CN115062402A - A data-driven method for extracting train level acceleration - Google Patents

Abstract

The invention relates to a data-driven train level acceleration extraction method, and belongs to the technical field of computers and information science. The method firstly optimizes the learning sequence of the samples by utilizing different loss change modes of the noise and the normal samples in the training process. Then introducing self-sampling learning into AdaBoost and gradient boosting learning respectively to provide an SSAdaBoost and SSGB steady learning method, establishing a target train performance estimation model, mining special train performance knowledge in noisy operation data, and fitting a mapping function between acceleration and characteristics such as speed, a rank sequence (implicit delay characteristic) and gradient. And finally, controlling the characteristic influence of a level sequence, time delay and the like by using a 'query sample', establishing a performance table, and realizing the extraction of the quantitative relation between the target characteristic and the label. The problems that the tracking difficulty of the recommended speed is high and the extra energy consumption is caused by frequent switching of the vehicle control level in the existing method are solved. The extraction performance is similar to the actual performance, and the method can be used for establishing performance constraint matched with the controlled train and improving the optimization effect of the recommendation speed.

Description

A data-driven method for extracting train level acceleration

technical field

The invention relates to a data-driven train level acceleration extraction method, and belongs to the technical field of computer and information science.

Background technique

The development of rail transit is an effective way to solve the travel problem of residents in modern big cities, and it is also an important way to build a green city and a smart city. Common rail transits include traditional railways (national railways, intercity railways and municipal railways), subways, light rails and trams, and new types of rail transits include maglev rail systems, monorail systems (straddle rail systems and suspended rail systems) and Passenger automatic rapid transit system, etc. With the diversified development of trains and railway technology, rail transit has shown more and more types, not only in long-distance land transportation, but also widely used in short- and medium-distance urban public transportation. It is of great significance to monitor its real-time operating status and ensure its safe operation.

The train operation data mainly includes the speed, recommended speed and level collected by the on-board system, as well as the position, slope, curve and target point position fed back by the CBTC, which are stored in formatted logs cycle by cycle. In the motion state estimation problem, the speed difference between adjacent sampling points is calculated to obtain the acceleration, which is used as a supervised learning label to mine the change pattern of the train motion state under the influence of the current speed, slope and level.

When the vehicle is running, affected by various factors such as sensor error, communication packet loss, processing delay, etc., the sampled data is prone to fluctuations, jumps, missing, etc., resulting in errors in the calculation of acceleration and the introduction of label noise. The recommended speed is the ideal running speed that satisfies constraints such as speed limit and timetable, and is optimized for train operation efficiency, energy consumption, comfort and other indicators. Existing recommendation speed optimization methods are generally based on continuous or discrete assumptions, focusing on the energy consumption level of the control target. For the stage speed control system commonly used by trains, the discrete assumption can consider the train delay, acceleration dispersion and other issues to a certain extent, but the existing methods fail to analyze and extract the train performance, and the optimization process is still independent of the controlled object. The recommended acceleration may not match the feasible region of the train acceleration, resulting in frequent level switching during operation, resulting in unexpected energy consumption.

SUMMARY OF THE INVENTION

The purpose of the present invention is that the existing method does not consider the matching of the recommended acceleration and the train level acceleration, so that the tracking of the recommended speed is difficult, the control level is frequently switched, and additional energy consumption is generated. A data-driven extraction method of train level acceleration is proposed.

The design principle of the present invention is as follows: firstly, the reliability and learning difficulty of the samples are estimated by using the different loss change patterns exhibited by the noise samples and the normal samples in the training process, and the learning sequence of the samples is optimized. Then, the self-sampling learning is introduced into AdaBoost and gradient boosting learning respectively, and SSAdaBoost and SSGB robust learning methods are proposed to establish the target train performance estimation model, mine the special car performance knowledge in the noisy running data, and fit the acceleration, speed, and level sequence (implicitly). The mapping function between features such as delay characteristics), slope and so on. Finally, a correlation extraction method based on feature control is proposed. The "query sample" is used to control the influence of features such as stage sequence, delay, and slope, and the corresponding acceleration values of the lower stages in different speed ranges are inquired in turn to establish a train-specific performance table. Quantitative relationship extraction between target features and labels.

The technical scheme of the present invention is achieved through the following steps:

Step 1, for noisy tabular data, a self-sampling learning method SSL is proposed.

Step 1.1, using the variation pattern of the loss function value during the training process of the boosting learning model, including the absolute value, the variation direction and the variation speed, of the noise-containing samples and the normal samples, to estimate the reliability of the samples and the learning difficulty.

Step 1.2, adjust the training sample subset at each iteration, and optimize the sample learning order while eliminating noise points.

Step 2, introduce self-sampling learning into AdaBoost, and establish SSAdaBoost and SSGB methods.

Step 2.1, establish the target train performance estimation model, mine the special car performance knowledge in the noisy running data, and fit the mapping function between acceleration and speed, level sequence (implicit delay characteristic), gradient and other features.

Step 3: Use the "query sample" to control the influence of features such as level sequence, delay, and gradient, and sequentially query the acceleration values corresponding to the lower levels in different speed ranges.

Step 3.1: Establish a train-specific performance table to extract the quantitative relationship between target features and labels.

beneficial effect

变化为viwi，有效限制了权重增长。通过对训练集进行筛选，学习器可在更稳定、低噪的样本上训练，相比SPLBoost，自采样学习可避免对非噪声困难样本的误判与抛弃，且无需样本学习顺序的先验知识。方法与原有AdaBoost相比，增加了两个采样控制超参数，计算成本(M×O(3kn+3n))与原方法(M×O(3kn+2n))接近。Compared with the original AdaBoost, the sample weight adjustment of SSAdaBoost is given by

Changed to viwi, which effectively limits the weight growth. By screening the training set, the learner can be trained on more stable and low-noise samples. Compared with SPLBoost, self-sampling learning can avoid the misjudgment and abandonment of non-noise difficult samples, and does not require prior knowledge of the sample learning order. . Compared with the original AdaBoost, the method adds two sampling control hyperparameters, and the computational cost (M×O(3kn+3n)) is close to the original method (M×O(3kn+2n)).

With the increase of train running speed, the actual acceleration of the same level may change. The basic rules are: 1) The actual acceleration of the train generated by the traction level decreases with the increase of speed, especially in the high-speed stage; 2) The actual deceleration of the train caused by the coasting stage increases with the speed increase, which is mainly affected by the resistance related to the speed; 3) The actual deceleration of the train caused by the braking stage increases with the speed increase, and the variation range is relatively large. Small. Before the trains of the same model go online, the first car will be adjusted for representative vehicles. On the controllable test line, the actual acceleration of each speed range will be tested, and a complete performance table will be drawn.

Description of drawings

FIG. 1 is a schematic diagram of a data-driven method for extracting train level acceleration according to the present invention.

Figure 2 shows the error rate statistics of the self-sampling boosting learning UCI task model.

Figure 3 shows the experimental results of self-sampling boosting learning on large-scale real tasks.

Figure 4 shows the comparison of the simulation effects of the knowledge extraction performance table.

Detailed ways

In order to better illustrate the purpose and advantages of the present invention, the embodiments of the method of the present invention will be described in further detail below with reference to examples.

The experiments mainly verify the robustness and modeling performance of SSAdaBoost and SSGB self-sampling learning algorithms and the application effect of the model interpretation method based on "query samples" for feature control in the task of train performance extraction.

(1) 70 sets of UCI public data sets with different noise levels and 3 sets of large-scale real task data sets are used to evaluate the learning performance of the proposed SSAdaBoost and SSGB self-sampling improvement. The comparison algorithms include AdaBoost, LogitBoost, GentleBoost, RBoost, CB- Six high-performance boosting learning algorithms, including AdaBoost and GradientBoost, cover the existing advanced robust boosting learning algorithms. The robustness of the proposed algorithm will directly affect the recognition effect of the model on the knowledge of train performance in noisy running data, which is the premise of accurate performance extraction and performance-adaptive recommended speed optimization.

In the experiment, each 5% interval is divided, the noise level is set from 0% to 30%, and 7 sub-data sets are established for each task, so the total number of data sets in the UCI data modeling experiment is 70. Noise is only added to the training set and validation set, and no noise is added to each test set to ensure the reliability of the test results. The 560 sets of experimental results are divided according to tasks, and the curve of the modeling error rate of different algorithms with increasing noise is drawn for easy observation.

The experiment takes the supervised two-classification task as the scene. Due to the label imbalance problem in some task data, the evaluation indicators use Error Rate, Accuracy, F1 Score and are used to test the performance significance between algorithms. The Nemenyi Test (Nemenyi Test), the calculation method is shown in formulas 1 to 4:

In the matrix, TP (True Positive) represents the number of positive examples predicted as positive examples; FN (False Negative) represents the number of positive examples predicted as negative examples; FP (False Positive) represents the number of negative examples predicted as positive examples ; TN (True Negative) represents the number of negative examples predicted as negative examples. According to the above symbol definitions, the calculation method of the evaluation index of the experiment is as follows:

The Nemenyi test method belongs to the post-Hoc Test method, which is used to test the significance of the algorithm performance difference in pairs. represent similar performance. Among them, the critical difference (Critical Difference, CD) is determined by the number of algorithms K and the number of experimental groups N of a single algorithm:

获得，详细的参数设置方法参见实验过程。where the critical value q _α is based on the Studentized Range statistic divided by

obtained, and the detailed parameter setting method can be found in the experimental procedure.

During the experiment, the depth of the weak learner tree is set to 2, the number of weak learners is 50, the hyperparameter of SSAdaBoost is the decay coefficient δ, the adjustment range is [0.9, 1), and the adjustment range of the sampling ratio μ is [0, 0.4] .

(2) Verify the application effect of the proposed model interpretation method based on "query samples" for feature control in the task of train performance extraction. Because the actual acceleration at different levels in each speed range of the train is not easy to measure, it is impossible to obtain the real performance table and compare the extracted performance table, and the "Conversion Method" is used for the experiment. Based on the physical simulation model, the extraction performance table is used as the train stage acceleration parameter to analyze the operation simulation effect. The more similar the simulation results are to the actual operation results, the closer the extraction performance is to the actual performance. At the same time, a simulation model is established by using the train ex-factory performance table, which is used to compare and analyze the accuracy of the extraction performance table. The experimental data is the train operation data of Xi'an Metro Airport and Hefei Metro Line 3.

In the experiment, the "transformation method" is used to analyze the simulation accuracy of the physical model based on the extraction performance table, and the performance extraction effect is equivalently evaluated. The performance extraction effect is affected by the train performance estimation model, and the acceleration estimation model built by SSGB needs to be evaluated. The output of the model is a continuous variable, and the mean squared error (MSE), a typical indicator of regression tasks with high sensitivity, is used for evaluation:

为模型输出的估计标签值，n为测试样本数量。MSE可对模型在测试集上的估计误差进行衡量，结果可代表平均绝对误差、均方根误差等相似指标的评价效果。In the formula, _yi represents the true label value of the ith test sample,

is the estimated label value output by the model, and n is the number of test samples. MSE can measure the estimation error of the model on the test set, and the result can represent the evaluation effect of similar indicators such as mean absolute error and root mean square error.

By comparing the similarity between the simulation curve obtained based on the extraction performance and the real curve, the performance extraction effect is analyzed, and the curve area error rate AUC _err , the end point speed error _verr , the end point position error s _err and other indicators are used:

At the same time, the curve is drawn qualitatively to analyze the similarity of the curve, and the comparison objects include the real speed-position curve, the simulation curve based on the extraction performance table, and the simulation curve based on the factory performance table.

The hardware environment of this experiment is MSI Prestige desktop computer, the CPU model is Intel Core i7-10700K octa-core 16-thread processor, the CPU frequency is 3.8GHz, the physical memory is 32G, the memory frequency is 2400MHz, the graphics card is GeForce RTX 2080SUPER, equipped with 8GB Independent video memory, the operating system is windows 10, 64 bit.

The specific process of this experiment is as follows:

Step 1: A self-sampling learning (SSL) method is proposed for noisy tabular data.

Step 1.1, using the variation pattern of the loss function value during the training process of the boosting learning model, including the absolute value, the variation direction and the variation speed, of the noise-containing samples and the normal samples, to estimate the reliability of the samples and the learning difficulty.

是第i个样本的特征向量，y_i∈{-1,1}是第i个样本的标签。一般的有监督学习过程，将在该训练集上通过优化以下问题构建分类模型：Specifically, taking the binary classification task as an example, suppose the training set contains n samples (x ₁ , y ₁ ),...,(x _n ,y _n ), where

is the feature vector of the ith sample, and y _i ∈ {-1,1} is the label of the ith sample. A general supervised learning process will build a classification model on this training set by optimizing the following problems:

为简化描述，下文将F_m(x_i,Θ)写作F_m。自采样学习关注样本训练效果及训练速度，其形式为：where L(y _i , f( _xi , Θ)) represents the loss function corresponding to the loss algorithm, and the strong learner in boost learning is composed of multiple weak learners, namely

To simplify the description, F _m ( _xi ,Θ) is hereinafter written as F _m . Self-sampling learning focuses on the sample training effect and training speed, and its form is:

Among them, λ is the self-sampling coefficient, which determines the size of the training samples for the next iteration, and α is the balance coefficient, which controls the strength of the two regularities.

Step 1.2, adjust the training sample subset at each iteration, and optimize the sample learning order while eliminating noise points.

Step 2, introduce self-sampling learning into AdaBoost, and establish SSAdaBoost and SSGB methods.

The self-sampling learning is introduced into AdaBoost, and the SSAdaBoost method is established. AdaBoost builds an incremental logistic regression model based on exponential loss, and builds a new weak learner f(x) in iterations by optimizing the following problems:

The SSAdaBoost training target adds self-sampling regularization to constrain the weak learner training sample set:

Among them, α is the balance coefficient, λ is the sampling rate coefficient, and the above formula can be solved by alternately updating F(x) and the self-sampling weight v. The method takes the strong learner as the optimization object, and updating F(x) means training a new weak learner f(x), which is weighted and combined into the strong learner.

变为一个常数值，此时优化问题(11)变为加权指数损失最小化问题：In each iteration, v is first fixed so that the subcomponents of the optimization objective are

becomes a constant value, and the optimization problem (11) becomes a weighted exponential loss minimization problem:

Step 2.1, establish the target train performance estimation model, mine the special car performance knowledge in the noisy running data, and fit the mapping function between acceleration and speed, level sequence (implicit delay characteristic), gradient and other features.

This optimization problem is similar to the original function of AdaBoost (10), and the same solution method can be used to establish an incremental logistic regression model to minimize E(ve ^-yF(x) ) by the quasi-Newton method. Expand the above formula to the second order:

Regarding f(x)∈{-1,1} and c>0, the point-by-point minimization of the above formula can be obtained:

where v _i w _i is the sample weight, and the second-order approximate minimum value of the above formula is still the weighted least squares of f(x _i )∈{-1,1}. The calculation mode in the original algorithm can be maintained to train a new weak learner f based on the sample weight, and the objective function L(F+cf) is convex with respect to c, and the weak learner weight c can be obtained directly by setting its derivative to 0 :

如果弱学习器的判别准确率低于50％，则c可能为负值。in

If the discriminative accuracy of the weak learner is lower than 50%, c may be negative.

Then fix c and f, and solve v in problem (11):

表征的学习效果及以

表征的学习速率来完成样本可靠性估计与选择，提升训练集质量。From the form of the self-sampling weight solution, it can be seen that the proposed method can be

The learning effect of representation and

The learning rate of the representation is used to complete the estimation and selection of sample reliability and improve the quality of the training set.

The self-sampling learning method is introduced into the gradient boosting learning, and the SSGB method is proposed. The characteristic of gradient boosting learning is to minimize the target loss in a similar way to gradient descent. The weak learner training mode and weight are:

The derivation and optimization process of SSGB is very similar to SSAdaBoost, and its hyperparameters can also be determined by the same strategy.

Step 3: Use the "query sample" to control the influence of features such as level sequence, delay, and gradient, and sequentially query the acceleration values corresponding to the lower levels in different speed ranges.

Assuming that the model fits the function mapping between the feature (x _a , x _b , x _c ) and the label y, and it is known that x _b = c _b , x _c = c _c , it has no effect on the label y (for example, x _b represents the correction Slope, c _b = 0, y represents acceleration, the specific relationship between the correction slope and acceleration is unknown, but x _b = 0 has no effect on other features or acceleration), then define (q _a , c _b , c _c ) is a query sample, which can be used to extract the corresponding label y value when x _a =q _a .

Step 3.1: Establish a train-specific performance table to extract the quantitative relationship between target features and labels.

Test results: Based on the real data of Xi'an Airport Line, combined with the automatic driving system and the proposed element to improve the learning and accurate motion simulation, the test performance is adaptive to recommend the speed optimization effect. The results show that SSAdaBoost significantly improves the robustness of the model compared with the existing methods, and can accurately model the delayed and noisy running data, and the performance of the train extracted by the method is similar to the actual one.

The above-mentioned specific descriptions further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above-mentioned descriptions are only specific embodiments of the present invention, and are not intended to limit the protection of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims (6)

1. The data-driven train level acceleration extraction method is characterized by comprising the following steps of:

step 1, aiming at noisy table type data, a self-sampling learning method SSL is provided: estimating the reliability and the learning difficulty of the samples by using the change modes of the loss function values of the noise-containing samples and the normal samples in the training process of the learning model, including the absolute values, the change directions, the change speeds and the like, adjusting the training sample subset in each iteration, and optimizing the learning sequence of the samples while eliminating noise points;

step 2, introducing self-sampling learning into AdaBoost, establishing an SSAdaBoost and SSGB method, establishing a target train performance estimation model, mining special train performance knowledge in noisy operation data, and fitting a mapping function between acceleration and characteristics such as speed, a level sequence (implicit delay characteristic) and gradient;

and 3, controlling the characteristic influences of the level sequence, the time delay, the gradient and the like by using the 'query sample', sequentially querying the corresponding acceleration values of each level in different speed intervals, establishing a train exclusive performance table, and realizing the extraction of the quantitative relation between the target characteristic and the label.

2. The data-driven train-level acceleration extraction method of claim 1, characterized in that: in step 1, the present invention is directed to noisy tabular data.

3. The data-driven train-level acceleration extraction method of claim 1, characterized in that: in step 2, the method for establishing the SSAdaBoost and the SSGB is realized by introducing self-sampling learning into the AdaBoost and gradient boosting learning, the self-sampling learning focuses on the training effect and the training speed of the sample, and the form is as follows:

wherein, lambda is a self-sampling coefficient, the scale of the training sample of the next iteration is determined, alpha is a balance coefficient, and the action intensity of two types of regularization is controlled.

4. The data-driven train-level acceleration extraction method of claim 1, characterized in that: in step 2, adding a self-sampling regular into an SSAdaBoost training target, and constraining a weak learner training sample set:

wherein alpha is a balance coefficient, lambda is a sampling rate coefficient, the above formula can be solved by alternately updating F (x) and self-sampling weight v, the method takes a strong learner as an optimization object, updating F (x) means training a new weak learner f (x) and performing weighted combination into the strong learning, and the input of the method is a training sample set { (x) ₁ ，y ₁ )，…，(x _n ,y _n ) Training iteration times M, sampling proportion mu, balance coefficient alpha being 0.5, decay coefficient delta and outputting as a strong classifier F _M (x)。

5. The data-driven train-level acceleration extraction method of claim 1, characterized in that: in step 3, a self-sampling learning method is introduced into gradient boosting learning, and an SSGB method is provided, wherein the gradient boosting learning is characterized in that target loss is minimized in a gradient descending similar mode, and a weak learner training mode and weight are as follows:

the input of the method is a training sample set { (x) ₁ ，y ₁ )，…，(x _n ，y _n ) The training iteration times M, the sampling proportion mu, the balance coefficient alpha equal to 0.5 and the decay coefficient delta, and the strong classifier F is output _M (x)。

6. The data-driven train-level acceleration extraction method of claim 1, characterized in that: in step 3, the hypothesis model fits the features (x) _a ，x _b ，x _c ) Mapping with a function of tag y, and knowing x _b ＝c _b ，x _c ＝c _c When it is time, it has no effect on tag y (e.g., x) _b Representing the correction gradient, c _b With 0, y representing acceleration, the specific relationship to acceleration when correcting for non-zero slope is unknown, but x _b When 0 has no influence on other characteristics or acceleration), then (q) is defined _a ，c _b ，c _c ) For query samples, it can be used to extract x _a ＝q _a The corresponding tag y value.

Images