CN-121246828-B - Front vehicle speed parallel prediction method based on deep confidence and gating circulation unit

Abstract

The invention belongs to the field of intelligent transportation and machine learning, and particularly relates to a front vehicle speed parallel prediction method based on deep confidence and a gating circulation unit. The method comprises the steps of firstly collecting the speed and the acceleration of a front vehicle to construct an original time sequence data set, setting rolling window segmentation and normalization to serve as training and verification data, secondly constructing and training a deep confidence network to extract multi-layer features of the data, thirdly inputting the extracted high-order features into a gating circulation unit after splicing, learning time sequence dependence and realizing speed prediction, fourthly designing an event trigger mechanism of multi-condition fusion, starting offline parameter learning and synchronously updating an online module when the data reach a trigger threshold value, and carrying out real-time speed prediction according to an updated model by online prediction. The invention realizes high-precision online speed prediction through deep feature extraction, time sequence modeling and event-driven updating mechanisms, improves the adaptability to dynamic scenes, and is suitable for scenes such as advanced auxiliary driving systems, internet of vehicles and the like.

Inventors

• ZHANG ZHE
• DONG HUAHUI
• ZHANG NIAONA
• LI SHAOSONG
• DAI XINJIE
• HU TONGYU

Assignees

• 长春工业大学

Dates

Publication Date

20260508

Application Date

20250919

Claims (5)

1. 1. The front vehicle speed parallel prediction method based on the deep confidence and the gating circulation unit is characterized by comprising the following steps of: Step one, acquiring historical speed and acceleration data of a front vehicle through a vehicle-mounted sensor, dividing a time sequence by adopting a rolling time window with the length of n, constructing speed and acceleration data at a plurality of continuous moments into time sequence samples, and preprocessing the sample data by adopting a normalization method to obtain an input data set for model training and prediction, wherein the input data is a two-dimensional time sequence matrix comprising two channels of speed and acceleration; Step two, constructing a Deep Belief Network (DBN) formed by stacking limited Boltzmann machines (RBM) layer by layer, performing feature learning on normalized input data through unsupervised layer by layer pre-training, performing iterative updating on network parameters in each layer of limited Boltzmann machines by adopting a contrast divergence algorithm to approximate probability distribution characteristics of the input data, performing layer by layer recursive training on the output of a previous layer of hidden layer as the input of a next layer, and outputting a high-dimensional feature vector X DBN after pre-training and sending the high-dimensional feature vector X DBN into the step three, wherein the input data is a two-dimensional matrix with the dimensions of (2, T) and corresponds to a front vehicle speed channel and an acceleration channel respectively; inputting the high-order features output by the deep confidence network into a gating circulation unit network, performing time sequence modeling on the feature sequence by utilizing a gating structure of the high-order features, and outputting a speed prediction result of a front vehicle at a future moment; And fourthly, constructing a parallel prediction framework formed by the offline learning module and the online prediction module, wherein the online prediction module predicts the speed of the front vehicle based on the real-time collected vehicle data, stores the real-time data into the data buffer module, and designs an event trigger mechanism based on multi-condition fusion, wherein the event trigger mechanism constructs a multi-condition fusion trigger function based on the acceleration of the front vehicle, the acceleration change rate, the traffic density and the road type, and the model parameters are updated by starting the offline learning module when the trigger conditions are met by weighting fusion and comparing the trigger conditions with a preset trigger threshold value, and the updated parameters are switched to the online prediction module in an asynchronous mode so as to realize the real-time prediction of the speed of the front vehicle.
2. 2. The method for parallel prediction of front vehicle speed based on deep confidence and gating cyclic unit according to claim 1, wherein in the first step, the collected front vehicle historical speed and acceleration construct an original data set, a rolling window is set to divide a time sequence and perform normalization preprocessing, and the process is as follows: The historical data of the front vehicle, including speed and acceleration, obtained through the network connection and the sensor, is set to be n, and a data sample X t of the historical front vehicle speed v h and the acceleration a h at the time t, which are divided by the rolling window, is represented as follows: (1) After the data set is selected, the data is normalized by adopting a maximum value-minimum value method, so that the scale of each feature is consistent, and the formula is as follows: (2) wherein X max is the maximum value of the input data, and X min is the minimum value of the input data.
3. 3. The front vehicle speed parallel prediction method based on the deep confidence and gating circulation unit according to claim 1, wherein the deep confidence network (DBN) with the convolution-limited boltzmann machine is constructed and trained in the second step, and the normalized traffic data is subjected to multi-level feature extraction to obtain a high-order semantic feature representation, and the process is as follows: Aiming at the nonlinear and multi-cause coupling of the vehicle speed, the invention designs a feature extraction method based on a depth confidence network, wherein the core of the depth confidence network is a limited Boltzmann machine (RBM), and in order to reduce the computational complexity and improve the expression capability of the local nonlinear and multi-source coupling of the time sequence, a convolution limited Boltzmann machine is designed at each RBM layer and expressed as follows: (3) in the formula, Is the input data of the network, is a two-dimensional matrix, and has the size of (2, T), wherein T is the sampling time interval; T and k are time index and characteristic index respectively, b is bias of hidden layer, h is neuron state of hidden layer; Is the bias of the c-th channel (c=1 is the bias of the velocity channel, c=2 is the bias of the acceleration channel); Representing an element in the input data (c=1 is a velocity element, c=2 is an acceleration element); Representing the local degree of matching of the kth filter at the time point t of the input sequence, wherein, Representing the cross-correlation operation, The learning process of the visible layer and the hidden layer is expressed as follows: (4) Wherein, the Is a sigmoid activation function, and the training goal is to optimize network parameters by maximizing the log likelihood of the data The log-likelihood function of RBM is expressed as: (5) Wherein, the Is an allocation function for ensuring normalization of probability distribution, and in order to maximize log likelihood we need to calculate its parameters And (3) updating parameters by adopting a contrast divergence mode, wherein the whole expression is as follows: (6) And The method comprises the steps of calculating expected values obtained by calculation from data and expected values obtained by calculation after model reconstruction once, calculating approximate values of parameter gradients through a contrast divergence algorithm, updating model parameters through a gradient descent rule, enabling a DBN to gradually learn probability distribution of input data, enabling hidden layer output after training of a first layer RBM to serve as visible layer input of a second layer RBM, recursively pre-training parameters layer by layer, enabling high-dimensional feature representation of DBN network output obtained after pre-training of a plurality of layers of RBM is achieved to be X DBN , and enabling trained feature X DBN to be fed into a third step.
4. 4. The front vehicle speed parallel prediction method based on the deep confidence and gating cycle unit according to claim 1, wherein the third step is to splice the extracted high-order features and input the spliced high-order features into the gating cycle unit (GRU), learn time sequence dependence and implement speed prediction, and the specific flow is as follows: The gating circulation unit network is a circulation neural network with simple structure, less parameters and high calculation efficiency, can effectively capture long-term dependence in a sequence, can effectively avoid gradient elimination and explosion problems in a traditional circulation neural network (RNN), is good at capturing dependence in a long-term sequence, and compared with a method of a long-term memory network (LSTM) and a transducer model, the lightweight structure of the GRU network effectively reduces calculation amount and further improves the real-time performance of the system; the specific GRU network design is as follows: (7) Wherein X DBN is a high-dimensional feature of the deep belief network output, z t is an update gate, r t is a reset gate; The final state is mapped to the predicted value by adopting the full connection layer, and the method is expressed as follows: (8) Wherein, the And Is the right of the full connection layer; is the hidden state of the last moment; And Ä represents element-by-element multiplication; The goal of the GRU network is to predict the speed of a preceding vehicle from time series data, so the whole process can be optimised by end-to-end training with the goal of minimising the following loss functions: (9) V t is the real vehicle speed acquired at the moment t; the invention adopts a self-Attention mechanism to generate a final front vehicle predicted value after GRU network learning, the Attention mechanism distributes enough Attention to key information in a weight distribution mode to highlight the influence of the important information, thereby improving the accuracy of a neural network model, and the Attention mechanism is expressed as follows: (10) Wherein e t represents the attention weight value at time t, the attention weight value is represented by the output vector from the GRU network layer U is a weight coefficient matrix; Is bias parameter, e j is attention score obtained by calculating relationship between input and query (query), alpha t is attention weight obtained by normalizing attention score of every input, and the invention adopts softmax function to obtain final prediction speed of front car at time t 。
5. 5. The method for parallel prediction of the front vehicle speed based on the deep confidence and the gating circulation unit according to claim 1 is characterized in that the fourth step is to design an event trigger mechanism of multi-condition fusion, and when the data reach a trigger threshold, a parameter learning module is started to execute parameter updating and asynchronously update on-line module parameters, and the on-line prediction carries out real-time speed prediction according to the parameter updating, wherein the method comprises the following specific procedures: Designing an event trigger mechanism of multi-condition fusion, starting a parameter learning module when the data reach a trigger threshold value, asynchronously updating parameters of an online module, and carrying out online prediction to carry out real-time speed prediction according to the parameters; in the prediction process of the speed of the front vehicle, in order to solve the problem of training efficiency and self-adaptive capacity of the deep learning model under large-scale data and multiple parameters, the invention designs a parameter parallel optimization mode; According to four single parameters of acceleration, acceleration change rate, traffic density and road grade of the front vehicle, designing a multi-condition fusion triggering condition: the single parameter acceleration trigger condition is expressed as: (11) The acceleration threshold value a th is used for representing the change degree of the longitudinal movement state of the vehicle, and the value of the acceleration threshold value a th is set according to the dynamics characteristic of the vehicle, and is preferably 0.5-2.0 m/s2; the single parameter acceleration rate of change trigger condition is expressed as: (12) the acceleration change rate threshold j th is used for representing the intensity of the acceleration change of the vehicle, and the value of the acceleration change rate threshold j th is set according to the driving behavior smoothness requirement, and is preferably 0.5-2.5 m/s3; the single parameter traffic density change rate triggering condition is expressed as: (13) The traffic density threshold ρ th is a normalized traffic density threshold and is used for representing traffic flow state change, and the value of the traffic density threshold is set according to traffic flow density distribution characteristics, preferably 0.2-0.5; the single parameter road class triggering condition is expressed as: (14) the road grade R perv is expressed by discrete variables, and is triggered when the road grade at the current moment is inconsistent with the road grade at the current moment; the thresholds can be adaptively adjusted according to different vehicle types, traffic environments and data sampling frequencies, and can be optimized and determined through historical data statistical analysis or model training processes; the single parameter is fused, a multi-condition fusion triggering strategy is designed, and the method is expressed as follows: (15) Wherein, the Weight coefficients for the one-shot conditions of equations 11-14, respectively; The weight represents a dynamic threshold, and the invention is designed as The trigger threshold is adjusted in real time through the design, the trigger threshold is reduced in a high traffic density scene, a conservative trigger strategy is kept in a low density scene, and therefore the sensitivity and the calculation efficiency of model update are balanced, ρ norm is normalized traffic density, is obtained according to a network environment, changes of the speed of a front vehicle, traffic density and road type are fully considered through a multi-condition fusion trigger mechanism, and are more suitable for a dynamic traffic scene, meanwhile, an asynchronous update rule is designed, an online prediction module is used for processing a real-time prediction request, and current stability parameters are used When the trigger mechanism triggers, the background executes an increment updating mechanism to generate new parameters The parameters obtained offline are switched to the online prediction module through atomic pointer exchange, the time consumption is very low and less than 1 mu s in the process, a hardware FPGA environment can be adopted in the actual quoting process, the real-time performance of the system is further improved, the real-time performance basically can reach <10ms, and the real-time performance requirement of an automatic driving control closed loop can be met.

Description

Front vehicle speed parallel prediction method based on deep confidence and gating circulation unit Technical Field The invention belongs to the field of intelligent transportation and machine learning, and particularly relates to a front vehicle speed parallel prediction method based on deep confidence and a gating circulation unit. Background In an automatic driving system, accurate and rapid front vehicle speed prediction can help controlled vehicles to plan a driving path in advance and optimize energy consumption, and can avoid potential collision risks in complex traffic scenes, and meanwhile, the overall efficiency of traffic flow is improved. The speed prediction of the front vehicle has good application prospect, but the technology also has a plurality of challenges. With the gradual complexity of intelligent driving scenes, the prediction model is difficult to capture deep structural features in data or has the problems of gradient disappearance, dimensional explosion and the like when processing complex nonlinear relations or time sequence problems. Meanwhile, an automatic driving system needs to conduct real-time prediction of the vehicle speed under the condition of limited computing resources, and the model is required to have high precision and quick response capability. Disclosure of Invention The invention provides a front vehicle speed parallel prediction method based on deep confidence and a gating circulation unit. The method provides an online and offline hybrid parallel architecture based on a deep confidence network and a gating neural unit, historical data is finely trained through an offline module, the online module predicts the speed of a front vehicle in real time, and a condition triggering mechanism is designed to complete parameter updating of the online module. The architecture can effectively improve the self-adaptability of the prediction model while ensuring high precision, can also meet the real-time requirement, and has engineering application capability. Drawings In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are needed in the embodiments will be briefly described below, it being understood that the following drawings only illustrate some embodiments of the present invention and therefore should not be considered as limiting the scope, and other related drawings may be obtained according to these drawings without inventive effort for a person skilled in the art. FIG. 1 is a schematic diagram of the overall structure of the present invention; FIG. 2 is a schematic diagram of an offline learning structure of the second and third steps of the present invention; FIG. 3 is a graph comparing the predicted results of three methods; FIG. 4 is a graph of mean absolute error versus root mean square error for three methods; FIG. 5 is a graph of calculated time versus three methods; Detailed Description The front vehicle speed parallel prediction method based on the depth confidence and the gating circulation unit comprises the following steps: Firstly, acquiring the speed and the acceleration of a front vehicle to construct an original data set, setting a rolling window to divide a time sequence and carrying out normalization preprocessing to be used as a training and verification data set of a prediction model; Step two, constructing and training a Deep Belief Network (DBN) with a convolution type limited Boltzmann machine, and extracting multi-level features of normalized traffic data to obtain high-order semantic feature representation; step three, inputting the spliced high-order features into a gate control circulation unit (GRU), learning time sequence dependence and realizing speed prediction; and fourthly, designing a multi-condition fusion event trigger mechanism, starting a parameter learning module when the data reach a trigger threshold value, asynchronously updating parameters of the online module, and carrying out online prediction to carry out real-time speed prediction according to the parameters. The specific method of the first step is as follows: Historical data of the front vehicle including speed and acceleration are obtained through the network and the sensors. Setting the length of the rolling window to be n, and representing the data samples of the historical front vehicle speed v h and the acceleration a h at the time t, which are divided by the rolling window, as follows: (1) Where T represents the time interval of the samples. After the data set is selected, the data is normalized by adopting a maximum-minimum method, the scale of each feature is ensured to be consistent, and the formula is expressed as follows: (2) Wherein X max is the maximum value of the input data, and X min is the minimum value of the input data. The specific method of the second step is as follows: Aiming at the nonlinear and multi-factor coupling of the vehicle speed, the invention designs a feature extrac