A strategy planning control method and system for applying locomotive auxiliary driving

Abstract

This application discloses a strategy planning and control method and system for locomotive assisted driving, relating to the technical field of rail transit control. The method includes: acquiring multiple conditional data associated with driving decisions; encoding the conditional data into Boolean conditional vectors; and generating the current driving decision through linear mapping; collecting test data and operational data from multiple locomotive segments; calculating the instantaneous traction force at each moment based on the operational data and traction characteristic function; training a time-series calibration model based on the operational data and instantaneous traction force; calculating the current instantaneous traction force; inputting the historical traction force sequence and the current instantaneous traction force into the trained time-series calibration model; and outputting the calibrated current traction force; constructing a locomotive dynamics model based on the calibrated current traction force; updating and predicting the locomotive speed at the next moment; and controlling the locomotive's motion state in conjunction with the current driving decision. This application can effectively improve the accuracy and continuity of the traction characteristic analysis process.

Description

Technical Field

[0001] This application relates to the technical field of rail transit control, and in particular to a strategy planning and control method and system for locomotive-assisted driving. Background Technology

[0002] In recent years, assisted driving technology for rail transit has become crucial for achieving intelligent and efficient train operation. This system integrates track data, locomotive status, and operational requirements to automatically plan the optimal target speed curve for safety, smoothness, punctuality, and energy efficiency, and then controls the locomotive to execute it precisely. An assisted driving system typically comprises three core modules: strategy, planning, and control. Specifically, the strategy layer generates current driving decisions based on driving conditions, the planning layer predicts future states using locomotive dynamics models, and the control layer ensures precise tracking of the speed curve.

[0003] In related technologies, the decision-making and planning of assisted driving systems have shortcomings. In actual locomotive operation, driving decisions are affected by a variety of conditions, with complex combinations of conditions that dynamically change with the scenario. Currently, commonly used methods are mostly hard-coded logic or logical lookup table methods. Hard-coding results in high logical coupling, and modifying or adding conditions can easily trigger other logical errors and create logical blind spots. The lookup table method requires explicitly storing the corresponding decision results for each combination of conditions, and the storage size grows exponentially with the increase of condition dimensions, making it difficult to adapt to dynamically changing driving scenarios. The locomotive dynamics model is the core of the planning layer, and one of its key parameters is traction force. Existing modeling mainly relies on experimental data provided by the "Train Traction Calculation Regulations" ("Train Traction Regulations") to establish the relationship between traction force and speed and class through static function fitting. However, this method only relies on the speed and class at the current moment to calculate the instantaneous traction force, ignoring the physical characteristic of traction force continuously changing over time, resulting in a significant deviation between the estimated value and the actual output. This inaccurate and noisy traction force input will distort the acceleration and velocity trajectory predicted by the dynamics model, thereby affecting the reliability of the planning curve and the control accuracy.

[0004] Therefore, there is an urgent need for a control method that can overcome the above-mentioned defects, which can not only achieve efficient generation and flexible configuration of decision rules, but also improve the accuracy and continuity of traction estimation, thereby improving the overall performance of the driver assistance system. Summary of the Invention

[0005] The purpose of this application is to provide a strategy planning and control method and system for locomotive assisted driving, which can effectively improve the accuracy and applicability of the assisted driving system control process.

[0006] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a strategy planning and control method for locomotive assisted driving, comprising: acquiring multiple conditional data associated with driving decisions, encoding the conditional data into Boolean conditional vectors, and generating a current driving decision through linear mapping; collecting test data and operational data of multiple locomotive segments, constructing a traction characteristic function based on the test data, calculating the instantaneous traction force at each moment based on the operational data and the traction characteristic function, wherein the operational data includes speed, class, and actual traction force; constructing a time-series calibration model, training the time-series calibration model based on the operational data and the instantaneous traction force to obtain a trained time-series calibration model; acquiring the current operational data and historical traction force sequence of the locomotive in real time, calculating the current instantaneous traction force in combination with the traction characteristic function, inputting the historical traction force sequence and the current instantaneous traction force into the trained time-series calibration model, and outputting the calibrated current traction force; constructing a locomotive dynamics model based on the calibrated current traction force, updating and predicting the locomotive speed at the next moment, and controlling the locomotive's motion state in combination with the current driving decision.

[0007] Secondly, this application provides a strategy planning and control system for locomotive assisted driving. The system includes: an assisted driving strategy layer, used to acquire multiple conditional data associated with driving decisions, encode the conditional data into Boolean conditional vectors, and generate the current driving decision through linear mapping; also used to collect locomotive test data and operating data; an assisted driving planning layer, used to construct a traction characteristic function based on the test data, calculate instantaneous traction force based on the operating data and the traction characteristic function; also used to construct and train a time-series calibration model, call the trained time-series calibration model in real time during locomotive operation, and output the calibrated current traction force; also used to update the locomotive dynamics model based on the calibrated current traction force and predict the speed at the next moment; and an assisted driving control layer, used to control the locomotive's motion state based on the speed at the next moment.

[0008] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a strategy planning and control method and system for locomotive assisted driving. By encoding multiple driving conditions into Boolean condition vectors and using a linear mapping matrix to generate the current driving decision, complex multi-condition logic judgments are transformed into data-driven parameterized mappings, improving the flexibility and scalability of the strategy system. A time-series calibration model is introduced and trained. By learning the dynamic correlation between historical traction force sequences and current instantaneous estimates, noise and unreasonable jumps in the static function estimation results can be filtered out, outputting a more physically continuous and accurate calibrated traction force, thus improving prediction fidelity and simulation realism. By training the time-series calibration model with data containing real traction force labels, it can learn and correct the inherent estimation bias of the static function, obtaining a traction force value that is closer to the actual locomotive output and has higher accuracy, thus obtaining a more accurate predicted speed. Based on the speed predicted by the above high-fidelity dynamic model, the solution environment for speed curve optimization is more consistent with the real world, improving the accuracy, response speed, and anti-interference capability of the assisted driving system control. This application addresses the technical challenges of existing driver assistance systems in terms of decision-making flexibility, model accuracy, and control robustness through data-driven generation of decision rules, temporal calibration of traction force, and closed-loop correction of control, thereby improving the overall performance and engineering applicability of driver assistance systems. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a basic module structure diagram for locomotive assisted driving.

[0011] Figure 2 This is a schematic diagram of the locomotive's traction characteristic curve.

[0012] Figure 3 A comparison chart showing the locomotive force estimated by the traction characteristic function in related technologies and the actual locomotive force.

[0013] Figure 4 for Figure 3 Enlarged view of section A.

[0014] Figure 5 This is a schematic diagram of the traction force on the locomotive traction characteristic curve, representing continuous data.

[0015] Figure 6 This is a flowchart of a strategy planning and control method for locomotive assisted driving in an embodiment of this application.

[0016] Figure 7 This is a comparison diagram of the conditional vectorization encoding methods in the embodiments of this application.

[0017] Figure 8 This is a schematic diagram illustrating the rule generation method in an embodiment of this application.

[0018] Figure 9 This is a comparison diagram of the rule generation methods in the embodiments of this application.

[0019] Figure 10 This is a flowchart illustrating the application of the timing calibration model in the embodiments of this application.

[0020] Figure 11 This is a schematic diagram of the control layer algorithm correction process in an embodiment of this application. Detailed Implementation

[0021] 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 some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] like Figure 1 The diagram shown is a basic modular structure diagram of locomotive assisted driving. The locomotive assisted driving system strictly follows the modular design principle during the design process and is mainly divided into four major modules: strategy, planning, control and locomotive model. The strategy module utilizes information such as route data, signal status, locomotive operating status, train schedule, and speed limits, combined with the locomotive's weight and vehicle formation, to form a set of strategy data, which is then transmitted to the planning module to plan a reasonable control speed curve. The planning module, focusing on safety, stability, punctuality, and energy conservation, employs various intelligent algorithms to solve for motion curves in different motion stages such as acceleration, deceleration, and cruising, as well as in different scenarios such as starting, phase crossing, cyclic air braking, and parking planning, to obtain an optimal planned speed curve. The control module, based on control principles and algorithms, and in conjunction with the planned speed curve, controls and manipulates the locomotive. The locomotive dynamics model is a crucial technical tool in the planning process. This model requires modeling and calculating train running resistance (unit basic resistance, gradient resistance, and curve resistance, etc.), air braking force (air braking friction coefficient, braking idle time, and effective braking time, etc.), and locomotive characteristics. The accuracy of the locomotive dynamics model directly affects the accuracy of the planned route.

[0024] like Figure 2 The diagram shown illustrates the locomotive traction characteristic curve. This curve is derived from experimental data using a linear fitting method, capturing the traction characteristics at discrete speeds with a minimum unit of 0.1. This is achieved through data collection... Figure 2 Points on the traction characteristic curve are piecewise fitted to multiple curves to obtain the corresponding traction characteristic function, as shown below: in, Indicates speed, The term "level" or "handle level" refers to the traction or braking power setting on the locomotive driver's controller (usually a lever that can be pushed back and forth). It is a direct command from the driver to the locomotive control system regarding the desired traction or braking force. Different positions of the lever correspond to different levels (e.g., level 1, level 2... highest traction level; or various levels of the braking zone). In this invention, for the convenience of high-precision modeling, these levels are treated as continuous or high-resolution discrete variables.

[0025] like Figure 3 The figure shows a comparison between the locomotive force estimated by the traction characteristic function and the actual locomotive force in related technologies. The horizontal axis represents speed, and the vertical axis represents traction force. It can be seen that the instantaneous force estimated by the traction characteristic function and the actual locomotive force collected are consistent in trend, but there is a certain estimation deviation. This estimation deviation is because the instantaneous force calculated by the traction characteristic function is based only on the current speed and current gear, and is not related to the traction force of the previous second. However, the change in traction force is a function of time, and its changes before and after are strongly correlated. The instantaneous traction force estimated by the characteristic function has noise, fluctuations, and discontinuities, inevitably leading to a certain estimation deviation. Figure 4 yes Figure 3 The enlarged view of section A contains 33 seconds of runtime data. It can be seen that there is a certain deviation between the estimated and actual values, and this deviation can be as high as 80 kN in one estimation.

[0026] like Figure 5 The diagram shows the traction force of continuous data on the locomotive traction characteristic curve. When estimating the instantaneous traction force, the current speed and gear level are used to estimate the instantaneous traction force. It can be seen that even a small jump in gear level can cause a large change in the estimated instantaneous traction force, which does not conform to the continuous change of traction force. Analyzing traction characteristics in this way affects the continuity and accuracy of the analysis results. To address the above problems, this application provides the following solution: like Figure 6As shown in the figure, this application provides a strategy planning and control method for locomotive assisted driving, the method including the following steps: S610. Obtain multiple conditional data associated with driving decisions, encode the conditional data into Boolean conditional vectors, and generate the current driving decision through linear mapping.

[0027] S620 collects test data and operational data from multiple locomotive segments, constructs a traction characteristic function based on the test data, and calculates the instantaneous traction force at each moment based on the operational data and the traction characteristic function. The operational data includes speed, class, and actual traction force.

[0028] S630. Construct a timing calibration model. Train the timing calibration model based on the running data and instantaneous traction force to obtain a trained timing calibration model.

[0029] S640 acquires the locomotive's current operating data and historical traction force sequence in real time, and calculates the current instantaneous traction force by combining the traction characteristic function. It then inputs the historical traction force sequence and the current instantaneous traction force into the trained time series calibration model and outputs the calibrated current traction force.

[0030] S650: Construct a locomotive dynamics model based on the calibrated current traction force, update and predict the locomotive speed at the next moment, and control the locomotive's motion state in conjunction with the current driving decision.

[0031] This application provides a strategy planning and control method for locomotive assisted driving. By encoding multiple driving conditions into Boolean condition vectors and generating the current driving decision using a linear mapping matrix, the complex multi-condition logic judgment is transformed into a data-driven parameterized mapping, improving the flexibility and scalability of the strategy system. A time-series calibration model is introduced and trained. By learning the dynamic correlation between historical traction force sequences and the current instantaneous estimate, noise and unreasonable jumps in the static function estimation results can be filtered out, resulting in a more physically continuous and accurate calibrated traction force output, thus improving prediction fidelity and simulation realism. By training the time-series calibration model with data containing real traction force labels, it can learn and correct the inherent estimation bias of the static function, obtaining a traction force value that is closer to the actual locomotive output and has higher accuracy, leading to more accurate speed prediction. Based on the speed predicted by the aforementioned high-fidelity dynamic model, the solution environment for speed curve optimization is more consistent with the real world, improving the accuracy, response speed, and anti-interference capability of the assisted driving system control. This application addresses the technical challenges of existing driver assistance systems in terms of decision-making flexibility, model accuracy, and control robustness through data-driven generation of decision rules, temporal calibration of traction force, and closed-loop correction of control, thereby improving the overall performance and engineering applicability of driver assistance systems.

[0032] During locomotive operation, drivers are constrained by numerous driving conditions and rules, such as the locomotive formation, whether it's a single locomotive, whether it's in a phase-separated zone, whether it's in a high-speed or low-speed zone, and whether it's in air or electric control. Even a single change in these conditions affects the next step of the operation. Different combinations of conditions correspond to different decisions, and as driving scenarios increase, the conditions affecting locomotive operation change, leading to dynamic changes in the corresponding decisions. Currently used methods involve nested logic and branch expansion, which easily create logical blind spots and have high coupling between logic components. Modifying or adding a condition can easily cause other logic to malfunction or collapse. Alternatively, logical lookup tables are used, which require displaying and storing the corresponding decision results for each combination of conditions. The storage size increases exponentially with the number of conditions, making this method particularly unsuitable for dynamically changing conditions. In this embodiment, a linear mapping rule generation method based on discrete Boolean conditions is used to confirm the driving strategy.

[0033] For example, the process of acquiring multiple conditional data associated with driving decisions and confirming the current driving decision based on a preset driving decision and the conditional data includes the following steps: Multiple conditional data points are encoded into Boolean conditional vectors, with each dimension of the Boolean conditional vector taking a value of 0 or 1. Multiple preset driving decisions are encoded into one-hot decision vectors, with only one dimension of the one-hot decision vector having a value of 1 and the rest being 0. A linear mapping matrix is ​​constructed to map the Boolean conditional vectors to one-hot decision vectors, thus establishing a mapping relationship between conditions and decisions. The output of the linear mapping matrix is ​​converted into a probability distribution, and the weights of the linear mapping matrix are obtained through gradient descent training using cross-entropy as the loss function. The current Boolean conditional vector is obtained by encoding the conditional data at the current time step. The current Boolean conditional vector is multiplied by the trained linear mapping matrix to obtain the predicted output vector. The index corresponding to the maximum value in the predicted output vector is taken as the current driving decision.

[0034] Specifically, it includes the following processes: 1. Conditional encoding vectorization: Under a certain driving state, all conditions affecting the current driving decision are encoded using one dimension for conditions with a single state, such as 1 representing excessive speed and 0 representing moderate speed. For conditions with multiple states, multiple dimensions are used for encoding, such as 01 representing the speeding zone, 10 representing the underspeeding zone, 11 representing the constant speed zone, and 00 representing other conditions. This encoding method forms the conditional vector c, as shown below: ; Where n represents the conditional dimension; each dimension c i It is a Boolean value, i {1,2,…,n},c iOnly 0 and 1 are used, satisfying: The condition vector c resides in a discrete Boolean space of dimension n. For example... Figure 7 As shown, the decision factors, states, state codes, and condition vectors are mutually corresponding.

[0035] 2. One-hot transformation of decision vectors: Under a certain driving state, the decision corresponding to the combination of conditions is represented as a decision vector, which is also in the discrete Boolean space, i.e.: ; Where m is the dimension of the decision vector, representing the number of decisions, and the decision vector is a one-hot vector, i.e.: ; Based on the above constraints, among the m components in the decision vector, only one component has a value of 1, while the other components are all 0.

[0036] 3. Map the condition vectors to the decision vectors using a linear mapping matrix M: ; The above formula represents the rule generation method, such as Figure 8 and Figure 9 The diagram illustrates the rule-based decision generation process. Both the condition vector *c* and the decision vector *d* are defined in a discrete Boolean space, a prerequisite for the linear mapping rule generation method based on discrete Boolean conditions. The 0 or 1 property of discrete Boolean space is suitable for logical condition encoding. However, this property also determines that discrete Boolean space is finite and exhaustive. In continuous space, each dimension of a vector is a continuous real number, and is infinite and uncountable. This property means that it is impossible to cover all sample spaces in continuous space; only sampling can be used to obtain a limited number of samples. For the unsampled samples, the model's generalization ability is highly dependent. When both the condition vector *c* and the decision vector *d* exist in discrete Boolean space, especially when *c* represents different combinations of conditions and *d* represents different decisions, commonly used distance concepts, such as Euclidean distance, become meaningless because the condition space and decision space, for example… Cannot be converted into a combination of conditions Combination of conditions The distance description between them is different from ordinary linear fitting because there is no concept of similarity or dissimilarity between the two different combinations of conditions. The linear mapping matrix M in the above formula is a structured representation of the rules. When exhaustively enumerating the condition vector c, the linear mapping matrix M obtained by taking all the condition-decision correspondences is a deterministic mapping, which has a natural advantage over the lookup table method. That is, the rule generation method of discrete Boolean space parameterizes the rules. All rules are implicitly encoded in the linear mapping matrix M. This compression of rules greatly facilitates engineering applications.

[0037] 4. Solution process: In this embodiment, a loss function is used, specifically a softmax-based cross-entropy loss function, as shown below: ; ; in, This represents the predicted output, which is a continuous vector in a continuous space. Therefore, the role of softmax is to... Transform it into a probability distribution vector in a continuous space, i.e. The optimization problem is transformed into approximating the conditional distribution vector to the decision vector, as shown in the following equation: ; Because in the decision vector d to be approximated, only one dimension has a value of 1, assume 1, j = t, and the values ​​in other dimensions are all 0, that is In other words, the loss value only considers the loss in this dimension. As gradient descent learning progresses, the loss value becomes smaller and smaller. It is getting closer and closer to 1, because If so, then the other dimensions are close to 0.

[0038] The above steps and methods in this application embodiment model multi-condition logic judgment as a condition vector c defined on a discrete Boolean space, and convert it into a decision vector d on the same space through a linear mapping matrix M, converting the control flow into a data flow, and perfectly decoupling the conditions, logic, and decisions from a completely coupled state, thereby achieving flexible configurability and scalability of conditions and decisions.

[0039] For example, in step S620 above, constructing the traction characteristic function based on the experimental data specifically includes: establishing a functional relationship between the locomotive traction force, locomotive speed, and vehicle class based on the experimental data of the traction characteristics using a piecewise polynomial fitting method. The traction characteristic function is shown below: ; in, Indicates the locomotive's traction force. Indicates speed, Indicates the level (or handle level).

[0040] The temporal calibration model in this embodiment is a Long Short-Term Memory (LSTM) network model, including a forget gate, an input gate, a memory gate, and an output gate. The inputs of the temporal calibration model include: a historical traction force sequence containing at least the traction force values ​​of the previous 5 seconds, and the current instantaneous traction force calculated by the traction characteristic function. The output of the temporal calibration model is the calibrated current traction force corresponding to the current moment.

[0041] Traction force, as a key dynamic variable in locomotive operation, is characterized by continuous variation, strong nonlinearity, and susceptibility to noise interference. The design structure of the LSTM model itself makes it highly suitable for learning and capturing slowly changing data sequences over time. Therefore, this patent uses an LSTM model to learn the process of force change. The model construction includes the following steps: First, let's define the instantaneous traction force at time t as... The instantaneous traction force contains a certain amount of noise, prior to time t. l The traction force per second is a historical traction force sequence. In the embodiments of this application, the following are used: l =5 seconds of historical traction sequence, i.e. We now use an LSTM model, utilizing 5 seconds of historical traction force, to correct the instantaneous traction force at time t, thus estimating the correct traction force at time t. See below: ; Among them, F LSTM Let t represent the mapping function of LSTM. This represents the traction force at time t after calibration.

[0042] The forget gate, input gate, and memory gate of the time-series calibration model are shown below: Forgotten Gate: ; Memory Gate: ; Input Gate: ; The renewal of memory cells is as follows: ; in, This involves concatenating vectors. For element-wise multiplication, , , This represents the weights of the forget gate, the remember gate, and the input gate. , , This indicates the corresponding bias.

[0043] Table 1 shows the dimensionality parameters of the LSTM model used in this patent. The input sequence has a length of 6, including the estimated traction force value for the first 5 seconds and the current instantaneous traction force. The input data has only one dimension: traction force, therefore the input dimension is 1. Similarly, only traction force is estimated, so the output sequence dimension is 1. The hidden state at each time step stores information from historical time steps after passing through the forget gate and the input gate. Therefore, the dimension of the hidden layer determines the richness and diversity of information from filtered historical time steps that can be stored at the current time step. Since the purpose of this embodiment is to simulate the change process of traction force using LSTM, the hidden layer state dimension and the memory cell dimension are both chosen to be 2.

[0044] Table 1 Table 2 shows the matrix parameter table of the LSTM model in this embodiment, which represents the matrix type and matrix dimension of each part of the model. Since the input dimension is 1 and the hidden layer dimension is 2, the weight dimension of the various gates in the middle is 3.

[0045] Table 2 During the training of the temporal calibration model, the hidden state from the previous time step and the current input are used as the common inputs to all gates. The current memory cell state and the hidden state are updated through the calculation of the forget gate, input gate, memory gate, and output gate. The training process uses mean squared error as the loss function. The mean squared error between the estimated traction force calculated by forward propagation and the corresponding actual traction force is used as the optimization objective. The weight matrices and bias parameters of the forget gate, input gate, memory gate, and output gate in the temporal calibration model are updated through the backpropagation algorithm to minimize the loss function. The weights of the forget gate, memory gate, output gate, and input gate are then calculated. , , , and the corresponding bias The loss function MSE (Mean Squared Error) is used to solve this problem. ; in, The MSE loss function is... The estimated total number of samples at time point, The actual traction force value at time t. Let t be the estimated traction force at time t.

[0046] For example, when training the time-series calibration model based on running data and instantaneous traction force, each segment of running data collected is continuous data with a duration of not less than 10 minutes, and the sampling interval for speed, class, and actual traction force is 1 second; the actual traction force sequence of 5 consecutive seconds in each segment of running data is combined with the instantaneous traction force of the 6th second to form a training sample, and the actual traction force of the 6th second is used as the label of the sample for model training.

[0047] Specifically, collection Time period Actual locomotive force The traction force estimated by the traction characteristic function The duration of the time period is Each time period The data collected internally must meet the following requirements: 1. It must be continuous; 2. The duration must be no less than 10 minutes; 3. Actual traction force. The sampling interval is 1 second; 4. Due to the speed and level The sampling interval is 1 second, therefore the traction characteristic function is calculated as follows: The period is also 1 second. That is, in each continuous data segment, multiple training samples are constructed in a sliding window manner (window length 6 seconds, the first 5 seconds are the historical actual traction force sequence, and the 6th second is the instantaneous traction force calculated by the traction characteristic function), and the actual traction force in the 6th second is used as the label of the sample.

[0048] Based on the above requirements, the following data can be obtained: The above data is processed to obtain the training set: ; in, For matrix multiplication, , , , , This represents the weights of the forget gate, memory gate, output gate, input gate, and fully connected layer.

[0049] After training the time-series calibration model, it is deployed by inputting the historical traction force sequence and the current instantaneous traction force into the trained model, which then outputs the calibrated current traction force. Based on the calibrated current traction force, a locomotive dynamics model is constructed to update and predict the locomotive speed for the next moment, which is then used to control the locomotive's motion.

[0050] like Figure 10 The diagram shown is a flowchart of the application of the timing calibration model in this embodiment. The construction and application of the model mainly include the following processes: Step S1: Fit the locomotive traction characteristic function based on the experimental data in the "Traction Regulations"; Step S2: Collect locomotive operation data over multiple time periods, mainly including speed, class, and locomotive force; Step S3: Calculate the instantaneous traction force of the data collected in Step S2; Step S4: Construct an LSTM model; Step S5: Integrate the collected operation data and instantaneous traction force data into an LSTM training dataset; Step S6: Train the LSTM model, obtain the model weights, and deploy the updated model; Step S7: Obtain the current speed and class of the moving locomotive; Step S8: Obtain the instantaneous traction force at the current moment based on the speed and class in S7; Step S9: Obtain the locomotive traction force of the moving locomotive over the past 5 seconds; Step S10: Use the trained LSTM model to calibrate the current instantaneous traction force using the traction force over the past 5 seconds; Step S11: Obtain the calibrated traction force at the current moment.

[0051] In some embodiments, the locomotive in this application is a rail train. The process of constructing a locomotive dynamics model based on the calibrated current traction force and updating the predicted locomotive speed for the next moment includes the following steps: 1. Calculate the unit traction force based on the calibrated current traction force, locomotive mass, and traction mass, as shown below: ; in, Indicates unit traction force. The current traction force after calibration. For locomotive quality, For traction mass; 2. The unit basic resistance, unit additional resistance, and unit braking force of the locomotive are calculated by combining the unit traction force to obtain the unit resultant force of the locomotive. The calculation method is as follows: ; in, Indicates the combined force of locomotive units. Indicates the basic resistance per unit. This indicates the unit additional resistance of the locomotive. Indicates the unit braking force of the locomotive; 3. Based on Newton's second law and the unit net force on the locomotive, calculate and update the velocity at the next moment, as shown below: ; in, Indicates the velocity at the next moment. Indicates the speed at the current moment. Indicates a time interval.

[0052] In the above process, the calculation method for the unit basic resistance of the computer vehicle is as follows: ; in, As a unit of basic resistance, , , This represents the coefficient of the basic resistance formula.

[0053] The calculation method for the unit additional resistance of a computer-controlled vehicle is as follows: ; in, This indicates the unit additional resistance of the locomotive. Indicates the additional resistance per unit ramp. This indicates the additional resistance of the unit curve.

[0054] The calculation method for the unit braking force of a computer-controlled vehicle is as follows: ; in, This indicates the conversion of brake shoe pressure. This indicates the conversion coefficient of friction. This represents the unit braking force of the locomotive. Based on this, the resultant force of the locomotive unit in the previous steps can be calculated. The calculation is used to calculate and predict the velocity at the next moment.

[0055] After predicting the speed for the next moment, the locomotive's motion state is controlled. This process includes: acquiring track data and operational constraints; generating the optimal target speed curve for the locomotive based on the next moment's speed, track data, and operational constraints; acquiring real-time locomotive operation data; using the optimal target speed curve as feedforward input; combining the current driving decision with the predicted output of the locomotive dynamics model; and performing online correction of the optimal target speed curve using a PID control algorithm. Specifically, this involves calculating the error between the real-time locomotive operation data and the predicted output of the locomotive dynamics model; generating a correction value through the PID controller; using this correction value to correct the optimal target speed curve; and calculating and outputting traction or braking commands in real-time based on the corrected target speed curve to control the locomotive's motion state.

[0056] During driving, based on the driving stage, the current driving conditions are obtained. These driving conditions are then encoded and vectorized offline (as shown above) to obtain a condition vector c. , obtain The index j with the largest predicted value is the one used in the middle. One decision.

[0057] like Figure 11 As shown, in the control layer, a PID algorithm is used as a prediction error compensator to perform online closed-loop compensation on the residuals of the prediction model, thereby correcting the planning curve in the planning layer. Specifically, the control layer corrects the planning curve. (See figure.) It refers to the predicted locomotive state at time t+1 on the planning curve, such as position, speed, operating condition, and gear. Yes The corrected locomotive state at time t+1 Corrections are made using a PID control algorithm.

[0058] The locomotive state after correction at time t. Given the actual locomotive state at time t, calculate the error between the two conditions to obtain the error. Using a PID control system, the error value is... Convert this to the amount that needs to be corrected to achieve the correction. By correcting the system, Make corrections to obtain Use the revised version This involves controlling the locomotive. Specifically, it calculates and outputs traction or braking commands in real time based on the corrected target speed curve to control the locomotive's motion.

[0059] This application also provides a strategy planning and control system for locomotive assisted driving, including: an assisted driving strategy layer, an assisted driving planning layer, and an assisted driving control layer.

[0060] The assisted driving strategy layer is used to collect the locomotive's test data and operational data. The assisted driving planning layer is used to construct a traction characteristic function based on the test data, calculate instantaneous traction force based on the operational data and the traction characteristic function; it is also used to construct and train a timing calibration model, call the trained timing calibration model in real time during locomotive operation, and output the calibrated current traction force; it is also used to update the locomotive dynamics model based on the calibrated traction force and predict the speed at the next moment. The assisted driving control layer is used to control the locomotive's motion state based on the speed at the next moment.

[0061] This application provides a strategy planning and control method and system for locomotive assisted driving, which introduces and trains a time-series calibration model. By learning the dynamic correlation between historical traction force sequences and current instantaneous estimates, it can filter out noise and unreasonable jumps in the static function estimation results, outputting a more continuous and accurate calibrated traction force in a physically meaningful way, thus improving prediction fidelity and simulation realism. By training the time-series calibration model with data containing real traction force labels, it can learn and correct the inherent estimation bias of the static function, obtaining a traction force value that is closer to the actual output of the locomotive and has higher accuracy, thereby obtaining a more accurate predicted speed. Based on the speed predicted by the above-mentioned high-fidelity dynamic model, the solution environment is more consistent with the real world when optimizing the speed curve, improving the accuracy, response speed, and anti-interference capability of the assisted driving system control. The control system in this application systematically solves the technical problems of existing assisted driving systems in terms of decision flexibility, model accuracy, and control robustness through data-driven generation of decision rules at the strategy layer, time-series calibration of traction force at the planning layer, and closed-loop correction at the control layer, significantly improving the overall performance and engineering applicability of the assisted driving system.

[0062] Furthermore, a precise traction model can accurately predict the speed and force changes resulting from any maneuvering command (such as traction, cruise, coasting, and braking), thus providing reliable environmental feedback for optimization algorithms (such as dynamic programming and reinforcement learning). Based on this, a "suggestion curve" can be generated to guide the driver's precise operation, providing the core control logic for the intelligent driving system. This not only facilitates post-event energy consumption statistical analysis but also enables pre-event assessment of the energy consumption of different driving strategies, providing crucial data support and decision-making basis for developing energy-saving operation plans and achieving energy-saving and emission-reducing green rail transit.

[0063] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0064] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0065] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0066] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0067] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0068] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0069] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0070] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A strategy planning and control method for locomotive assisted driving, characterized in that, The strategy planning and control method for applying locomotive assisted driving includes: Acquire multiple conditional data associated with driving decisions, encode the conditional data into Boolean conditional vectors, and generate the current driving decision through linear mapping; Multiple sections of locomotive test data and operation data are collected. A traction characteristic function is constructed based on the test data. The instantaneous traction force at each moment is calculated based on the operation data and the traction characteristic function. The operation data includes speed, class, and actual traction force. A time-series calibration model is constructed, and the time-series calibration model is trained based on the operating data and the instantaneous traction force to obtain a trained time-series calibration model; The system acquires the locomotive's current operating data and historical traction force sequence in real time, calculates the current instantaneous traction force using the traction characteristic function, inputs the historical traction force sequence and the current instantaneous traction force into the trained time series calibration model, and outputs the calibrated current traction force. Based on the calibrated current traction force, a locomotive dynamics model is constructed, the locomotive speed for the next moment is updated and predicted, and the locomotive's motion state is controlled in conjunction with the current driving decision.

2. The strategy planning and control method for locomotive assisted driving according to claim 1, characterized in that, The traction characteristic function is constructed based on the test data, specifically including: Based on the experimental data of traction characteristics, a functional relationship between locomotive traction force, locomotive speed, and vehicle class is established using a piecewise polynomial fitting method. The traction characteristic function is shown below: in, Indicates the locomotive's traction force. Indicates speed, Indicates the level or rank.

3. The strategy planning and control method for locomotive assisted driving according to claim 1, characterized in that, The time-series calibration model is a long short-term memory network model, including a forget gate, an input gate, a memory gate, and an output gate; The inputs to the time-series calibration model include: a historical traction force sequence containing at least the traction force values ​​of the previous 5 seconds, and the current instantaneous traction force calculated by the traction characteristic function. The output of the time-series calibration model is the calibrated current traction force corresponding to the current moment.

4. The strategy planning and control method for locomotive assisted driving according to claim 3, characterized in that, When training the time-series calibration model, the hidden state of the previous time step and the current input are used as the common input of each gate. The current memory cell state and the hidden state are updated through the calculation of the forget gate, input gate, memory gate and output gate. The training process uses mean squared error as the loss function. The mean squared error between the estimated traction force calculated by forward propagation and the corresponding actual traction force is used as the optimization objective. The weight matrices and bias parameters of the forget gate, input gate, memory gate, and output gate in the time-series calibration model are updated using the backpropagation algorithm to minimize the loss function. The loss function is calculated as follows: in, The MSE loss function is... The estimated total number of samples at time point, The actual traction force value at time t. Let t be the estimated traction force at time t.

5. The strategy planning and control method for locomotive assisted driving according to claim 3, characterized in that, When training the time-series calibration model based on the operational data and the instantaneous traction force, each segment of operational data is collected as continuous data with a duration of not less than 10 minutes, and the sampling interval for speed, class, and actual traction force is 1 second. The actual traction force sequence of 5 consecutive seconds in each segment of operational data is combined with the instantaneous traction force of the 6th second to form a training sample, and the actual traction force of the 6th second is used as the label of the sample for model training.

6. The strategy planning and control method for locomotive assisted driving according to claim 1, characterized in that, Based on the calibrated current traction force, a locomotive dynamics model is constructed, and the locomotive speed for the next moment is updated and predicted, including: Based on the calibrated current traction force, locomotive mass, and traction mass, the unit traction force is calculated as follows: in, The current traction force after calibration. For locomotive quality, For traction mass; The locomotive's unit basic resistance, unit additional resistance, and unit braking force are calculated using the computer, combined with the unit traction force, to obtain the locomotive's unit resultant force. The calculation method is as follows: in, Indicates the combined force of locomotive units. Indicates the basic resistance per unit. This indicates the unit additional resistance of the locomotive. Indicates the unit braking force of the locomotive; Based on Newton's second law and the unit net force on the locomotive, the velocity at the next moment is calculated and updated as follows: in, Indicates the velocity at the next moment. Indicates the speed at the current moment. Indicates a time interval.

7. The strategy planning and control method for locomotive assisted driving according to claim 6, characterized in that, The basic unit resistance of the computer vehicle is calculated as follows: in, As a unit of basic resistance, , , This represents the coefficient in the drag formula. Indicates speed; The additional resistance per unit of the computer vehicle is calculated as follows: in, This indicates the unit additional resistance of the locomotive. Indicates the additional resistance per unit ramp. This indicates the additional resistance of the unit curve; The unit braking force of the computer vehicle is calculated as follows: in, This indicates the conversion of brake shoe pressure. This indicates the conversion coefficient of friction. This indicates the unit braking force of the locomotive.

8. The strategy planning and control method for locomotive assisted driving according to claim 1, characterized in that, Controlling the locomotive's motion state in conjunction with the current driving decision includes: The system acquires track data and operational constraints, and generates the optimal target speed curve for the locomotive based on the locomotive speed at the next moment, the track data, and the operational constraints. The track data includes: track gradient information, curve information, signal status, and speed limit information. The operational constraints include: locomotive traction or braking characteristic constraints, operating condition transition time constraints, operating timetable constraints, and safe speed limit constraints. The system acquires real-time locomotive operation data, uses the optimal target speed curve as a feedforward input, and combines the current driving decision with the predicted output of the locomotive dynamics model to perform online correction of the optimal target speed curve through a PID control algorithm. Specifically, this includes: calculating the error between the real-time locomotive operation data and the predicted output of the locomotive dynamics model, generating a correction amount through a PID controller, and using the correction amount to correct the optimal target speed curve to obtain the corrected target speed curve. The locomotive's motion state is controlled by calculating and outputting traction or braking commands in real time based on the corrected target speed curve.

9. The strategy planning and control method for applying locomotive assisted driving according to claim 8, characterized in that, The process of acquiring multiple conditional data associated with driving decisions, encoding the conditional data into Boolean conditional vectors, and generating the current driving decision through linear mapping includes: Each of the aforementioned conditional data is encoded into a Boolean conditional vector, wherein each dimension of the Boolean conditional vector takes a value of 0 or 1. Multiple preset driving decisions are encoded into one-hot decision vectors, where only one dimension of each one-hot decision vector is 1 and the rest are 0. Construct a linear mapping matrix to map the Boolean condition vector to the one-hot decision vector, thereby establishing a mapping relationship between conditions and decisions; The output of the linear mapping matrix is ​​converted into a probability distribution, and the weights of the linear mapping matrix are obtained by training with gradient descent using cross-entropy as the loss function. The current Boolean condition vector is obtained by encoding the conditional data at the current moment. The current Boolean condition vector is multiplied by the trained linear mapping matrix to obtain the predicted output vector. The index corresponding to the maximum value in the predicted output vector is taken as the current driving decision.

10. A strategy planning and control system for locomotive assisted driving, characterized in that, The system includes: The assisted driving strategy layer is used to acquire multiple conditional data related to driving decisions, encode the conditional data into Boolean conditional vectors, and generate the current driving decision through linear mapping. It is also used to collect the locomotive's test data and operation data. The assisted driving planning layer is used to construct a traction characteristic function based on the test data, calculate instantaneous traction force based on the operating data and the traction characteristic function; it is also used to construct and train a time-series calibration model, call the trained time-series calibration model in real time during locomotive operation, output the calibrated current traction force; it is also used to update the locomotive dynamics model based on the calibrated current traction force and predict the speed at the next moment. The driver assistance control layer is used to control the locomotive's motion state based on the speed at the next moment.

Images