, while the second derivative component is used as the derailment coefficient margin consumption acceleration and the wheel load reduction rate margin consumption acceleration.

[0090] The working principle and beneficial effects of the above technical solution are as follows:

[0091] Modeling the time-varying sequence of the equivalent derailment coefficient as a linear time-varying system means approximating the variation of the equivalent derailment coefficient as a dynamic system whose state changes linearly with time. In this linear time-varying system, the equivalent derailment coefficient at the current moment is predicted by adding the equivalent derailment coefficient at the previous moment to the first derivative of the previous moment and multiplying by the sampling time interval. The first derivative itself also changes with time. The equivalent derailment coefficient is estimated online using a first strong tracking Kalman filter. The state vector of the first strong tracking Kalman filter has a 3-dimensional dimension, consisting of the equivalent derailment coefficient, the first derivative of the equivalent derailment coefficient, and the second derivative of the equivalent derailment coefficient in sequence. The state transition matrix is ​​set to a 3x3 matrix based on the constant acceleration motion model, and the observation matrix is ​​set to a 1x3 matrix. The first row and first column of the observation matrix are 1, corresponding to the direct observation of the equivalent derailment coefficient, and the remaining elements are 0.

[0092] Similarly, the sequence of equivalent wheel load reduction rate over time is modeled as a linear time-varying system, and the equivalent wheel load reduction rate is estimated online using a second strong tracking Kalman filter. The state vector of the second strong tracking Kalman filter has a 3-dimensional dimension, consisting of the equivalent wheel load reduction rate, its first derivative, and its second derivative, in that order. The state transition matrix is ​​set as a 3x3 matrix based on the constant acceleration motion model, and the observation matrix is ​​set as a 1x3 matrix. The first row and first column of the observation matrix are set to 1, corresponding to the direct observation of the equivalent wheel load reduction rate, and the remaining elements are 0.

[0093] The fading factor of the strong tracking Kalman filter is adjusted in real time based on the covariance matrix of the state estimation error. Specifically, the autocorrelation matrix of the state estimation residual sequence is calculated, and the ratio of the trace of this autocorrelation matrix to a preset residual threshold is used as the basic fading factor. When the equivalent derailment coefficient or equivalent wheel load reduction rate changes abruptly, the trace of the autocorrelation matrix of the state estimation residual sequence increases sharply, causing the basic fading factor to increase accordingly. The basic fading factor is multiplied by a preset amplification factor to obtain the final fading factor. This final fading factor is then multiplied into the state prediction error covariance matrix, increasing the filter gain and thus enhancing the filter's ability to track abrupt changes in state. When neither the equivalent derailment coefficient nor the equivalent wheel load reduction rate changes abruptly, the basic fading factor is set to 1, and the state prediction error covariance matrix is ​​not amplified.

[0094] The first strong-tracking Kalman filter outputs a 3-dimensional state estimate at each sampling time. The first dimension is the filtered estimate of the equivalent derailment coefficient, the second dimension is the first derivative of the equivalent derailment coefficient, and the third dimension is the second derivative of the equivalent derailment coefficient. The second dimension component is extracted from the state estimate as the derailment coefficient margin consumption rate, and the third dimension component is extracted as the derailment coefficient margin consumption acceleration.

[0095] The second strong tracking Kalman filter outputs a 3-dimensional state estimate at each sampling time. The first dimension is the filtered estimate of the equivalent wheel load reduction rate, the second dimension is the first derivative of the equivalent wheel load reduction rate, and the third dimension is the second derivative of the equivalent wheel load reduction rate. The second dimension component is extracted from the state estimate as the wheel load reduction rate margin consumption rate, and the third dimension component is extracted as the wheel load reduction rate margin consumption acceleration.

[0096] This embodiment, based on Embodiment 1, further defines the state vector dimension and fading factor adjustment mechanism of the strong tracking Kalman filter. The state vector simultaneously includes the equivalent derailment coefficient and the equivalent wheel load reduction rate, along with their derivatives, enabling the filter to simultaneously track the evolution trends of the two core parameters of derailment risk. The fading factor is adjusted in real time based on the state estimation error covariance matrix, automatically increasing the filter gain when the equivalent derailment coefficient or equivalent wheel load reduction rate undergoes abrupt changes. This allows the filter to complete the tracking of the abrupt state within several sampling periods, avoiding the tracking delay problem that occurs in traditional Kalman filters during state abrupt changes. This mechanism ensures that the first and second time-evolution curves accurately reflect the actual state changes during the accelerated fault deterioration phase, providing a reliable basis for the accurate calculation of the margin consumption rate.

[0097] Example 5

[0098] Based on Example 4, the real-time adjustment method of the fading factor in the strong tracking Kalman filter is as follows:

[0099] Calculate the autocorrelation matrix of the state estimation residual sequence, and use the ratio of the trace of the autocorrelation matrix to the preset residual threshold as the basic fading factor;

[0100] When the absolute value of the second derivative of the equivalent derailment coefficient exceeds the preset threshold for the second derivative of the derailment coefficient or the absolute value of the second derivative of the equivalent wheel load reduction rate exceeds the preset threshold for the second derivative of the wheel load reduction rate, the basic fading factor is multiplied by the preset amplification factor to obtain the final fading factor. At the same time, the derailment coefficient margin consumption rate is added to the derailment coefficient margin consumption acceleration and multiplied by the preset prediction time step to obtain the predicted derailment coefficient margin consumption rate. The wheel load reduction rate margin consumption rate is added to the wheel load reduction rate margin consumption acceleration and multiplied by the preset prediction time step to obtain the predicted wheel load reduction rate margin consumption rate. The predicted derailment coefficient margin consumption rate and the predicted wheel load reduction rate margin consumption rate are used to replace the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate in the bogie fault feature vector.

[0101] When the absolute value of the second derivative of the equivalent derailment coefficient does not exceed the preset threshold for the second derivative of the derailment coefficient and the absolute value of the second derivative of the equivalent wheel load reduction rate does not exceed the preset threshold for the second derivative of the wheel load reduction rate, the basic fading factor is directly used as the final fading factor, keeping the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate in the bogie fault feature vector unchanged.

[0102] The working principle and beneficial effects of the above technical solution are as follows:

[0103] The autocorrelation matrix of the state estimation residual sequence is calculated as follows: At each sampling time, the difference between the observed value and the predicted value of the strong tracking Kalman filter at that time is calculated to obtain the residual vector. The residual vectors of the current time and the previous N times are arranged in chronological order, and the autocorrelation matrix of the sequence is calculated. The trace of the autocorrelation matrix is ​​the sum of the diagonal elements of the matrix. The preset residual threshold is a pre-set positive number, for example, 0.1. The ratio of the trace of the autocorrelation matrix to the preset residual threshold is used as the basic fading factor. When the trace of the autocorrelation matrix is ​​greater than the preset residual threshold, the basic fading factor is greater than 1; when the trace of the autocorrelation matrix is ​​less than or equal to the preset residual threshold, the basic fading factor is less than or equal to 1.

[0104] A preset threshold for the second derivative of the derailment coefficient is used to determine whether the change in the equivalent derailment coefficient is in an accelerating phase. When the absolute value of the second derivative of the equivalent derailment coefficient exceeds the preset threshold, it indicates that the rate of change of the equivalent derailment coefficient itself is increasing rapidly, meaning that the deterioration of the derailment coefficient is accelerating. Similarly, a preset threshold for the second derivative of the wheel load reduction rate is used to determine whether the change in the equivalent wheel load reduction rate is in an accelerating phase. When the absolute value of the second derivative of the equivalent derailment coefficient exceeds the preset threshold, or the absolute value of the second derivative of the equivalent wheel load reduction rate exceeds the preset threshold, the basic fading factor is multiplied by a preset amplification factor to obtain the final fading factor. The preset amplification factor is a value greater than 1, for example, 2. The final fading factor is multiplied into the state prediction error covariance matrix to increase the filter gain and enhance the tracking ability of the accelerating deterioration state.

[0105] In the case of accelerated deterioration, a prediction correction mechanism is simultaneously activated. The preset prediction time step is a pre-defined time interval, for example, 10 seconds. The predicted derailment coefficient margin consumption rate is obtained by adding the derailment coefficient margin consumption acceleration to the product of the preset prediction time step. This predicted value represents the estimated derailment coefficient margin consumption rate after the preset prediction time step from the current moment. Similarly, the predicted wheel load reduction rate margin consumption rate is obtained by adding the wheel load reduction rate margin consumption acceleration to the product of the preset prediction time step. The predicted derailment coefficient margin consumption rate and the predicted wheel load reduction rate margin consumption rate are used to replace the derailment coefficient margin consumption rate in the bogie fault feature vector, respectively. This allows the derailment risk diagnosis module to calculate the remaining safe time based on the predicted margin consumption rate at future moments, thus reflecting the impact of accelerated deterioration on the remaining safe time in advance.

[0106] When the absolute value of the second derivative of the equivalent derailment coefficient does not exceed the preset threshold for the second derivative of the derailment coefficient, and the absolute value of the second derivative of the equivalent wheel load reduction rate does not exceed the preset threshold for the second derivative of the wheel load reduction rate, it indicates that the deterioration rate in both dimensions has not accelerated. In this case, the basic fading factor is directly used as the final fading factor, and no additional amplification is applied to the filter gain. At the same time, the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate in the bogie fault feature vector remain unchanged, and predicted values ​​are not used for replacement.

[0107] This embodiment, based on Embodiment 4, further defines a dual-threshold adjustment mechanism and a prediction correction mechanism for the fading factor. The basic fading factor is calculated by the ratio of the trace of the autocorrelation matrix of the state estimation residual to a preset residual threshold, enabling the filter to respond to general state estimation biases. By comparing the second derivative with a preset second derivative threshold for the derailment coefficient, it specifically identifies accelerated deterioration states of the equivalent derailment coefficient or equivalent wheel load reduction rate. During accelerated deterioration, the fading factor is further amplified, enhancing the filter's rapid tracking capability. Simultaneously, during accelerated deterioration, the margin consumption acceleration is used to predict the margin consumption rate at future moments. The predicted value replaces the current value in the calculation of the remaining safe time, allowing the derailment risk diagnosis results to reflect the impact of accelerated deterioration in advance. This mechanism significantly improves the response speed of the derailment risk diagnosis system to the accelerated deterioration process of the fault, providing drivers with more time for braking or evasive decision-making.

[0108] Example 6

[0109] Based on Example 1, the derailment risk diagnosis module takes the smaller of the safe remaining time in the derailment coefficient dimension and the safe remaining time in the wheel load reduction rate dimension as the derailment risk diagnosis result, and then corrects the derailment risk diagnosis result based on the ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate margin consumption rate in the bogie fault feature vector.

[0110] The correction method for the derailment risk diagnosis result is as follows: when the ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate margin consumption rate exceeds the preset normal ratio range, the derailment risk diagnosis result is divided by a penalty coefficient greater than 1. The magnitude of the penalty coefficient is positively correlated with the degree to which the ratio deviates from the preset normal ratio range.

[0111] The preset normal ratio range is determined based on the statistical distribution of the ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate consumption rate under the historical normal operating conditions of the bogie.

[0112] The working principle and beneficial effects of the above technical solution are as follows:

[0113] The ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate margin consumption rate in the bogie fault characteristic vector reflects the degree of imbalance in derailment risk across two dimensions. Under normal bogie operation, the effects of bearing wear, gear meshing status, and suspension system degradation on the derailment coefficient and wheel load reduction rate have a specific proportional relationship, and the two margin consumption rates typically remain within a stable ratio range. When a particular fault type dominates, this ratio deviates from the normal range. For example, bearing faults primarily affect the derailment coefficient, leading to an increase in the derailment coefficient margin consumption rate while the wheel load reduction rate margin consumption rate changes relatively little, resulting in a significantly larger ratio; conversely, suspension faults primarily affect the wheel load reduction rate, leading to an increase in the wheel load reduction rate margin consumption rate while the derailment coefficient margin consumption rate changes relatively little, resulting in a significantly smaller ratio.

[0114] The preset normal ratio range is determined as follows: Under the historical normal operating conditions of the bogie, the derailment coefficient margin consumption rate and wheel load reduction rate consumption rate are continuously collected for multiple time windows. The ratio of the two rates within each time window is calculated to obtain a ratio sequence. The mean and standard deviation of this ratio sequence are calculated. The mean minus twice the standard deviation is used as the lower limit of the preset normal ratio range, and the mean plus twice the standard deviation is used as the upper limit of the preset normal ratio range. For example, if the mean of the ratio sequence is 0.8 and the standard deviation is 0.1, then the preset normal ratio range is 0.6 to 1.0.

[0115] When the current ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate margin consumption rate exceeds the preset normal ratio range, it indicates that the current fault mode deviates from the historical normal mode, and a certain fault type dominates. In this case, the derailment risk diagnosis result is divided by a penalty coefficient greater than 1. The correction strength of the penalty coefficient is positively correlated with the degree to which the ratio deviates from the preset normal ratio range; that is, the further the ratio deviates from the range, the greater the correction magnitude. Specifically, the calculation method is as follows: calculate the deviation ratio of the current ratio from the center value of the preset normal ratio range, use the square of this deviation ratio as the base penalty value, and use the smaller of the base penalty value and the preset penalty upper limit value as the final penalty coefficient. For example, if the preset normal ratio range center value is 0.8, the current ratio is 1.2, the deviation ratio is 1.5, the square of the deviation ratio is 2.25, the base penalty value is 2.25, and assuming the preset penalty upper limit value is 5.0, then the final penalty coefficient is 2.25. Dividing the derailment risk diagnosis result by 2.25 shortens the remaining safe time, resulting in a revised derailment risk diagnosis result. This revised result is more conservative and reflects the additional risks brought about by the dominant fault type.

[0116] This embodiment, building upon Embodiment 1, further introduces a correction mechanism based on the ratio of margin consumption rates. By calculating the ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate margin consumption rate and comparing it with a preset normal ratio range determined based on historical statistical distributions, abnormal fault modes dominated by bearing failure or suspension failure can be identified. When an abnormal ratio is detected, a penalty coefficient is used to conservatively correct the derailment risk diagnosis results, making the derailment risk diagnosis results more stringent under the dominant fault type. This correction mechanism avoids situations where the remaining safety time in one dimension is normal while the remaining safety time in another dimension is small and not given sufficient attention, thus improving the early warning sensitivity of the derailment risk diagnosis system under unbalanced fault modes.

[0117] Example 7

[0118] Based on Example 6, the penalty coefficient is calculated as follows:

[0119] Calculate the current ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate margin consumption rate, and calculate the deviation ratio of the current ratio from the center value of the preset normal ratio range.

[0120] The square of the deviation multiplier is used as the base penalty value, and the smaller of the base penalty value and the preset penalty upper limit value is used as the penalty coefficient.

[0121] The working principle and beneficial effects of the above technical solution are as follows:

[0122] The deviation ratio between the current ratio and the center value of the preset normal ratio range is calculated as follows: Divide the current ratio by the center value of the preset normal ratio range to obtain the first ratio; divide the center value of the preset normal ratio range by the current ratio to obtain the second ratio; and take the larger of the first and second ratios as the deviation ratio. For example, if the center value of the preset normal ratio range is 0.8 and the current ratio is 1.2, then the first ratio is 1.2 divided by 0.8, which equals 1.5, and the second ratio is 0.8 divided by 1.2, which is approximately 0.67. The larger value, 1.5, is taken as the deviation ratio. If the current ratio is 0.5, then the first ratio is 0.5 divided by 0.8, which equals 0.625, and the second ratio is 0.8 divided by 0.5, which equals 1.6. The larger value, 1.6, is taken as the deviation ratio. This deviation ratio is always greater than or equal to 1; the larger the deviation ratio, the further the current ratio deviates from the normal center value.

[0123] The square of the deviation ratio is used as the base penalty value. The square of the deviation ratio causes the penalty value to increase rapidly with the degree of deviation; for example, the square is 2.25 when the deviation ratio is 1.5, and 4.0 when the deviation ratio is 2.0. A preset penalty upper limit is a pre-defined maximum penalty coefficient, for example, 5.0. The smaller value between the base penalty value and the preset penalty upper limit is used as the penalty coefficient. That is, when the base penalty value exceeds the preset penalty upper limit, the penalty coefficient is the preset penalty upper limit, avoiding an overly conservative derailment risk diagnosis result due to an excessively large penalty coefficient. For example, when the base penalty value is 2.25, which is less than 5.0, the penalty coefficient is 2.25; when the base penalty value is 6.0, which is greater than 5.0, the penalty coefficient is 5.0.

[0124] This embodiment, based on Embodiment 6, further defines the calculation method for the penalty coefficient. The base penalty value is calculated by squaring the deviation ratio, causing the penalty coefficient to increase rapidly with the degree of ratio deviation. This aligns with the practical requirement that the more severe the fault dominance, the more conservative the remaining safe time should be. A preset upper limit for the penalty prevents the penalty coefficient from becoming too large, thus rendering the derailment risk diagnosis results worthless.

[0125] Example 8

[0126] Based on Embodiment 1, a fault type identification module is also included. The fault type identification module is used to input the bogie fault feature vector into a pre-trained fault classifier when the derailment risk diagnosis result is less than a preset time threshold. The fault classifier outputs the fault type identification result of the current bogie.

[0127] The fault classifier uses a support vector machine, and the kernel function of the support vector machine is a radial basis function. The width parameter of the radial basis function is determined based on the average Euclidean distance between the bogie fault feature vector samples.

[0128] The fault type identification results include bearing fault categories, gear fault categories, suspension fault categories, and no fault categories.

[0129] The working principle and beneficial effects of the above technical solution are as follows:

[0130] The fault type identification module is triggered when the derailment risk diagnosis result is less than a preset time threshold. That is, when the system determines that the derailment risk has reached the warning level, the fault type identification function is automatically activated. The fault type identification module inputs the bogie fault feature vector at the current moment into a pre-trained support vector machine classifier. The bogie fault feature vector contains two dimensions: the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate. This vector serves as the input feature to the support vector machine classifier.

[0131] The kernel function of the Support Vector Machine (SVM) classifier is the Radial Basis Function (RBF) kernel. The mathematical form of the RBF kernel depends on a width parameter, which controls the range of influence of the kernel function. The width parameter is determined as follows: 1000 bogie fault feature vector samples are randomly selected from the training sample set. The Euclidean distance between every two samples is calculated, and the average of all Euclidean distances is obtained. The reciprocal of this average Euclidean distance is used as the width parameter of the RBF kernel. For example, if the average Euclidean distance is 0.5, then the width parameter is 2.0.

[0132] The training process of the Support Vector Machine (SVM) classifier is as follows: A training sample set is collected, with each sample containing a bogie fault feature vector and its corresponding fault type label. There are four fault type labels: bearing fault, gear fault, suspension fault, and no fault. In the bogie fault feature vector corresponding to the bearing fault category, the ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate margin consumption rate is greater than 1.2. In the bogie fault feature vector corresponding to the gear fault category, both margin consumption rates are at a moderate level, and the ratio is close to 1.0. In the bogie fault feature vector corresponding to the suspension fault category, the ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate margin consumption rate is less than 0.7. In the bogie fault feature vector corresponding to the no fault category, both margin consumption rates are close to 0. A one-to-many strategy is used to decompose the four-class classification problem into four binary classification sub-problems, and a binary classification SVM is trained for each category. During training, the sequential minimum optimization algorithm is used to solve the optimization problem of the SVM, obtaining the decision function for each binary classification SVM.

[0133] The support vector machine classifier outputs the fault type identification result for the current bogie. The output result is one of the following: bearing fault category, gear fault category, suspension fault category, or no fault category. For example, when the derailment factor margin consumption rate is 0.12 per hour, the wheel load reduction rate margin consumption rate is 0.08 per hour, and the ratio is 1.5, the support vector machine classifier outputs the bearing fault category.

[0134] This embodiment, based on Embodiment 1, further integrates a fault type identification module. When the derailment risk diagnosis result triggers an early warning, the fault type identification module automatically activates, inputting the bogie fault feature vector into a support vector machine classifier and outputting a bearing fault category, gear fault category, suspension fault category, or no fault category. This module enables the derailment risk diagnosis system not only to output the remaining safe time but also to identify the specific fault type causing the derailment risk. Maintenance personnel can quickly locate the faulty component based on the fault type identification result; for example, if the fault is identified as a bearing fault, the bearing condition is checked first; if the fault is identified as a suspension fault, the suspension system condition is checked first. This significantly shortens the fault investigation time, improves maintenance efficiency, and achieves a complete diagnostic closed loop from risk warning to fault location.

[0135] Example 9

[0136] Based on Example 8, the training samples of the fault classifier are amplified using synthetic minority class oversampling technology to balance the number of samples for each type of fault.

[0137] The amplification method of synthetic minority class oversampling technology is as follows: for fault types with fewer than a preset threshold number of samples, two adjacent samples of the fault type are selected in the bogie fault feature vector space, and new samples are generated by random interpolation on the line connecting the two adjacent samples. The interpolation operation is repeated until the number of samples of the fault type reaches the preset threshold number of samples.

[0138] The working principle and beneficial effects of the above technical solution are as follows:

[0139] In actual engineering data acquisition, the occurrence probabilities of bearing failures, gear failures, and suspension failures differ, leading to a severe imbalance in the number of training samples collected for each type of failure. For example, there might be 500 suspension failure samples, 300 gear failure samples, and 200 bearing failure samples, while the number of samples in the no-failure category is 10,000. The preset threshold is determined based on the maximum number of samples for each failure type. For instance, if the preset threshold of 10,000 samples in the no-failure category is used, the number of samples for suspension failures, gear failures, and bearing failures needs to be increased to 10,000.

[0140] The amplification method of synthetic minority class oversampling is as follows: For fault types with fewer than a preset threshold number of samples, all sample points of that fault type are first mapped to the bogie fault feature vector space. The bogie fault feature vector space is a two-dimensional space, where the horizontal axis represents the derailment coefficient margin consumption rate and the vertical axis represents the wheel load reduction rate margin consumption rate. In this two-dimensional space, each sample point corresponds to a coordinate position. A sample point is randomly selected from the sample points of that fault type as a reference point. The Euclidean distance between the reference point and all other sample points is calculated, and the sample point with the smallest Euclidean distance is selected as the neighboring point. The line connecting the reference point and the neighboring point is used as an interpolation segment, and a new sample point is generated by randomly selecting a position on the interpolation segment. The coordinates of the new sample point are equal to the coordinates of the reference point plus the coordinates of the neighboring point minus the coordinates of the reference point multiplied by a random number, where the random number ranges from 0 to 1. For example, if the reference point coordinates are (0.12, 0.08), the adjacent point coordinates are (0.14, 0.09), and the random number is 0.6, then the new sample point coordinates are (0.12 + 0.02 * 0.6, 0.08 + 0.01 * 0.6). Add the newly generated sample point to the training sample set. Repeat the above operations of randomly selecting a reference point, finding adjacent points, and random interpolation until the number of samples of this fault type reaches a preset threshold.

[0141] During the amplification process, each new sample point generated by interpolation lies on the line connecting two real sample points. Therefore, the feature distribution of the new sample points maintains the same local geometric structure as the feature distribution of the real sample points. For example, the original sample points for the bearing failure category are concentrated in the region of derailment coefficient margin consumption rate of 0.10 to 0.15 per hour and wheel load reduction rate margin consumption rate of 0.05 to 0.10 per hour. The new sample points generated by the synthetic minority class oversampling technique are also distributed in the same region, and no unreasonable sample points outside this region will be generated.

[0142] This embodiment, building upon Embodiment 8, further introduces a synthetic minority class oversampling technique to amplify the training samples. By increasing the number of minority class fault type samples to the same level as the majority class fault types, the support vector machine classifier is prevented from biasing towards the fault-free category or suspension fault category with a larger number of samples during training. Neighboring samples are selected in the two-dimensional bogie fault feature vector space, and new samples are generated by random interpolation along the connecting lines. This ensures that the feature distribution of the amplified samples is consistent with the feature distribution of the original samples, avoiding overfitting caused by simply copying samples. The support vector machine classifier trained after sample balancing has a balanced recognition accuracy for all fault types, avoiding the problem of missed detection of minority class fault types due to sample imbalance, and further improving the reliability of the derailment risk diagnosis system under various fault scenarios.

[0143] Example 10

[0144] Based on Example 1, it also includes a prediction uncertainty quantification module. The prediction uncertainty quantification module is used to output the first prediction interval of the equivalent derailment coefficient and the second prediction interval of the equivalent wheel load reduction rate while the state mapping model outputs the equivalent derailment coefficient and the equivalent wheel load reduction rate.

[0145] The first and second prediction intervals were obtained through multiple forward inferences after randomly discarding some neurons in the state mapping model.

[0146] When calculating the derailment risk diagnosis results, the derailment risk diagnosis module uses the upper boundary of the first prediction interval to replace the equivalent derailment coefficient and the upper boundary of the second prediction interval to replace the equivalent wheel load reduction rate, thus obtaining a conservative estimate of the derailment risk diagnosis results.

[0147] The working principle and beneficial effects of the above technical solution are as follows:

[0148] The prediction uncertainty quantification module is used to evaluate the reliability of the output results of the state mapping model. When outputting the equivalent derailment coefficient and equivalent wheel load reduction rate, the state mapping model suffers from a certain degree of prediction error due to noise in the input signal, uncertainty in model parameters, and limitations in the coverage of training data. The prediction uncertainty quantification module estimates this uncertainty using the Monte Carlo dropout method, outputting a first prediction interval for the equivalent derailment coefficient and a second prediction interval for the equivalent wheel load reduction rate. The first prediction interval is a numerical range, for example, 0.75 to 0.85, representing a 95% probability that the true value of the equivalent derailment coefficient falls within this range. The second prediction interval is calculated similarly.

[0149] The Monte Carlo dropout method is implemented as follows: After training, each neuron in the dual-channel neural network of the state mapping model has fixed weight parameters. During the inference phase, some neurons in the network are randomly dropped, with a dropout rate ranging from 10% to 30%. After each random dropout, the network's connection structure changes, essentially forming a different sub-network. The time-frequency plot of the current input vibration acceleration signal and the time series of the bearing temperature signal are input into this sub-network for forward inference, yielding an equivalent derailment coefficient sample value and an equivalent wheel load reduction rate sample value. This process of randomly dropping neurons and forward inference is repeated T times, where T ranges from 100 to 500. The T equivalent derailment coefficient sample values ​​obtained from the T inferences are collected, and the 2.5% and 97.5% quantiles of these sample values ​​are calculated. These two quantiles are used as the lower and upper boundaries of the first prediction interval. Similarly, collect T equivalent wheel load reduction rate sample values ​​and calculate their 2.5% and 97.5% quantiles as the lower and upper boundaries of the second prediction interval. For example, among the equivalent derailment coefficient sample values ​​obtained from 100 inferences, the 2.5th sample value after sorting is 0.75 and the 97.5th sample value is 0.85, then the first prediction interval is 0.75 to 0.85.

[0150] The derailment risk diagnosis module employs a conservative estimation strategy when calculating derailment risk diagnosis results. Specifically, the upper boundary of the first prediction interval is used to replace the current equivalent derailment coefficient, and the upper boundary of the second prediction interval is used to replace the current equivalent wheel load reduction rate. For example, if the current equivalent derailment coefficient is 0.80 and the first prediction interval is 0.75 to 0.85, then 0.85 is used as the equivalent derailment coefficient in the calculation. If the current equivalent wheel load reduction rate is 0.30 and the second prediction interval is 0.28 to 0.33, then 0.33 is used as the equivalent wheel load reduction rate in the calculation. These upper boundary values ​​are used to calculate the safe remaining time in both the derailment coefficient and wheel load reduction rate dimensions to obtain the derailment risk diagnosis results. Because the upper boundary of the prediction interval is used, the calculated safe remaining time is shorter than that calculated using point estimates. Therefore, the obtained derailment risk diagnosis results are more conservative, allowing for earlier warnings when uncertainty is high, avoiding warning delays due to prediction errors.

[0151] This embodiment, based on Embodiment 1, further integrates a prediction uncertainty quantification module. By repeatedly discarding some neurons in the state mapping model using the Monte Carlo dropout method followed by multiple forward inferences, a prediction interval for the equivalent derailment coefficient and equivalent wheel load reduction rate is obtained. This prediction interval reflects the degree of uncertainty in the model's output. The derailment risk diagnosis module uses the upper boundary of the prediction interval to replace the point estimate for calculation, obtaining a conservative estimate of the derailment risk diagnosis result. When the prediction uncertainty of the state mapping model is high, for example, if the time-frequency diagram of the input signal differs significantly from the training samples or if there are abnormal fluctuations in the bearing temperature signal, the prediction interval will widen accordingly, and the upper boundary value will be significantly higher than the point estimate, resulting in a shorter derailment risk diagnosis result and more stringent conditions for triggering the warning. This mechanism effectively avoids the risk of overestimating the remaining safety time due to model prediction errors, improving the safety margin of the derailment risk diagnosis system under atypical operating conditions.

[0152] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A train derailment risk diagnosis system based on bogie fault characteristic analysis, characterized in that, include: The fault feature extraction module is used to collect vibration acceleration signals and bearing temperature signals of the bogie axle box. The vibration acceleration signals and bearing temperature signals are input into a pre-established state mapping model. The state mapping model outputs the corresponding equivalent derailment coefficient and equivalent wheel load reduction rate based on the current vibration acceleration signal and the current bearing temperature signal. The fault feature calculation module is used to fit the equivalent derailment coefficient obtained from multiple consecutive sampling times into a first time evolution curve, calculate the first derivative of the first time evolution curve at the current time as the derailment coefficient margin consumption rate, fit the equivalent wheel load reduction rate obtained from multiple consecutive sampling times into a second time evolution curve, calculate the first derivative of the second time evolution curve at the current time as the wheel load reduction rate margin consumption rate, and combine the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate into a bogie fault feature vector. The derailment risk diagnosis module is used to obtain the safe remaining time in the derailment coefficient dimension by dividing the difference between the derailment coefficient safety boundary and the current equivalent derailment coefficient by the derailment coefficient margin consumption rate in the bogie fault feature vector when the derailment coefficient margin consumption rate is greater than zero. When the derailment coefficient margin consumption rate is not greater than zero, the safe remaining time in the derailment coefficient dimension is recorded as the preset maximum safe duration. The module also obtains the safe remaining time in the wheel load reduction rate dimension by dividing the difference between the wheel load reduction rate safety boundary and the current equivalent wheel load reduction rate by the wheel load reduction rate margin consumption rate in the bogie fault feature vector. When the wheel load reduction rate margin consumption rate is not greater than zero, the safe remaining time in the wheel load reduction rate dimension is recorded as the preset maximum safe duration. The smaller of the safe remaining time in the derailment coefficient dimension and the safe remaining time in the wheel load reduction rate dimension is taken as the derailment risk diagnosis result. The diagnostic output module is used to output a derailment warning signal when the derailment risk diagnosis result is less than a preset time threshold, and to display the bogie fault feature vector and the derailment risk diagnosis result together on the driver's cab display screen.

2. The train derailment risk diagnosis system based on bogie fault characteristic analysis according to claim 1, characterized in that, The state mapping model in the fault feature extraction module adopts a dual-channel neural network architecture, which includes a vibration feature processing channel and a temperature feature processing channel. The vibration feature processing channel is composed of multiple convolutional layers and multiple pooling layers stacked together. The input of the vibration feature processing channel is the time-frequency plot of the vibration acceleration signal, and the output of the vibration feature processing channel is the vibration deep feature vector. The temperature feature processing channel is composed of multiple fully connected layers stacked together. The input of the temperature feature processing channel is the time series of the bearing temperature signal, and the output of the temperature feature processing channel is the deep temperature feature vector. The dual-channel neural network architecture also includes a feature fusion layer and a regression output layer. The feature fusion layer concatenates the deep vibration feature vector and the deep temperature feature vector and inputs them into the regression output layer. The regression output layer outputs the equivalent derailment coefficient and the equivalent wheel load reduction rate. During the training process of the dual-channel neural network architecture, the measured derailment coefficient and the measured wheel load reduction rate used to calculate the output error are obtained by a wheel-rail force measuring device installed on the bogie wheelset.

3. The train derailment risk diagnosis system based on bogie fault characteristic analysis according to claim 2, characterized in that, The training process of the dual-channel neural network architecture introduces an adversarial domain adaptation mechanism, which is used to eliminate the distribution differences of vibration acceleration signals under different line conditions and the distribution differences of bearing temperature signals under different ambient temperatures. The adversarial domain adaptation mechanism includes a feature extractor, a domain discriminator, and a regressor. The feature extractor maps the input signal to a common feature space. The domain discriminator is used to determine which line conditions or which ambient temperature range the features in the common feature space originate from. The regressor is used to regress the equivalent derailment coefficient and the equivalent wheel load reduction rate from the features in the common feature space. During the training process of the dual-channel neural network architecture, the optimization objective of the feature extractor is to minimize the prediction error of the regressor and maximize the discrimination error of the domain discriminator. The optimization objective of the domain discriminator is to minimize the domain discrimination error. Through adversarial training, the feature distribution in the common feature space is decoupled from the line conditions and ambient temperature.

4. The train derailment risk diagnosis system based on bogie fault characteristic analysis according to claim 1, characterized in that, The fault feature calculation module fits the equivalent derailment coefficient obtained from multiple consecutive sampling times into the first time evolution curve by modeling the sequence of the equivalent derailment coefficient changing with time as a linear time-varying system, and using a strong tracking Kalman filter to estimate the state of the linear time-varying system online. The equivalent derailment coefficient is processed using a first strong tracking Kalman filter. The state vector of the first strong tracking Kalman filter includes the equivalent derailment coefficient, the first derivative of the equivalent derailment coefficient, and the second derivative of the equivalent derailment coefficient. The equivalent wheel load reduction rate is processed using a second strong tracking Kalman filter. The state vector of the second strong tracking Kalman filter includes the equivalent wheel load reduction rate, the first derivative of the equivalent wheel load reduction rate, and the second derivative of the equivalent wheel load reduction rate. The fading factor of the strong tracking Kalman filter is adjusted in real time according to the covariance matrix of the state estimation error, so that the strong tracking Kalman filter can maintain its adaptive tracking capability to the real state when the equivalent derailment coefficient changes abruptly. The first derivative component in the state estimate output by the strong tracking Kalman filter is used as the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate, while the second derivative component is used as the derailment coefficient margin consumption acceleration and the wheel load reduction rate margin consumption acceleration.

5. The train derailment risk diagnosis system based on bogie fault characteristic analysis according to claim 4, characterized in that, The real-time adjustment method for the fading factor in a strong tracking Kalman filter is as follows: Calculate the autocorrelation matrix of the state estimation residual sequence, and use the ratio of the trace of the autocorrelation matrix to the preset residual threshold as the basic fading factor; When the absolute value of the second derivative of the equivalent derailment coefficient exceeds the preset threshold for the second derivative of the derailment coefficient or the absolute value of the second derivative of the equivalent wheel load reduction rate exceeds the preset threshold for the second derivative of the wheel load reduction rate, the basic fading factor is multiplied by the preset amplification factor to obtain the final fading factor. At the same time, the derailment coefficient margin consumption rate is added to the derailment coefficient margin consumption acceleration and multiplied by the preset prediction time step to obtain the predicted derailment coefficient margin consumption rate. The wheel load reduction rate margin consumption rate is added to the wheel load reduction rate margin consumption acceleration and multiplied by the preset prediction time step to obtain the predicted wheel load reduction rate margin consumption rate. The predicted derailment coefficient margin consumption rate and the predicted wheel load reduction rate margin consumption rate are used to replace the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate in the bogie fault feature vector. When the absolute value of the second derivative of the equivalent derailment coefficient does not exceed the preset threshold for the second derivative of the derailment coefficient and the absolute value of the second derivative of the equivalent wheel load reduction rate does not exceed the preset threshold for the second derivative of the wheel load reduction rate, the basic fading factor is directly used as the final fading factor, keeping the derailment coefficient margin consumption rate and the wheel load reduction rate margin consumption rate in the bogie fault feature vector unchanged.

6. The train derailment risk diagnosis system based on bogie fault characteristic analysis according to claim 1, characterized in that, The derailment risk diagnosis module takes the smaller of the safe remaining time in the derailment coefficient dimension and the safe remaining time in the wheel load reduction rate dimension as the derailment risk diagnosis result. It also corrects the derailment risk diagnosis result based on the ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate margin consumption rate in the bogie fault feature vector. The correction method for the derailment risk diagnosis result is as follows: when the ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate margin consumption rate exceeds the preset normal ratio range, the derailment risk diagnosis result is divided by a penalty coefficient greater than 1. The magnitude of the penalty coefficient is positively correlated with the degree to which the ratio deviates from the preset normal ratio range. The preset normal ratio range is determined based on the statistical distribution of the ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate consumption rate under the historical normal operating conditions of the bogie.

7. The train derailment risk diagnosis system based on bogie fault characteristic analysis according to claim 6, characterized in that, The penalty coefficient is calculated as follows: Calculate the current ratio of the derailment coefficient margin consumption rate to the wheel load reduction rate margin consumption rate, and calculate the deviation ratio of the current ratio from the center value of the preset normal ratio range. The square of the deviation multiplier is used as the base penalty value, and the smaller of the base penalty value and the preset penalty upper limit value is used as the penalty coefficient.

8. The train derailment risk diagnosis system based on bogie fault characteristic analysis according to claim 1, characterized in that, It also includes a fault type identification module, which is used to input the bogie fault feature vector into a pre-trained fault classifier when the derailment risk diagnosis result is less than a preset time threshold. The fault classifier outputs the fault type identification result of the current bogie. The fault classifier uses a support vector machine, and the kernel function of the support vector machine is a radial basis function. The width parameter of the radial basis function is determined based on the average Euclidean distance between the bogie fault feature vector samples. The fault type identification results include bearing fault categories, gear fault categories, suspension fault categories, and no fault categories.

9. The train derailment risk diagnosis system based on bogie fault characteristic analysis according to claim 8, characterized in that, The training samples for the fault classifier are amplified using synthetic minority class oversampling technology to balance the number of samples for each type of fault. The amplification method of synthetic minority class oversampling technique is as follows: for fault types with fewer than a preset threshold number of samples, two adjacent samples of the corresponding fault type are selected in the bogie fault feature vector space, and new samples are generated by random interpolation on the line connecting the two adjacent samples. The interpolation operation is repeated until the number of samples of the corresponding fault type reaches the preset threshold number of samples.

10. The train derailment risk diagnosis system based on bogie fault characteristic analysis according to claim 1, characterized in that, It also includes a prediction uncertainty quantification module, which is used to output the first prediction interval of the equivalent derailment coefficient and the second prediction interval of the equivalent wheel load reduction rate while the state mapping model outputs the equivalent derailment coefficient and the equivalent wheel load reduction rate. The first and second prediction intervals were obtained through multiple forward inferences after randomly discarding some neurons in the state mapping model. When calculating the derailment risk diagnosis results, the derailment risk diagnosis module uses the upper boundary of the first prediction interval to replace the equivalent derailment coefficient and the upper boundary of the second prediction interval to replace the equivalent wheel load reduction rate, thus obtaining a conservative estimate of the derailment risk diagnosis results.

Images