Cloud control self-driving automobile control method considering automobile cloud communication time delay

Abstract

The invention discloses a cloud control self-driving automobile control method considering automobile cloud communication time delay. By constructing a time delay prediction model based on historical communication delay data, the back-and-forth communication time delay at the current moment is accurately predicted, and forward compensation is carried out on the vehicle state by adopting an iterative prediction formula in combination with a vehicle transverse dynamic model. According to the method, a process disturbance correction item is introduced, system modeling errors caused by environment disturbance are effectively corrected, and the state prediction precision is improved. And deducing a future vehicle state through the predicted time delay step number in combination with a historical vehicle state and control input so as to realize dynamic compensation of the time delay influence. The method effectively solves the problems of control lag and state information lag caused by communication delay in vehicle-cloud cooperative control, improves the control real-time performance and precision of a cloud control automatic driving system, enhances the robustness and safety of the system in a complex environment, and is suitable for automatic driving scenes with high dynamic and high safety requirements.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of cloud control automatic driving car control, and particularly relates to a cloud control automatic driving car control method considering vehicle-cloud communication time delay. BACKGROUND

[0002] The intelligent networked cloud control system is a very complex system, which connects people, vehicles, roads and clouds into one body by using new generation communication technology, and performs sensing, decision making and control. The system is conducive to realizing vehicle-road cooperation, better guaranteeing automobile driving safety, and improving vehicle passing efficiency. For the control system, compared with the vehicle end control system, the cloud end control system is conducive to transferring all control modules from the vehicle end to the cloud end, thereby reducing the cost of the vehicle end control system, and conducive to centralized arrangement of the control system, so that the control effect is better.

[0003] At present, some scholars have carried out research on the intelligent networked cloud control system, but there is no great progress. Most of the current research is carried out in simulation, and lacks real vehicle test, so it cannot be applied. One of the important reasons affecting the actual application is the influence of communication time delay on path tracking control results. This problem is difficult to break through.

[0004] In existing research, the description of time delay is relatively vague, and the characteristics of time delay are not clear. Some teams assume normal distribution as the probability distribution density of time delay size, some scholars assume time delay as a fixed value or a discrete variable, and some scholars assume that the time delay is discretized and set the probability of time delay. The above methods all use various models to simulate the distribution of time delay, without considering the real situation of time delay and without measuring the time delay. Therefore, a new method is needed to process time delay data.

[0005] At the same time, the current research lacks consideration of time delay in controller design. There are many controllers such as PID, LQR and MPC, which have been widely used in different fields. However, the existing controllers are not perfect, and their control performance is poor in harsh environments such as large network time delay, and even loss of control occurs. Therefore, there is an urgent need to design a new controller to achieve better tracking effect.

[0006] In the prior art, Chinese patent CN115598983A discloses a method and device for considering time-varying delay in horizontal and longitudinal cooperative control of unmanned vehicles, which comprises the following steps: step 1, online acquisition of bottom-layer time-varying CAN communication delay estimation value τCAN; step 2, according to a series of reference point information in a predicted time domain Np found, using a prediction model, combining a horizontal and longitudinal integrated evaluation function J and horizontal and longitudinal joint constraints s.t. to solve the optimal control problem rolling, and outputting the control quantity of the vehicle. However, the method has the following disadvantages: first, the method is mainly applicable to the traditional architecture of deploying the controller at the end of the vehicle, and does not involve the design and optimization of the cloud control architecture; second, the delay model is limited to the bottom-layer CAN bus, and cannot cover the multi-stage communication delay under complex network conditions. SUMMARY

[0007] In view of the above disadvantages of the prior art, the present application provides a cloud control automatic driving vehicle control method considering vehicle-cloud communication delay, which solves the problem of poor path tracking control effect of the vehicle under time delay, and achieves the effect of improving the practicality of vehicle cloud control function and user experience.

[0008] The object of the present application can be achieved by the following technical solutions:

[0009] The present application provides a cloud control automatic driving vehicle control method considering vehicle-cloud communication delay, comprising the following steps:

[0010] The vehicle state information at the current time is collected by the automatic driving vehicle and uploaded to the cloud control platform through the communication network;

[0011] The cloud control platform constructs a time delay prediction model based on historical communication delay data to predict the round-trip communication delay at the current time;

[0012] Based on the historical vehicle state information and the historical control input sequence, the vehicle dynamics model is used to compensate the vehicle state information at the current time according to the round-trip communication delay, and the compensated vehicle state information after the round-trip communication delay is obtained;

[0013] The cloud control platform designs a linear quadratic regulator controller based on the compensated vehicle state information after the round-trip communication delay, and calculates the optimal control input with the compensated vehicle state information as the input;

[0014] The optimal control input is sent to the automatic driving vehicle through the network, and is received and executed by the automatic driving vehicle at the future time under the effect of the downlink communication delay, realizing the feedforward compensation and control optimization of the bidirectional communication delay.

[0015] Further, the vehicle state information includes lateral vehicle speed v, yaw angular velocity r and sending time t.

[0016] Further, the historical communication delay data comprises round-trip communication delays between the autonomous vehicle and the cloud control platform at a plurality of historical time points, the round-trip communication delay being a time difference between the autonomous vehicle sending vehicle state information and receiving control input sent by the cloud control platform.

[0017] Further, the autonomous vehicle sends vehicle state information to the cloud control platform at a fixed time interval τ.

[0018] Further, the time delay prediction model is a U-Mixer framework, the U-Mixer framework comprising a normalization processing module, an embedding module, a Unet encoder-decoder structure composed of a plurality of multilayer perceptron mechanisms, a stationarity correction module, and a prediction output module.

[0019] The historical communication delay data is used to construct a time delay prediction model to predict the round-trip communication delay at the current time point, specifically comprising:

[0020] The historical communication delay data is constructed into a time series W in chronological order of time nodes;

[0021] The time series data W is normalized to calculate the normalized data:

[0022]

[0023] Wherein, μ and σ are the mean and standard deviation of the historical communication delay data, respectively.

[0024] The normalized data W is divided into a patch sequence W according to a preset patch length P and a step S; p Wherein, the number of patches is L is the length of the normalized data W .

[0025] The patch sequence W p is input into the Unet encoder-decoder structure composed of a plurality of multilayer perceptron mechanisms to obtain an encoded feature representation;

[0026] The encoded feature representation is corrected for stationarity to generate a corrected feature representation;

[0027] The corrected feature representation is mapped by the prediction output module to output the predicted round-trip communication delay at the current time point:

[0028]

[0029] Wherein, Y ′ is the predicted round-trip communication delay, For the corrected feature representation, W1, W2, b1, b2 are the weight matrix and bias vector of the two full connection layers in the prediction output module respectively, and σ is the activation function.

[0030] According to the prediction, the round-trip communication delay Y at the current time is predicted ′ With the vehicle state sampling interval τ, the step number of the predicted round-trip communication delay is calculated:

[0031]

[0032] Wherein, n is the step number of the predicted round-trip communication delay, Indicates rounding down.

[0033] Further, the training process of the delay prediction model specifically includes:

[0034] Based on the historical communication delay time sequence W={w1, w2,…, w T},w t Indicates the communication delay at time t. The sliding difference method is used to detect the mutation point of the communication delay, and the time sequence is divided into a mutation sample segment and a non-mutation sample segment. The mutation sample segment is a time interval that satisfies |w t -w t-t |>θ, and the non-mutation sample segment is a time interval that satisfies |w t -w t-1 |≤θ, and θ is a mutation threshold value, which is adaptively determined through historical data statistics.

[0035] In the mutation and non-mutation sample segments, sample pairs are constructed to form: query samples W q , positive samples W + and negative samples W - , wherein the query samples W q and the positive samples W + come from any non-mutation sample segment, and the negative samples W - come from any mutation sample segment.

[0036] The training sample pair is input into the U-Mixer model, and the hidden feature representation f(W) of the sample is extracted by the Unet encoder-decoder structure.

[0037] According to the hidden feature representation f(W) of each sample, the mutation-aware contrast loss function is calculated:

[0038]

[0039] Wherein, f(W q ), f(W + ), f(W j -) respectively represent the feature representation of the query sample, the positive sample and the jth negative sample; sim represents the cosine similarity, τ is the temperature coefficient, and L is the mutation-aware contrast loss function;

[0040] The parameters of the time delay prediction model are updated according to the calculated mutation-aware contrast loss function.

[0041] Further, based on the historical vehicle state information and the historical control input sequence, the vehicle dynamics model is used to compensate the vehicle state information at the current time according to the round-trip communication delay, and specifically includes:

[0042] The vehicle state is forward predicted by using the vehicle dynamics discrete model, and the current received vehicle state information is X(k), the historical control input sequence is U(k-1), U(k-2), …, the historical vehicle state information is X(k-1), X(k-2), …, and the predicted round-trip communication delay step is n;

[0043] According to the historical vehicle state information and the historical control input sequence, the process disturbance mean μ caused by environmental disturbance is estimated w As a system modeling error correction term, the specific calculation method is:

[0044] w(j)=X(j+1)-AX(j)-BU(j),j=k-m,…,k-1

[0045]

[0046] Wherein, m is the length of the historical time window for estimating the process disturbance, w(j) is the model prediction error at the jth time, X(j+1) and X(j) are the vehicle state information received by the cloud platform at the j+1th time and the jth time respectively, U(j) is the control input sent by the cloud platform at the jth time, A is the state transition matrix, and B is the control input matrix;

[0047] Taking the current state X(k) as the initial value, the iterative prediction formula based on the vehicle lateral dynamics model is used to gradually forward predict the vehicle state at the future k+n time:

[0048]

[0049] Wherein, U(k+i) is the control input at the k+i time, which is obtained by calculating the U(k+i) linear quadratic regulator controller; based on the iterative prediction formula, the compensation vehicle state information after the round-trip communication delay is obtained

[0050] Further, the current received vehicle state information X(k) when k = 1, the historical vehicle state information X(k-1), X(k-2), … are all set to X(1), and the historical control input sequence U(k-1), U(k-2), … are all set to zero.

[0051] Further, the state transition matrix A is:

[0052]

[0053] Wherein, a, b are the front and rear wheelbase of the car, C f , C r are the front and rear wheel cornering stiffness of the car, m is the weight of the car, u is the longitudinal speed of the car, I is the moment of inertia of the car;

[0054] The control input matrix B is:

[0055]

[0056] The control input U = δ, δ is the front wheel steering angle of the car.

[0057] Further, the compensation vehicle state information based on the round-trip communication delay is designed to calculate the optimal control input by taking the compensation vehicle state information as the input, and the linear quadratic regulator controller is specifically included:

[0058] Obtain the compensation vehicle state information Set the linear quadratic performance index function:

[0059]

[0060] Wherein, J is the linear quadratic performance index function, Q, R are the state weight matrix and the control input weight matrix respectively, U(k) is the optimal control input corresponding to the current received vehicle state information X(k) to be obtained;

[0061] The linear system optimal state feedback law is solved to minimize the linear quadratic performance index function J:

[0062]

[0063] Based on the Riccati equation solution P, the feedback gain matrix K is calculated:

[0064]

[0065] According to the feedback gain matrix K, the optimal control input at time k is calculated:

[0066]

[0067] wherein, U(k) is the optimal control input corresponding to the current received vehicle state information X(k) for compensating the vehicle state information.

[0068] Compared with the prior art, the present application has the following advantages:

[0069] (1) The present application solves the problem of control lag and control precision decline caused by unpredictable and uncontrollable network communication delay in the existing cloud control automatic driving system by constructing a complete control process of "vehicle-cloud control platform-network" cooperation, introducing a round-trip communication delay modeling and compensation mechanism. By predicting the current communication delay using historical delay data, and combining state compensation and feedforward control design, the controller can predict the vehicle execution time and generate optimal control commands in advance, thereby improving the timeliness and stability of control response. This method first combines the combination mechanism of state compensation+control feedforward+delay modeling in design, which has stronger anti-delay disturbance ability compared with the traditional scheme considering only local state feedback, and improves the robustness and safety of the control system under the cloud control architecture.

[0070] (2) The present application solves the problem of strong dynamic and unpredictable communication delay in vehicle-cloud cooperative automatic driving by introducing a U-Mixer delay prediction model. Traditional methods are difficult to capture the communication mutation characteristics, and have low prediction accuracy, which leads to control lag, vehicle deviation increase, and affects the stability and safety of the system. The present application extracts delay sequence features through Unet structure, combines normalization, patch division and stationarity correction to realize high-precision delay prediction. At the same time, a mutation perception training mechanism is introduced to improve the model's ability to adapt to network sudden fluctuations. The prediction results are used for state compensation and control calculation to realize feedforward compensation of bidirectional communication delay, significantly improving the real-time and accuracy of control, and enhancing the response performance and control stability of the vehicle under complex network.

[0071] (3)In the present application, the vehicle state is compensated by the predicted communication delay, combined with the iterative prediction formula of the vehicle lateral dynamics model and the disturbance correction term, which mainly solves the technical problems that due to the non-negligible two-way delay of vehicle-cloud communication, the vehicle state information obtained by the cloud control platform lags behind the actual state, the control command action time delay, thereby causing the control precision to decline, the system stability to deteriorate and even causing safety risks. The traditional cloud control method usually assumes that the communication delay can be ignored or fixed, and cannot effectively cope with the influence of dynamic delay on system state observation and control input, especially in high-frequency control or complex dynamic environment, the lag state may cause the control command to no longer adapt to the current vehicle state, thereby causing system performance degradation. By introducing the prediction result of the communication delay in the state compensation process, and combining the vehicle dynamics model, the current state is forward predicted, and the information time lag caused by the delay is compensated. At the same time, by introducing the disturbance mean estimated by the historical state and control input as a correction term, the deviation between the model and the actual can be effectively reduced, and the prediction accuracy and system robustness can be improved. This way not only ensures the prediction accuracy, but also improves the adaptability to environmental disturbances and model uncertainties. Significantly improve the vehicle state perception accuracy in the presence of communication delay, provide more close to the actual state of the future input for the controller, help to generate more reasonable and timely response control command. Compared with the system without compensation, it can effectively reduce the state estimation error and control deviation caused by the delay, and improve the stability and control performance of the system. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 The cloud control automatic driving car control method flow chart of the embodiment of the present application;

[0073] Figure 2 The network connection diagram of the communication delay data of the embodiment of the present application;

[0074] Figure 3 The structure diagram of the U-Mixer framework of the embodiment of the present application. DETAILED DESCRIPTION

[0075] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor shall fall within the scope of protection of the present application.

[0076] The present embodiment provides a cloud control automatic driving car control method considering vehicle-cloud communication delay, as shown in Figure 1 The steps include:

[0077] Step S1: collecting vehicle state information at the current time by the autonomous vehicle, and uploading to the cloud control platform through the communication network; wherein the vehicle state information includes lateral vehicle speed v, yaw rate r and sending time t.

[0078] The communication relationship among the autonomous vehicle, the communication network and the cloud control platform is shown in Figure 2 The autonomous vehicle sends vehicle state information to the cloud control platform at a fixed time interval τ.

[0079] Step S2: the cloud control platform constructs a time delay prediction model based on historical communication delay data to predict the round-trip communication delay at the current time;

[0080] The historical communication delay data includes the round-trip communication delay between the autonomous vehicle and the cloud control platform at multiple historical times, and the round-trip communication delay is the time difference between the autonomous vehicle sending the vehicle state information and receiving the control input sent by the cloud control platform.

[0081] The time delay prediction model is a U-Mixer framework, as shown in Figure 3 The U-Mixer framework includes a normalization processing module, an embedding module, a Unet encoder-decoder structure composed of multiple layers of perception, a stationarity correction module and a prediction output module.

[0082] Step S2 specifically includes:

[0083] The historical communication delay data is constructed into a time sequence W in the order of time nodes;

[0084] The time sequence data W is normalized to calculate the normalized data:

[0085]

[0086] Wherein μ and σ are the mean and standard deviation of the historical communication delay data, respectively;

[0087] The normalized data W is divided into a patch sequence W according to the preset patch length P and the step S; p Wherein the number of patches is L is the length of the normalized data W .

[0088] The patch sequence W p is input into the Unet encoder-decoder structure composed of multiple layers of perception to obtain the encoded feature representation;

[0089] The encoded feature representation is corrected for stationarity to generate the corrected feature representation;

[0090] The corrected feature representation is mapped by a prediction output module to output a predicted round-trip communication delay at the current time:

[0091]

[0092] where Y ′ is the predicted round-trip communication delay, is the corrected feature representation, W1, W2, b1, b2 are weight matrices and bias vectors of two full connection layers in the prediction output module respectively, and σ is an activation function;

[0093] According to the predicted round-trip communication delay Y ′ at the current time and a vehicle state sampling interval τ, the number of steps of the predicted round-trip communication delay is calculated:

[0094]

[0095] where n is the number of steps of the predicted round-trip communication delay, represents a downward rounding.

[0096] The training process of the delay prediction model specifically includes:

[0097] Based on a historical communication delay time sequence W={w1,w2,…,w T}, w t denotes a communication delay at time t, a sliding difference method is used to detect a mutation point of the communication delay, and the time sequence is divided into a mutation sample segment and a non-mutation sample segment, wherein the mutation sample segment is a time interval satisfying |w t -w t-1 |>θ, the non-mutation sample segment is a time interval satisfying |w t -w t-1 |≤θ, and θ is a mutation threshold value, which is adaptively determined through historical data statistics;

[0098] In the mutation and non-mutation sample segments, sample pairs are sampled to form: a query sample W q , a positive sample W + and a negative sample W - , wherein the query sample W q and the positive sample W + come from any non-mutation sample segment, and the negative sample W - comes from any mutation sample segment;

[0099] The training sample pairs are input into a U-Mixer model, and hidden feature representations f(W) of the training sample pairs are extracted by an Unet encoder-decoder structure;

[0100] According to the hidden feature representations f(W) of the respective samples, a mutation-aware contrastive loss function is calculated:

[0101]

[0102] wherein f(W q ), f(W + ), f(W j - ) represent the feature representation of the query sample, the positive sample and the jth negative sample respectively; sim represents the cosine similarity, τ is the temperature coefficient, and L is the mutation-aware contrastive loss function.

[0103] Updating the parameters of the time delay prediction model according to the calculated mutation-aware contrastive loss function.

[0104] Step S2 realizes accurate prediction of the round-trip time delay of vehicle-cloud communication by introducing the time delay prediction model U-Mixer framework, and solves the problem of high dynamicity and unpredictability of the round-trip communication time delay caused by unstable communication network in the vehicle-cloud collaborative autonomous driving system. The existing technology usually estimates the communication time delay by using methods such as moving average, filtering or static modeling, which cannot effectively capture the mutation behavior and non-stationary characteristics in the communication delay, resulting in large prediction error, ultimately affecting the timeliness and accuracy of the control command in the autonomous driving system, easily causing problems such as control lag and increased vehicle state deviation, and reducing the safety and control stability of vehicle driving. The U-Mixer time delay prediction model of the present application can effectively capture the evolution trend of the round-trip communication time delay by time modeling and pattern extraction on the historical communication delay sequence, significantly improving the prediction accuracy. The model uses the Unet encoder-decoder structure combined with normalization processing, patch division, stationarity correction and output mapping module to construct a complete communication time delay modeling and prediction process.

[0105] And, by constructing mutation and non-mutation sample segments and introducing a mutation-aware contrast learning mechanism, the problem of traditional communication time delay prediction model in the face of communication mutation events (such as sudden network congestion, base station switching, link jitter) is solved. The prediction accuracy and model generalization ability are poor. Traditional methods are mostly based on stationary time series modeling, assuming that the communication delay changes slowly and continuously, and using regression, moving average or LSTM method for prediction. These methods are still acceptable when the time delay changes smoothly, but in the actual vehicle cloud communication environment, the communication delay is sudden, irregular and highly non-stationary. The traditional model is difficult to distinguish between "normal fluctuation" and "mutation event", and is prone to failure in mutation scenarios, resulting in a serious deviation of the predicted value from the true value, thereby affecting the accuracy of vehicle state compensation and control. In order to solve the above problems, the present application proposes to divide the historical communication time delay into mutation sample segment and non-mutation sample segment, and introduce a mutation detection mechanism (based on sliding difference) to clearly identify the communication mutation interval. By constructing negative samples and positive samples from mutation segments and non-mutation segments, and inputting them into the contrast learning module with the query sample to form a triple, the model can explicitly learn and distinguish the feature representation of mutation and non-mutation communication patterns.

[0106] The mutation-aware contrast loss function is used to strengthen the model's ability to distinguish different communication delay patterns, so that it can learn the rules of historical communication behavior while remaining sensitive to mutation events, improving the prediction adaptability of sudden abnormalities. Compared with conventional prediction loss functions (such as MSE), this contrast learning mechanism emphasizes the structural modeling of time delay change trends, rather than just fitting the numerical size, thereby improving the anti-mutation interference ability and generalization robustness of the model. This mechanism significantly enhances the prediction stability and anti-interference ability of the model in extreme communication environments, so that even in the case of sudden and severe changes in communication delay, it can still output relatively accurate and reliable delay estimation results to provide more reliable data support for vehicle control logic.

[0107] Step S3: Based on the historical vehicle state information and the historical control input sequence, the vehicle dynamics model is used to compensate the vehicle state information at the current time according to the round-trip communication time delay to obtain the compensated vehicle state information after the round-trip communication time delay; specifically including:

[0108] The vehicle dynamics discrete model is used to predict the vehicle state forward, assuming that the current received vehicle state information is X(k), the historical control input sequence is U(k-1), U(k-2), …, the historical vehicle state information is X(k-1), X(k-2), …, and the predicted round-trip communication time delay step is n;

[0109] According to the historical vehicle state information and the historical control input sequence, the process disturbance mean μ caused by environmental disturbance is estimated w As a system modeling error correction term, the specific calculation method is:

[0110] w(j) = X(j+1) - AX(j) - BU(j), j = k-m,..., k-1

[0111]

[0112] wherein m is the length of the historical time window for estimating process disturbance, w(j) is the model prediction error at the jth moment, X(j+1) and X(j) are the vehicle state information received by the cloud platform at the j+1th moment and the jth moment respectively, U(j) is the control input sent by the cloud platform at the jth moment, A is a state transition matrix, and B is a control input matrix;

[0113] The state transition matrix A is:

[0114]

[0115] wherein a and b are the front and rear wheelbase of the automobile respectively, C f and C r are the front and rear wheel cornering stiffness of the automobile respectively, m is the weight of the automobile, u is the longitudinal speed of the automobile, and I is the moment of inertia of the automobile.

[0116] The control input matrix B is:

[0117]

[0118] The control input U = δ, and δ is the front wheel steering angle of the automobile.

[0119] With the current state X(k) as the initial value, the following iterative prediction formula is used to gradually forward predict the vehicle state at the k+nth moment in the future:

[0120]

[0121] wherein, U(k+i) is the control input at the k+i moment, which is obtained by calculation through the U(k+i) linear quadratic regulator controller; based on the iterative prediction formula, the compensated vehicle state information after the round-trip communication time delay is obtained

[0122] The current received vehicle state information X(k), when k = 1, the historical vehicle state information X(k-1), X(k-2), … are all set to X(1), and the historical control input sequence U(k-1), U(k-2), … are all set to zero.

[0123] Step S3 solves the problem of control lag and state information lag caused by communication delay in the vehicle-cloud system by predicting the vehicle state forwardly based on the vehicle dynamics model combined with the communication delay prediction result, thereby improving the control real-time performance and accuracy of the cloud control automatic driving system. In the vehicle-cloud collaborative architecture, there is a non-negligible communication round-trip delay between the vehicle and the cloud control platform. If the delay is not compensated, the control instruction calculated by the cloud control platform based on the received vehicle state information will lag behind the real state of the vehicle, resulting in invalid control instruction or reduced accuracy, which may cause the vehicle trajectory to deviate or safety hazards, especially at high speed or in complex dynamic environment. Therefore, accurately predicting the state information of the vehicle after the delay becomes the key to ensuring the effectiveness of the control. To solve this problem, the present application uses a discrete lateral dynamics model of the vehicle, takes the vehicle state X(k) at the current time as the initial value, and performs n-step (prediction communication delay step) forward state deduction on the historical control input sequence U(k-1), U(k-2)... and the historical vehicle state X(k-1), X(k-2)... through linear system iterative prediction, to obtain the vehicle state after the end of the delay The dynamic compensation of the state is realized. Wherein, in order to enhance the fitting ability of the dynamics model in complex environment, a process disturbance correction term μ_w is introduced, and the system modeling error is compensated by estimating the mean value of multiple historical time modeling errors w(j). This disturbance modeling method can effectively capture the influence of non-modeling factors such as road friction change, side wind interference and sensor error on vehicle motion, further improve the prediction accuracy, and overcome the defect that the traditional model cannot correct the environmental disturbance in real time. The method obtains the control input sequence U(k+i) at the prediction time based on the linear quadratic regulator (LQR), ensures that the control input meets the optimization performance index requirement in the prediction process, maintains the system stability and control effect. The state lag problem caused by communication delay is solved, so that the control decision is based on the future state of the vehicle rather than the delayed state; the real-time performance and accuracy of the control are improved, the control error is significantly reduced, and the driving safety is enhanced; the dynamics model and disturbance estimation are integrated, the robustness of the state prediction in the actual road environment is improved; the communication delay prediction model is linked, a closed-loop state compensation mechanism is formed, and the practicality and stability of the vehicle-cloud collaborative control system are enhanced.

[0124] Step S4: The cloud control platform designs a linear quadratic regulator controller based on the compensated vehicle state information after the round-trip communication delay, and calculates the optimal control input by taking the compensated vehicle state information as the input; specifically including:

[0125] Obtaining the compensated vehicle state information Setting a linear quadratic performance index function:

[0126]

[0127] Wherein, J is a linear quadratic performance index function, Q and R are state weight matrix and control input weight matrix respectively, U(k) is the optimal control input corresponding to the current received vehicle state information X(k) to be obtained;

[0128] To minimize the linear quadratic performance index function J, the optimal state feedback law of the linear system is solved:

[0129]

[0130] Based on the Riccati equation solution P, the feedback gain matrix K is calculated:

[0131]

[0132] According to the feedback gain matrix K, the optimal control input at time k is calculated:

[0133]

[0134] Wherein, To compensate for the vehicle state information, U(k) is the optimal control input corresponding to the current received vehicle state information X(k).

[0135] Step S5: The optimal control input is sent to the autonomous vehicle through the network, and is received and executed by the autonomous vehicle at the future time under the action of the downlink communication delay, realizing the feedforward compensation of the two-way communication delay and the control optimization.

[0136] The above functions, if realized in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of software products. The computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and various program code storage media.

[0137] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements shall be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A cloud-controlled autonomous driving vehicle control method considering vehicle-cloud communication delay, characterized in that: The following steps are involved: The autonomous vehicle collects the current vehicle status information and uploads it to the cloud control platform through the communication network; The cloud control platform builds a delay prediction model based on historical communication delay data to predict the round-trip communication delay at the current moment; Based on historical vehicle state information and historical control input sequences, the vehicle dynamics model is used to compensate the current vehicle state information according to the round-trip communication delay to obtain the compensated vehicle state information after the round-trip communication delay. The cloud control platform designs a linear quadratic regulator controller based on the compensated vehicle state information after the round-trip communication delay, and uses the compensated vehicle state information as input to calculate the optimal control input. The optimal control input is sent to the autonomous vehicle through the network. Under the influence of downlink communication delay, it is received and executed by the autonomous vehicle at a future time, realizing feedforward compensation and control optimization of the two-way communication delay.

2. The cloud-controlled autonomous driving vehicle control method considering vehicle-cloud communication delay according to claim 1 is characterized in that: The vehicle state information includes the lateral vehicle speed v, the yaw angular velocity r and the sending time t.

3. The cloud-controlled autonomous driving vehicle control method considering vehicle-cloud communication delay according to claim 1 is characterized in that: The historical communication delay data includes the round-trip communication delay between the autonomous driving vehicle and the cloud control platform at multiple historical moments. The round-trip communication delay is the time difference between the autonomous driving vehicle sending vehicle status information and receiving the control input sent by the cloud control platform.

4. The cloud-controlled autonomous driving vehicle control method considering vehicle-cloud communication delay according to claim 1 is characterized in that: The autonomous vehicle sends vehicle status information to the cloud control platform at a fixed time interval τ.

5. The cloud-controlled autonomous driving vehicle control method considering vehicle-cloud communication delay according to claim 1 is characterized in that: The delay prediction model is a U-Mixer framework, which includes a normalization processing module, an embedding module, a Unet encoder-decoder structure composed of a multi-layer perceptron, a stationarity correction module, and a prediction output module; The method of constructing a delay prediction model based on historical communication delay data to predict the round-trip communication delay at the current moment specifically includes: Construct a time series W based on the order of time nodes using historical communication delay data; Normalize the time series data W and calculate the normalized data: Where μ and σ are the mean and standard deviation of historical communication delay data respectively; The normalized data Divide into patch sequences W according to the preset patch length P and stride S p , where the number of patches is L is the normalized data length; The patch sequence W p Input the Unet encoder-decoder structure composed of a multi-layer perceptron to obtain the encoded feature representation; Perform stationarity correction on the encoded feature representation to generate a corrected feature representation; The corrected feature representation is mapped through the prediction output module, and the predicted round-trip communication delay at the current moment is output: Among them, Y ′ is the predicted round-trip communication delay, is the corrected feature representation, W1, W2, b1, b2 are the weight matrices and bias vectors of the two fully connected layers in the prediction output module, and σ is the activation function; According to the predicted round-trip communication delay Y at the current moment ′ The number of steps of the predicted round-trip communication delay is calculated based on the vehicle state sampling interval τ: Where n is the number of steps of the predicted round-trip communication delay, Indicates rounding down.

6. The cloud-controlled autonomous driving vehicle control method considering vehicle-cloud communication delay according to claim 5 is characterized in that: The training process of the delay prediction model specifically includes: Based on the historical communication delay time series W={w1,w2,…,w T },w t Refers to the communication delay at time t. The sliding difference method is used to detect the mutation point of the communication delay, and the time series is divided into mutation sample segments and non-mutation sample segments. Among them, the mutation sample segment satisfies |w t -w t-1 |>θ, the non-mutation sample segment satisfies |w T -w T-1 |≤θ time interval, θ is the mutation threshold, which is determined by historical data statistics and adaptively; In the mutation and non-mutation sample segments, sampling is used to construct training sample pairs to form: query sample W Q , positive sample W + With negative samples W - , where the query sample W q With positive sample W + From any non-mutated sample segment, negative sample W - From any mutation sample segment; The training sample pairs are input into the U-Mixer model, and their hidden feature representation f(W) is extracted by the Unet encoder-decoder structure; According to the hidden feature representation f(W) of each sample, the mutation-aware contrast loss function is calculated: Among them, f(W q )、f(W + ), Represent the feature representations of query samples, positive samples, and the j-th negative sample respectively; sim represents cosine similarity, τ is the temperature coefficient, and L is the mutation-aware contrast loss function; The parameters of the delay prediction model are updated according to the calculated mutation-aware contrast loss function.

7. The cloud-controlled autonomous driving vehicle control method considering vehicle-cloud communication delay according to claim 1, characterized in that: The method of compensating the current vehicle state information based on the historical vehicle state information and the historical control input sequence using the vehicle dynamics model according to the round-trip communication delay specifically includes: The vehicle state is forward predicted using the vehicle dynamics discrete model. Let the currently received vehicle state information be X(k), the historical control input sequence be U(k-1), U(k-2), ..., the historical vehicle state information be X(k-1), X(k-2), ..., and the predicted round-trip communication delay number be n. Estimate the mean value μ of the process disturbance caused by environmental disturbance based on historical vehicle state information and historical control input sequence w As a system modeling error correction term, the specific calculation method is: w(j)=X(j+1)-AX(j)-BU(j),j=km,…,k-1 Where m is the length of the historical time window used to estimate the process disturbance, w(j) is the model prediction error at the jth moment, X(j+1) and X(j) are the vehicle state information received by the cloud platform at the j+1th moment and the jth moment, respectively, U(j) is the control input sent by the cloud platform at the jth moment, A is the state transfer matrix, and B is the control input matrix. Taking the current state X(k) as the initial value, the iterative prediction formula based on the vehicle lateral dynamics model is used to gradually predict the vehicle state at the k+nth moment in the future: in, U(k+i) is the control input at time k+i, which is calculated by the U(k+i) linear quadratic regulator controller; based on the iterative prediction formula, the compensated vehicle state information after the round-trip communication delay is obtained.

8. The cloud-controlled autonomous driving vehicle control method considering vehicle-cloud communication delay according to claim 7, characterized in that: The currently received vehicle state information X(k), when k=1, sets the historical vehicle state information X(k-1), X(k-2),... to X(1), and sets the historical control input sequences U(k-1), U(k-2),... to zero.

9. The cloud-controlled autonomous driving vehicle control method considering vehicle-cloud communication delay according to claim 7, characterized in that: The state transfer matrix A is: Among them, a and b are the front and rear wheelbases of the car respectively, C f 、C r are the front and rear wheel cornering stiffness of the car, m is the weight of the car, u is the longitudinal speed of the car, and I is the moment of inertia of the car; The control input matrix B is: The control input U=δ, where δ is the front wheel steering angle of the vehicle.

10. The cloud-controlled autonomous driving vehicle control method considering vehicle-cloud communication delay according to claim 1, characterized in that: The method of designing a linear quadratic regulator controller based on the compensated vehicle state information after the round-trip communication delay and calculating the optimal control input using the compensated vehicle state information as input specifically includes: Get compensation vehicle status information Set the linear quadratic performance index function: Where J is a linear quadratic performance index function, Q and R are the state weight matrix and the control input weight matrix, respectively, and U(k) is the optimal control input corresponding to the currently received vehicle state information X(k). With the goal of minimizing the linear quadratic performance index function J, the optimal state feedback law of the linear system is solved: Based on the solved Riccati equation solution P, the feedback gain matrix K is calculated: According to the feedback gain matrix K, the optimal control input at time k is calculated: in, To compensate for the vehicle state information, U(k) is the optimal control input corresponding to the currently received vehicle state information X(k).

Images