CN116901981A - Online self-learning Markov vehicle speed prediction method - Google Patents

Abstract

The invention discloses an online self-learning Markov vehicle speed prediction method, which comprises the following steps: (10) constructing an offline database: and selecting standard working conditions, and integrating and forming an off-line database of the vehicle running speed. (20) offline training of a state transition probability matrix: defining the acceleration of the vehicle as a Markov state, determining a prediction range, counting state transition events, and calculating a transition probability matrix. (30) designing a state transition matrix real-time updating algorithm: and deducing a recursive form of the state transition matrix, defining a self-learning factor, and utilizing vehicle acceleration historical information to realize online updating of the state transition matrix. (40) on-line prediction: and acquiring acceleration historical data of the vehicle, updating a transition probability matrix, and predicting the speed of the vehicle. The online self-learning Markov prediction method provided by the invention can improve the accuracy of speed prediction and the calculation timeliness; meanwhile, the method is simple and easy to realize, has strong working condition self-adaption and has good engineering application prospect.

Description

An online self-learning Markov vehicle speed prediction method

Technical field

The invention belongs to the field of automatic driving, and in particular relates to an online self-learning Markov vehicle speed prediction method.

Background technique

With the development of vehicle-road-network technology, ground vehicles can obtain surrounding traffic information, which can be used to design predictive motion control, realize advanced assisted driving functions and even high-level driverless technology. Among them, vehicle speed prediction has become a prerequisite for several functions such as predictive energy management, adaptive cruise and automatic lane changing, and its prediction accuracy directly affects the safety and fuel economy of vehicle control. Therefore, accurate vehicle speed prediction methods are of great significance to achieve safe and efficient intelligent driving.

At present, vehicle speed prediction technology has been widely studied, including polynomial fitting methods, traditional Markov chains and a type of deep learning methods based on neural networks. However, the above methods have problems such as poor prediction accuracy, poor adaptability to working conditions, waste of offline training resources, and low online calculation timeliness. Existing vehicle speed prediction methods cannot take into account high accuracy, adaptability and high real-time performance, and need further improvement.

Contents of the invention

The purpose of the present invention is to provide an online self-learning Markov vehicle speed prediction method that effectively improves prediction accuracy and ensures real-time performance because current speed prediction methods cannot take into account high accuracy, adaptability and high real-time performance.

The technical solution to achieve the purpose of the present invention is:

An online self-learning Markov vehicle speed prediction method, including the following steps:

(10) Offline database construction: Select standard working conditions, integrate the data of the standard working conditions, and build an offline database of vehicle driving speeds.

(20) Offline training of state transition probability matrix: use vehicle acceleration as a state in a Markov event, define its state grid, determine the prediction range, obtain acceleration data according to the offline database, count state transition events, and establish a state transition matrix , calculate and store the transition probability matrix.

(30) Design a real-time update algorithm for the state transfer matrix: derive the recursive form of the state transfer matrix, define the self-learning factor, use the vehicle's historical acceleration transfer information, and combine with the offline state transfer matrix to achieve online update of the state transfer matrix.

(40) Online prediction: Obtain the acceleration history data of the vehicle at the current moment, identify the state transition event with reference to the acceleration state grid, update the transition probability matrix online, predict the vehicle acceleration, and then obtain the predicted vehicle speed.

Compared with the prior art, the significant advantages of the present invention are:

1. High computational efficiency: Based on the training of the offline state transfer matrix, the online self-learning speed prediction method only needs to replace and update the offline data in real time, without the need for cumbersome online algorithms, which greatly improves the online calculation timeliness, making the Methods can adapt to new environments and changes more quickly.

2. Strong adaptability: The traditional speed prediction method collects a large amount of data for offline training through offline training models, and uses offline data for prediction. This often has the shortcomings of lag and low accuracy, while the online self-learning model of the present invention The vehicle speed prediction method performs real-time learning and prediction on new data based on the comprehensive utilization of traditional Markov chains and vehicle historical data. By comparing with actual observation results, the system can adjust and improve the prediction model. This self-optimizing The ability allows the prediction model to be continuously upgraded to adapt to different driving environments.

Description of the drawings

Figure 1 is a schematic flow chart of an online self-learning Markov vehicle speed prediction method of the present invention.

Figure 2 is a combination diagram of offline working conditions.

Figure 3 is the flow chart of offline training of transition probability matrix.

Figure 4 is a multi-step Markov transition probability matrix diagram.

Figure 4(a) is the acceleration transition probability matrix diagram 1 second after the current time.

Figure 4(b) is an acceleration transfer probability matrix diagram 3 seconds from the current moment.

Figure 4(c) is the acceleration transfer probability matrix diagram 7 seconds after the current time.

Figure 4(d) is an acceleration transfer probability matrix diagram 10 seconds from the current moment.

Figure 5 is a real-time update flow chart of the transition probability matrix.

Figure 6 is the speed prediction trajectory and root mean square error diagram of the traditional Markov model.

Figure 6(a) is the speed prediction trajectory error diagram of the traditional Markov model.

Figure 6(b) is the root mean square error diagram of the speed prediction data of the traditional Markov model.

Figure 7 is the online self-learning speed prediction trajectory and root mean square error diagram.

Figure 7(a) shows the online self-learning speed prediction trajectory error.

Figure 7(b) is the root mean square error diagram of online self-learning speed prediction data.

Table 1 shows the performance comparison between the present invention and the two existing methods.

Detailed ways

The present invention will be described below in conjunction with the accompanying drawings, so that those skilled in the art can better understand the present invention from the description.

Figure 1 shows a flow chart of an online self-learning Markov speed prediction method of the present invention.

In this embodiment, as shown in Figure 1, the steps of an online self-learning Markov speed prediction method include:

(10) Offline database construction: Select standard working conditions, integrate the data of the standard working conditions, and build an offline database of vehicle driving speeds.

In order to more comprehensively cover various typical road sections and make the prediction results more accurate, this embodiment integrates several typical working conditions and obtains offline data from them.

Figure 2 shows the offline working condition combination diagram.

(20) Offline training of state transition probability matrix: use vehicle acceleration as a state in a Markov event, define its state grid, determine the prediction range, obtain acceleration data according to the offline database, count state transition events, and establish a state transition matrix , calculate and store the transition probability matrix.

Figure 3 shows the (20) state transition probability matrix offline training steps including:

(21) State grid definition: Taking vehicle acceleration as the state in the Markov event, that is, the acceleration state space can be defined as X={a ₁ , a ₂ ,..., a _p }, a ₁ and a _p represents the minimum acceleration and maximum acceleration respectively, and p is the size of the state space. Define the acceleration state grid, the acceleration a(-4≤a≤4)m/s ² , and divide the variables with a grid size of 0.1m/s ² .

(22) Determine the prediction time step: take the prediction time range as 10 seconds.

(23) Statistical state transition events: at each prediction time step, traverse the values in each acceleration state grid according to the offline acceleration trajectory. Count the number of samples that each current acceleration value a _i transfers to its corresponding acceleration value a _j at the next moment, recorded as p is the size of the state space, k is the future immediate index, k∈{1, 2,...,10}, and Q is the acceleration curve test length. Count the number of samples in the offline trajectory that are equal to the value of a _i , recorded as/>

(24) Calculation of state transition matrix: number of samples counted according to the state events The state transition matrix under k at a certain time in the future can be calculated as:

The (24) state transition matrix calculation steps include:

(241) State transition matrix establishment: initialize transition probability matrix The size is (Nacc, Nacc, 10), Nacc is the length of the acceleration grid, initialized/> Matrix, used to count the number of occurrences of acceleration a _i at each time step, initialized/> Matrix used to count the number of transitions from one acceleration value to another.

(242) Frequency matrix establishment: According to the number of samples The ratio of the two to the test length Q of the acceleration curve respectively represents the frequency of the two events appearing in the curve, namely:

Create frequency matrix

(243) State transition matrix formula establishment: According to the event frequency The state transition matrix at a certain point in the future can also be calculated as:

(244) Transition probability matrix storage: According to the acceleration a _i , a _j , record the position index i, j in the acceleration state grid, and store the data in the corresponding position of the state transition matrix.

Figure 4 shows the multi-step Markov transition probability matrix diagram initially obtained according to the offline training of the state transition probability matrix in (20). The P _ij shown in the figure represents the acceleration transferred from the current moment a _i to the next moment a The probability of _j .

Figure 4(a) is the acceleration transfer probability matrix diagram 1 second after the current time.

Figure 4(b) is the acceleration transfer probability matrix diagram 3 seconds from the current moment.

Figure 4(c) is the acceleration transfer probability matrix diagram 7 seconds after the current time.

Figure 4(d) is the acceleration transfer probability matrix diagram 10 seconds from the current moment.

In order to better demonstrate that the present invention can adapt to more road conditions than traditional methods, this embodiment chooses to conduct experiments under the US06 working condition that is not integrated into the offline database.

(30) Design a real-time update algorithm for the state transfer matrix: derive the recursive form of the state transfer matrix, define the self-learning factor, use the vehicle's historical acceleration transfer information, and combine with the offline state transfer matrix to achieve online update of the state transfer matrix.

Figure 5 shows the (30) design state transition matrix real-time update algorithm steps including:

(31) Define the self-learning factor: according to the event frequency Deriving its recursive form, we can get:

in Represents the number of samples in the offline data in which the acceleration is transferred from a _i to the acceleration value a _j after k time, The same can be said/>

According to the recursive forms of the two equations, the state transition matrix at a certain moment can be calculated as:

If there is no corresponding acceleration value a _i in the offline acceleration trajectory, then Right now:

Among them, eta is the introduced Markov learning coefficient, which can be freely adjusted in practical applications to achieve better prediction results.

(32) State transition matrix online update: for each time step, obtain the latest series of acceleration data, and determine the position index in the acceleration state grid based on the current acceleration value and the acceleration values of the previous k time steps. j and i, then update the corresponding frequency matrix, and finally calculate the transition probability matrix.

(40) Online prediction: Obtain the acceleration history data of the vehicle at the current moment, identify the state transition event with reference to the acceleration state grid, update the transition probability matrix online, predict the vehicle acceleration, and then obtain the predicted vehicle speed.

The (40) online prediction step includes:

(41) Acceleration value and velocity value acquisition: Obtain the velocity value and acceleration value at the current moment, and find the closest index a _i in the acceleration grid by approximating the current acceleration value.

(42) Initialize acceleration and velocity prediction vectors: Create two zero matrices to store acceleration and velocity predictions for k time steps in the future.

(43) Prediction result calculation: According to the transition probability matrix, obtain the transition probability vector corresponding to the i-th row and k-th column, and use this probability vector Multiply and sum the elements of the acceleration grid one by one to obtain the acceleration prediction value/> Right now:

Then by adding the first k predicted acceleration values and adding the current speed value v, the predicted speed value in the next k time steps is obtained. Right now:

If any of the speed predictions are less than 0, set them to 0 to ensure that the speed values are non-negative.

(44) Prediction error calculation: Calculate the root mean square error between the predicted speed value and the true speed value, and store it at the corresponding moment in the improved prediction error vector. Plot the prediction results.

In order to clearly quantify the speed prediction error, the root mean square error (RMSE) is introduced as a performance evaluation index: the smaller the root mean square error, the higher the prediction accuracy.

Figure 6 shows the speed prediction trajectory and root mean square error plot of the traditional Markov model.

Figure 6(a) shows the speed prediction trajectory error diagram of the traditional Markov model.

Figure 6(b) shows the root mean square error plot of the speed prediction data of the traditional Markov model.

Figure 7 shows the speed prediction trajectory and root mean square error diagram of the online self-learning type.

Figure 7(a) shows the online self-learning speed prediction trajectory error.

Figure 7(b) shows the root mean square error diagram of the online self-learning speed prediction data.

The red lines in Figures 6(a) and 7(a) represent the trajectory of the predicted speed at each moment using this method, and the thick lines represent the actual vehicle speed curve. It can be seen from the figures that the speed predicted by the speed prediction method of the present invention is The drawn speed prediction trajectory is more consistent with the actual vehicle speed trajectory. The green shaded area in Figures 6(b) and 7(b) represents the root mean square error at each moment. It can be seen from the figure that here In this embodiment, the present invention has higher accuracy than the traditional Markov model speed prediction.

In order to further verify the advantages of the present invention, in this embodiment, a more advanced long short-term memory network (LSTM, Long Short-Term Memory) is also used to compare with the prediction results of the present invention. Table 1 is a comparison table of the results of the three methods:

Table 1

Table 1 is a comparison list between the present invention and the traditional Markov and long-short-term memory network speed prediction methods in terms of root mean square error and calculation time. From Table 1, we can derive the mean mean square error using the traditional Markov prediction method. The root mean square error is 2.16m/s. The root mean square error of the long short-term memory network and the self-learning speed prediction method are 1.86m/s and 1.82m/s respectively. The prediction accuracy is higher than the traditional Markov speed prediction method respectively. The increases are 13.9% and 15.7%, indicating that the prediction accuracy of the present invention is better than the two existing prediction technologies. From this, we can also conclude that when the accuracy is higher than the traditional Markov prediction method, the calculation time of the present invention is 0.6ms. The calculation time of the long short-term memory network prediction method requires 24 ms, and the calculation time of the prediction results of the present invention is much shorter than that of the long short-term memory network prediction method, indicating that the calculation efficiency of the present invention is high. embodies the advantages of the present invention.

Claims (6)

1. An online self-learning Markov vehicle speed prediction method is characterized by comprising the following steps:

(10) And (3) offline database construction: and selecting a standard working condition, integrating the data of the standard working condition, and constructing a vehicle running speed off-line database.

(20) Offline training of state transition probability matrix: and taking the acceleration of the vehicle as a state in the Markov event, defining a state grid of the vehicle, determining a prediction range, acquiring acceleration data according to the offline database, counting a state transition event, establishing a state transition matrix, and calculating and storing a transition probability matrix.

(30) Designing a state transition matrix real-time updating algorithm: and deducing a recursive form of the state transition matrix, defining a self-learning factor, and utilizing vehicle historical acceleration transition information and combining the offline state transition matrix to realize online updating of the state transition matrix.

(40) On-line prediction: and acquiring acceleration historical data of the vehicle at the current moment, identifying a state transition event by referring to the acceleration state grid, updating a transition probability matrix on line, predicting the acceleration of the vehicle, and further acquiring the predicted speed of the vehicle.

2. The speed prediction method according to claim 1, wherein the (10) offline database construction step includes:

(11) Standard working condition integration: and integrating experimental data of several working conditions according to the urban working condition FUDS, the comprehensive working condition NEDC, the high-speed working condition HWFET, the urban working condition NYCC and the comprehensive working condition WLTP.

(12) Offline data acquisition: and according to the integration of the experimental data of the working conditions, acquiring a speed-time curve and an acceleration-time curve.

3. The speed prediction method according to claim 1, wherein the (20) state transition probability matrix offline training step comprises:

(21) State grid definition: the acceleration of the vehicle is used as the state in the Markov event, an acceleration state grid is defined, and the acceleration (-4 a is less than or equal to 4) m/s ² With a mesh size of 0.1m/s ² The variables are divided.

(22) Prediction range determination: the predicted time range was taken to be 10 seconds.

(23) Counting state transition events: traversing the values in each acceleration state grid under each predicted time step, and counting each current acceleration value a according to the offline acceleration track _i Transition to its corresponding next time acceleration value a _j Is recorded as the number of samples ofi, j e {1, 2....p }, p is the state space size, k is an index of future instants, k e {1,2,.. The. Q is the acceleration curve test length. Statistics of the sum a in the offline track _i The number of samples of equal value, noted +.>

(24) Calculating a state transition matrix: based on the number of samples of state transition event, the transition probability matrix at a future time can be calculated from the number of samplesAnd->The ratio of the two is obtained.

4. A speed prediction method according to claim 3, wherein the (24) state transition matrix calculation step comprises:

(241) And (3) establishing a state transition matrix: initializing state transition matricesInitializing matrix->For counting acceleration a per time step _i The number of occurrences, initialize the matrix->For counting the number of transitions from one acceleration value to another.

(242) And (3) establishing a frequency matrix: according to the number of samplesThe ratio of the two to the test length Q of the acceleration curve respectively represents the frequency of occurrence of two events in the curve, which is marked as +.>Establishing a frequency matrix->

(243) And (3) establishing a state transition matrix formula: according to the event frequencyThe state transition matrix at a future time can be defined by +.>The ratio of the two is obtained.

(244) And (5) storing a transition probability matrix: according to the acceleration a _i 、a _j And recording the position indexes i and j of the acceleration state grid, and storing the data in the corresponding positions of the state transition matrix.

5. The method of speed prediction according to claim 1, wherein the step of (30) designing a state transition matrix real-time update algorithm comprises:

(31) Defining a self-learning factor: according to the state transition matrixDeriving the state transition matrix recursion form, and obtaining:

wherein the method comprises the steps ofRepresenting the current acceleration from a in the offline data _i Acceleration value a after transition to its corresponding k time _j Sample number of>Wherein eta is an introduced Markov learning coefficient, and can be freely adjusted in practical application to realize a better prediction effect.

(32) On-line updating of the state transition matrix: and for each time step, acquiring the latest series of acceleration data, determining a position index in an acceleration state grid according to the current acceleration value and the acceleration value of the next predicted moment, updating a corresponding frequency matrix, and finally calculating a transition probability matrix.

6. The speed prediction method according to claim 1, wherein the (40) online prediction step includes:

(41) Acceleration value and speed value acquisition: and acquiring a speed value and an acceleration value at the current moment, and finding the closest index in the acceleration grid by performing approximate processing on the current acceleration value.

(42) Initializing acceleration and speed prediction vectors: two zero matrices are created for storing the acceleration and velocity predictions, respectively, over a future prediction time horizon.

(43) Calculating a prediction result: according to the state transition matrix, a corresponding transition probability vector is obtained, the probability vector is multiplied with elements of the acceleration grid one by one and summed to obtain an acceleration predicted value, then the velocity predicted value in the next k time steps is obtained by adding the previous predicted acceleration values and adding the current velocity value, and if any velocity predicted value is smaller than 0, the velocity predicted value is set to 0 to ensure that the velocity value is non-negative.

(44) Calculating a prediction error: and drawing a prediction result by solving root mean square error of the predicted speed value and the real speed value and storing the root mean square error in the improved prediction error vector at the corresponding moment.