A locomotive wheel set turning and wheel set axle sequence planning method and system

Abstract

A locomotive wheel reprofiling wheel set axle sequence planning method and system, comprising a multi-source data integration platform, a wheel key size prediction module, a reprofiling axle sequence sorting module and a traction signal output unit, each module cooperates to complete intelligent sorting of the reprofiling wheel set and output of target axle sequence data. The technical scheme of the application breaks through the bottleneck of traditional experience prediction, realizes accurate prediction of the size of the locomotive wheel, builds a multi-source data integration platform, realizes unified control of the related data of the locomotive wheel, constructs a scientific reprofiling axle sequence sorting system, solves the problem of non-standard manual sorting, and establishes an innovative linkage control mechanism to promote the upgrading of the automatic positioning operation.

Description

Technical Field

[0001] This invention relates to the fields of railway locomotive maintenance, deep learning, and intelligent control technology, and in particular to a method and system for planning the axle sequence of locomotive wheelset repair without removing wheels. Background Technology

[0002] In the field of railway locomotive maintenance, the planning of the axle rotation sequence and the automatic wheelset placement are two core preliminary processes in locomotive wheelset refining without wheel removal. Their rationality and accuracy directly determine the overall quality and maintenance efficiency of the refining operation. However, currently, the wheelset refining sequence in the industry largely relies on the experience of professional engineers to formulate work plans, while the wheelset placement operation still uses the traditional manual control mode, resulting in significant technical bottlenecks and operational pain points.

[0003] The current wheelset turning sequence relies entirely on the experience and judgment of professional engineers, without establishing a quantitative turning standard based on multi-source data. The turning process only refers to some alarm information and experience-based size estimates, without comprehensively considering key indicators such as platform alarm information, dynamic detection alarm information, flaw detection results, and size prediction data of locomotive wheels. As a result, the rationality of the wheelset turning sequence cannot be guaranteed.

[0004] Especially in controlling the wheel diameter difference of locomotives with the same axle, bogie, section, and wheel, experience-based sorting is prone to causing wheelsets to fail to meet the wheel diameter difference requirements of locomotive assembly after wheel turning, leading to rework, significantly delaying the turning operation, and increasing maintenance costs. At the same time, the manual sorting standards of different locomotive depots are not uniform, making it impossible to form a unified turning operation standard across the entire railway system, which is not conducive to the overall improvement of the industry's maintenance level.

[0005] Locomotive axle turning sequence planning is highly dependent on wheel size information. If the key dimensions of the wheel (tread wear, flange thickness, wheel diameter) exceed the limit, or the wheel diameter difference exceeds the standard threshold, the wheelset needs to be turned first. Even if the dimensions are within the acceptable range, the wheel diameter parameter is still the core basis for the turning sequence of other wheelsets.

[0006] However, the measurement of key wheel dimensions suffers from significant periodic limitations, with data gaps existing on dimensional changes within measurement intervals. Current dimensional prediction relies heavily on manual experience, which not only considers a single dimension and lacks comprehensive coverage of influencing factors, making a holistic assessment difficult, but also demonstrates a high degree of coupling between wheel wear patterns and operating parameters such as locomotive type, axle sequence layout, and route of operation. Wear characteristics vary significantly under different operating conditions, and manual prediction methods cannot accurately capture these complex relationships, leading to substantial prediction deviations.

[0007] Furthermore, although some locomotive depots have equipped themselves with dynamic wheel size measurement equipment, the measurement data is highly volatile due to insufficient equipment accuracy, and the reference value of a single sample data point is very limited. This not only makes it impossible to accurately assess key wheel dimensions, but also makes it difficult to effectively predict potential risks such as exceeding size limits or wheel diameter differences, and fails to provide reliable data support for turning and repair work planning.

[0008] In the wheelset placement stage before locomotive overhaul, the operation control of external traction equipment has long relied on manual issuance of operating commands, and an automated signal transmission and execution mechanism deeply linked to the overhaul sequencing results has not yet been established. According to research, most locomotive depots still use manual remote control to operate dual-purpose road-rail vehicles to complete the wheelset placement operation; in some operation scenarios, it is even necessary to manually drive dual-purpose road-rail vehicles to perform traction tasks.

[0009] This operating mode is not only inefficient, but also prone to human error leading to deviations in wheelset positioning accuracy, directly affecting the clamping accuracy of subsequent turning equipment and thus reducing the quality of the turning operation. Furthermore, the manual control mode requires multiple operators to work together, further hindering efficiency and posing safety risks due to human error. Summary of the Invention

[0010] To overcome the existing shortcomings, this invention proposes a method and system for planning the axle sequence of locomotive wheelset repair without wheel removal.

[0011] A locomotive wheelset axle sequence planning system for non-dismounting wheelset repair includes a multi-source data integration platform, a wheel key dimension prediction module, a wheelset repair sequence sorting module, and a traction signal output unit. These modules work together to intelligently sort the wheelsets to be repaired and output the target axle sequence data.

[0012] The multi-source data integration platform is used to integrate 6A / 5T alarm platform data, dynamic detection data, flaw detection results and wheel size measurement data through standardized interfaces, and classify and archive them according to the hierarchy of "locomotive depot-locomotive type-routes-axle sequence-end position" to form a standardized data resource pool;

[0013] The wheel key dimension prediction module is built on deep learning technology and adopts a combination of classification modeling, LSTM neural network, transfer learning and Optuna hyperparameter optimization scheme to accurately predict wheel flange thickness, tread wear and wheel diameter, filling the data gap in the measurement interval.

[0014] The wheel repair sequence sorting module takes wheelsets as the smallest unit of analysis and combines wheelset health status classification (healthy, early warning, alarm) and wheel diameter parameters to realize intelligent planning of wheelset repair sequence. The sorting logic prioritizes alarm wheelsets and sorts them in ascending order of wheel diameter within the same bogie.

[0015] The traction signal output unit, based on the RS-485 serial communication protocol, converts the sorting results into standardized control signals to drive the dual-purpose road-rail vehicle to achieve automated wheelset placement and receive positioning status signals transmitted back from the equipment.

[0016] The data access of the multi-source data integration platform includes:

[0017] The 6A alarm platform provides data on tread level 2 alarms, tread level 1 alarms, and tread early warnings.

[0018] Data from the 5T alarm platform, including polygonal level 2 alarm, polygonal level 1 alarm, polygonal early warning, tread level 2 alarm, tread level 1 alarm, and tread early warning.

[0019] Severe, moderate, and mild alarms for scratch detection, wheel shape detection, and flaw detection based on dynamic detection data;

[0020] Non-compliant data in flaw detection results and crack damage reports;

[0021] The results of manual and dynamic measurements of wheel size data.

[0022] The construction of the wheel critical dimension prediction module includes:

[0023] Based on the time-series dataset within the turning cycle, the data includes wheel flange thickness sequence, tread wear sequence, and post-turning mileage sequence;

[0024] An LSTM neural network is used to mine the long-term dependency between size and mileage. Its hyperparameters are optimized using the Optuna framework. The optimization objects include sliding window length, number of hidden layers, number of LSTM layers, learning rate, frozen layer range, Dropout rate, and L2 regularization coefficient.

[0025] A transfer learning pre-training-fine-tuning paradigm is introduced, which pre-trains using similar wheel data and then fine-tunes using target wheel data to address the problem of small sample data.

[0026] The sorting logic of the turning shaft sequence sorting module includes:

[0027] Wheelset health status determination: If any wheelset has an alarm, the wheelset will be alarmed; if there is no alarm but there is a warning, then a warning will be issued; if all wheels are healthy, the wheelset is healthy.

[0028] Wheel diameter correction: When the rim thickness is below the threshold, the expected wheel diameter loss is calculated using the rim refining recovery ratio parameter model. Corrected wheel diameter = original wheel diameter - expected wheel diameter loss.

[0029] Sorting rules:

[0030] Scenario 1 (Alarm wheelsets exist): Prioritize sorting alarm wheelsets by wheel diameter from smallest to largest, then process the warning and healthy wheelsets within the same bogie;

[0031] Scenario 2 (Warning wheelsets only): Sort directly by wheel diameter from smallest to largest;

[0032] Scenario 3 (All healthy wheelsets): Sort by wheel diameter from smallest to largest.

[0033] The communication protocol of the traction signal output unit is RS-485 serial port, baud rate 9600bps, 8 data bits, no parity, and 1 stop bit;

[0034] The system sends signals to the dual-purpose road and rail vehicle including the tractor address, function code, locomotive type, current axle sequence, entry point, target axle sequence, traction confirmation, and number of data items, encoded in hexadecimal.

[0035] The feedback signals from the dual-purpose road-rail vehicle include the tractor address, function code, positioning status, and number of data points, enabling closed-loop control.

[0036] The system is also applicable to maintenance operations in urban rail transit, high-speed trains, and rolling stock, and can be promoted across different fields by adapting to different models and data interfaces.

[0037] A method and system for planning the axle sequence for locomotive wheelset repair without wheel removal, the method comprising the following steps:

[0038] Step 1: Construction and Calculation of Wheel Critical Dimension Prediction Model

[0039] Collect manual and dynamic measurement data, including data types such as tread wear, wheel flange thickness and corresponding post-turning mileage; classify the datasets according to the hierarchy of "locomotive depot-machine type-routes-axle sequence-end position", and form a time-series dataset with the current turning cycle start node as the benchmark (post-turning mileage set to 0);

[0040] A prediction model is built based on the technical solution of "classification modeling + LSTM + transfer learning + Optuna":

[0041] LSTM neural networks are used to explore the long-term dependency between wheel size and mileage. The LSTM model controls the information flow through forget gates, input gates, and output gates. Its internal computation process is defined by the following formula:

[0042]

[0043] In the formula: W f and b f —respectively, the Gate of Oblivion f t The weight matrix and bias terms;

[0044] W I and bI —These are the input gates I t The weight matrix and bias terms;

[0045] W C and b C —Memory cell C t The weight matrix and bias terms;

[0046] W O and b O —These are output gates O t The weight matrix and bias terms;

[0047] X t — Enter at the current moment;

[0048] h t-1 —The state was hidden a moment ago;

[0049] tanh — activation function;

[0050] t —Candidate state of memory cells;

[0051] *——Hada code multiplication.

[0052] Introducing a transfer learning pre-training-fine-tuning paradigm, the model is first pre-trained using similar wheel data (30-100 pre-training rounds), and then fine-tuned to adapt to the target wheel (50-200 fine-tuning rounds).

[0053] Optimize hyperparameters using the Optuna framework (e.g., sliding window length 3-8, number of hidden layers 30-100), and improve model accuracy by combining K-fold cross-validation.

[0054] The predicted values ​​of the output rim thickness and wheel diameter are controlled with a root mean square error (RMSE) within 0.05 mm.

[0055] Step 2: Multi-source data acquisition and preprocessing

[0056] Multi-source data is accessed through standardized interfaces: 6A / 5T alarm information, dynamic detection alarm data, flaw detection results, and historical wheel size data; outliers are cleaned using the 3σ principle and integrated into a standardized data resource pool.

[0057] Step 3: Multi-source injury data fusion analysis and health status assessment

[0058] Based on the data from step 2, a fusion analysis was performed within a specified time frame; in accordance with the limitations of the "Railway Technical Management Regulations," the health status of each wheel was divided into three levels: "healthy," "early warning," and "alarm."

[0059] Alarms: Tread wear >7.0mm, flange thickness <23.0mm or >34.0mm, wheel diameter difference exceeding the standard for the same bogie, etc.

[0060] Warning: Tread wear 5.0-7.0mm, rim thickness 25.0-33.8mm, etc.

[0061] Health: Meets standards.

[0062] Step 4: Wheelset data conversion and wheel diameter correction

[0063] Change the unit of analysis from wheel to wheelset:

[0064] The wheel diameter is the minimum value of the two wheels.

[0065] Health status determination: If any round of alarms occurs, the corresponding round will also alarm; if there are no alarms but a warning is issued, a warning will be issued; if all are healthy, the status is healthy.

[0066] Wheel diameter correction: When the rim thickness is lower than the threshold (default 25.0mm, adjustable to 28.0mm), the expected wheel diameter loss is calculated by the rim turning recovery ratio parameter model. Corrected wheel diameter = original wheel diameter - loss.

[0067] This parametric model is constructed based on geometric analysis of tread repair and big data regression.

[0068] Step 5: Calculation of the turning and repair shaft sequence sorting model and generation of the optimal shaft sequence

[0069] Based on the refining plan, select the wheelsets that need refining and apply sorting logic:

[0070] Scenario 1 (alarm wheelsets exist): Alarm wheelsets are sorted by wheel diameter from smallest to largest. Within the same bogie, the smaller wheel diameter alarm axle is processed first, followed by the warning and healthy wheelsets in sequence.

[0071] Scenario 2 (Warning wheelsets only): Sort directly by wheel diameter from smallest to largest;

[0072] Scenario 3 (All healthy wheelsets): Sort by wheel diameter from smallest to largest;

[0073] Output the optimal turning sequence list to ensure it meets the locomotive assembly requirements.

[0074] Step 6: Traction signal output and execution status reception

[0075] Control signals are sent via RS-485 serial communication (baud rate 9600bps, 8 data bits, no parity, 1 stop bit):

[0076] The system sends the following signal format: tractor address, function code 0xa1 (enable), locomotive model, current axle sequence, entry point, target axle sequence, traction confirmation (0x01 execute), number of data items 0x08, and termination code 0x0d;

[0077] The dual-purpose road-rail vehicle transmits signals including positioning status (0x01 completed), enabling automated closed-loop control of wheelset placement.

[0078] The hyperparameter optimization in step 1 adopts the Optuna framework, and the optimization objects include sliding window length, number of hidden layers, number of LSTM layers, learning rate, frozen layer range, dropout rate, and L2 regularization coefficient. The relationship between learning rate and RMSE is verified by curve.

[0079] The wheel flange resurfacing recovery ratio parameter model in step 4 is constructed based on big data of resurfacing records across the entire road, and optimized using geometric analysis and GA algorithm to ensure wheel diameter correction accuracy.

[0080] The method is applicable to railway locomotive depots, urban rail transit, and high-speed trains, and can be used across different scenarios by adapting data interfaces. Attached Figure Description

[0081] Figure 1 This is a flowchart of the LSTM analysis process.

[0082] Figure 2 For the prediction scheme process.

[0083] Figure 3 Flowchart for constructing the recovery ratio parameter model for wheel flange refinishing.

[0084] Figure 4 This is a diagram of the YOLOv8n-SMC network structure.

[0085] Figure 5 The figure shows the preliminary analysis results of wheel diameter loss during turning, corresponding to the rim thickness recovery amount.

[0086] Figure 6 The final analysis result of the wheel diameter loss corresponding to the rim thickness recovery is shown in the figure. Detailed Implementation

[0087] The following describes in detail, with reference to the accompanying drawings and specific embodiments, a locomotive wheel alignment sequence planning method and system provided by the present invention.

[0088] according to Figure 1-2 As shown, a locomotive wheelset axle sequence planning system for non-removable wheel turning includes a multi-source data integration platform, a wheel key dimension prediction module, a turning axle sequence sorting module, and a traction signal output unit. These modules work together to intelligently sort the turning wheelsets and output the target axle sequence data.

[0089] The multi-source data integration platform is used to integrate 6A / 5T alarm platform data, dynamic detection data, flaw detection results and wheel size measurement data through standardized interfaces, and classify and archive them according to the hierarchy of "locomotive depot-locomotive type-routes-axle sequence-end position" to form a standardized data resource pool;

[0090] The wheel key dimension prediction module is built on deep learning technology and adopts a combination of classification modeling, LSTM neural network, transfer learning and Optuna hyperparameter optimization scheme to accurately predict wheel flange thickness, tread wear and wheel diameter, filling the data gap in the measurement interval.

[0091] The wheel repair sequence sorting module takes wheelsets as the smallest unit of analysis and combines wheelset health status classification (healthy, early warning, alarm) and wheel diameter parameters to realize intelligent planning of wheelset repair sequence. The sorting logic prioritizes alarm wheelsets and sorts them in ascending order of wheel diameter within the same bogie.

[0092] The traction signal output unit, based on the RS-485 serial communication protocol, converts the sorting results into standardized control signals to drive the dual-purpose road-rail vehicle to achieve automated wheelset placement and receive positioning status signals transmitted back from the equipment.

[0093] The data access of the multi-source data integration platform includes:

[0094] The 6A alarm platform provides data on tread level 2 alarms, tread level 1 alarms, and tread early warnings.

[0095] Data from the 5T alarm platform, including polygonal level 2 alarm, polygonal level 1 alarm, polygonal early warning, tread level 2 alarm, tread level 1 alarm, and tread early warning.

[0096] Severe, moderate, and mild alarms for scratch detection, wheel shape detection, and flaw detection based on dynamic detection data;

[0097] Non-compliant data in flaw detection results and crack damage reports;

[0098] The results of manual and dynamic measurements of wheel size data.

[0099] The construction of the wheel critical dimension prediction module includes:

[0100] Based on the time-series dataset within the turning cycle, the data includes wheel flange thickness sequence, tread wear sequence, and post-turning mileage sequence;

[0101] An LSTM neural network is used to mine the long-term dependency between size and mileage. Its hyperparameters are optimized using the Optuna framework. The optimization objects include sliding window length, number of hidden layers, number of LSTM layers, learning rate, frozen layer range, Dropout rate, and L2 regularization coefficient.

[0102] A transfer learning pre-training-fine-tuning paradigm is introduced, which pre-trains using similar wheel data and then fine-tunes using target wheel data to address the problem of small sample data.

[0103] The sorting logic of the turning shaft sequence sorting module includes:

[0104] Wheelset health status determination: If any wheelset has an alarm, the wheelset will be alarmed; if there is no alarm but there is a warning, then a warning will be issued; if all wheels are healthy, the wheelset is healthy.

[0105] Wheel diameter correction: When the rim thickness is below the threshold, the expected wheel diameter loss is calculated using the rim refining recovery ratio parameter model. Corrected wheel diameter = original wheel diameter - expected wheel diameter loss.

[0106] Sorting rules:

[0107] Scenario 1 (Alarm wheelsets exist): Prioritize sorting alarm wheelsets by wheel diameter from smallest to largest, then process the warning and healthy wheelsets within the same bogie;

[0108] Scenario 2 (Warning wheelsets only): Sort directly by wheel diameter from smallest to largest;

[0109] Scenario 3 (All healthy wheelsets): Sort by wheel diameter from smallest to largest.

[0110] The communication protocol of the traction signal output unit is RS-485 serial port, baud rate 9600bps, 8 data bits, no parity, and 1 stop bit;

[0111] The system sends signals to the dual-purpose road and rail vehicle including the tractor address, function code, locomotive type, current axle sequence, entry point, target axle sequence, traction confirmation, and number of data items, encoded in hexadecimal.

[0112] The feedback signals from the dual-purpose road-rail vehicle include the tractor address, function code, positioning status, and number of data points, enabling closed-loop control.

[0113] The system is also applicable to maintenance operations in urban rail transit, high-speed trains, and rolling stock, and can be promoted across different fields by adapting to different models and data interfaces.

[0114] A method and system for planning the axle sequence for locomotive wheelset repair without wheel removal, the method comprising the following steps:

[0115] Step 1: Construction and Calculation of Wheel Critical Dimension Prediction Model

[0116] Collect manual and dynamic measurement data, including data types such as tread wear, wheel flange thickness and corresponding post-turning mileage; classify the datasets according to the hierarchy of "locomotive depot-machine type-routes-axle sequence-end position", and form a time-series dataset with the current turning cycle start node as the benchmark (post-turning mileage set to 0);

[0117] For this invention, the key dimensions of the wheel are the rim thickness, tread wear, and wheel diameter. Since the wheel diameter is not measured in routine measurements, the wheel diameter value is usually estimated using the tread wear, and the calculation method is as follows:

[0118]

[0119] In the formula: d n The current wheel diameter (mm), d t The wheel diameter was measured after the most recent turning; t n This represents the current tread wear (mm).

[0120] Currently, available historical wheel measurement data includes both manual and dynamic measurement data. Manual measurement data is obtained during locomotive maintenance using a fourth type of inspection device or wheel diameter difference measurement, with a measurement cycle of 70,000-250,000 km. Dynamic measurement data is obtained during routine locomotive maintenance by the Railway Locomotive and Rolling Stock Inspection System (TDDS), with a measurement cycle of 1,000-3,000 km. However, some locomotives operate outdoors for extended periods, or maintenance workshops lack TDDS equipment, resulting in limited dynamic wheel measurement data and a scarcity of historical measurement data. The corresponding post-repair mileage can be found in the database.

[0121] First, a dataset is constructed, including flange thickness F, tread wear T, and the corresponding post-return mileage S. Locomotive wheels undergo multiple returning operations throughout their lifespan, thus forming several independent returning cycles. Returning operations cause abrupt changes in key dimensions such as flange thickness, tread wear, and wheel diameter, which is fundamentally different from the gradual wear during normal operation. Therefore, when performing dimensional predictions, cross-cycle modeling is unnecessary; predictions only need to be limited to the current returning cycle.

[0122] First, the post-turning measurement data at the start of the current turning cycle is selected as the modeling benchmark. Specific parameters include the rim thickness f. l tread wear t l wheel diameter d l And the post-turning travel mileage s of that node. l The value is assigned to 0. Since wheel turning restores the wheel to its standard tread profile and effectively eliminates defects such as polygonal wear and abnormal wear, using this as the starting data for prediction ensures the accuracy and reliability of subsequent analysis. Subsequently, the manually measured rim thickness f is used... p tread wear t p wheel diameter d p and the corresponding mileage after turning s p , and the rim thickness f obtained by dynamic measurement a tread wear ta wheel diameter d a and the corresponding mileage after turning s a The data is then integrated into the dataset; next, it is sorted in ascending order based on the post-turning mileage, and the remaining parameters in the dataset are sorted synchronously with the mileage sequence, ultimately forming a sequence containing the flange thickness F={f1, f2, ..., f n}, Tread wear sequence T={t1, t2, ..., t n}, and the corresponding post-turning travel mileage sequence S={s1, s2, ..., s n A standardized time-series dataset.

[0123] Flange thickness and tread wear are core dimensions of locomotive wheels. Their variation patterns are influenced by multiple factors, including locomotive type, axle sequence layout, operating routes, and braking conditions, often exhibiting significant nonlinear evolution trends. Under certain special operating conditions, locomotive wheels may even show non-monotonic changes in flange thickness. Traditional linear fitting or manual experience-based judgment methods are insufficient to accurately capture these complex dimensional variation characteristics, leading to significant deviations between predicted results and actual conditions.

[0124] This invention addresses the large-scale application needs of railway locomotives nationwide, fully considering the nonlinear evolution trend of wheel flange thickness and tread wear. Furthermore, the flange thickness exhibits certain non-monotonic variation characteristics.

[0125] This invention addresses the large-scale application needs of railway locomotives nationwide, fully considering the nonlinear evolution trends of wheel flange thickness and tread wear. Furthermore, wheel flange thickness exhibits certain non-monotonic variation characteristics. A predictive model is constructed based on a "classification modeling + LSTM + transfer learning + Optuna" technical solution. The specific implementation process is as follows: First, based on the wheel's full life-cycle operating data, predictive modeling of wheel flange thickness and tread wear is carried out according to the hierarchical classification logic of "locomotive depot - locomotive type - route - axle sequence - end position," thereby mitigating the differential impact of wheel flange wear under different operating conditions. Second, relying on LSTM neural networks to deeply mine the variation patterns in time-series data, combined with transfer learning technology to overcome the application limitations of small sample data, and simultaneously introducing the Optuna framework and cross-validation mechanism to optimize model hyperparameters, significantly improving the stability and generalization ability of the predictive model. Finally, the model performance is verified using measured data with a test interval of no less than one maintenance cycle to ensure that the model meets the requirements of field applications.

[0126] LSTM is a key improvement on traditional RNNs. By introducing three gated units—forget gate, input gate, and output gate—and a memory cell structure, it significantly enhances the model's ability to handle long-term dependencies. Specifically, the forget gate filters out historical information to be retained, the input gate controls the updating of new information, and the output gate determines the output of the current state. These three work together to ensure the accurate transmission of information in the time series, thus effectively capturing long-term dependencies in the sequence data. The analysis process is shown below. LSTM is well-suited for tasks such as sequence data prediction, demonstrating superior performance in areas such as power generation prediction and financial time series analysis. The internal computation process of the LSTM model is defined by the following formula:

[0127]

[0128] In the formula: W f and b f —respectively, the Gate of Oblivion f t The weight matrix and bias terms;

[0129] W I and b I —These are the input gates I t The weight matrix and bias terms;

[0130] W C and b C —Memory cell C t The weight matrix and bias terms;

[0131] W O and b O —These are output gates O t The weight matrix and bias terms;

[0132] X t — Enter at the current moment;

[0133] h t-1 —The state was hidden a moment ago;

[0134] tanh — activation function;

[0135] t —Candidate state of memory cells;

[0136] *——Hada code multiplication.

[0137] Taking the prediction of wheel flange thickness as an example, using the LSTM model, the wheel flange thickness sequence F={f1, f2, ..., f n} and the corresponding post-turning travel mileage sequence S={s1, s2, ..., s nConduct systematic and in-depth analysis to uncover potential correlations between data. Based on domain prior knowledge and model tuning experience, adjust the sliding window length S of the LSTM model. op Hidden layer range (Hop), number of LSTM layers (H) op Learning rate lr, frozen layer range F op Dropout rate D op and L2 regularization coefficient L op It was used as a hyperparameter optimization target, and the corresponding search range was set.

[0138] LSTM model analysis requires a certain amount of data support; however, the amount of dynamic measurement data on the dimensions of sample wheels is limited. Relying solely on historical usage data of a single wheel (covering multiple historical cycles) is insufficient to meet the data requirements for model training. Furthermore, the mileage of each wheel is relatively short, lacking sufficient data to reach the target mileage. To address these issues, wheel usage data from the same depot, same aircraft type, same route, same axle sequence, and same end position are introduced to expand the training set data size.

[0139] Transfer learning is a technique that allows a model to transfer knowledge learned in a source task to a target task. Its core principle is to pre-train the model to master general features, and then fine-tune it to adapt to the new task. Transfer learning has significant advantages in small-data scenarios. It can effectively improve prediction accuracy while reducing the risk of overfitting, even with limited data.

[0140] This model employs the pretrain-finnetune paradigm in transfer learning to analyze dynamic measurement data of wheel flange thickness. Specifically: First, from the qualified wheel data, historical usage data of the remaining wheels (excluding the target sample data) are selected as the source dataset, and based on this, T... pt The model is pre-trained based on the number of wheels. This process fully mines and accumulates the common features inherent in the usage data of wheels in similar usage environments. Subsequently, using the target sample wheel data as the transfer target, the pre-trained model is switched to fine-tuning mode to complete model adaptation for the target task. In this mode, T-processing is performed on the target wheel data. ft Fine-tuning the number of wheels allows the model to accurately adapt to the unique flange thickness variation characteristics of the target wheel, thereby improving the model's performance on the target task. Specifically, the number of pre-training wheels T in transfer learning... pt and fine-tuning round number T ft As a hyperparameter optimization target.

[0141] For prediction sample scenarios with different wheel flange thicknesses and tread wear, the sequence data volume is n, and K-fold cross-validation is used. The dataset for each wheel is randomly divided into... mutually exclusive subsets, in which One subset is used as the training set for the prediction model, and one subset is used as the validation set. The validation process involves a total of [number] loops. This ensures that each subset is used as a validation set once.

[0142] The hyperparameter optimization scheme of the model adopts the Optuna automated hyperparameter optimization framework as its core, and performs parameter tuning for the wheel key dimension prediction model. In specific operation, Optuna is used to optimize nine pre-defined hyperparameters. The validation set divided by K-fold cross-validation is used as the evaluation benchmark data, and the average validation loss is used as the evaluation index. Then, through 100 trials, the hyperparameter combination that minimizes the objective function is searched.

[0143] Based on experience, the hyperparameter optimization space for the LSTM model is set as follows: sliding window length Sop = 3-8, number of hidden layers Hop = 30-100, number of LSTM layers Nop = 1-4, learning rate lr = 0.0001-0.01, frozen layer range Fop = 0-2, dropout rate Dop = 0-0.5, and L2 regularization coefficient Lop = 0.00001-0.001. Furthermore, the number of pre-training epochs Tpt in transfer learning is 30-100, and the number of fine-tuning epochs T... ft It ranges from 50 to 200.

[0144] Analysis shows that, using wheel usage data from Type I locomotives in Locomotive Depot A, the hyperparameter optimization result through the Optuna framework is: the optimal parameter S of the LSTM model. op =3,H op =68, N op =1, lr=0.009878, F op =0, D op =0.27, L op =0.000967. Furthermore, the optimal result T for transfer learning... pt =32, T ft =193.

[0145] The wheel flange thickness data of Type I locomotives in Locomotive Depot A (a total of 7 wheels) was used as the test sample, with test intervals exceeding one maintenance cycle (77,000 km). Since the wheel flange prediction is relatively precise, the root mean square error (RMSE) was chosen as the evaluation standard for the algorithm's prediction accuracy, effectively highlighting more serious error items. The calculation method is shown in the following formula, and the analysis results are shown in the table below. The prediction analysis results show that the RMSE for all sample wheels is 0.05 mm. Among them, wheel 7 has a larger absolute error; investigation revealed that the material of wheel 7 differs from the other wheels.

[0146]

[0147] In the formula: f ti — Actual flange thickness (mm);

[0148] f pi —Predicted rim thickness (mm).

[0149]

[0150] Table 1. RMSE Analysis Results of Test Data (mm)

[0151] To verify the impact of key parameters on prediction performance, sensitivity analysis was conducted using the seven wheels in Table 1 as samples. First, the number of frozen layers, F, was analyzed. op The effect of F when other parameters are fixed: op When =0, RMSE = 0.05mm; F op When =1, RMSE = 3.18 mm; F op When F = 2, RMSE = 2.76 mm. This indicates that the amount of frozen layer significantly affects the model's prediction accuracy. op The optimal performance is achieved when the learning rate (lr) is 0. Subsequently, the relationship between the learning rate (lr) and the RMSE is analyzed, and the results are shown below. With other parameters fixed, a larger lr corresponds to a smaller RMSE. The optimal lr for this prediction task is 0.009878, with a corresponding RMSE of 0.05 mm.

[0152] To verify the performance differences of different types of classic recurrent neural networks in the task of predicting wheel flange thickness, a comparative experiment was conducted on the above seven wheel data. The results show that: when using the basic RNN model, the prediction RMSE is 0.43mm; after switching to the GRU model, the RMSE drops to 0.27mm, and the accuracy is significantly improved; when using the LSTM model alone, the prediction effect is further optimized, with an RMSE of only 0.22mm, but it is still inferior to the 0.05mm of the "LSTM + transfer learning + Optuna" combination scheme.

[0153] In summary, the combination of "classification modeling + LSTM + transfer learning + Optuna" can make full use of existing data resources, and its prediction accuracy meets the requirements of field application. It can provide reliable data support for professional engineers to formulate reasonable wheel refinishing plans.

[0154] Step 2: Multi-source data acquisition and preprocessing

[0155] Multi-source data is accessed through standardized interfaces: 6A / 5T alarm information, dynamic detection alarm data, flaw detection results, and historical wheel size data; outliers are cleaned using the 3σ principle and integrated into a standardized data resource pool.

[0156]

[0157] Table 2 Platform Alarm Information Types

[0158] Table 3 Types of Dynamic Detection Alarm Information

[0159] Step 3: Multi-source injury data fusion analysis and health status assessment

[0160] Based on the data from step 2, a fusion analysis was performed within a specified time frame; in accordance with the limitations of the "Railway Technical Management Regulations," the health status of each wheel was divided into three levels: "healthy," "early warning," and "alarm."

[0161] Alarms: Tread wear >7.0mm, flange thickness <23.0mm or >34.0mm, wheel diameter difference exceeding the standard for the same bogie, etc.

[0162] Warning: Tread wear 5.0-7.0mm, rim thickness 25.0-33.8mm, etc.

[0163] Health: Meets standards.

[0164] Step 4: Wheelset data conversion and wheel diameter correction

[0165] Change the unit of analysis from wheel to wheelset:

[0166] The wheel diameter is the minimum value of the two wheels.

[0167] Health status determination: If any round of alarms occurs, the corresponding round will also alarm; if there are no alarms but a warning is issued, a warning will be issued; if all are healthy, the status is healthy.

[0168] Wheel diameter correction: When the rim thickness is lower than the threshold (default 25.0mm, adjustable to 28.0mm), the expected wheel diameter loss is calculated by the rim turning recovery ratio parameter model. Corrected wheel diameter = original wheel diameter - loss.

[0169] This parametric model is constructed based on geometric analysis of tread repair and big data regression.

[0170] Table 4. Health Analysis of Wheel Diameter, Wheel Diameter Difference, and Coaxial Flange Thickness Difference

[0171] Within a specified timeframe, systematic integration and analysis will be conducted, and a tiered assessment mechanism will be used to classify the health status of locomotive wheels into three levels: "healthy," "early warning," and "alarm," with the specific tiering standards shown in the table below.

[0172] Table 5. Simplified Classification of Locomotive Wheel Health Status Assessment Levels

[0173] Based on individual wheels, data such as wheel diameter, rim thickness, and health status are collected.

[0174] Wheel diameter correction driven by rim thickness

[0175] Iterate through the flange thickness of all wheels and determine if it is below a threshold (the default value is equal to the lower limit of flange thickness warning required by technical regulations, 25.0 mm; to strengthen management, each locomotive depot will raise the threshold above the technical regulations, generally setting it to 28.0 mm). If it is below the threshold, calculate the corrected wheel diameter.

[0176] Corrected wheel diameter = Original wheel diameter – Expected wheel diameter loss

[0177] Analysis of Expected Wheel Diameter Loss. During wheel reworking, restoring a certain flange thickness typically requires several times the wheel diameter, based on the vehicle's geometric profile. Research shows that the wear characteristics of the wheel flange and tread vary depending on factors such as the wheel profile type, current flange thickness, and track conditions. This results in different wheel diameter values ​​being consumed to restore a certain flange thickness.

[0178] To analyze the wheel diameter loss caused by restoring the current flange thickness to different target flange thicknesses, a two-dimensional dataset was constructed. Based on the analysis conclusions of the tread turning geometry model and the on-site turning data, the analysis examined the turning diameter loss caused when the current flange thickness data range is [value missing] with a step size of 0.1 mm, and the target flange thickness data range is [value missing] with a step size of 0.1 mm. The analysis results were then presented as a two-dimensional data table and embedded into the wheel precision turning model. Figure 3 The flowchart shown is a construction flowchart of the rim resurfacing ratio parameter model.

[0179] A study was conducted on the geometric model of tread resurfacing. The tread profile and profile type used for individual wheels before resurfacing were recorded. Discrete data points were combined into a digital profile model using a two-dimensional array interpolation method. Subsequently, using standard tread profiles JM3-34, JM3-30, JM3-29, JM3-28, JM3-27, JM3-26, and JM3-25, the wheel diameter required to fully restore the profile shape was calculated and the data was recorded. Field surveys revealed that, influenced by factors such as the type of resurfacing machine program and locomotive operating conditions, the application unit required the use of JM3 tread profiles with a diameter greater than 30mm. Therefore, a multinomial regression analysis model was used to predict the wheel diameter loss caused by using standard profiles JM3-32 and JM3-31. The calculation method for single wheels was extended to other vehicle models and units, and the analysis results were statistically calculated.

[0180] Secondly, based on the big data of the entire road turning and repair records, we analyzed the wheel diameter measurement data before and after each turning and repair, and calculated the wheel diameter loss caused by turning and repair. Based on the wheel flange thickness before turning and the wheel flange thickness after turning and repair, we performed binary classification on the dataset, and recorded the research conclusions of the geometric model into the corresponding classification unit as the turning and repair wheel diameter loss of that classification unit. Since the "reason for wheel turning" dimension was already included during data collection, data with "near limit of flange thickness" and "profile modification" as the reasons for wheel turning were first extracted from each wheel turning dataset classification unit. These two reasons primarily involve wheel diameter loss due to flange thickness restoration. Outlier data was cleaned using the 3σ principle. If the number of data points in a classification unit met the requirements, the median of the remaining data in that unit was calculated as the wheel diameter loss for that classification unit. If the number of data points in a classification unit was insufficient, the mean, median, and mode of the dataset were calculated for all data points in that unit. If the values ​​of the three parameters were equal, that value was extracted as the wheel diameter loss for that classification unit. At this point, the analysis of the key items in the dataset was completed, such as... Figure 4 As shown.

[0181] Based on existing data analysis results, a multivariate custom function regression algorithm was adopted, and the hyperparameters of the custom function were tuned using GA and cross-validation methods to obtain a complete table of parameters for the expected wheel diameter loss, such as... Figure 5 As shown.

[0182] When estimating the corrected wheel diameter, the expected wheel diameter loss during turning can be obtained by referring to the above graph based on the current flange thickness and the engineer's planned target flange thickness. Different current flange thicknesses require different amounts of wheel diameter to restore a 1mm flange thickness; analysis shows that the smaller the current flange thickness, the greater the required wheel diameter. Currently, this method has been adapted to a precise wheel turning model, enabling relatively accurate estimation of the corrected wheel diameter.

[0183] The smallest unit of analysis is changed from the wheel to the wheelset. The wheelset diameter is taken as the smallest diameter of the two wheels in the wheelset; the health status of the wheelset is determined according to the following rules: if the health status of any wheel in the wheelset is "alarm", then the health status of the wheelset is determined to be "alarm"; if neither wheel in the wheelset is in an "alarm" state, but at least one wheel is in a "warning" state, then the health status of the wheelset is determined to be "warning"; if both wheels are in a "healthy" state, then the health status of the wheelset is determined to be "healthy".

[0184] Based on the turning plan developed by professional engineers, the sequence analysis is performed only on the axle sequence (wheelsets) that need to be turned.

[0185] Step 5: Calculation of the turning and repair shaft sequence sorting model and generation of the optimal shaft sequence

[0186] Based on the refining plan, select the wheelsets that need refining and apply sorting logic:

[0187] Scenario 1 (alarm wheelsets exist): Alarm wheelsets are sorted by wheel diameter from smallest to largest. Within the same bogie, the smaller wheel diameter alarm axle is processed first, followed by the warning and healthy wheelsets in sequence.

[0188] 1) For axle sequences with alarms, sort them by wheel pair diameter from smallest to largest.

[0189] 2) First, repair the alarm axle with the smallest wheel diameter. If there are still alarm axles in the same bogie, repair them according to wheel diameter, prioritizing the repair of axles with smaller wheel diameters.

[0190] 3) Within the same bogie, determine whether there is a warning axle. If there are multiple warning axles, sort them by wheel diameter and repair the warning axle with the smaller wheel diameter first.

[0191] 4) Within the same bogie, the healthy wheelsets are turned over first, starting with the healthy axles with smaller wheel diameters;

[0192] 5) After completing the turning of the first bogie, turn the second bogie according to the wheelset diameter of the alarm axle. The turning sequence is the same as that of the first bogie.

[0193] 6) Continue to complete the turning and repair of the remaining bogies of the same locomotive, following the same turning and repair sequence as the first bogie.

[0194] Scenario 2 (Warning wheelsets only): Sort directly by wheel diameter from smallest to largest;

[0195] 1) For axle sequences with warnings, sort them by wheelset diameter from smallest to largest.

[0196] 2) First, repair the warning axle with the smallest wheel diameter. If there are still warning axles in the same bogie, repair them according to wheel diameter, prioritizing the repair of axles with smaller wheel diameters.

[0197] 3) Within the same bogie, the healthy wheelsets are re-machined, starting with the healthy axles with smaller wheel diameters;

[0198] 4) After completing the turning of the first bogie, turn the second bogie according to the wheelset diameter of the warning axle. The turning sequence is the same as that of the first bogie.

[0199] 5) Continue to complete the turning and repair of the remaining bogies of the same locomotive, following the same turning and repair sequence as the first bogie.

[0200] Scenario 3 (All healthy wheelsets): Sort by wheel diameter from smallest to largest;

[0201] 1) For the axles that need to be turned over on the whole vehicle, sort them in order of wheel pair diameter from smallest to largest.

[0202] 2) First, turn over the wheelset with the smallest wheel diameter. Within the same bogie, turn over the axles with the smallest wheel diameter first.

[0203] 3) After completing the turning of the first bogie, turn the second bogie according to the wheelset diameter of the warning axle. The turning sequence is the same as that of the first bogie.

[0204] 4) Continue to complete the turning and repair of the remaining bogies of the same locomotive, following the same turning and repair sequence as the first bogie.

[0205] Output the optimal turning sequence list to ensure it meets the locomotive assembly requirements.

[0206] Step 6: Traction signal output and execution status reception

[0207] Control signals are sent via RS-485 serial communication (baud rate 9600bps, 8 data bits, no parity, 1 stop bit):

[0208] The system sends the following signal format: tractor address, function code 0xa1 (enable), locomotive model, current axle sequence, entry point, target axle sequence, traction confirmation (0x01 execute), number of data items 0x08, and termination code 0x0d;

[0209] The dual-purpose road-rail vehicle transmits signals including positioning status (0x01 completed), enabling automated closed-loop control of wheelset placement.

[0210] The hyperparameter optimization in step 1 uses the Optuna framework, and the optimization targets include the sliding window length, the number of hidden layers, the number of LSTM layers, the learning rate, the range of frozen layers, the dropout rate, and the L2 regularization coefficient. The relationship between the learning rate and RMSE is verified by curve analysis, such as... Figure 6 As shown.

[0211] The wheel flange resurfacing recovery ratio parameter model in step 4 is constructed based on big data of resurfacing records across the entire road, and optimized using geometric analysis and GA algorithm to ensure wheel diameter correction accuracy.

[0212] The method is applicable to railway locomotive depots, urban rail transit, and high-speed trains, and can be used across different scenarios by adapting data interfaces.

[0213] This invention breaks through the bottleneck of traditional experience-based prediction, enabling accurate prediction of locomotive wheel dimensions. It abandons the single-source manual experience-based judgment mode, relying on multi-source data integration, combining manual measurement data with dynamic measurement data, and combining deep learning technology to accurately capture the complex correlation patterns of wheel wear under different working conditions. This fills the data gaps in dimensional changes within measurement intervals, providing reliable data support for wheel turning and repair planning, and effectively solving the technical pain points of traditional predictions being limited by a single dimension and having large biases.

[0214] A multi-source data integration platform was established to achieve unified management and control of locomotive wheel-related data. This platform integrates various data sources, including 6A / 5T alarms, dynamic detection, flaw detection results, and wheel size measurements, breaking down data silos between different stages and forming a standardized data resource pool. This provides data support for subsequent accurate prediction and intelligent sorting, improving the overall management efficiency of pre-repair processes.

[0215] A scientific wheelset reshaping sequence system is constructed to address the issue of inconsistent manual reshaping. Based on standardized data output from a data integration platform, a quantitative reshaping standard is established. This standard comprehensively considers key indicators such as platform alarms, dynamic detection, flaw detection results, and wheel dimensions to achieve intelligent planning of wheelset reshaping. This avoids rework caused by experience-based reshaping and provides a technical foundation for standardized reshaping operations across the entire railway network.

[0216] An innovative linkage control mechanism promotes the automation upgrade of the placement operation. A signal linkage mechanism between the sorting results and the traction equipment is established to generate standardized control signals and drive the traction equipment to operate automatically. This supports data interaction between multiple types of equipment, replacing the traditional manual operation mode, effectively improving placement accuracy and work efficiency, and reducing the safety risks associated with human operation.

[0217] Example 1

[0218] This embodiment uses a typical railway locomotive depot maintenance scenario as a background to elaborate on the specific implementation process of the system of the present invention in locomotive wheel-turning repair without wheel removal. The embodiment focuses on a Type I freight locomotive in Locomotive Depot A, which undertakes heavy-haul transportation tasks and requires frequent and precise wheel-turning repairs. Traditional operation methods rely on manual experience, resulting in problems such as insufficient rationality in wheelset sequencing, large deviations in dimensional prediction, and low efficiency in wheel-turning operations. This system, through technological innovations such as multi-source data fusion, deep learning prediction, intelligent sequencing decision-making, and automated control, achieves intelligent management and control of the entire wheel-turning repair process, significantly improving operation quality and efficiency.

[0219] The system in this embodiment is deployed in the lathe repair workshop of Locomotive Depot A, and is used for operations on Type I eight-axle freight locomotives. The system hardware configuration includes a multi-source data acquisition server, a deep learning computing workstation, a sorting decision terminal, and a signal control unit, with data interaction between the modules achieved through industrial Ethernet. The software platform is developed based on the Python and TensorFlow frameworks, supporting real-time data processing and iterative model updates.

[0220] The data environment is constructed using a layered architecture: the foundation layer connects to the real-time data stream from the 6A / 5T alarm platform, including nine types of alarm information such as tread alarms and polygon alarms; the process layer integrates three types of inspection data from the dynamic detection system—scratching, wheel shape, and flaw detection—as well as ultrasonic / eddy current flaw detection results; the application layer integrates historical wheel size data, including manual measurement records and TDDS dynamic measurement data. All data is classified and stored according to five levels of tags: "locomotive depot-machine type-routes-axle sequence-end position," forming a standardized data resource pool.

[0221] It is particularly noteworthy that this embodiment innovatively designs a data compensation mechanism for situations where data is missing. For wheels that have experienced long-term outdoor operations resulting in missing dynamic measurement data, the system uses transfer learning technology to call upon historical data from similar vehicle models for compensation and prediction, ensuring the integrity of the model input. This design breaks through the traditional system's reliance on data completeness and expands the system's applicability.

[0222] During the data acquisition and preprocessing phase, after system startup, multi-source data from the most recent maintenance cycle (approximately 100,000 kilometers) of the Type I locomotive is first collected via the data interface. Specifically, this includes: acquiring 3 level-one tread alarms from the 6A system, 5 polygon warnings from the 5T system, 12 scratch alarm records from the dynamic detection system, integrating 2 ultrasonic flaw detection reports, and acquiring 28 manually measured data points and 156 dynamically measured data points. Data cleaning employs an automatic filtering algorithm based on the 3σ principle to remove abnormal records with excessive deviations from the mean, ensuring the reliability of the input data.

[0223] In the data preprocessing stage, the system innovatively introduces a time-series alignment algorithm. For heterogeneous data from different sources, the system normalizes the mileage based on the previous overhaul date, constructing a unified time-series dataset. Taking the left wheel of the third axle as an example, starting from the previous overhaul, the system integrates the rim thickness and tread wear data at different mileages, forming a time-series sequence containing 24 data points. This processing method overcomes the limitations of traditional single-point data analysis, laying a solid foundation for subsequent deep learning-based time-series prediction.

[0224] For the implementation of key wheel dimension prediction, the dimension prediction module adopts an innovative "hierarchical transfer learning" strategy. First, based on the entire railway network's big data on locomotives, a three-level classification is performed according to locomotive type, route, and axle order to construct a basic prediction model. Then, for the target wheel, an LSTM neural network is used for personalized training. The LSTM model captures long-term dependencies through a gating mechanism.

[0225] The model training employs an innovative two-stage optimization method: the first stage uses large datasets of similar vehicle models for pre-training to grasp common wear patterns; the second stage uses target wheel data for fine-tuning to adapt to personalized features. Hyperparameter optimization introduces the Optuna framework, obtaining the optimal parameter combination through 100 iterations: sliding window length 5, hidden layers 68, and learning rate 0.009878.

[0226] Taking the right wheel of the third axle as an example, with its mileage of 85,000 kilometers as input, the model outputs a predicted rim thickness of 28.12 mm and a predicted tread wear of 6.8 mm. Verification through actual testing shows that the absolute error between the predicted and actual values ​​is only 0.03 mm, significantly better than the 0.25 mm error level of the traditional linear regression method. This predictive accuracy provides reliable data support for subsequent refinishing decisions.

[0227] Secondly, regarding multi-source data fusion and health status assessment, the health status assessment module employs an innovative weighted fusion algorithm. The system first sets the threshold ranges for each indicator according to the technical specifications, and then calculates a comprehensive health score based on factors such as alarm level and degree of dimensional exceedance. Specifically, the weights for tread wear alarms are set to 0.3, flange thickness alarms to 0.25, dynamic detection alarms to 0.2, and flaw detection results to 0.25. The health index for each wheel is obtained through weighted summation.

[0228] Taking the left wheel of the 5th axle as an example, its predicted tread wear is 7.2mm (0.2mm over the limit), and the rim thickness is 33.9mm (close to the upper limit). Dynamic detection shows slight scratches, and the flaw detection results are qualified. After weighted calculation, the health index is 72 points, corresponding to the "warning" level. The system automatically generates a health status report and marks the main risk point as tread wear near the limit. This multi-indicator fusion assessment method overcomes the limitations of traditional single-indicator judgment and achieves a comprehensive and accurate assessment of wheel condition.

[0229] The sorting module employs an innovative "dynamic priority" algorithm. First, a basic priority is determined based on the wheelset's health status: alarm wheelsets have a priority of 3, warning wheelsets have a priority of 2, and healthy wheelsets have a priority of 1. Then, a wheel diameter correction coefficient is introduced to adjust the priority of wheelsets with flange thicknesses nearing the limit. Wheel diameter correction is a key innovation of this invention, achieved through a flange refinishing recovery ratio parameter model.

[0230] The parametric model, trained using historical revetting data, can accurately predict wheel diameter loss corresponding to different flange thickness recovery amounts. Taking the left wheel of the 5th axle as an example, the current flange thickness is 33.9 mm, the target value is 34.0 mm, and the wheel diameter loss is found to be 1.2 mm according to the parameter table. Based on this, the corrected wheel diameter is calculated, and the sorting priority of this wheelset is adjusted.

[0231] The final sorting results employ a hierarchical decision-making mechanism: first, wheelsets are classified according to their health status, and then within the same level, they are arranged in ascending order of corrected wheel diameter. This sorting strategy ensures that high-risk wheelsets are processed first, while also taking into account the control requirements for wheel diameter differences, thus fundamentally preventing rework.

[0232] In the implementation of automated axle placement control, the traction control unit adopts an innovative signal encoding and decoding scheme. The system converts the sorting results into standardized control commands, which are then sent to the road-rail dual-purpose vehicle via an RS-485 bus. The command format is specially designed, containing multiple types of information such as locomotive identification, target position positioning, and motion control. Taking the control of the placement of the third axle as an example, the command sequence is: start code 0x02, function code 0xa1, locomotive type 0x01, current axle sequence 0x01, target axle sequence 0x03, check code 0x08, and stop code 0x0d.

[0233] After receiving instructions, the dual-purpose road-rail vehicle parses the instructions through its onboard controller and drives the actuators to complete the traction operation. Throughout the process, the system receives position feedback signals from the equipment in real time, forming a closed-loop control. This linkage mechanism based on a standard protocol achieves millimeter-level positioning accuracy, an order of magnitude improvement over manual operation.

[0234] To verify the system's effectiveness, a three-month comparative test was conducted at Locomotive Depot A. The test group used the system for lathe repairs, while the control group used the traditional manual method. Statistical results showed that: in terms of work efficiency, the average time for a single lathe repair in the test group decreased from 4.2 hours to 2.8 hours, a reduction of 33%; in terms of work quality, the wheel diameter difference pass rate increased from 82% to 98%, and the rework rate decreased from 15% to 2%; in terms of economy, due to reduced rework and shorter operation time, the average annual maintenance cost per locomotive was reduced by approximately 45,000 yuan.

[0235] Of particular note is the improvement in prediction accuracy. By comparing the predicted data and measured values ​​for seven wheels, the RMSE index of this system is 0.05mm, significantly better than the 0.43mm of the traditional empirical prediction method. The improved prediction accuracy directly leads to precise control of the turning amount, saving an average of 15% of turning material per wheel.

[0236] This embodiment, through multi-level technological innovation, has formed a complete intelligent locomotive turning and repair solution. At the data level, a fusion architecture for multi-source heterogeneous data has been constructed, breaking through the limitations of information silos; at the prediction level, a deep learning-based size prediction model has been developed, solving the problem of nonlinear prediction; at the decision-making level, an intelligent ranking mechanism integrating multiple indicators has been established, enabling scientific decision-making; and at the execution level, a standardized signal linkage scheme has been created, completing the automation implementation. These innovations support and synergize with each other, jointly promoting a paradigm shift in locomotive turning and repair operations from experience-driven to data-driven.

[0237] The system is now ready for nationwide deployment. By adapting to different locomotive models and operational specifications, it can be rapidly expanded to urban rail transit, high-speed rail, and other fields. With the deepening of the railway industry's intelligent transformation, this technical solution is expected to become a technical standard in locomotive operation and maintenance, providing crucial support for the industry's high-quality development.

[0238] Finally, it should be noted that the above embodiments are only used to describe the technical solutions of the present invention and not to limit the technical methods. The present invention can be extended to other modifications, variations, applications and embodiments, and therefore all such modifications, variations, applications and embodiments are considered to be within the scope of the present invention.

Claims

1. A method and system for planning the axle sequence for locomotive wheelset repair without wheel removal, characterized in that, The system includes a multi-source data integration platform, a wheel key dimension prediction module, a wheel turning sequence sorting module, and a traction signal output unit. These modules work together to intelligently sort the wheelsets to be turned and output the target wheel sequence data. The multi-source data integration platform is used to integrate 6A / 5T alarm platform data, dynamic detection data, flaw detection results and wheel size measurement data through standardized interfaces, and classify and archive them according to the hierarchy of "locomotive depot-machine type-routes-axle sequence-end position" to form a standardized data resource pool; The wheel key dimension prediction module is built on deep learning technology and adopts a combination of classification modeling, LSTM neural network, transfer learning and Optuna hyperparameter optimization scheme to accurately predict wheel flange thickness, tread wear and wheel diameter, filling the data gap in the measurement interval. The wheel repair sequence sorting module takes wheelsets as the smallest unit of analysis and combines wheelset health status classification (healthy, early warning, alarm) and wheel diameter parameters to realize intelligent planning of wheelset repair sequence. The sorting logic prioritizes alarm wheelsets and sorts them in ascending order of wheel diameter within the same bogie. The traction signal output unit, based on the RS-485 serial communication protocol, converts the sorting results into standardized control signals to drive the dual-purpose road-rail vehicle to achieve automated wheelset placement and receive positioning status signals transmitted back from the equipment.

2. The locomotive wheelset axle sequence planning method and system for non-removal wheel turning and repair according to claim 1, characterized in that, The data access of the multi-source data integration platform includes: The 6A alarm platform provides data on tread level 2 alarms, tread level 1 alarms, and tread early warnings. Data from the 5T alarm platform, including polygonal level 2 alarm, polygonal level 1 alarm, polygonal early warning, tread level 2 alarm, tread level 1 alarm, and tread early warning. Severe, moderate, and mild alarms for scratch detection, wheel shape detection, and flaw detection based on dynamic detection data; Non-compliant data in flaw detection results and crack damage reports; The results of manual and dynamic measurements of wheel size data.

3. The locomotive wheelset axle sequence planning method and system for non-removal wheel turning and repair according to claim 1, characterized in that, The construction of the wheel critical dimension prediction module includes: Based on the time-series dataset within the turning cycle, the data includes wheel flange thickness sequence, tread wear sequence, and post-turning mileage sequence; An LSTM neural network is used to mine the long-term dependency between size and mileage. Its hyperparameters are optimized using the Optuna framework. The optimization objects include sliding window length, number of hidden layers, number of LSTM layers, learning rate, frozen layer range, Dropout rate, and L2 regularization coefficient. A transfer learning pre-training-fine-tuning paradigm is introduced, which pre-trains using similar wheel data and then fine-tunes using target wheel data to address the problem of small sample data.

4. The locomotive wheelset axle sequence planning method and system for non-removal wheel turning and repair according to claim 1, characterized in that, The sorting logic of the turning shaft sequence sorting module includes: Wheelset health status determination: If any wheelset has an alarm, the wheelset will be alarmed; if there is no alarm but there is a warning, then a warning will be issued; if all wheels are healthy, the wheelset is healthy. Wheel diameter correction: When the rim thickness is below the threshold, the expected wheel diameter loss is calculated using the rim refining recovery ratio parameter model. Corrected wheel diameter = original wheel diameter - expected wheel diameter loss. Sorting rules: Scenario 1 (Alarm wheelsets exist): Prioritize sorting alarm wheelsets by wheel diameter from smallest to largest, then process the warning and healthy wheelsets within the same bogie; Scenario 2 (Warning wheelsets only): Sort directly by wheel diameter from smallest to largest; Scenario 3 (All healthy wheelsets): Sort by wheel diameter from smallest to largest.

5. The locomotive wheelset axle sequence planning method and system for non-removal wheel turning and repair according to claim 1, characterized in that, The communication protocol of the traction signal output unit is RS-485 serial port, baud rate 9600bps, 8 data bits, no parity, and 1 stop bit; The system sends signals to the dual-purpose road and rail vehicle including the tractor address, function code, locomotive type, current axle sequence, entry point, target axle sequence, traction confirmation, and number of data items, encoded in hexadecimal. The feedback signals from the dual-purpose road-rail vehicle include the tractor address, function code, positioning status, and number of data points, enabling closed-loop control.

6. The locomotive wheelset axle sequence planning method and system for non-removal wheel turning and repair according to claim 1, characterized in that, The system is also applicable to maintenance operations in urban rail transit, high-speed trains, and rolling stock, and can be promoted across different fields by adapting to different models and data interfaces.

7. The locomotive wheelset axle sequence planning method and system for non-removal wheel turning and repair according to claim 1, characterized in that, The method includes the following steps: Step 1: Construction and Calculation of Wheel Critical Dimension Prediction Model Collect manual and dynamic measurement data, including data types such as tread wear, wheel flange thickness and corresponding post-turning mileage; classify the datasets according to the hierarchy of "locomotive depot-machine type-routes-axle sequence-end position", and form a time-series dataset with the current turning cycle start node as the reference (post-turning mileage set to 0); A prediction model is built based on the technical solution of "classification modeling + LSTM + transfer learning + Optuna": LSTM neural networks are used to explore the long-term dependency between wheel size and mileage. The LSTM model controls the information flow through forget gates, input gates, and output gates. Its internal computation process is defined by the following formula:

8. W f and b f are weight matrices and bias terms for the forget gate f t respectively. W I and b I - are, respectively, a weight matrix and a bias term for the input gate I t ​ W C and b C —Memory cell C t The weight matrix and bias terms; W O and b O —These are output gates O t The weight matrix and bias terms; X t — Enter at the current moment; h t-1 —The state was hidden a moment ago; tanh — activation function; t —Candidate state of memory cells; *——Hada code multiplication.

9. Introduce the transfer learning pre-training-fine-tuning paradigm: first, pre-train the model using similar wheel data (30-100 pre-training rounds), and then fine-tune it to fit the target wheel (50-200 fine-tuning rounds). Optimize hyperparameters using the Optuna framework (e.g., sliding window length 3-8, number of hidden layers 30-100), and improve model accuracy by combining K-fold cross-validation. The predicted values ​​of the output rim thickness and wheel diameter are controlled with a root mean square error (RMSE) within 0.05 mm.

10. Step 2: Multi-source data acquisition and preprocessing Multi-source data is accessed through standardized interfaces: 6A / 5T alarm information, dynamic detection alarm data, flaw detection results, and historical wheel size data; outliers are cleaned using the 3σ principle and integrated into a standardized data resource pool.

11. Step 3: Multi-source injury data fusion analysis and health status assessment Based on the data from step 2, a fusion analysis was performed within a specified time frame; in accordance with the limitations of the "Railway Technical Management Regulations," the health status of each wheel was divided into three levels: "Healthy," "Warning," and "Alarm." Alarms: Tread wear >7.0mm, flange thickness <23.0mm or >34.0mm, wheel diameter difference exceeding the standard for the same bogie, etc. Warning: Tread wear 5.0-7.0mm, rim thickness 25.0-33.8mm, etc. Health: Meets standards.

12. Step 4: Wheelset Data Conversion and Wheel Diameter Correction Change the unit of analysis from wheel to wheelset: The wheel diameter is the minimum value of the two wheels. Health status determination: If any round of alarms occurs, the corresponding round will also alarm; if there are no alarms but a warning is issued, a warning will be issued; if all are healthy, the status is healthy. Wheel diameter correction: When the rim thickness is lower than the threshold (default 25.0mm, adjustable to 28.0mm), the expected wheel diameter loss is calculated by the rim turning recovery ratio parameter model. Corrected wheel diameter = original wheel diameter - loss. This parametric model is constructed based on geometric analysis of tread repair and big data regression. Step 5: Calculation of the turning and repair shaft sequence sorting model and generation of the optimal shaft sequence Based on the refining plan, select the wheelsets that need refining and apply sorting logic: Scenario 1 (alarm wheelsets exist): Alarm wheelsets are sorted by wheel diameter from smallest to largest. Within the same bogie, the smaller wheel diameter alarm axle is processed first, followed by the warning and healthy wheelsets in sequence. Scenario 2 (Warning wheelsets only): Sort directly by wheel diameter from smallest to largest; Scenario 3 (All healthy wheelsets): Sort by wheel diameter from smallest to largest; Output the optimal turning sequence list to ensure it meets the locomotive assembly requirements.

13. Step 6: Traction signal output and execution status reception Control signals are sent via RS-485 serial communication (baud rate 9600bps, 8 data bits, no parity, 1 stop bit): The system sends the following signal format: tractor address, function code 0xa1 (enable), locomotive model, current axle sequence, entry point, target axle sequence, traction confirmation (0x01 execute), number of data items 0x08, and termination code 0x0d; The dual-purpose road-rail vehicle transmits signals including positioning status (0x01 completed), enabling automated closed-loop control of wheelset placement.

14. The method according to claim 7, characterized in that, The hyperparameter optimization in step 1 adopts the Optuna framework, and the optimization objects include sliding window length, number of hidden layers, number of LSTM layers, learning rate, frozen layer range, dropout rate, and L2 regularization coefficient. The relationship between learning rate and RMSE is verified by curve.

15. The method according to claim 7, characterized in that, The wheel flange resurfacing recovery ratio parameter model in step 4 is constructed based on big data of resurfacing records across the entire road, and optimized using geometric analysis and GA algorithm to ensure wheel diameter correction accuracy.

16. The method according to claim 7, characterized in that, The method is applicable to railway locomotive depots, urban rail transit, and high-speed trains, and can be used across different scenarios by adapting data interfaces.

Images