Tracked vehicle steering curvature identification method based on machine learning

Abstract

The invention relates to a method for identifying the steering curvature of a tracked vehicle based on machine learning, belongs to the field of vehicle driving control, and solves the problem of difficulty in real-time calculation of the steering curvature in the prior art. The method comprises the following specific steps: acquiring driving data of a tracked vehicle to form a feature vector set; a historical slip rate set is formed by calculating the theoretical slip rates of the two sides of the corresponding track; based on the feature vector set and the historical slip rate set, training an offline two-stage machine learning model to obtain a two-side slip rate estimation model of the tracked vehicle; by means of the online identification process, real-time estimation values of the slip rates of the two sides are obtained through the two-side slip rate estimation model according to vehicle driving data collected in real time; and the real-time vehicle speed and the yaw velocity of the tracked vehicle are calculated based on the real-time estimation value of the slip rate of the two sides, and then real-time estimation of the steering curvature of the tracked vehicle is achieved.

Description

Technical Field

[0001] The present invention relates to the field of vehicle driving control, and in particular to a method for identifying the turning curvature of a tracked vehicle based on machine learning. Background Art

[0002] The estimation of slip rate and steering curvature of tracked vehicles is a complex problem, which is strongly related to the dynamic characteristics of the vehicle itself and the interaction between the vehicle and the ground. The interaction between the vehicle and the ground is difficult to accurately evaluate through model-based methods. In the actual calculation process, the theoretical motion state of the vehicle is often accurately calculated first, and then the actual state transition of the vehicle on the road is measured. Finally, the slip rate and steering curvature are accurately estimated based on the deviation between the two. Among them, when calculating the theoretical motion state of the vehicle, the theoretical state change of the vehicle without considering the slip and slip between the track and the ground can be inferred by measuring the underlying data related to the vehicle's driving. This process is relatively simple and reliable. When measuring the actual state transition of the vehicle on the road, it can be done manually or with the help of the vehicle's GPS sensor information. However, in this process, manual measurement is difficult to cope with high-intensity and long-term vehicle driving processes, and it is difficult to achieve real-time calculation; the accuracy of the vehicle's GPS position and orientation information is seriously affected by signal conditions, and it is difficult to be applied to various driving conditions. Summary of the invention

[0003] In view of the above analysis, an embodiment of the present invention aims to provide a method for identifying the turning curvature of a tracked vehicle based on machine learning, so as to solve the problem of difficulty in real-time calculation of the turning curvature in the prior art.

[0004] The purpose of the present invention is mainly achieved through the following technical solutions:

[0005] A method for identifying the turning curvature of a tracked vehicle based on machine learning comprises the following steps:

[0006] Collecting the driving data of the tracked vehicle to form a characteristic vector set τ1 of the tracked vehicle, wherein the driving data includes the left and right track winding speeds, the left and right track driving torques, the left and right driving wheel speeds, the yaw angular velocity, the longitudinal acceleration, and the lateral acceleration;

[0007] By calculating the theoretical slip rates on both sides of the corresponding crawler, a historical slip rate set τ2 is formed, wherein the historical slip rate set τ2 includes the slip rates of the crawlers on both sides;

[0008] Based on the feature vector set τ1 and the historical slip rate set τ2, an offline two-stage machine learning model is trained to obtain a slip rate estimation model for both sides of the tracked vehicle;

[0009] By using the online identification process, the real-time collected vehicle driving data is passed through the two-side slip rate estimation model to obtain the real-time estimation value of the two-side slip rate;

[0010] Based on the real-time estimated value of the slip rate on both sides and the real-time measured track winding speed, the real-time vehicle speed of the tracked vehicle is obtained; based on the structural parameters of the tracked vehicle and the vehicle speed, the real-time yaw angular velocity is obtained; based on the vehicle speed and the yaw angular velocity, the real-time estimated value of the steering curvature of the tracked vehicle is obtained.

[0011] Furthermore, the two-stage machine learning model includes an offline clustering model and an offline regression model, and the specific steps of obtaining the two-side slip rate estimation model include:

[0012] Based on the feature vector set τ1, a steering action feature sub-vector set τ3, a steering motivation feature sub-vector set τ4 and a steering acceleration feature sub-vector set τ5 are obtained;

[0013] Offline clustering models are established based on the three sub-vector sets, and three cluster labels of steering action degree label ty1, steering motivation degree label ty2 and steering acceleration degree label ty3 are obtained after training;

[0014] The feature vector set τ1 is combined with the three cluster labels to form an input training set, and the historical slip rate set τ2 is used as the corresponding output to train an offline regression model of the slip rates on both sides.

[0015] Furthermore, the expressions of the corresponding sub-vectors in the three sub-vector sets are respectively:

[0016]

[0017]

[0018] in, are the i-th subvectors of the training sets τ3, τ4, and τ5 respectively; v nl and v nr are the left and right track winding speeds respectively; M l and M r are the left and right track driving torques respectively; n l and n r are the speeds of the left and right driving wheels respectively; ω yaw is the yaw angular velocity; a1 is the longitudinal acceleration; a2 is the lateral acceleration; the superscript i represents the i-th sampling result.

[0019] Furthermore, the specific steps of training the offline clustering model include:

[0020] The input data sets of the offline clustering model are respectively formed based on the three sub-vector sets;

[0021] Initialize the particle swarm based on the particle swarm optimization algorithm and calculate the fitness of each particle according to the current position of the particle;

[0022] Particles update themselves by tracking individual and global extrema while recording the positions of said particles;

[0023] The weights and constraints in the particle swarm optimization algorithm are associated with the particle fitness and time, the updated particles are obtained and the fitness is calculated, and the entire particle swarm is iterated to find the best particle position in combination with the population fitness variance;

[0024] The optimal position is used as the initial center, the input data set is divided, and the K-means clustering algorithm is used for training respectively to obtain the steering action degree label ty1, the steering motivation degree label ty2 and the steering acceleration degree label ty3.

[0025] Furthermore, the calculation formula of the fitness of the particle is:

[0026]

[0027] Among them, f i is the fitness of the ith particle; m is the number of clusters set for initialization; C i , C j are the i-th and j-th clusters respectively; avg(C) represents the average distance operation of the clusters;

[0028] represents the center of cluster i; |C i | represents the number of data in cluster i; dis(x, y) represents the Euclidean distance, and

[0029] j is a number from 1 to m that is not equal to i.

[0030] Furthermore, the specific steps of training the offline regression model include:

[0031] Using the three cluster labels in combination with the feature vector set τ1, a training set τ6 is obtained as an input sample of an offline regression model of slip rates on both sides, and the historical slip rate set τ2 is used as the corresponding true value output;

[0032] Using the DBSCAN clustering algorithm to divide the input samples into normal samples and abnormal samples;

[0033] The sample weights are initialized based on the AdaBoost regression model, and the sample weights are updated by using a weight update strategy in combination with the abnormal samples to train the offline regression model.

[0034] Furthermore, the expression of the offline regression model output is:

[0035]

[0036] Among them, s l 、s r are the estimated values ​​of the left and right track slip rates, respectively; They represent the learners for estimating the left and right track slip rates respectively; They represent the weight coefficients of the left and right track slip rate estimation learners respectively; m represents the number of iterations.

[0037] Furthermore, the subset expression of the training set τ6 is:

[0038]

[0039] in, is the i-th subset of the training set τ6; v nl and v nr are the left and right track winding speeds respectively; M l and M r are the left and right track driving torques respectively; n l and n r are the speeds of the left and right driving wheels respectively; ω yaw is the yaw angular velocity; a1 is the longitudinal acceleration; a2 is the lateral acceleration; the superscript i represents the i-th sampling result.

[0040] Furthermore, the specific steps of the online identification process include:

[0041] Extracting real-time feature vectors using real-time collected tracked vehicle driving data;

[0042] Input the real-time feature vector into the trained offline clustering model for clustering, assign labels to the data points of the clustering results, and combine with the feature vector to form an input vector of the offline regression model;

[0043] The input vector is input into the trained offline regression model to obtain a real-time estimated value of the slip rate on both sides of the tracked vehicle.

[0044] Furthermore, a real-time estimated value of the steering curvature is obtained based on the following formula:

[0045]

[0046] Among them, ρ i is the turning curvature of the i-th sampling result; ω yaw is the yaw angular velocity; v i is the vehicle speed of the i-th sampling result; v nl and vnr Respectively, the left and right track winding speeds; s l 、s r are the estimated values ​​of the left and right track slip rates respectively; B is the center distance of the tracks on both sides; the superscript i represents the i-th sampling result.

[0047] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0048] 1. The present invention proposes to obtain the vehicle speed and yaw rate of the tracked vehicle by accurately estimating the real-time value of the slip rate on both sides of the tracked vehicle, combining the structural parameters of the tracked vehicle and the measured track winding speed, and based on the vehicle speed and yaw rate, to achieve accurate real-time estimation of the steering curvature of the tracked vehicle, thereby solving the problem of difficulty in real-time calculation of the steering curvature.

[0049] 2. A machine learning model based only on the vehicle's underlying driving data and inertial navigation output acceleration is used to obtain the mapping relationship between the vehicle's driving data and the slip rate. The vehicle's real-time driving parameters are input into the model through a linear estimation process to achieve real-time estimation of the slip rate. This method does not require the use of high-precision GPS data and can accurately estimate the slip rate under any signal conditions, thereby improving the applicability and accuracy of estimation under various working conditions.

[0050] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0052] Figure 1 is a flow chart of an embodiment of the present invention;

[0053] Figure 2 A schematic diagram of a design of a turning curvature recognition method according to an embodiment of the present invention;

[0054] Figure 3 Schematic diagram of an offline clustering model according to an embodiment of the present invention;

[0055] Figure 4 This is a schematic diagram of clustering results according to an embodiment of the present invention;

[0056] Figure 5 Schematic diagram of an offline regression model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0058] A specific embodiment of the present invention discloses a method for identifying the turning curvature of a tracked vehicle based on machine learning, such as Figure 1 As shown, the following steps are included:

[0059] Step S1, collecting driving data of the tracked vehicle to form a characteristic vector set τ1 of the tracked vehicle, wherein the driving data includes left and right track winding speeds, left and right track driving torques, left and right driving wheel speeds, yaw angular velocity, longitudinal acceleration, and lateral acceleration;

[0060] Step S2, by calculating the theoretical slip rates on both sides of the corresponding crawler, a historical slip rate set τ2 is formed;

[0061] Step S3, based on the feature vector set τ1 and the historical slip rate set τ2, training an offline two-stage machine learning model to obtain a slip rate estimation model for both sides of the tracked vehicle;

[0062] Step S4, using an online identification process, the real-time collected vehicle driving data is passed through the two-side slip rate estimation model to obtain a real-time estimation value of the two-side slip rate;

[0063] Step S5, obtaining the real-time vehicle speed of the tracked vehicle based on the real-time estimated value of the slip rate on both sides and the real-time measured track winding speed; obtaining the real-time yaw angular velocity based on the structural parameters of the tracked vehicle and the vehicle speed; obtaining the real-time estimated value of the steering curvature of the tracked vehicle based on the vehicle speed and the yaw angular velocity.

[0064] Through the above method, the real-time value of the slip rate on both sides of the tracked vehicle is estimated. Combined with the structural parameters of the tracked vehicle and the measured track winding speed, the vehicle speed and yaw angular velocity of the tracked vehicle are obtained. Based on the vehicle speed and yaw angular velocity, the real-time estimation of the steering curvature of the tracked vehicle is finally achieved, thereby solving the problem of difficulty in real-time calculation of the steering curvature.

[0065] Exemplarily, in step S1, various sensors (including inertial sensors, wheel speed sensors, etc.) and inertial navigation are installed at corresponding positions of the vehicle, and data such as the wheel speed of the vehicle is collected during the driving process of the vehicle to form a feature vector set τ1 of the vehicle, where the expression of the i-th feature vector is:

[0066]

[0067] The superscript i indicates the i-th sampling result, which reflects the sampling situation at a certain time point under a certain type of road condition; the collected data includes the left and right track winding speed v nl and v nr , left and right track driving torque M l and M r , left and right driving wheel speed n l and n r , yaw angular velocity ω yaw , longitudinal acceleration a1 and lateral acceleration a2.

[0068] It should be noted that in order to ensure that the final curvature is closer to the curvature of the actual road, try to keep the vehicle on the center line of the road when driving, and select areas with stable and good GPS signals to collect multi-dimensional driving data of vehicles on different road surfaces.

[0069] It should be noted that the calculation of the theoretical slip rate on both sides of the track in step S2 may be related to factors such as the track traction force, the ground shear deformation coefficient, the track ground contact length, etc., resulting in different calculation formulas.

[0070] For example, the theoretical slip rate of the tracks on both sides is calculated by the following formula:

[0071] Where K is the ground shear deformation coefficient; L is the ground contact length of the track; M is the track traction; R is the driving wheel radius; S c is a constant related to ground conditions; G is the vehicle weight; is the ground friction angle; during the slip rate calculation process on the left and right sides, the values ​​of L, M and R may be different.

[0072] The subset expression of the historical slip rate set τ2 is:

[0073] Furthermore, the slip rate estimation model on both sides in step S3 is as follows: Figure 2As shown in the figure, it includes two offline training and one online identification process. In the offline training part, the vehicle driving parameters are first measured by the vehicle sensors and inertial navigation to form a historical feature vector set. According to different feature attributes, the feature vector set is divided into three sub-vector sets; then the three sub-vector sets are clustered using the improved particle swarm optimization K-means algorithm, and labels are assigned according to their respective clustering results to form a new data set with the original feature vector set; using this data set, the relationship between the vehicle driving feature vector and the three steering degree labels and the slip rate is found by training the regression model; in the online identification part, the feature vector is formed by combining the various data collected by the vehicle sensors and inertial navigation during the vehicle driving process through the trained offline clustering model and offline regression model, and the current slip rate on both sides of the track is estimated online in real time, and the real-time estimated value of the vehicle steering curvature (i.e., road curvature) is obtained by combining the calculated relationship between the vehicle speed and the slip rate of the track and the winding speed of the tracks on both sides.

[0074] Exemplarily, the specific implementation steps of the two-side slip rate estimation model include:

[0075] Step S31, according to different characteristic attributes, the feature vector set τ1 is processed and divided into a steering action feature sub-vector set τ3, a steering motivation feature sub-vector set τ4 and a steering acceleration feature sub-vector set τ5;

[0076] Specifically, the left and right track winding speeds and the left and right driving wheel speeds are subtracted and combined with the vehicle yaw angular velocity to form a characteristic sub-vector set τ3 that represents the vehicle steering action; the left and right track driving torques are subtracted to form a characteristic sub-vector set τ4 that represents the vehicle steering motivation; the longitudinal and lateral accelerations of the vehicle are extracted from the vehicle characteristic parameter vector set τ1 to form a steering acceleration characteristic sub-vector set τ5, which represents the steering acceleration characteristics of the vehicle steering process. The corresponding sub-vector expressions are as follows:

[0077]

[0078] Step S32: establishing offline clustering models based on the three sub-vector sets respectively, and obtaining three clustering labels through training;

[0079] Exemplarily, when a car is driving, taking into account the impact of different steering attributes on the steering performance of the entire vehicle, as well as the correlation coupling relationship between the steering attributes and the slip rate, during offline clustering processing, the three sub-vector sets are clustered separately using the improved particle swarm optimization K-means algorithm to obtain relevant classifications and labels of steering action, steering motivation, and steering acceleration.

[0080] Specifically, the global search capability of the improved particle swarm optimization algorithm (PSO) is used to find a good particle swarm as the center of the K-means algorithm, and then the K-means algorithm is used to search the center to obtain the best clustering result. Figure 3 As shown, the specific steps include:

[0081] Step S321, establishing input data sets of offline clustering models based on the three sub-vector sets respectively;

[0082] For example, the input vector set α=[α1, α2...α k ] is used as the model input data set, α represents the steering action feature sub-vector set τ3, the steering motivation feature sub-vector set τ4 or the steering acceleration feature sub-vector set τ5, and the corresponding sub-vectors are expressed as:

[0083]

[0084] Step S322, initializing the particle swarm and calculating the fitness of each particle according to the current position of the particle;

[0085] For example, n particles are randomly generated, each particle is an m*k dimensional vector, where m is the number of clusters set for initialization, and k is the dimension of each data in the data set. A random data point is selected as the center, and each particle divides the data set of the input model according to the minimum distance principle based on the selected center. The calculation formula for the current fitness of each particle is:

[0086]

[0087] Among them, f i is the fitness of the ith particle; C i is the i-th cluster; avg(C) represents the average distance operation of the cluster; represents the center of cluster i; |C| represents the number of data in the current cluster; j is a number from 1 to m that is not equal to i; dis(x, y) represents the Euclidean distance, and

[0088]

[0089] Step S323: The particle updates itself by tracking the individual extreme value and the global extreme value, and finds the corresponding particle position at the same time.

[0090] Specifically, in each iteration, if the current fitness value is better than the individual extreme value or the global extreme value, the fitness value of the particle is updated to the extreme value and the corresponding position is recorded, otherwise it remains unchanged. The calculation formula for iterative update of the entire particle swarm is:

[0091] v i (t) = ω i*v i (t-1)+C1*r1*(P i (t-1)-X i (t-1))+C2*r2*(g(t-1)-X i (t-1)),

[0092] X i (t) = X i (t-1)+v i (t),

[0093] Among them, v i (t) represents the speed of particle i in the t-th iteration process, which indicates the direction and distance of the particle in the solution space and determines how the particle moves in the solution space to find a better solution; X i (t) represents the position of particle i in the t-th iteration process, which directly corresponds to a potential solution in the data set; P i (t) and g(t) represent the individual extreme position and global optimal position of particle i in the t-th iteration process respectively; ω i is the inertia weight of particle i; C1, C2 are the constraint factors of the particle; r1, r2 are random numbers uniformly distributed between 0 and 1.

[0094] Step S324, associating the weights and constraint factors in the PSO algorithm with the particle fitness and time, obtaining updated particles and calculating the fitness, and iterating the entire particle swarm in combination with the population fitness variance to find the optimal particle position;

[0095] Exemplary, the calculation formulas for the improved weights and constraint factors are:

[0096]

[0097] Among them, g B es t is the global extreme value; tmax is the maximum time allowed, which is related to the maximum number of iterations; time is the current time.

[0098] A new iterative update formula is formed based on the improved weights and constraint factors. The updated particles are used as cluster centers to re-cluster the data set and recalculate the fitness of each particle. The current population fitness variance θ is calculated using the following formula:

[0099]

[0100] Among them, f avg is the fitness f i The average value of .

[0101] The variance θ is higher than the set threshold or the maximum number of iterations N is not reachedmax , then readjust the weights and constraint factors, and repeat the iterative process until the optimal position is obtained using the particle swarm algorithm.

[0102] Through the above improvements to PSO, weights and constraint factors can be adjusted dynamically, which can promote the convergence and search efficiency of the algorithm and thus find the optimal solution more quickly.

[0103] Step S325: Use the obtained optimal position as the initial center, re-divide the input data set, and use the K-means clustering algorithm to train them respectively to obtain three related classification clusters. By assigning cluster labels to the classification clusters, the different degrees of the three attributes are characterized to obtain the steering action degree label ty1, the steering motivation degree label ty2 and the steering acceleration degree label ty3, as shown in FIG. Figure 4 As shown in Figure 2. Sort the data in the cluster by the mean value, and assign labels to each cluster and the data points in it.

[0104] Specifically, the best position is used as the initial center to establish the category matrix M n*k , the data set is re-divided and clustered according to the nearest distance principle of the K-means algorithm, and the category matrix is ​​updated based on the following formula:

[0105]

[0106] According to the division results, the average value of each cluster is calculated as the new cluster center. The calculation formula is:

[0107]

[0108] The iteration is considered to be completed when the category matrix does not change or the maximum number of iterations is exceeded.

[0109] Step S33, the vehicle's feature vector set τ1 is combined with three cluster labels to form a training set τ6 of the offline regression model of the slip rate on both sides, and the historical slip rate set τ2 is used as the corresponding true value output to train and obtain the offline regression model of the slip rate on both sides.

[0110] Exemplarily, assuming that the training set τ6 contains N training subsets, the expression of the i-th training subset is:

[0111]

[0112] Furthermore, the AdaBoost regression model improved based on the DBSCAN algorithm is used for training to find the relationship between the vehicle driving feature vector and the three-category degree labels and the slip rate, and the estimated results of the slip rates on both sides are obtained, such as Figure 5 shown.

[0113] It should be noted that the track slip rates on the left and right sides of the tracked vehicle are different, so the offline regression models are trained separately. For example, taking the left side as an example, the specific steps include:

[0114] Step S331, using three cluster labels combined with the vehicle's feature vector set τ1 to obtain a training set τ6, using the data of the training set τ6 as input samples of an offline regression model of the slip rates on both sides, and using the historical slip rate set τ2 as the true value output;

[0115] For example, the feature vector and common vector dimension on the left side of the slip rate regression training set τ6 are extracted to form a feature data set τ l , the expression is:

[0116] τ l =[ω yaw ,v nl ,M l ,n l ,a1,a2,ty1,ty2,ty3],

[0117] The subscript l represents the relevant feature vector of the left track.

[0118] Step S332: using the DBSCAN clustering algorithm to divide all the input samples into normal samples and abnormal samples;

[0119] For example, the abnormal samples are identified according to the DBSCAN clustering algorithm, that is, by examining the data set τ l Any point in Search by the minimum number you set All density-reachable objects in the ε neighborhood of About the ε neighborhood and the minimum number of classes, and at the next iteration, examine other points in the ε neighborhood for expansion until a complete class is found. The number of points contained in the ε neighborhood of is less than the minimum number, that is, is the boundary point, It will be temporarily marked as an outlier. After the iteration is completed, the remaining points that do not belong to any class are outliers, which are marked, and all samples are divided into normal samples and abnormal samples. Among them, ε is a manually set parameter.

[0120] Step S333: Initialize sample weights based on the AdaBoost regression model, use the weight update strategy in combination with the abnormal samples to update the weights, and train to obtain an offline regression model.

[0121] For example, let the initial sample weight be Set the sampling generalization of the i-th data sample to And extract n (n is not greater than N) samples from the training set with replacement, and train the base learner G according to the training data m Get the regression model, base learner G m It can be a decision tree, Bayesian classifier, etc.

[0122] For example, a decision tree is selected to learn the regression model That is, traverse all the eigenvalues ​​of all the features, divide the data set once for each eigenvalue, then calculate the sum of the total variance of the two subsets after the division, find the division eigenvalue that can minimize the sum of the total variance, and then divide it again to get two trees, and recursively perform the same operation on the two trees. Then, the maximum deviation of the learner's prediction value on the training sample is calculated based on the following formula

[0123]

[0124] Calculate the relative error for each training sample The formula is:

[0125]

[0126] Calculate the loss value for each training sample The formula is:

[0127]

[0128] Calculate the error value of the learner The formula is:

[0129]

[0130] It should be noted that if e m >1 / 2, the effect of the learner is worse than random guessing. In this case, delete the learner and iterate again. Therefore, the correct iteration condition requires 0 <e m <1 / 2.

[0131] Calculate the weight coefficient α of the learner m The formula is:

[0132]

[0133] Specifically, according to the weight update strategy formula, combined with the abnormal samples identified by the DBSCAN clustering algorithm, the weight of the sample is updated:

[0134]

[0135] Where m represents the number of iterations; ω m,i is the weight of the i-th sample at the m-th iteration; lossm,i is the loss value of the i-th sample at the m-th iteration; e m is the prediction error rate of the learner; RE i is the relative error of abnormal samples; is the input sample data for the left slip rate estimation; the superscript 1 represents the relevant parameters for the left slip rate estimation.

[0136] After completing the number of iterations m, the weighted average of the prediction results of the m learners is calculated as the output. After the training of both sides is completed, the final strong regression function is obtained through the output of the offline regression model:

[0137]

[0138] Based on this, the final offline regression model is obtained through training, which can achieve a more accurate estimation of the vehicle slip rate by combining the characteristic parameters of the tracked vehicle during driving with the degree of slip obtained by classification.

[0139] In the online identification process of step S4, the real-time collected tracked vehicle driving data is used for feature extraction to form a feature vector; the feature vector is used as input, and the data is clustered through the offline clustering model of the slip rate on both sides obtained by training, and labels are assigned to the data points. The feature vector is combined to form an offline regression data set τ of the slip rate. on , where the subset expression is:

[0140]

[0141] The data set τ on An offline regression model of the slip rates on both sides is input, and the real-time slip rates on both sides of the tracked vehicle are estimated and output by the offline regression model of the slip rates on both sides.

[0142] It should be noted that after the online identification process is completed, in order to improve the accuracy of the model, the estimated values ​​of the track slip rates on both sides obtained by the model are recorded. At the same time, the theoretical true slip rate of the tracks on both sides is calculated, and the estimated value is compared with the true value to obtain the relative error data set γ. When the mean and variance of the error set are less than the set threshold, the model is considered accurate. Otherwise, the model is retrained by calculating the true slip rate until it is accurate.

[0143] Furthermore, in step S5, the estimated value of the slip rate is used to estimate the vehicle speed and yaw rate through the vehicle kinematics relationship, and the calculation formula is:

[0144]

[0145] In the formula, are the estimated values ​​of the track slip rates on both sides. When the slip rate is greater than 0, the track slips; when the slip rate is less than 0, the track slips. nl 、v nr are the directly measured winding speeds of the left and right tracks, B is the center distance of the tracks on both sides, v is the estimated result of the vehicle speed, ω yaw is the yaw angular velocity.

[0146] In addition, the turning curvature is the inverse of the turning radius. The real-time turning curvature of the current tracked vehicle can be estimated by the estimated vehicle speed and the actual measured yaw rate. The calculation formula is:

[0147]

[0148] Among them, ρ i is the turning curvature of the i-th sampling result.

[0149] It should be noted that during the movement of a tracked vehicle, a large amount of slip and sliding motion will occur between the track and the ground, which makes the tracked vehicle dynamics model containing a large number of unstructured road parameters very complex and not suitable for research with high real-time requirements. Therefore, it is difficult to accurately estimate the vehicle speed. However, through the above method, the real-time slip rate of the tracks on both sides can be estimated, and the vehicle speed and yaw angular velocity can be obtained using the structural parameters of the tracked vehicle and the measured track winding speed, and then the real-time steering curvature can be identified.

[0150] Compared with the prior art, the present embodiment provides a method for identifying the steering curvature of a tracked vehicle based on machine learning. By estimating the real-time value of the slip rate on both sides of the tracked vehicle, the vehicle speed and yaw rate of the tracked vehicle are obtained by combining the structural parameters of the tracked vehicle and the measured track winding speed. Based on the vehicle speed and yaw rate, the accurate real-time estimation of the steering curvature of the tracked vehicle is achieved, thereby solving the problem of the difficulty in real-time calculation of the steering curvature. On the one hand, the applicability and accuracy of the slip rate estimation method under various working conditions are guaranteed. A machine learning model based only on the underlying driving data of the vehicle and the output acceleration of the inertial navigation is adopted. The mapping relationship between the vehicle driving data and the slip rate is obtained by constructing a two-stage machine learning model to achieve accurate estimation. On the other hand, in the online estimation process, the slip rate can be estimated by simply inputting the real-time driving parameters of the vehicle into the model. This process does not require the use of high-precision GPS data, and can achieve accurate estimation of the slip rate under any conditions.

[0151] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0152] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for identifying the turning curvature of tracked vehicles based on machine learning, It is characterized in that The steps include: Collect the driving data of the tracked vehicle to form the feature vector set τ of the tracked vehicle driving 1 , the driving data includes left and right track winding speeds, left and right track driving torques, left and right driving wheel speeds, yaw angular velocity, longitudinal acceleration, and lateral acceleration; By calculating the theoretical slip rates on both sides of the corresponding track, the historical slip rate set τ is constructed. 2 , where the historical slip rate set τ 2 Including the slip rate of the tracks on both sides; Based on the feature vector set τ 1 and the historical slip rate set τ 2 , train an offline two-stage machine learning model to obtain a two-side slip rate estimation model for tracked vehicles; By using the online identification process, the real-time collected vehicle driving data is passed through the two-side slip rate estimation model to obtain the real-time estimation value of the two-side slip rate; Based on the real-time estimated value of the slip rate on both sides and the real-time measured track winding speed, the real-time vehicle speed of the tracked vehicle is obtained; based on the structural parameters of the tracked vehicle and the vehicle speed, the real-time yaw angular velocity is obtained; based on the vehicle speed and the yaw angular velocity, the real-time estimated value of the steering curvature of the tracked vehicle is obtained.

2. The method for identifying the turning curvature of a tracked vehicle based on machine learning according to claim 1, It is characterized in that The two-stage machine learning model includes an offline clustering model and an offline regression model. The specific steps of obtaining the two-side slip rate estimation model include: Based on the feature vector set τ 1 Get the turning action feature sub-vector set τ 3 , steering motivation feature sub-vector set τ 4 and the steering acceleration characteristic subvector set τ 5 ; Based on the three sub-vector sets, offline clustering models are established respectively, and the steering action degree label ty is obtained after training. 1 , Turn to motivation level labelty 2 and steering acceleration degree labelty 3 Three cluster labels of; The feature vector set τ 1 The three cluster labels are combined to form an input training set, and the historical slip rate set τ 2 As the corresponding output, an offline regression model of the slip rate on both sides is trained.

3. The method for identifying the turning curvature of a tracked vehicle based on machine learning according to claim 2, It is characterized in that The expressions of the corresponding sub-vectors in the three sub-vector sets are respectively: in, The training set τ 3 , τ 4 , τ 5 The i-th subvector of n1 and v nr are the left and right track winding speeds respectively; M l and M r are the left and right track driving torques respectively; n l and n r are the speeds of the left and right driving wheels respectively; ω yaw is the yaw angular velocity; a 1 is the longitudinal acceleration; a 2 is the lateral acceleration; the superscript i represents the i-th sampling result.

4. The method for identifying the turning curvature of a tracked vehicle based on machine learning according to claim 2, It is characterized in that The specific steps of training the offline clustering model include: The input data sets of the offline clustering model are respectively formed based on the three sub-vector sets; Initialize the particle swarm based on the particle swarm optimization algorithm and calculate the fitness of each particle according to the current position of the particle; Particles update themselves by tracking individual and global extrema while recording the positions of said particles; The weights and constraints in the particle swarm optimization algorithm are associated with the particle fitness and time, the updated particles are obtained and the fitness is calculated, and the entire particle swarm is iterated to find the best particle position in combination with the population fitness variance; The optimal position is used as the initial center, the input data set is divided, and the K-means clustering algorithm is used to train them respectively to obtain the steering action degree label ty 1 , the steering motivation degree label ty 2 and the steering acceleration degree label ty 3 .

5. The method for identifying the turning curvature of a tracked vehicle based on machine learning according to claim 4, It is characterized in that The calculation formula of the fitness of the particle is: Among them, f i is the fitness of the ith particle; m is the number of clusters set for initialization; C i , C j are the i-th and j-th clusters respectively; avg(C) represents the average distance operation of the clusters; represents the center of cluster i; |C i | represents the number of data in cluster i; dis(x, y) represents the Euclidean distance, and j is a number from 1 to m that is not equal to i.

6. The method for identifying the turning curvature of a tracked vehicle based on machine learning according to claim 2, It is characterized in that The specific steps of training the offline regression model include: Using the three cluster labels combined with the feature vector set τ 1 , get the training set τ 6 , as the input sample of the offline regression model of the slip rate on both sides, the historical slip rate set τ 2 As the corresponding true value output; Using the DBSCAN clustering algorithm to divide the input samples into normal samples and abnormal samples; The sample weights are initialized based on the AdaBoost regression model, and the sample weights are updated by using a weight update strategy in combination with the abnormal samples to train the offline regression model.

7. The method for identifying turning curvature of a tracked vehicle based on machine learning according to claim 6, It is characterized in that The expression of the offline regression model output is: Among them, s l 、s r are the estimated values ​​of the left and right track slip rates, respectively; They represent the learners for estimating the left and right track slip rates respectively; They represent the weight coefficients of the left and right track slip rate estimation learners respectively; m represents the number of iterations.

8. The method for identifying the turning curvature of a tracked vehicle based on machine learning according to claim 6, It is characterized in that The training set τ 6 The subset expression of is: in, is the training set τ 6 The i-th subset of nl and v nr are the left and right track winding speeds respectively; M l and M r are the left and right track driving torques respectively; n l and n r are the speeds of the left and right driving wheels respectively; ω yaw is the yaw angular velocity; a 1 is the longitudinal acceleration; a 2 is the lateral acceleration; the superscript i represents the i-th sampling result.

9. The method for identifying turning curvature of a tracked vehicle based on machine learning according to claim 2, It is characterized in that The specific steps of the online identification process include: Extracting real-time feature vectors using real-time collected tracked vehicle driving data; Input the real-time feature vector into the trained offline clustering model for clustering, assign labels to the data points of the clustering results, and combine with the feature vector to form an input vector of the offline regression model; The input vector is input into the trained offline regression model to obtain a real-time estimated value of the slip rate on both sides of the tracked vehicle.

10. A method for identifying turning curvature of a tracked vehicle based on machine learning according to any one of claims 1 to 9, It is characterized in that The real-time estimated value of the steering curvature is obtained based on the following formula: Among them, ρ i is the turning curvature of the i-th sampling result; ω yaw is the yaw angular velocity; v i is the vehicle speed of the i-th sampling result; v nl and v nr Respectively, the left and right track winding speeds; s l 、s r are the estimated values ​​of the left and right track slip rates respectively; B is the center distance of the tracks on both sides; the superscript i represents the i-th sampling result.

Images