Train traction braking performance prediction method

Abstract

The invention discloses a method for predicting traction and braking performance of a train. The method comprises the following steps: constructing a dynamic model and a curve model; pSO, GA or LSTM is selected to process coasting stage data to obtain basic resistance parameters; a sliding window averaging method is adopted to process the traction stage data and the braking stage data to obtain train acceleration data, and the full traction force and braking force data of the train are obtained according to the ATO current corresponding traction force and braking force conversion relation, the train acceleration data, the basic resistance parameters and the dynamic model; processing full traction force and braking force data by adopting a sliding window Z score processing technology; according to the full traction force data, obtaining traction characteristic curve parameters by adopting a segmented particle swarm algorithm or a weighted genetic algorithm; according to the full braking force data, the full-connection neural network and equilibrium loss fitting, braking characteristic curve parameters are obtained; and obtaining a traction characteristic curve and a braking characteristic curve of the train according to the traction characteristic curve parameters, the braking characteristic curve parameters and the curve model.

Description

Technical Field

[0001] The present invention relates to the technical field of urban rail transit, and more specifically to a method for predicting train traction and braking performance. Background Art

[0002] With the rapid development of urban rail transit, the traction and braking performance of urban rail trains has become increasingly important in ensuring train safety, improving operating efficiency and optimizing passenger experience. Traction performance directly affects the start, acceleration and smooth running of the train, while braking performance determines the response speed and accuracy of the train during deceleration, parking and emergency braking. In order to cope with the increasingly complex operating environment and growing traffic demand, traditional traction and braking control systems have been unable to meet the requirements of modern urban rail trains for efficiency, safety, energy saving and other aspects.

[0003] The traction and braking performance of urban rail trains are affected by many factors, including the train's load, speed, track conditions, air resistance, and external environment. These factors present a complex nonlinear relationship. Especially when the train is running at high speed, the distribution of traction and braking force becomes more difficult to control. Existing physical models or experience-based control strategies are difficult to accurately capture these nonlinear characteristics, resulting in problems such as excessive energy consumption of the train, inaccurate braking distance, or reduced ride comfort. In addition, modern urban rail trains usually use a multi-mode braking system that combines electric braking and mechanical braking. How to optimize the distribution of braking force under different working conditions and balance energy consumption and safety is also a major technical challenge.

[0004] Based on this background, the present invention proposes a method for predicting the traction and braking performance of trains. The method combines big data, machine learning, deep learning and other technologies. By dynamically analyzing the real-time operation data of the train, an intelligent prediction model is constructed. The method can predict the demand for traction and braking force under different speeds, loads and environmental conditions. The method is not only suitable for the optimization of traction and braking performance in the train design stage, but also can provide accurate performance evaluation and control solutions for trains in actual operation, effectively improving the reliability, economy and safety of train operation. Summary of the invention

[0005] The object of the present invention is to provide a method for predicting train traction braking performance to solve at least one of the problems existing in the prior art.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The present invention provides a method for predicting train traction and braking performance, comprising:

[0008] S1. Acquire train operation data, and construct a dynamic model including basic train resistance, additional train resistance, train traction force and braking force, and a curve model; the train operation data includes coasting phase data, traction phase data, and braking phase data;

[0009] S2. Process the data of the coasting stage using a particle swarm optimization algorithm, a genetic algorithm or a long-short sequence neural network method to obtain the basic resistance parameters of the train;

[0010] S3, using a sliding window average method to process the traction phase data and the braking phase data to obtain train acceleration data, and obtaining the full traction force data and full braking force data of the train according to the ATO current corresponding to the traction force conversion relationship, the ATO current corresponding to the braking force conversion relationship, the train acceleration data, the basic resistance parameters and the dynamic model;

[0011] S4, using a sliding window Z score processing technology to clean up abnormal values ​​and noise in the full traction force data and the full braking force data, and outputting smooth full traction force data and full braking force data;

[0012] S5, using a particle swarm algorithm or a weighted genetic algorithm to process the smoothed full traction force data to obtain traction characteristic curve parameters; using a fully connected neural network and balanced loss to process the smoothed full braking force data to obtain braking characteristic curve parameters;

[0013] S6. Obtaining a traction characteristic curve and a braking characteristic curve of the train according to the traction characteristic curve parameters, the braking characteristic curve parameters and the curve model.

[0014] Optionally, the kinetic model is

[0015]

[0016] Where M is the mass of the train; a(v x ) is the train speed v x The acceleration of the train at x ) is the train speed v x The traction or braking force at that time; f base (v x ) is the train speed v x W(x) is the unit basic resistance when the train is at position x; k1, k2 and k3 are the basic resistance parameters of the train; k3 is the friction resistance of the train, k2 is the resistance coefficient of the train, k1 is the aerodynamic resistance coefficient of the train; W i Add resistance to the ramp; w i is the additional resistance per unit slope; g is the acceleration due to gravity, w r is the additional resistance of the unit curve; R is the curve radius of the line; A is an empirical constant; F(vfront ) is the speed of the train v front Braking force at time, v front 、F back and v x are the speeds of the train at the corresponding front, back, and x positions, respectively, and F front is the braking force of the train at the corresponding position front; F back is the braking force of the train at the corresponding position back.

[0017] Optionally, the curve model is

[0018]

[0019] Among them, F(v) is the traction force when the train speed is v; v is the running speed of the train; K1 is the maximum traction force constant of the train, m, n, p and q are the constant coefficients of the traction characteristic curve fitting of the train in the constant torque area; B(v) is the braking force when the train speed is v; K2 is the maximum braking force constant of the train; o, l and h are the constant coefficients of the braking characteristic curve fitting; the traction characteristic curve of the train is divided into constant power area and constant torque area according to the running characteristics; the train running speed is less than v 转 When the train is in the constant torque zone, the maximum traction force of the train remains constant; in the constant power zone, when the train speed increases until it exceeds v 转 After that, it enters the constant torque area. In the constant torque area, the traction power of the train remains unchanged, v 终 is the maximum traction speed under the current speed level; when the train running speed is less than v 转折 When the train speed increases until it exceeds v 转折 After that, the train speed and the maximum power have a quadratic linear relationship, v 终点 It is the maximum braking speed at the current speed level.

[0020] Optionally, S2 further includes

[0021] S21. When the basic resistance parameter range of the train is known, perform preliminary analysis, rapid calibration or parameter initialization tasks, and use particle swarm optimization algorithm to process the data of the coasting stage to obtain the basic resistance parameters of the train;

[0022] S22. When it is necessary to accurately identify the basic resistance parameters of the train under various speed and load conditions, a genetic algorithm is used to process the data of the coasting stage to obtain the basic resistance parameters of the train;

[0023] S23. When the basic resistance parameters of the train are changing dynamically, or it is necessary to predict the basic resistance parameters under future speed changes, the long-short sequence neural network method is used to process the data of the coasting phase to obtain the basic resistance parameters of the train.

[0024] Optionally, the specific formula for processing data in the idling phase using the particle swarm optimization algorithm is:

[0025]

[0026] Among them, w start_PSO and w end_PSO are the inertia weights when the POS algorithm starts and ends exploration; f mean is the average fitness of the population using the POS algorithm, f min_PSO is the current optimal fitness using the POS algorithm, f is the fitness of the current individual, and F is the traction or braking force of the train;

[0027] The specific formula for processing data in the coasting stage using the genetic algorithm is:

[0028]

[0029] Among them, f(k) is the fitness value, F predicted is the resistance value calculated by following the parameters identified by the GA optimization process, F actual is the actual measured resistance value; v jdata For the jth data The train speed for each data point; For individual i indiv The weight of For individual i indiv The fitness value, f avg is the average fitness of the population using the GA algorithm, f min_GA is the optimal fitness of the population using the GA algorithm, f is the fitness of the current individual; w start_GA is the initial value of the inertia weight during GA algorithm exploration; w end_GA It is the end value of the inertia weight during GA algorithm exploration;

[0030] The specific formula for processing data in the idling phase using the long-short sequence neural network method is:

[0031]

[0032] in, is the mean of parameters k1, k2 and k3 in each round of iteration; is the standard deviation of parameters k1, k2 and k3 in each round of iteration; For the i para The parameter is in the jthsamp The value in the sample; n num is the sample size; For the i para The average value of parameters; α is the weight coefficient, which is used to balance the MSE loss and parameter constraint loss, L MSE is the mean square error, L constraint is the constraint error.

[0033] Optionally, the method of using a sliding window average method to process the traction phase data and the braking phase data to obtain the train acceleration data includes calculating the train acceleration in the window row by row and smoothing the train acceleration in the window. The specific formula is:

[0034]

[0035] in, Indicates the speed of the last line of the current window; Indicates the speed of the first line of the current window; △t is a fixed time interval; For window i win The average acceleration within The acceleration calculated for the current row; is the acceleration corresponding to each row in the window; N is the preset number of rows.

[0036] Optionally, the conversion relationship between the ATO current and the traction force and the conversion relationship between the ATO current and the braking force are:

[0037]

[0038] Among them, F all (v x ) is the train speed v x Full traction or full braking force, a all (v x ) is the train speed v x The acceleration of full traction force or full braking force when x ) is the train speed v x ATO current at all is the ATO current at full traction or full braking force; the ATO current when the train has no traction or the ATO current when the train has no braking force is 4; g is the acceleration due to gravity.

[0039] The calculation formula of the smoothed full traction force data or full braking force data is:

[0040]

[0041] in, The speed for cleaning outliers and noise is vx Full traction or braking force at The speed for cleaning outliers and noise is v x The acceleration at full traction or full braking force.

[0042] Optionally, a sliding window Z score processing technique is used to clean up outliers and noise in the full traction force data and the full braking force data, and smooth full traction force data and full braking force data are output, further including using Z score processing for the acceleration data of the full traction force in the constant power area, and using the sliding window automatic Z score processing for the acceleration data of the full traction force in the constant power area and the constant torque area; the acceleration data of the full braking force is processed by local weighted regression, interpolation outlier processing and moving average + spline interpolation method, and the specific formula is as follows

[0043]

[0044] in, is the i-th acceleration data of the converted full traction force acc Acceleration The Z score of the data; μ is the mean of the full traction acceleration data; σ is the standard deviation; W is the window size; represents the jth win-acc acceleration data points; For the i win The mean of the data in the window; σ iwin For the i win The standard deviation of the data within the window; Indicates the kth data-point The Z score of the data point; τ is the threshold of the Z score; is the kth data-point acceleration data; W' is the number of non-abnormal points in the window; D represents the current acceleration data and the acceleration data to be estimated a d The distance between each acceleration data For the acceleration data to be estimated a d Weight Calculated by cubic weight function; abnormal acceleration data Apply interpolation to replace it with adjacent acceleration data and A reasonable valuation at a given point in time On the basis of adjacent points The data is used to interpolate the current acceleration data; is the moving average acceleration data in the sliding window, is the jth nor Acceleration; Sner (v inter ) is the spline function; v inter is the velocity data of the interpolation point to be solved, v ner is the velocity of the interpolated adjacent points; b ner 、c ner d ner and e ner The spline coefficients fitted according to the data points and continuity conditions respectively.

[0045] Optionally, the particle swarm algorithm or weighted genetic algorithm is used to process the smoothed full traction force data to obtain traction characteristic curve parameters, including

[0046] When performing preliminary analysis, fast calibration or parameter initialization tasks, the particle swarm algorithm processes the smoothed full traction force data to obtain the traction characteristic curve parameters; the specific formula for processing the smoothed full traction force data using the particle swarm algorithm is:

[0047]

[0048] in, Represents particle i particle The speed at time t, For particle i particle The individual optimal position, g best is the global optimal position, w is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers between [0, 1]; Represents particle i particle The position at time t; is the actual full traction value, is the predicted value; is the second-order derivative of the fitting curve; λ is the weight coefficient of the penalty term; x1 and x n are the upper and lower limits of the turning point of the curve respectively; It is used to ensure that when the second-order derivative is negative, the penalty is not increased; when the second-order derivative is positive, the penalty is increased;

[0049] When it is necessary to identify the traction characteristic curve parameters of the train under various speed and load conditions, the traction characteristic curve parameters are obtained by using the genetic algorithm to process the smoothed full traction force data; the specific formula for using the genetic algorithm to process the smoothed full traction force data is:

[0050]

[0051] Where α is a random number between [0, 1], pa1 and pa2 are the first parent individual and the second parent individual, ch1 and ch2 are the first offspring individual and the second offspring individual, respectively; x′ indiv For individual xindiv The gene value after mutation.

[0052] Optionally, the loss function of the full braking force data smoothed by using a fully connected neural network and balanced loss processing is:

[0053] Loss=MSE+λ1×transition value constraint+λ2×transition slopeconstraint+λ3×(k penalty +a penalty +b penalty +c penalty )

[0054] Among them, λ1, λ2 and λ3 are the first control weight, the second control weight and the third control weight respectively; transitionvalue constraint is v 转折 The transition slope constraint is v 转折 The slope at time k penalty is the range constraint of K2, a penalty is the range constraint of o, b penalty is the range constraint of l, c penalty is the range constraint for h.

[0055] The beneficial effects of the present invention are as follows:

[0056] The present invention focuses on the accuracy of train traction and braking performance prediction and discloses a method for predicting train traction and braking performance. The present invention generates traction characteristic curves and braking characteristic curves based on train operation data, effectively reduces prediction errors, improves the safety and energy saving of train operation, and is suitable for the design and actual operation evaluation of urban rail trains. Compared with the train controller modeling of the physical model, the present invention is more accurate and efficient in predicting the train traction and braking performance, and the present invention can also evaluate the predicted results. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0058] Figure 1 Flowchart showing a method for predicting train traction and braking performance.

[0059] Figure 2 The basic resistance-speed curve diagram is shown, which is a result of processing the coasting phase data using the particle swarm algorithm, genetic algorithm and long-short sequence neural network method to obtain the basic resistance parameter results of the train.

[0060] Figure 3aA tractive effort-speed graph is shown without cleaning the full tractive effort data of outliers and noise.

[0061] Figure 3b A traction force-speed graph is shown after cleaning the outliers and noise for the full traction force data.

[0062] Figure 4a A braking force-speed graph is shown without full traction data cleaned of outliers and noise.

[0063] Figure 4b A braking force-speed graph is shown after cleaning outliers and noise for the full traction data.

[0064] Figure 5a A traction force-speed curve diagram of traction characteristic curve parameters obtained by using the particle swarm algorithm to process and smooth the full traction force data is shown.

[0065] Figure 5b A traction force-speed curve diagram showing the traction characteristic curve parameter results obtained by using a weighted genetic algorithm to process the smoothed full traction force data.

[0066] Figure 6 Braking force-speed graph showing the results of the fully connected neural network and balanced loss fitting braking characteristic curve parameters. DETAILED DESCRIPTION

[0067] In order to explain the present invention more clearly, the present invention is further described below in conjunction with embodiments and drawings. Similar components in the drawings are represented by the same reference numerals. It should be understood by those skilled in the art that the content specifically described below is illustrative rather than restrictive, and the protection scope of the present invention should not be limited thereto.

[0068] The urban rail train traction and braking performance prediction method provided by the present invention uses actual urban rail train operation data and line data to build a simulation test environment, and realizes model construction and solution under the Python platform.

[0069] In one embodiment, Figure 1 As shown, a method for predicting train traction braking performance includes the following steps:

[0070] S1. Acquire train operation data, and construct a dynamic model including basic train resistance, additional train resistance, train traction force and braking force, and a curve model; the train operation data includes coasting phase data, traction phase data, and braking phase data;

[0071] S2. Processing the data in the coasting phase by a particle swarm optimization algorithm (PSO), a genetic algorithm (GA) or a long short-term memory (LSTM) method to obtain basic resistance parameters of the train, and calculating basic resistances corresponding to different speeds of the train according to the basic resistance parameters;

[0072] S3, using a sliding window average method to process the traction phase data and the braking phase data to obtain train acceleration data, and obtaining the full traction force data and full braking force data of the train according to the ATO current corresponding to the traction force conversion relationship, the ATO current corresponding to the braking force conversion relationship, the train acceleration data, the basic resistance parameters and the dynamic model;

[0073] S4, using a sliding window Z score processing technique to clean up abnormal values ​​and noise in the full traction force data and the full braking force data, and outputting the flat full traction force data and the full braking force data;

[0074] S5, using a particle swarm algorithm or a weighted genetic algorithm to process the smoothed full traction force data to obtain traction characteristic curve parameters; using a fully connected neural network and balanced loss to process the smoothed full braking force data to obtain braking characteristic curve parameters;

[0075] S6. Obtaining a traction characteristic curve and a braking characteristic curve of the train according to the traction characteristic curve parameters, the braking characteristic curve parameters and the curve model.

[0076] The present invention further includes S7, evaluating the prediction results according to the basic resistance, traction characteristic curve and braking characteristic curve corresponding to different speeds of the train to provide a reference for performance evaluation and optimization of the train.

[0077] The present invention focuses on the accuracy of train traction and braking performance prediction and discloses a method for predicting train traction and braking performance. The present invention generates traction characteristic curves and braking characteristic curves based on train operation data, effectively reduces prediction errors, improves the safety and energy saving of train operation, and is suitable for the design and actual operation evaluation of urban rail trains. Compared with the train controller modeling of the physical model, the present invention is more accurate and efficient in predicting the train traction and braking performance, and the present invention can also evaluate the predicted results.

[0078] In one embodiment, step S1 further comprises

[0079] S11. Construction of single-mass dynamics model of urban rail transit trains. In urban rail transit, trains generally adopt a dispersed power marshaling method to ensure good consistency of traction and braking of the trains. Therefore, when modeling the dynamics of urban rail transit trains, a single-mass model can be used. The dynamics model formula is as follows:

[0080]

[0081] Where M is the mass of the train; a(v x ) is the train speed v x The acceleration of the train at time, in m / s 2 ; F(v x ) is the train speed v x The traction force or braking force at that time is in KN; f base (v x ) is the train speed v x The unit basic resistance when the train is at position x, in KN / N; W(x) is the total additional resistance when the train is at position x, in KN; k1, k2 and k3 are the basic resistance parameters of the train; k3 is the friction resistance of the train, k2 is the resistance coefficient of the train, and k1 is the aerodynamic resistance coefficient of the train; W i is the additional resistance of the ramp, in KN; w i is the additional resistance of the unit slope, in N / KN; g is the acceleration due to gravity, w r is the additional resistance of the unit curve, in N / KN; R is the curve radius of the line, in m; A is an empirical constant, usually 600; F(v front ) is the speed of the train v front Braking force at time, v front 、F back and v x are the speeds of the train at the corresponding front, back, and x positions, respectively, and F front is the braking force of the train at the corresponding position front; D back is the braking force of the train at the corresponding position back.

[0082] S12, the traction characteristic curve of the train is divided into constant power area and constant torque area according to the running characteristics; the train running speed is less than v 转 When the train is in the constant torque zone, the maximum traction force of the train remains constant; in the constant power zone, when the train speed increases until it exceeds v 转 After that, it enters the constant torque area. In the constant torque area, the traction power of the train remains unchanged, v 终 It is the maximum traction speed under the current speed level; the direction of the braking force is opposite to the direction of the train's travel, and the effect is opposite to the traction force. The braking characteristic curve of the train, when the train speed is less than v 转折When the train speed increases until it exceeds v 转折 After that, the train speed and the maximum power have a quadratic linear relationship, v 终点 It is the maximum braking speed under the current speed level. The output of braking force can be adjusted according to the actual needs of the train during travel. The specific formula of the traction braking characteristic curve is as follows

[0083]

[0084]

[0085] Wherein, F(v) is the traction force when the train speed is v; v is the train speed; K1 is the maximum traction force constant of the train, m, n, p and q are the constant coefficients of the traction characteristic curve fitting of the train in the constant torque area; B(V) is the braking force when the train speed is v; K2 is the maximum braking force constant of the train; o, l and h are the constant coefficients of the braking characteristic curve fitting.

[0086] In one embodiment, step S2 analyzes the actual running data during the coasting phase, processes the data during the coasting phase based on the particle swarm algorithm, the genetic algorithm and the long-short sequence neural network method to obtain the basic resistance parameters of the train, and accurately calculates the corresponding basic resistance at different speeds; Figure 2 The basic resistance-speed curve diagram shown is a diagram of the basic resistance parameters of the train obtained by processing the coasting phase data based on the particle swarm algorithm, genetic algorithm and long-short sequence neural network method. The horizontal axis is the speed of the train, in km / h; the vertical axis is the basic resistance of the train, in KN; the root mean square error (RMSE) is used to judge the parameter fitting effect. It can be seen from the figure that the RMSE of the three algorithms are all less than 0.035, and the maximum error of the resistance curve is 1%.

[0087] In one embodiment, step S2 includes

[0088] S21. When the basic resistance parameter range of the train is known and preliminary analysis, rapid calibration or parameter initialization tasks are performed and results need to be obtained quickly, the particle swarm optimization algorithm is used to process the data of the coasting stage to obtain the basic resistance parameters of the train;

[0089] Specifically, PSO gradually approaches the optimal solution by iteratively updating the position and velocity of particles. In the prediction of train resistance parameters, PSO is used to find the three parameters k1, k2, and k3 in the resistance formula. In order to enhance the global search and local search capabilities of PSO, this paper introduces an adaptive inertia weight adjustment mechanism. In the standard algorithm of PSO, the inertia weight w is an important parameter for controlling the particle speed. The larger the inertia weight, the more the algorithm tends to perform a global search, and when the weight is smaller, the algorithm pays more attention to local search. In order to dynamically balance the two, the inertia weight w is dynamically adjusted with the fitness value.

[0090]

[0091] Where g is the acceleration due to gravity, usually 9.81; Q is the preset value, usually 1000 for the conversion between KN and N; v is the speed of the train, M is the mass of the train; w start_PSO and w end_PSO are the inertia weights when the POS algorithm starts and ends exploration; f mean is the average fitness of the population using the POS algorithm, f min_PSO is the current optimal fitness using the POS algorithm, f is the fitness of the current individual, and F is the traction or braking force of the train;

[0092] S22. When it is necessary to accurately identify the basic resistance parameters of the train under various speed and load conditions, a genetic algorithm is used to process the data of the coasting stage to obtain the basic resistance parameters of the train;

[0093] Specifically, the train resistance parameter identification method based on GA iteratively optimizes the fitness function and continuously updates individual parameters to finally obtain the optimal solution. This includes using the tournament selection method to select individuals with higher fitness for reproduction, using crossover and mutation operations to introduce diversity, and adaptively updating weights to ensure the accuracy of the identification results:

[0094]

[0095] Among them, f(k) is the fitness value, F predicted is the resistance value calculated by following the parameters identified by the GA optimization process, F actual is the actual measured resistance value; For the The train speed of each data point, the smaller the value of the fitness function, the more the individual is adapted to the solution space of the current problem; For individual i indiv The weight of For individual i indiv The fitness value, f avg is the average fitness of the population using the GA algorithm, f min_GAis the optimal fitness of the population using the GA algorithm, f is the fitness of the current individual; w start_GA is the initial value of the inertia weight during GA algorithm exploration; w end_GA It is the end value of the inertia weight during GA algorithm exploration.

[0096] S23. When the basic resistance parameters of the train are changing dynamically, or it is necessary to predict the basic resistance parameters under future speed changes, the long-short sequence neural network method is used to process the data of the coasting phase to obtain the basic resistance parameters of the train.

[0097] Specifically, the resistance parameters of the train show a complex nonlinear relationship with the speed. LSTM can predict the resistance parameters k1, k2 and k3 by learning the historical speed and acceleration data. During the LSTM training process, the model will output different resistance prediction values ​​according to the parameter values ​​k1, k2 and k3 of each iteration. In order to avoid the influence of noise in the model training process and ensure the stability of the prediction results, the mean and mean square error of the parameter values ​​predicted by LSTM are optimized. In order to ensure that the prediction results of the LSTM model are reasonable in the physical sense, parameter constraint loss is added to the loss function. The total loss function includes the mean square error (MSE) loss: used to calculate the error between the predicted acceleration and the actual acceleration; the constraint loss is used to limit the value range of k1, k2 and k3 to ensure that the parameters meet the physical requirements.

[0098]

[0099] in, is the mean of parameters k1, k2 and k3 in each round of iteration; is the standard deviation of parameters k1, k2 and k3 in each round of iteration; For the i para The parameter is in the jth samp The value in the sample; n num is the sample size; For the i para The average value of parameters; α is the weight coefficient, which is used to balance the MSE loss and parameter constraint loss, L MSE is the mean square error, L constraint is the constraint error.

[0100] In summary, the three-dimensional particle swarm algorithm with adaptive inertia weight, the adaptive weight genetic algorithm, and the LSTM algorithm with mean square error and multi-objective loss function are used to identify the parameters of the basic resistance of the train, and the basic resistance of the urban rail train is predicted. The three methods provided can adapt to different operating environments and different operating data and accurately predict the basic resistance of the train.

[0101] In one embodiment, step S3 performs data preprocessing on the ATO (Automatic Train Operation) running curve data of the actual train, and the acceleration calculation adopts the sliding window average method to obtain a more stable and smooth acceleration result. The method includes two steps: line-by-line acceleration calculation and window acceleration smoothing. The specific formula is as follows

[0102]

[0103] in, Indicates the speed of the last line of the current window (i win +N line speed); Indicates the speed of the first line of the current window; △t is a fixed time interval; For window i win The average acceleration within The acceleration calculated for the current row; is the acceleration corresponding to each row in the window; N is the preset number of rows, for example, N=5 means that the acceleration of every 5 rows is averaged.

[0104] In one embodiment, the current range of the ATO analog output is generally selected to be 0-20mA, but only the range of 4-18mA is used as a linear set value, and its value is 0 to 100% of the set value. The output in the range of 0-2mA is considered invalid. The range of 2-4mA is used for the set value of 0% (the 0% set value means that the train is coasting and no acceleration or deceleration should occur at all), and the range of 18-20mA is used for the set value of 100%. According to the ATO current corresponding to the traction force conversion relationship, the ATO current corresponding to the braking force conversion relationship, the train acceleration data and the dynamic model, the full traction force data and the full braking force data of the train are obtained;

[0105] The conversion relationship between ATO current and traction force and the conversion relationship between ATO current and braking force are as follows:

[0106]

[0107] Among them, F all (v x ) is the train speed v x The full traction force or full braking force (corresponding to ATO current of 18mA or -18mA), a all (v x ) is the train speed v x The acceleration of full traction force or full braking force (corresponding to ATO current of 18mA or -18mA), I(v x ) is the train speed v x ATO current atall is the ATO current of full traction force or full braking force (18mA or -18mA), 4 is the ATO current when the train has no traction force or the ATO current when the train has no braking force; g is the acceleration due to gravity.

[0108] The calculation formula of the smoothed full traction force data or full braking force data is:

[0109]

[0110] in, The speed for cleaning outliers and noise is v x Full traction or braking force (corresponding ATO current is 18mA or -18mA), The speed for cleaning outliers and noise is v x The acceleration of full traction force or full braking force when the ATO current is 18mA or -18mA.

[0111] In one embodiment, step S4 uses a sliding window Z score processing technology to clean up outliers and noise in the full traction force data and the full braking force data, and outputs smooth full traction force data and full braking force data, further including using Z score processing for the acceleration data of the full traction force in the constant power area, and using the sliding window automatic Z score processing for the acceleration data of the full traction force in the constant power area and the constant torque area; using local weighted regression, interpolation outlier processing and moving average + spline interpolation method for the acceleration data of the full braking force, the specific formula is as follows

[0112]

[0113] in, is the i-th acceleration data of the converted full traction force acc Acceleration The Z score of the data; μ is the mean of the full traction acceleration data; σ is the standard deviation; W is the window size; Indicates the first acceleration data points; For the i win The mean of the data in the window; For the i win The standard deviation of the data within the window; Indicates the kth data-point The Z score of the data point; τ is the threshold of the Z score; is the kth data-point acceleration data; W' is the number of non-abnormal points in the window; D represents the current acceleration data and the acceleration data to be estimated a dThe distance between them is usually the Euclidean distance as the non-abnormal data point in the window; each acceleration data For the acceleration data to be estimated a d Weight Calculated by cubic weight function; abnormal acceleration data Apply interpolation to replace it with adjacent acceleration data and A reasonable valuation at a given point in time On the basis of adjacent points The data is used to interpolate the current acceleration data; is the moving average acceleration data in the sliding window, is the jth nor Acceleration; S ner (v inter ) is the spline function; v inter is the velocity data of the interpolation point to be solved, v ner is the velocity of the interpolated adjacent points; b ner 、c ner d ner and e ner The spline coefficients fitted according to the data points and continuity conditions respectively.

[0114] Specifically, Figure 3a A traction force-speed graph showing that full traction force data has not been cleaned of outliers and noise, Figure 3b A traction force-speed curve diagram is shown after cleaning outliers and noise from full traction force data;

[0115] Figure 3a and Figure 3b After comparison, it is clear that Figure 3b The curve has less noise and is smoother, which is suitable for parameter identification of traction characteristic curve. Figure 3a and Figure 3b The horizontal axis is the speed, in km / h; the vertical axis is the traction, in KN. Figure 4a Braking force-speed graph showing full traction data not cleaned of outliers and noise; Figure 4b A braking force-speed graph is shown after cleaning outliers and noise for the full traction data. Figure 4a and Figure 4b After comparison, it is clear that Figure 4b The curve has less noise and is smoother, which is suitable for parameter identification of the braking characteristic curve. Figure 4a and Figure 4bThe horizontal axis is speed, in km / h; the vertical axis is braking force, in KN. By combining a variety of data optimization processing technologies such as sliding window Z score, outliers and noise are cleaned up, and smooth data to be identified is output, thereby ensuring the accuracy of the data and providing good input data for subsequent basic resistance, traction, and braking performance predictions.

[0116] In one embodiment, the full traction force data and full braking force data after cleaning outliers and noise are analyzed, a particle swarm algorithm or a weighted genetic algorithm is selected to identify traction characteristic curve parameters, and a fully connected neural network (FNN) and balanced loss are used to fit the braking characteristic curve parameters.

[0117] In order to accurately predict the traction curve of the train, such as Figure 5a and Figure 5b Particle swarm optimization algorithm (PSO) and genetic algorithm (GA) were used to optimize the key parameters of traction in the curve model. Figure 5a A traction force-speed curve diagram of traction characteristic curve parameters obtained by using the particle swarm algorithm to process and smooth the full traction force data is shown. Figure 5b The traction force-speed curve diagram is shown, which shows the traction characteristic curve parameter results obtained by using the weighted genetic algorithm to process the smoothed full traction force data. The RMSE of the identification results of the two algorithms are both less than 1%. Figure 5a and Figure 5b The horizontal axis is speed, the unit is km / h; the vertical axis is traction, the unit is KN.

[0118] In one embodiment, when performing preliminary analysis, fast calibration or parameter initialization tasks, the particle swarm algorithm processes the smoothed full traction force data to obtain the traction characteristic curve parameters; the specific formula for processing the smoothed full traction force data using the particle swarm algorithm is:

[0119]

[0120] in, Represents particle i particle The speed at time t, For particle i particle The individual optimal position, g best is the global optimal position, w is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers between [0, 1]; Represents particle i particle The position at time t; is the actual full traction value, is the predicted value; is the second-order derivative of the fitting curve; λ is the weight coefficient of the penalty term; x1 and x n are the upper and lower limits of the turning point of the curve respectively; It is used to ensure that when the second-order derivative is negative (i.e. concavity), the penalty is not increased. When the second-order derivative is positive (convexity), the penalty is increased.

[0121] Specifically, the basic idea of ​​the segmented particle swarm algorithm (PSO) is to find the optimal solution by simulating the collaborative search behavior of individuals in the group in the search space, and predict the straight and curved parts of the traction characteristic curve respectively. In the PSO algorithm, each individual (i.e., particle) represents a potential solution, and its quality is evaluated according to the fitness function. The algorithm updates its state by adjusting the speed and position of the particle, gradually approaching the optimal solution. The fitness function of the algorithm is based on the residual sum of squares (RSS), that is, the error between the fitted value and the actual value; in order to ensure the smoothness and concavity of the curve, a curvature penalty term is introduced. This penalty term calculates the second-order derivative of the fitted curve and increases the penalty when the second-order derivative is positive (i.e., the curve is convex), so that the algorithm generates a concave curve. Using dynamic weights and learning factors, in each round of iteration, the inertia weight and learning factor are dynamically adjusted according to the number of iterations. The inertia weight is gradually reduced from the initial value of 0.9 to 0.4, and the learning factor is also adjusted linearly with the number of iterations. This dynamic adjustment mechanism helps to explore a larger search space in the early stage and converge to the global optimal solution in the later stage.

[0122] In one embodiment,

[0123] When it is necessary to identify the traction characteristic curve parameters of the train under various speed and load conditions, the traction characteristic curve parameters are obtained by using the genetic algorithm to process the smoothed full traction force data; the specific formula for using the genetic algorithm to process the smoothed full traction force data is:

[0124]

[0125] Where α is a random number between [0, 1], pa1 and pa2 are the first parent individual and the second parent individual, ch1 and ch2 are the first offspring individual and the second offspring individual, respectively; x′ indiv For individual x indiv The gene value after mutation; σ is the standard deviation, which is used to control the amplitude of mutation.

[0126] Specifically, the genetic algorithm (GA) is an optimization algorithm that simulates the biological evolution process. It uses selection, crossover and mutation operations to evolve individuals in the population, thereby gradually approaching the optimal solution. The GA algorithm has global search capabilities and can handle nonlinear complex optimization problems. Similar to PSO, the fitness function of GA is also based on the residual sum of squares and contains turning point penalty and curvature penalty terms. In order to ensure that the constant K in the low-speed interval and the polynomial curve in the high-speed interval are continuous at the turning point, a turning point penalty term is introduced, requiring that the difference between the value of K and the polynomial here is less than 1, otherwise the difference will be squared. Penalty. Curvature penalty is also used to optimize the generation of concave curves. In each generation of GA, the roulette wheel selection operation is first used to select individuals with higher fitness for reproduction. Then, a uniform crossover operation is performed, that is, the genetic information of the two parents is combined into a new individual through random weights. In addition, the parameters of each individual are randomly fine-tuned through the Gaussian mutation operation to increase the diversity of the population. The mutation rate is set to 0.1 to ensure moderate diversity.

[0127] In one embodiment, in order to accurately predict the braking force curve of the train, a fully connected neural network (FNN) is used to identify the parameters of the train braking force model. Figure 6 As shown in the figure, the braking force-speed curve of the braking characteristic curve parameter results of the fully connected neural network and the balanced loss fitting. Among them, the horizontal axis is the speed, the unit is km / h; the vertical axis is the braking force, the unit is KN; according to Figure 6 As shown, the RMSE of the fully connected neural network and balanced loss fitting key braking parameters is less than 0.8%.

[0128] Specifically, we designed a neural network with multiple hidden layers, which outputs four key parameters: K2, o, l, and h. The model is trained to fit the braking force curve of the train at different speeds.

[0129] Extract speed and brake force data. In order to adapt to the input of the neural network, the speed data is converted into a tensor and the corresponding relationship between brake force and speed is maintained. The designed fully connected neural network contains three hidden layers, each followed by a ReLU activation function. The input of the network is speed, and the output is the parameters K2, o, l, h of the brake force model.

[0130] In order to ensure the fitting accuracy of the model in each speed range, a custom loss function is defined, which not only takes into account the mean square error (MSE), but also adds the constraint of smooth transition at the speed turning point (v = 85km / h). These constraints include:

[0131] ①k and polynomial f+gv+hv 2The values ​​at the turning points should be equal.

[0132] ② Smooth transition of the slope, that is, when v=85, the slope should remain consistent.

[0133] ③ Constrain the parameters K2, o, l and h to keep them within a reasonable range.

[0134] The model was trained using the Adam optimizer, with a learning rate of 0.07, and the training process was performed for 1000 epochs. In each epoch, the parameters K2, o, l, and h output by the neural network were used to calculate the braking force curve, and the loss function performed gradient backpropagation and parameter update based on the difference between the actual and predicted braking forces.

[0135] The loss function of the full braking force data smoothed by using a fully connected neural network and balanced loss processing is

[0136] Loss=MSE+λ1×transition value constraint+λ2×transition slopeconstraint+λ3×(k penalty +a penalty +b penalty +c penalty )

[0137] Among them, θ1, λ2 and λ3 are the first control weight, the second control weight and the third control weight respectively; transitionvalue constraint is v 转折 The transition slope constraint is v 转折 The slope at time k penalty is the range constraint of K2, a penalty is the range constraint of o, b penalty is the range constraint of l, c penalty is the range constraint for h.

[0138] The present invention focuses on the accuracy of train traction and braking performance prediction and discloses a method for predicting train traction and braking performance. The present invention generates traction characteristic curves and braking characteristic curves based on train operation data, effectively reduces prediction errors, improves the safety and energy saving of train operation, and is suitable for the design and actual operation evaluation of urban rail trains. Compared with the train controller modeling of the physical model, the present invention is more accurate and efficient in predicting the train traction and braking performance, and the present invention can also evaluate the predicted results.

[0139] In one embodiment, step S7 verifies and evaluates the accuracy and effectiveness of the predicted traction characteristic curve and braking characteristic curve. The proposed evaluation method includes:

[0140] ① Error calculation and deviation point statistics.

[0141] Error calculation: The traction characteristic curve and braking characteristic curve of the predicted train are discretized to obtain multiple discrete points. The relative error is then calculated for each discrete point and compared with the reference curve.

[0142] Deviation point statistics: Set a deviation threshold (such as 2%), count the number of points whose errors exceed the threshold, and then calculate the proportion of deviation points (set to no more than 1% or lower), which is considered to meet the prediction accuracy requirements.

[0143] ②Comparison of model simulation operation curves.

[0144] The curve model is used in combination with the line control strategy, and the operating curve is simulated considering the train and line conditions to obtain the speed-time curve and the speed-displacement curve. The obtained curve is compared with the operating curve of the reference model or the actual operating curve. The prediction accuracy is within a certain threshold range (99%) and is considered to meet the prediction accuracy requirements.

[0145] The present invention can evaluate the prediction results, effectively reduce the prediction error, improve the safety and energy saving of train operation, and is suitable for the design and actual operation evaluation of urban rail trains.

[0146] In the description of the present invention, it should be noted that the terms "upper", "lower", etc. indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific position, be constructed and operated in a specific position, and therefore cannot be understood as a limitation on the present invention. Unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be a connection between the two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0147] It should also be noted that, in the description of the present invention, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0148] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not limitations on the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the embodiments here. All obvious changes or modifications derived from the technical solution of the present invention are still within the protection scope of the present invention.

Claims

1. A method for predicting train traction and braking performance, characterized in that: include: S1. Obtain train operation data and construct a dynamic model including basic train resistance, additional train resistance, train traction and braking force, and a curve model; The train operation data includes coasting phase data, traction phase data and braking phase data; S2. Process the data of the coasting stage using a particle swarm optimization algorithm, a genetic algorithm or a long-short sequence neural network method to obtain the basic resistance parameters of the train; S3, using a sliding window average method to process the traction phase data and the braking phase data to obtain train acceleration data, and obtaining the full traction force data and full braking force data of the train according to the ATO current corresponding to the traction force conversion relationship, the ATO current corresponding to the braking force conversion relationship, the train acceleration data, the basic resistance parameters and the dynamic model; S4, using a sliding window Z score processing technology to clean up abnormal values ​​and noise in the full traction force data and the full braking force data, and outputting smooth full traction force data and full braking force data; S5, using a particle swarm algorithm or a weighted genetic algorithm to process the smoothed full traction force data to obtain traction characteristic curve parameters; using a fully connected neural network and balanced loss to process the smoothed full braking force data to obtain braking characteristic curve parameters; S6. Obtaining a traction characteristic curve and a braking characteristic curve of the train according to the traction characteristic curve parameters, the braking characteristic curve parameters and the curve model.

2. The method according to claim 1, characterized in that: The kinetic model is Where M is the mass of the train; a(v x ) is the train speed v x The acceleration of the train at x ) is the train speed v x The traction or braking force at that time; f base (v x ) is the train speed v x W(x) is the unit basic resistance when the train is at position x; k1, k2 and k3 are the basic resistance parameters of the train; k3 is the friction resistance of the train, k2 is the resistance coefficient of the train, k1 is the aerodynamic resistance coefficient of the train; W i Add resistance to the ramp; w i is the additional resistance per unit slope; g is the acceleration due to gravity, w r is the additional resistance of the unit curve; R is the curve radius of the line; A is an empirical constant; F(v front ) is the speed of the train v front Braking force at time, v front 、F back and v x are the speeds of the train at the corresponding front, back, and x positions, respectively, and F front is the braking force of the train at the corresponding position front; F back is the braking force of the train at the corresponding position back.

3. The method according to claim 2, characterized in that The curve model is Among them, F(v) is the traction force when the train speed is v; v is the running speed of the train; K1 is the maximum traction force constant of the train, m, n, p and q are the constant coefficients of the traction characteristic curve fitting of the train in the constant torque area; B(v) is the braking force when the train speed is v; K2 is the maximum braking force constant of the train; o, l and h are the constant coefficients of the braking characteristic curve fitting; the traction characteristic curve of the train is divided into constant power area and constant torque area according to the running characteristics; the train running speed is less than v 转 When the train is in the constant torque zone, the maximum traction force of the train remains constant; in the constant power zone, when the train speed increases until it exceeds v 转 After that, it enters the constant torque area. In the constant torque area, the traction power of the train remains unchanged, v 终 is the maximum traction speed under the current speed level; when the train running speed is less than v 转折 When the train speed increases until it exceeds v 转折 After that, the train speed and the maximum power have a quadratic linear relationship, v 终点 It is the maximum braking speed at the current speed level.

4. The method according to claim 3, characterized in that The S2 further comprises S21. When the basic resistance parameter range of the train is known, perform preliminary analysis, rapid calibration or parameter initialization tasks, and use particle swarm optimization algorithm to process the data of the coasting stage to obtain the basic resistance parameters of the train; S22. When it is necessary to accurately identify the basic resistance parameters of the train under various speed and load conditions, a genetic algorithm is used to process the data of the coasting stage to obtain the basic resistance parameters of the train; S23. When the basic resistance parameters of the train are changing dynamically, or it is necessary to predict the basic resistance parameters under future speed changes, the long-short sequence neural network method is used to process the data of the coasting phase to obtain the basic resistance parameters of the train.

5. The method according to claim 4, characterized in that The specific formula for processing data in the idling phase using the particle swarm optimization algorithm is: Among them, w start_PSO and w end_PSO are the inertia weights when the POS algorithm starts and ends exploration; f mean is the average fitness of the population using the POS algorithm, f min_PSO is the current optimal fitness using the POS algorithm, f is the fitness of the current individual, and F is the traction or braking force of the train; The specific formula for processing data in the coasting stage using the genetic algorithm is: Among them, f(k) is the fitness value, F predicted is the resistance value calculated by following the parameters identified by the GA optimization process, F actual is the actual measured resistance value; For the jth data The train speed for each data point; For individual i indiv The weight of For individual i indiv The fitness value, f avg is the average fitness of the population using the GA algorithm, f min_GA is the optimal fitness of the population using the GA algorithm, f is the fitness of the current individual; w start_GA is the initial value of the inertia weight during GA algorithm exploration; w end_GA It is the end value of the inertia weight during GA algorithm exploration; The specific formula for processing data in the idling phase using the long-short sequence neural network method is: in, is the mean of parameters k1, k2 and k3 in each round of iteration; is the standard deviation of parameters k1, k2 and k3 in each round of iteration; For the i para The parameter is in the jth samp The value in the sample; n num is the sample size; For the i para The average value of parameters; α is the weight coefficient, which is used to balance the MSE loss and parameter constraint loss, L MSE is the mean square error, L constraint is the constraint error.

6. The method according to claim 5, characterized in that The method of using the sliding window average method to process the traction phase data and the braking phase data to obtain the train acceleration data includes calculating the train acceleration in the window line by line and smoothing the train acceleration in the window. The specific formula is: in, Indicates the speed of the last line of the current window; Indicates the speed of the first line of the current window; △t is a fixed time interval; For window i win The average acceleration within The acceleration calculated for the current row; is the acceleration corresponding to each row in the window; N is the preset number of rows.

7. The method according to claim 6, characterized in that The conversion relationship between the ATO current and the traction force and the conversion relationship between the ATO current and the braking force are as follows: Among them, F all (v x ) is the train speed v x Full traction or full braking force, a all (v x ) is the train speed v x The acceleration of full traction force or full braking force when x ) is the train speed v x ATO current at all is the ATO current of full traction or full braking force; the ATO current when the train has no traction or the ATO current when the train has no braking force is 4; g is the acceleration due to gravity; The calculation formula of the smoothed full traction force data or full braking force data is: in, The speed for cleaning outliers and noise is v x Full traction or braking force at The speed for cleaning outliers and noise is v x The acceleration at full traction or full braking force.

8. The method according to claim 7, characterized in that The sliding window Z score processing technology is used to clean up the outliers and noise in the full traction force data and the full braking force data, and output the smooth full traction force data and the full braking force data, further including using Z score processing for the acceleration data of the full traction force in the constant power area, and using the sliding window automatic Z score processing for the acceleration data of the full traction force in the constant power area and the constant torque area; the acceleration data of the full braking force is processed by local weighted regression, interpolation outlier processing and moving average + spline interpolation method, and the specific formula is as follows in, is the i-th acceleration data of the converted full traction force acc Acceleration The Z score of the data; μ is the mean of the full traction acceleration data; σ is the standard deviation; W is the window size; represents the jth win-acc acceleration data points; For the i win The mean of the data in the window; For the i win The standard deviation of the data in the window; Indicates the kth data-point The Z score of the data point; τ is the threshold of the Z score; is the kth data-point acceleration data; W' is the number of non-abnormal points in the window; D represents the current acceleration data and the acceleration data to be estimated a d The distance between each acceleration data For the acceleration data to be estimated a d Weight Calculated by cubic weight function; abnormal acceleration data Apply interpolation to replace it with adjacent acceleration data and A reasonable valuation at a given point in time On the basis of adjacent points The data is used to interpolate the current acceleration data; is the moving average acceleration data in the sliding window, is the jth nor Acceleration; S ner (v inter ) is the spline function; v inter is the velocity data of the interpolation point to be solved, v ner is the velocity of the interpolated adjacent points; b ner 、c ner ,d ner and e ner The spline coefficients fitted according to the data points and continuity conditions respectively.

9. The method according to claim 8, characterized in that The particle swarm algorithm or weighted genetic algorithm is used to process the smoothed full traction force data to obtain the traction characteristic curve parameters, including When performing preliminary analysis, fast calibration or parameter initialization tasks, the particle swarm algorithm processes the smoothed full traction force data to obtain the traction characteristic curve parameters; the specific formula for processing the smoothed full traction force data using the particle swarm algorithm is: in, Represents particle i particle The speed at time t, For particle i particle The individual optimal position, g best is the global optimal position, w is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers between [0, 1]; Represents particle i particle The position at time t; is the actual full traction value, is the predicted value; is the second-order derivative of the fitting curve; λ is the weight coefficient of the penalty term; x1 and x n are the upper and lower limits of the turning point of the curve respectively; It is used to ensure that when the second-order derivative is negative, the penalty is not increased; when the second-order derivative is positive, the penalty is increased; When it is necessary to identify the traction characteristic curve parameters of the train under various speed and load conditions, the traction characteristic curve parameters are obtained by using the genetic algorithm to process the smoothed full traction force data; the specific formula for using the genetic algorithm to process the smoothed full traction force data is: Where α is a random number between [0, 1], pa1 and pa2 are the first parent individual and the second parent individual, ch1 and ch2 are the first offspring individual and the second offspring individual, respectively; x′ indiv For individual x indiv The gene value after mutation.

10. The method according to claim 9, characterized in that The loss function of the full braking force data smoothed by using the fully connected neural network and balanced loss processing is: Loss=MSE+λ1×transition valueconstraint+λ2×transition slope constraint+λ3×(k penalty +a penalty +b penalty +c penalty ) Among them, λ1, λ2 and λ3 are the first control weight, the second control weight and the third control weight respectively; transitionvalue constraint is v 转折 The transition slope constraint is v 转折 The slope at time k penalty is the range constraint of K2, a penalty is the range constraint of o, b penalty is the range constraint of l, c penalty is the range constraint for h.

Images