Wheel-rail equivalent conicity limit value evaluation method based on measured data

Abstract

The application discloses a wheel-rail equivalent conicity limit value evaluation method based on measured data, relates to the technical field of rail transit operation and safety evaluation, and solves the technical problem that the prior art cannot dynamically generate and update safety limit value standards that adapt to expected changes; the application comprises the following steps: synchronously collecting data, solving dynamic equivalent conicity, and extracting a high-order dynamic feature vector; inputting a historical sequence of the feature vector into a trained prediction model, and outputting a vehicle stability margin prediction value of a future target service node; performing statistics on a stability margin prediction value set based on a stream-based approximate quantile algorithm, obtaining a statistical lower bound value that meets a preset reliability target, and deducing a dynamic safety limit value of the high-order dynamic feature; outputting a limit value evaluation table, and iteratively updating the prediction model and the limit value by using new data; the application directly predicts future stability through an end-to-end model, so that the limit value evaluation is based on potential risks in the future rather than only the current state, and the safety management is advanced to the front.

Description

Technical Field

[0001] This invention relates to the field of rail transit operation and maintenance and safety assessment technology, specifically to a method for evaluating wheel-rail equivalent conicity limits based on measured data. Background Technology

[0002] Equivalent taper is a core parameter for evaluating wheel-rail geometry matching and is related to the lateral stability of vehicles. Currently, the industry faces the following main technical limitations in equivalent taper management:

[0003] 1. The widely adopted equivalent taper safety limits are mostly fixed thresholds determined based on historical experience or simulations of typical operating conditions. Such one-size-fits-all limits cannot adapt to the differences in wheel-rail contact states at different wear stages, such as new trains, stable periods, and deterioration periods, nor do they take into account the different requirements for dynamic performance under different track conditions, such as straight lines and curves, resulting in either conservative or overly aggressive limits.

[0004] 2. Existing maintenance procedures mostly rely on current or historical equivalent taper values ​​for exceeding limits, which is a post-event evaluation. They lack predictive assessments of the equivalent taper development trend and its impact on future operational stability, thus failing to achieve preventative safety management.

[0005] 3. Current limits are mostly derived from theoretical calculations or individual typical cases, rather than from the performance statistical distribution of a large group of vehicles under actual complex operating conditions. Therefore, their reliability and confidence level are unclear.

[0006] Some existing technologies aim to improve the accuracy of state perception or prediction. For example, Chinese patent document CN118551644B proposes to invert the current equivalent taper and wear state through vibration signals, but its ultimate goal is state monitoring; Chinese patent document CN120850467B proposes the concept of composite equivalent taper to more accurately describe contact geometric nonlinearity, but its ultimate goal is to calculate a more accurate "current state value". It is evident that existing technologies still cannot dynamically generate and update safety limit standards to adapt to expected changes. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a method for evaluating wheel-rail equivalent taper limits based on measured data. This method solves the technical problem that existing technologies cannot dynamically generate and update safety limit standards to adapt to expected changes based on data-driven and group statistics.

[0008] A method for evaluating the wheel-rail equivalent taper limit based on measured data includes the following steps:

[0009] S1. Data fusion and feature generation: Simultaneously collect the measured wheel-rail profile and lateral dynamic response of the target vehicle, calculate the dynamic equivalent taper sequence based on the dynamic response and profile data, and extract a high-order dynamic feature vector containing gradient sensitive features and wave energy spectrum features.

[0010] S2, Future Stability Margin Prediction: Input the historical sequence of the high-order dynamic feature vector generated in step 1 into the trained hybrid prediction model, and output the predicted value of vehicle stability margin at the future target service node.

[0011] S3. Dynamic safety boundary statistical inference: For a group of vehicles under the same evaluation conditions, obtain a set of predicted stability margin values ​​at the same target service node based on the predicted vehicle stability margin values ​​output in step 2; calculate the statistical lower bound value of the set that satisfies the preset reliability target based on the streaming approximate quantile algorithm; and deduce the dynamic safety limit value of the key features in the higher-order dynamic feature vector in reverse based on the comparison relationship between the statistical lower bound value and the preset minimum safety margin.

[0012] S4. Evaluation Result Generation and Iteration: Output a structured evaluation table containing the dynamic safety limits under different evaluation conditions, used to evaluate the vehicle wheelset status and generate maintenance decisions; collect new data generated after maintenance evaluation, incrementally train the hybrid prediction model, and repeat step S3 to update the dynamic safety limits.

[0013] Further, in step S1, the extraction of high-order dynamic feature vectors includes:

[0014] Calculate the average gradient of the equivalent taper curve within the lateral displacement range of the first preset wheelset, and use it as a gradient-sensitive feature;

[0015] Time-frequency analysis was performed on the dynamic equivalent taper sequence to extract the frequency band energy ratio that matches the main frequency of the vehicle's serpentine motion, which was used as the wave energy spectrum feature.

[0016] The gradient-sensitive feature, the wave energy spectrum feature, and the statistical feature of the dynamic equivalent taper sequence are fused to generate the higher-order dynamic feature vector.

[0017] Further, in step S2, the hybrid prediction model sequentially includes:

[0018] A one-dimensional convolutional neural network module is used to extract the local coupling relationship of multiple features at the same time step in the historical sequence of the high-order dynamic feature vector;

[0019] A bidirectional long short-term memory network module is used to model the bidirectional temporal dependencies of the sequence processed by the one-dimensional convolutional neural network module.

[0020] The attention mechanism module is used to assign weights to the temporal states output by the bidirectional long short-term memory network module in order to focus on key historical stages.

[0021] A fully connected output layer is used to directly output the stability margin prediction value based on the weighted context information.

[0022] Further, in step S3, the calculation of the statistical lower bound based on the streaming approximate quantile algorithm specifically involves:

[0023] The GK algorithm is used to perform a single traversal of the stability margin prediction value set in a streaming process to calculate an approximate value of the specified target percentile, which is the statistical lower bound value.

[0024] Furthermore, in step S3, the reverse derivation of the dynamic security limit specifically includes:

[0025] Select a subgroup of vehicles whose stability margin prediction value is not lower than the statistical lower bound value;

[0026] The distribution of high-order dynamic feature vector values ​​of the vehicle subgroup at the target service node is statistically analyzed.

[0027] The statistical upper bound of the key feature dimension in the value distribution is determined as the dynamic safety limit under the evaluation conditions and service stage.

[0028] Furthermore, step S3 also includes:

[0029] Based on a higher reliability requirement than the preset reliability target, or through a confidence interval estimation method, a more stringent repair recommendation limit than the dynamic safety limit is determined.

[0030] Further, in step S4, the incremental training of the hybrid prediction model specifically involves:

[0031] The high-order dynamic feature vector sequence in the new round of data and its corresponding actual vehicle stability performance are used as new training samples.

[0032] Without forgetting the original knowledge, the parameters of the hybrid prediction model are updated using an incremental learning algorithm.

[0033] The beneficial effects of this invention include:

[0034] By constructing an end-to-end predictive model, the system can directly output predictions of stability for future service stages based on current data. This expands the basis for limit assessment from the current state to potential future risks, enabling early warning of structural safety. Furthermore, the limits are generated based on statistical inference of the future performance distribution of a group, possessing a clear probabilistic statistical foundation. Compared to traditional methods relying on empirical thresholds, its reliability and objectivity are significantly improved. The generated limits are coupled with specific service stages, track conditions, and vehicle types, allowing for dynamic adjustment as wear progresses and the operating environment changes, enhancing the relevance and rationality of limit settings. Finally, by combining a closed-loop mechanism of dynamic limit feedback iterative optimization, the predictive model and limit standards can be continuously optimized using newly collected data, forming an adaptive system with self-evolution capabilities. In addition, the output dynamic safety limits and maintenance recommendation limits, serving as the basis for judgment from state prediction to maintenance action, together constitute key decision outputs in the predictive health management (PHM) system, helping to effectively achieve early intervention and closed-loop management of wheel-rail system performance degradation. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the overall process of a wheel-rail equivalent taper limit evaluation method based on measured data, which is an embodiment of the present invention.

[0036] Figure 2 This is a schematic diagram of the high-order dynamic feature generation process involved in an embodiment of the present invention.

[0037] Figure 3 This is a schematic diagram of the structure of the hybrid prediction model involved in the embodiments of the present invention.

[0038] Figure 4 This is a schematic diagram illustrating the generation of dynamic limits based on streaming quantile statistics and reverse derivation in an embodiment of the present invention.

[0039] Figure 5 This is a graph comparing the effects of the dynamic limit of the present invention and the traditional fixed limit in simulation applications in the embodiments of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to represent selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0041] The following is in conjunction with the appendix Figures 1-5 Specific embodiments of the present invention will be described in detail;

[0042] A method for evaluating the wheel-rail equivalent taper limit based on measured data, such as... Figure 1 As shown, it includes the following steps:

[0043] S1: Data fusion and high-order dynamic feature generation extract features that are strongly correlated with vehicle lateral stability from the raw data.

[0044] S11: Simultaneous acquisition of multi-source data:

[0045] Using a mobile or fixed profiler, the coordinates of discrete points on the wheel track and the top surface of the rail are collected as wheel-rail profile data; accelerometers and gyroscopes are installed in the axle box or frame to collect lateral acceleration. and head angular velocity Through integration and filtering, the real-time lateral displacement of the wheelset (or frame) relative to the track is calculated. As a dynamic response; real-time speed is obtained from the train operation monitoring system. Mileage location The curve radius R and superelevation h of the current segment are matched from the digital twin database of the railway line as the line-operating condition matching data. All data are aligned with GPS timestamps and mileage markers, and the sampling frequency is no less than 100Hz.

[0046] S12: Dynamic equivalent taper Sequence solution:

[0047] For each sampling time Based on the current lateral movement Based on the measured wheel and rail profile point sets, the wheel-rail contact geometry is solved using a numerical iterative method. The iterative process is as follows:

[0048] a. Initial estimation: Assuming rigid contact between wheel and rail, based on Find a point on the vehicle's outline; the angle between its tangent and the horizontal plane is the initial contact angle. .

[0049] b. Find the common point: Find a point on the rail profile such that the angle between the tangent at that point and the horizontal plane is also 0. And the difference in the Y-coordinate between this point and the wheel contact point is equal to .

[0050] c. Normal force iteration: Considering contact elastic deformation, calculate the contact ellipse and normal force according to Hertz contact theory, and fine-tune the contact point position to balance the normal force. Repeat steps bc until the change in contact point coordinates is less than a threshold (e.g., 0.01 mm).

[0051] Calculate the rolling radius of the left and right wheels based on the final contact point position. , ,get Traverse the entire analysis segment (such as an interval) to obtain... sequence.

[0052] Extracting high-order dynamic feature vectors The generation process is as follows Figure 2 As shown, it includes:

[0053] Basic statistical characteristics: (mean) (Standard deviation).

[0054] Gradient-sensitive features :exist Each transverse amount corresponding to the sequence Below, a scatter plot of the equivalent taper versus the lateral displacement can be plotted. This applies to all lateral displacements within the range of 2mm to 4mm. The data points are used for linear fitting. The slope of the fitted line is... This feature directly quantifies the rate of change of the equivalent taper in the critical lateral movement zone. The larger the value, the more sensitive the wheel-rail contact geometry is to small lateral movements, and the more easily it can induce hunting instability.

[0055] Wave energy spectrum characteristics :right The sequence is subjected to wavelet packet transform (WPT). The 'db4' wavelet is selected, and a 3-level decomposition is performed to obtain 8 (2^3) frequency band sub-signals. The sensitive frequency band for vehicle serpentine motion is then calculated. (For example, the energy of the sub-signal corresponding to 1-2Hz, which can be determined based on vehicle model modal analysis) .calculate Total signal energy Ratio: This feature characterizes the proportion of wave energy that induces lateral instability during operation.

[0056] Finally, the above features are fused to generate a high-order dynamic feature vector for each analysis segment:

[0057]

[0058] in This represents the average speed for this section.

[0059] S2: Establish a CNN-BiLSTM-Attention hybrid model as a hybrid prediction model from historical features to future stability results, and predict the future stability margin.

[0060] The model input X is a sequence of high-order feature vectors from N consecutive detections of a wheelset (e.g., the most recent 6 detections, with an interval of approximately 25,000 kilometers between each detection), in the form of... F_high is an N × D matrix (where D is the dimension of F_high).

[0061] The model output Y represents the stability margin for the next detection (future target point). . Defined as the critical speed at which a vehicle loses stability due to snake-like behavior. With the line's maximum permissible operating speed The ratio, i.e. Target value This can be obtained through multibody dynamics simulation or historical safety event data.

[0062] Hybrid prediction models such as Figure 3 As shown, it includes an input layer, a one-dimensional convolutional neural network, a bidirectional long short-term memory network, an attention mechanism, a feature fusion and output layer.

[0063] The input layer is used to receive an input (N, D) dimensional matrix.

[0064] The one-dimensional convolutional neural network (1D-CNN) module is used to learn within the same time step. Various features (such as) With Grad s The local coupling relationship between them specifically involves two 1D convolutional layers:

[0065] Conv1D_1: Number of filters = 32, kernel size = 3, activation function = 'ReLU', padding = 'same'. Output shape: (N, 32).

[0066] Conv1D_2: Number of filters = 64, kernel size = 3, activation function = 'ReLU', padding = 'same'. Output shape: (N, 64).

[0067] Then, Global Average Pooling 1D is performed on each feature channel, compressing the 64-dimensional features at each time step into a single scalar, which is then concatenated back across N time steps. The output shape becomes (N, D), but each feature has been encoded in a higher order.

[0068] The Bidirectional Long Short-Term Memory (BiLSTM) module is used to capture the long-term temporal dependence of the feature sequence in both directions and to model the wear and tear cumulative effect. Specifically, it consists of two BiLSTM layers:

[0069] BiLSTM_1: Number of elements = 128 The output shape is a 128-dimensional splice in both the forward and backward directions: (N, 256).

[0070] BiLSTM_2: Number of elements = 64 The output shape is a bidirectional final state of only the last time step: (128).

[0071] The attention mechanism module assigns weights to the state at each time step of the BiLSTM_1 output, focusing on the most critical historical stages for predicting the future. Specifically, it includes:

[0072] Let the output of BiLSTM_1 be Calculate attention weights:

[0073]

[0074] Where W, b, and v are learnable parameters;

[0075] Next, calculate the weighted context vector:

[0076]

[0077] The feature fusion and output layer is used to process the output of BiLSTM_2. Attention context vector And spliced ​​together with known future operating conditions (such as the planned average speed for the next cycle).

[0078] After passing through two fully connected layers (64 and 32 neurons respectively, with the activation function 'ReLU'), the final stability margin prediction value is output by the linear output layer. .

[0079] Model training: Using historical data, the Adam optimizer is used for training with mean squared error (MSE) as the loss function.

[0080] S3. Statistically infer the dynamic safety boundary and derive the limit value in reverse, transforming the population prediction results into characteristic limit value standards.

[0081] Constructing the future performance distribution of the population:

[0082] For a specific evaluation object, this embodiment takes a certain vehicle model in the 100,000-200,000 km service stage under the condition of R=7000m curve and speed of 250km / h as the evaluation object, and gathers the future of all M wheelsets within the evaluation object. Values, forming a set .

[0083] Streaming quantile statistics based on GK algorithm: Calculating the set Statistical lower bound That is, set of quantiles (e.g.) This indicates the 1st percentile. This aims to achieve a 99% vehicle safety reliability target. The GK algorithm (Greenwald-Khanna) process includes:

[0084] a. Initialization: Create an empty summary structure S to store tuples. ,in These are sample values, i.e., predicted values ​​of the stability margin. This represents the minimum span of the data points represented by the tuple. This represents the maximum uncertainty of the rank of the tuple.

[0085] b. Streaming insertion: For each value in set P :

[0086] i. Find S Location;

[0087] ii. Insert the new tuple at the appropriate position. ,in Let T be the initial rank uncertainty, T be the total number of samples processed, and ε be the allowable error parameter;

[0088] iii. When the summary size exceeds the threshold, a compression operation is triggered, merging adjacent tuples to control memory usage while ensuring accuracy, thus maintaining a space complexity of O(n log n). .

[0089] c. Quantile lookup: For quantiles... Calculate the target rank Traversal Summary Find satisfaction The tuple, its corresponding That is Approximate values ​​of quantiles. Here.

[0090] The statistical lower bound value is obtained using the GK algorithm. Right now .

[0091] A specific calculation example is as follows:

[0092] Set target quantile , The data stream represents the stability margin predictions of the hybrid forecasting model. The first 8 data points are: [1.45, 1.38, 1.50, 1.42, 1.39, 1.47, 1.41, 1.44]. The goal is to be able to quickly find a reliable approximation of the global 1st percentile from the current summary at any given time.

[0093] Create an empty summary structure S to store tuples. ,in These are sample values, i.e., predicted values ​​of the stability margin. This represents the minimum span of the data points represented by the tuple. This represents the maximum uncertainty of the rank of the tuple, with an initial value of [value]. T represents the total amount of data that has been observed so far.

[0094] Process the data stream sequentially and maintain a summary S.

[0095] Insert 1.45 (T=1):

[0096] New tuple: (1.45, 1, floor(2*0.001*1) = 0); Summary S: [(1.45, 1, 0)];

[0097] Insert 1.38 (T=2):

[0098] Insert and maintain ascending order: [(1.38, 1, 0), (1.45, 1, 0)];

[0099] Insert 1.50 (T=3):

[0100] Insert the tail: [(1.38, 1, 0), (1.45, 1, 0), (1.50, 1, 0)];

[0101] Insert 1.42 (T=4):

[0102] Insert between 1.38 and 1.45: [(1.38, 1, 0), (1.42, 1, 0), (1.45, 1, 0), (1.50, 1, 0)];

[0103] Try compression (when T=4):

[0104] Check if (1.42,1,0) and (1.45,1,0) can be merged. The compression condition is: Calculate the compression condition: 1 + 1 + 0 = 2, and 2εT = 2*0.001*4 = 0.008, floor(0.008)=0.

[0105] 2 > 0, the condition is not met, and compression is not possible. This indicates that the algorithm maintains high accuracy when the amount of data is small.

[0106] Continue inserting 1.39, 1.47, 1.41, 1.44 until (T=8). After insertion, summary S retains all values, but the order is sorted. This is just an example, and all intermediate states are not listed here.

[0107] As the amount of data T increases, the threshold of 2εT also increases, allowing compression to control the summary size.

[0108] Suppose that when T=1000, there are two adjacent tuples in the summary: (1.390, 15, 4) and (1.395, 12, 5).

[0109] Calculate the compression conditions:

[0110] ;

[0111] .

[0112] Since 32 > 2, the compression condition is still not met. The intervals represented by these two values ​​are still statistically significant and cannot be merged.

[0113] Suppose that when T=20000, there are two adjacent tuples in the summary: (1.380, 150, 15) and (1.382, 130, 18).

[0114] Calculate the compression conditions:

[0115] ;

[0116] .

[0117] Since 298 > 40, the condition is still not met.

[0118] The above iterative judgment process shows that the GK algorithm does not consistently merge adjacent values, but only does so when it is certain that merging will not cause future quantile query errors to exceed εT. Values ​​like 1.380 and 1.382, even if they appear close, represent a large number of data points ( (Very large), and uncertain The smaller values ​​might introduce significant errors when merged, so they were chosen to be retained. Ultimately, the summary retains... These are all key markers in the data distribution.

[0119] Assuming a stable summary S is obtained after processing the predictions for T = 100,000 wheel pairs, the safety limit needs to be assessed, i.e., by querying the 1st percentile. .

[0120] First, calculate the target rank r:

[0121] ;

[0122] The search target is the value at the 1000th position after sorting from smallest to largest.

[0123] Traverse and query within summary S: Suppose summary S contains the following section:

[0124] ... (1.280, 980, 20) -> (1.290, 25, 5) -> (1.310, 30, 8) ...

[0125] Compute the cumulative minimum rank up to the tuple (1.280, 980, 20). .

[0126] For the tuple (1.290, 25, 5):

[0127] That ;

[0128] That ;

[0129] Judgment: Since the target rank r = 1000 satisfies 980 ≤ 1000 ≤ 1009, it falls within the rank coverage of this tuple.

[0130] Therefore, the algorithm returns = 1.290 as the 1st percentile Approximate value.

[0131] Reverse deriving dynamic safety limits, such as Figure 4 As shown, it includes:

[0132] Determine safety criteria: require statistical lower bounds The This is a pre-set minimum safety margin.

[0133] Screening and statistics:

[0134] a. Filter all from set P The wheelsets constitute a safe subgroup. .

[0135] b. Statistics The current high-order feature vectors corresponding to all wheelsets in the current service stage (100,000-200,000 km) Key features and .

[0136] c. Calculate separately In the group and The statistical upper bound is specifically taken as the 95th percentile in this embodiment. .

[0137] Generate limits: and These are defined as the dynamic safety limits for the object being evaluated.

[0138] Generate recommended limits for turning and repair: To allow for operational margins, a more stringent target can be set based on the existing limits, such as requiring... Alternatively, the dynamic security limit can be multiplied by a discount factor. For example, 0.9, to obtain the recommended limit for repair.

[0139] S4. Generate evaluation results and collect feedback results for closed-loop iteration:

[0140] Output Limit Assessment Table: Apply the above process to all defined assessment objects, including different vehicle models, routes, and phases, to generate a structured limit assessment table.

[0141] Specific structured limit representations are shown in Table 1, which illustrates some examples of dynamic limit assessments generated based on the method of this embodiment. The safety criterion is determined as follows: that is, with a 99% confidence level, the 1st percentile of the predicted stability margin for the target vehicle group at the future target service node is guaranteed. Not less than 1.25 (i.e.) (This is determined based on statistical data). Under this unified standard, based on the specific data of different evaluation units, the following differentiated dynamic characteristic limits are derived in reverse:

[0142] Table 1 provides examples of dynamic limit assessments:

[0143]

[0144] In the table, the evaluation object represents the minimum context in which the limit applies. Based on the service stage (initial break-in period, stable wear period, and accelerated deterioration period), the evaluation object IDs are designated with the first letter A, B, and C, respectively. In field use, the corresponding object must first be located based on the wheelset's current cumulative mileage and the current / planned route.

[0145] The safety limit serves as a yellow alert, representing the theoretical safety boundary. Exceeding this value indicates that the pair has entered a high-risk group, and its future stability is highly likely (>1%) to be lower than the target, requiring close monitoring.

[0146] The recommended turning limit serves as a red alert, indicating a preventative action boundary. Reaching this value means that, based on statistical forecasts, performance is approaching a safe boundary, and turning is recommended during the next planned maintenance. It is typically stricter than the safe limit, allowing a time window for planning and execution.

[0147] The establishment of the basis reflects the data-driven nature and ensures the objectivity and reproducibility of the limits.

[0148] This section uses B02 [Stable Wear Range (100,000-250,000 km), Large Radius Curve (R=7000m)] as an example to illustrate the generation process of the table's content:

[0149] From the database, all historical wheel pair data that meet this condition are selected, assuming a total of N=2900 valid wheel pair detection cycle samples. The recent feature sequence of each sample is input into a pre-trained CNN-BiLSTM-Attention prediction model to obtain its stability margin prediction value for the next cycle (e.g., after 50,000 kilometers). This forms a set P.

[0150] Apply the GK algorithm (assuming ε = 0.001) to set P (containing 2900 values) to find its 1st percentile. ), and obtained the results = 1.26. This is the lower bound of the statistics.

[0151] Safety determination and reverse derivation:

[0152] Preset minimum safety margin target = 1.25.

[0153] Decision: If P1 (1.26) ≥ (1.25) indicates that the current group as a whole meets 99% of the safety requirements.

[0154] Screening for a safe subgroup: Identifying all [subgroups] from 2900 samples. Samples with a value ≥ 1.26 constitute a safe subgroup, assuming there are 2800 samples.

[0155] Upper bound of statistical characteristics: Calculate the current (stable wear stage) performance of 2800 safe samples. and Values. Their 95th percentiles are taken as statistical upper bounds. Assume we get:

[0156] The 95th percentile is 0.045. The 95th percentile is 0.105.

[0157] Generate security limits: i.e. ≤ 0.045, ≤ 0.105.

[0158] Generate recommended turning limits: To allow for margin, multiply the safety limit by a discount factor. The suggested repair value is 0.045 × 0.89 ≈ 0.040.

[0159] Combine the above calculation results with the sample quantity N. Enter the value, etc., into row "B02" of the evaluation form.

[0160] Field Application and Data Feedback: Using Limit Tables for Field Condition Assessment: When a certain wheelset characteristic value ≤0.040, the vehicle is in the safe zone; when 0.040 ≤ When the value is ≤ 0.045, the vehicle enters the warning zone. The system will suggest that it be turned over during the next scheduled maintenance. At this time, it will be marked and monitored more closely, and its subsequent actual operational stability data will be collected to form a new ( , Data pair. When When the value is >0.045, the vehicle enters a high-risk area and requires close monitoring; repairs may need to be arranged immediately.

[0161] Incremental training and limit update of the model:

[0162] The Elastic Weight Consolidation (EWC) algorithm is used for incremental training of the hybrid prediction model. EWC calculates the importance matrix of the parameters and penalizes important old parameters when optimizing the loss for new data, thus preventing catastrophic forgetting.

[0163] Using the updated model and all data including the new data, repeat steps S2 and S3, periodically regenerating the dynamic limit table on a monthly or quarterly basis to complete the closed-loop iteration.

[0164] Specific performance data are as follows: Figure 5As shown in the diagram, a comparative simulation experiment visually demonstrates the significant advantages of the dynamic limit assessment method proposed in this invention compared to the traditional fixed limit strategy. The horizontal axis represents operating mileage, and the vertical axis represents the cumulative number of risk events, including lateral instability alarms.

[0165] The red curve represents the application effect of the traditional fixed limit strategy. This traditional fixed limit strategy refers to determining an equivalent cone upper limit value based on long-term experience and theoretical analysis to ensure that the vehicle does not experience dangerous vibrations such as structural swaying instability. With increasing operating mileage, the number of risk events increases rapidly in an almost linear fashion, accumulating to approximately 11 risk events by 400,000 kilometers. The blue curve represents the application effect of the dynamic limit strategy provided in this embodiment. The curve grows very gently, accumulating to only about 3 risk events by the same operating mileage.

[0166] This embodiment evaluates an 8-car trainset with 16 wheelsets; total operating mileage: 400,000 km = 400,000 km. Total wheelset kilometers = Total operating mileage × Number of wheelsets per train = 400,000 km × 16 wheelsets = 6,400,000 wheelset kilometers; converting this to the standard unit "million wheelset kilometers": Standard workload = 6,400,000 wheelset kilometers ÷ 1,000,000 = 6.4 million wheelset kilometers.

[0167] Based on this, the risk event rate for the two strategies is calculated as follows: Risk event rate = (cumulative number of risk events) / (total operational workload).

[0168] Traditional fixed limit strategy (red curve): Risk event rate = 11 times / 6.4 million wheel pairs·km ≈ 1.72 times / million wheel pairs·km.

[0169] The dynamic limit strategy (blue curve) provided in this embodiment is: Risk event rate = 3 times / 6.4 million wheel pairs·km ≈ 0.47 times / million wheel pairs·km.

[0170] The calculation results show that the strategy in this embodiment achieves a significant improvement in security compared to the traditional strategy:

[0171] The instability risk rate of traditional strategies is as high as 1.72 times per million wheel pairs·km.

[0172] The strategy of this invention reduces the risk rate to 0.47 times per million wheel pairs per kilometer.

[0173] Risk reduction = [(1.72 - 0.47) / 1.72] × 100% ≈ 72.7%.

[0174] As can be seen, under the condition of an operating mileage of 400,000 kilometers, the dynamic limit assessment method proposed in this embodiment can reduce the lateral instability risk event rate from approximately 1.72 times / million wheelset·km to approximately 0.47 times / million wheelset·km, a risk reduction of up to approximately 72.7%. This data intuitively and powerfully demonstrates that this method can effectively resolve the contradiction between over-maintenance and under-maintenance under fixed limits, achieving proactive risk prevention while providing scientific support for optimizing maintenance decisions, thus achieving the optimal balance between safety and economy.

[0175] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

Claims

1. A method for evaluating the wheel-rail equivalent taper limit based on measured data, characterized in that, Includes the following steps: S1. Data fusion and feature generation: Simultaneously collect the measured wheel-rail profile and lateral dynamic response of the target vehicle, calculate the dynamic equivalent taper sequence based on the dynamic response and profile data, and extract a high-order dynamic feature vector containing gradient sensitive features and wave energy spectrum features. S2, Future Stability Margin Prediction: Input the historical sequence of the high-order dynamic feature vector generated in step S1 into the trained hybrid prediction model, and output the predicted value of vehicle stability margin at the future target service node. S3. Dynamic safety boundary statistical inference: For a group of vehicles under the same evaluation conditions, obtain the set of stability margin prediction values ​​at the same target service node based on the vehicle stability margin prediction values ​​output in step S2. The statistical lower bound of the set satisfying the preset reliability target is calculated based on the streaming approximate quantile algorithm; Based on the comparison between the statistical lower bound and the preset minimum safety margin, the dynamic safety limit of the key features in the higher-order dynamic feature vector is derived in reverse, including: requiring the statistical lower bound to be greater than or equal to the preset minimum safety margin, and screening out the vehicle subgroup whose stability margin prediction value is not lower than the statistical lower bound. The distribution of high-order dynamic feature vector values ​​of the vehicle subgroup at the target service node is statistically analyzed; the 95th percentile of the key feature value distribution in the high-order dynamic feature vector is taken as the statistical upper bound of the key feature dimension, and determined as the dynamic safety limit under the evaluation conditions and service stage. S4. Evaluation Result Generation and Iteration: Output a structured evaluation table containing the dynamic safety limits under different evaluation conditions, used to evaluate the vehicle wheelset status and generate maintenance decisions; collect new data generated after maintenance evaluation, incrementally train the hybrid prediction model, and repeat step S3 to update the dynamic safety limits.

2. The method for evaluating the wheel-rail equivalent taper limit based on measured data according to claim 1, characterized in that, In step S1, the extraction process of the high-order dynamic feature vector includes: Calculate the average gradient of the equivalent taper curve within the lateral displacement range of the first preset wheelset, and use it as a gradient-sensitive feature; Time-frequency analysis was performed on the dynamic equivalent taper sequence to extract the frequency band energy ratio that matches the main frequency of the vehicle's serpentine motion, which was used as the wave energy spectrum feature. The gradient-sensitive feature, the wave energy spectrum feature, and the statistical feature of the dynamic equivalent taper sequence are fused to generate the higher-order dynamic feature vector.

3. The method for evaluating the wheel-rail equivalent taper limit based on measured data according to claim 1, characterized in that, In step S2, the hybrid prediction model includes, in sequence: A one-dimensional convolutional neural network module is used to extract the local coupling relationship of multiple features at the same time step in the historical sequence of the high-order dynamic feature vector; A bidirectional long short-term memory network module is used to model the bidirectional temporal dependencies of the sequence processed by the one-dimensional convolutional neural network module. The attention mechanism module is used to assign weights to the temporal states output by the bidirectional long short-term memory network module in order to focus on key historical stages. A fully connected output layer is used to directly output the stability margin prediction value based on the weighted context information.

4. The method for evaluating the wheel-rail equivalent taper limit based on measured data according to claim 1, characterized in that, In step S3, the calculation of the statistical lower bound based on the streaming approximate quantile algorithm specifically involves: The GK algorithm is used to perform a single traversal of the stability margin prediction value set in a streaming process to calculate an approximate value of the specified target percentile, which is the statistical lower bound value.

5. The method for evaluating the wheel-rail equivalent taper limit based on measured data according to claim 1, characterized in that, Step S3 also includes: Based on a higher reliability requirement than the preset reliability target, or through a confidence interval estimation method, a more stringent repair recommendation limit than the dynamic safety limit is determined.

6. The method for evaluating the wheel-rail equivalent taper limit based on measured data according to claim 1, characterized in that, In step S4, the incremental training of the hybrid prediction model specifically involves: The new data generated after the maintenance assessment is used as the new round of data. The high-order dynamic feature vector sequence and its corresponding actual vehicle stability performance are extracted from the new round of data as new training samples. Without forgetting the original knowledge, the parameters of the hybrid prediction model are updated using an incremental learning algorithm.

Images