Front vehicle speed parallel prediction method based on deep confidence and gating cycle unit

Abstract

The invention belongs to the field of intelligent traffic 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 following steps of: 1, acquiring the speed and acceleration of a front vehicle to construct an original time sequence data set, and setting rolling window segmentation and normalization as training and verification data; 2, constructing and training a deep belief network, and performing multi-layer feature extraction on the data; 3, splicing the extracted high-order features, inputting the spliced high-order features into a gating circulation unit, learning time sequence dependence and realizing speed prediction; and 4, designing a multi-condition fusion event triggering mechanism, starting offline parameter learning and synchronously updating an online module when data reaches a triggering threshold value, and performing real-time speed prediction according to the updated model in online prediction. According to the method, high-precision online speed prediction is realized through deep feature extraction, time sequence modeling and event-driven updating mechanisms, the adaptability to dynamic scenes is improved, and the method is suitable for scenes such as an advanced auxiliary driving system and the Internet of Vehicles.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent transportation and machine learning, and specifically relates to a front vehicle speed parallel prediction method based on deep belief and gated recurrent unit. BACKGROUND

[0002] In an autonomous driving system, accurate and rapid prediction of the speed of a preceding vehicle can not only help the controlled vehicle to plan a driving path in advance and optimize energy consumption, but also avoid potential collision risks in complex traffic scenarios and improve the overall efficiency of traffic flow. The speed prediction of the preceding vehicle has good application prospects, but the technology also has many challenges. With the gradual complexity of intelligent driving scenarios, the prediction model has difficulty in capturing deep structure features in the data when dealing with complex nonlinear relationships or time series problems, or problems such as gradient disappearance and dimension explosion occur. At the same time, the autonomous driving system needs to perform real-time prediction of the vehicle speed under the condition of limited computing resources, which requires the model not only to have high precision, but also to have rapid response capability. SUMMARY

[0003] The present application provides a front vehicle speed parallel prediction method based on deep belief and gated recurrent unit. The method proposes an online and offline hybrid parallel architecture based on deep belief network and gated neural unit, which finely trains historical data through an offline module, realizes real-time prediction of the speed of the preceding vehicle through an online module, and designs a condition triggering mechanism to complete parameter updating of the online module. The architecture can effectively improve the adaptability of the prediction model while ensuring high precision, and can also meet the real-time requirement, having engineering application capability. BRIEF DESCRIPTION OF DRAWINGS

[0004] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0005] Figure 1 It is a schematic diagram of the overall structure of the present application;

[0006] Figure 2 It is a schematic diagram of the offline learning structure of step two and step three in the present application;

[0007] Figure 3 It is a comparison diagram of the prediction results of the three methods;

[0008] Figure 4 It is a comparison diagram of the mean absolute error and the root mean square error of the three methods;

[0009] Figure 5A comparison chart of the computation time for the three methods; Detailed Implementation

[0010] The parallel prediction method for the speed of the vehicle ahead based on deep belief and gated recurrent units includes the following steps:

[0011] Step 1: Collect the speed and acceleration of the vehicle in front to construct the original dataset, set a rolling window to segment the time series and perform normalization preprocessing, which will serve as the training and validation dataset for the prediction model.

[0012] Step 2: Construct and train a deep belief network (DBN) with a convolutional restricted Boltzmann machine to extract multi-level features from the normalized traffic data and obtain high-order semantic feature representations.

[0013] Step 3: The extracted high-order features are concatenated and input into a gated recurrent unit (GRU) to learn temporal dependencies and achieve velocity prediction;

[0014] Step 4: Design a multi-condition fusion event triggering mechanism. When the data reaches the trigger threshold, the parameter learning module is activated and the online module parameters are updated synchronously. The online prediction is then used to make real-time speed predictions.

[0015] The specific method for step one is as follows:

[0016] Historical data of the preceding vehicle, including speed and acceleration, is obtained through network connectivity and sensors. The scrolling window length is set to n, and the historical preceding vehicle speed v at time t is segmented by the scrolling window. h and acceleration a h The data sample is represented as follows:

[0017] Where T represents the sampling time interval. After selecting the dataset, this invention uses the maximum-minimum method to normalize the data, ensuring that the scale of each feature is consistent, as expressed in the following formula:

[0018] Among them, X max X represents the maximum value of the input data. min This represents the minimum value of the input data.

[0019] The specific method for step two is as follows:

[0020] To address the nonlinearity of vehicle speed and multi-factor coupling, this invention designs a feature extraction method based on deep belief networks. The core of the deep belief network is a restricted Boltzmann machine (RBM). To reduce computational complexity and improve the ability to express temporal local nonlinearity and multi-source coupling, this invention designs a convolutional restricted Boltzmann machine in each RBM layer, as follows:

[0021] In the formula, v = X' = [v h ,a h ] T The input data to the network is a two-dimensional matrix with dimensions (2, T), where T is the sampling time interval, and h represents the information matrix of the entire hidden layer. t and k are the time index and feature index, respectively, b is the bias of the hidden layer, h is the neuron state of the hidden layer, and c... c This 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), v c,t This represents the elements in the input data (c=1 is the velocity element, c=2 is the acceleration element), (W) k ★v) t W represents the local matching degree of the k-th filter at time point t of the input sequence. k This is the weight matrix, and ★ indicates the cross-correlation operation. The learning process of the visible and hidden layers is represented as follows:

[0022] Where σ is the sigmoid activation function. The training objective is to optimize the network parameters Θ = W by maximizing the log-likelihood of the data. k ,b k ,c c The log-likelihood function of an RBM is expressed as:

[0023] in, This is the assignment function, used to ensure the normalization of the probability distribution. To maximize the log-likelihood, we need to calculate its gradient with respect to the parameter Θ, using contrastive divergence for parameter updates, as shown in the following overall expression:

[0024] v (0) and v (1) These represent the expected value calculated from the data and the expected value calculated after one model reconstruction, respectively. Approximate values ​​of the parameter gradients are calculated using a divergence algorithm, followed by updating the model parameters using gradient descent rules, allowing the DBN to gradually learn the probability distribution of the input data. The hidden layer output after training the first RBM layer is used as the visible layer input for the second RBM layer, and the pre-trained parameters are recursively applied layer by layer. After completing the pre-training of multiple RBM layers, the high-dimensional feature representation of the DBN network output is denoted as X. DBN This feature contains deep abstract information about the input data. After training, the DBN network waits to receive data from the next time window before starting a new learning cycle. Each time window outputs a response X.DBN The trained feature X DBN The data is then fed into step three to achieve modeling and prediction of the time series data.

[0025] The specific method for step three is as follows:

[0026] After extracting high-order features from a deep belief network and concatenating the data, the data is fed into a GRU network to learn the temporal dependencies of these features. The GRU network is a simple, low-parameter, and computationally efficient recurrent neural network that effectively captures long-term dependencies in sequences and avoids the vanishing and exploding gradient problems common in traditional recurrent neural networks (RNNs). It excels at capturing dependencies in long-term sequences. Compared to methods such as Long Short-Term Memory (LSTM) networks and transformer models, the lightweight structure of the GRU network effectively reduces computational cost, thereby improving the real-time performance of the system.

[0027] The specific GRU network design is as follows:

[0028] Among them, z t To update the gate, r t It's a reset door, ξ t This represents the system's output state. A fully connected layer is used to map the final state to the predicted value, as shown below:

[0029] Where, ω y and θ t The weights ξ of the fully connected layer t-1 It is the hidden state from the previous moment, ω z U z ,θ z ,ω r U r ,θ r ,ω h U h ,θ h ,ω y ,θ t These are the weight parameters that need to be trained. This represents element-wise multiplication. The goal of a GRU network is to predict, for example, the speed of the car in front, from time-series data. Therefore, the entire process can be optimized through end-to-end training, with the objective of minimizing the following loss function:

[0030] Where n is the length of the time window, V t It is the actual vehicle speed collected at time t. This represents the predicted vehicle speed of the GRU network. After the GRU network learns, this invention employs a self-attention mechanism to generate the final predicted value of the vehicle ahead. The attention mechanism, through weight allocation, assigns sufficient attention to key information, highlighting the influence of important information, thereby improving the accuracy of the neural network model. The attention mechanism is represented as follows:

[0031] Where e t The attention weights at time t are represented by the output vector ξ from the GRU network layer. t Let u be the weight coefficient matrix and ψ be the bias parameter. j This is the attention score, typically obtained by calculating the relationship between the input and the query. α t The attention weights are obtained by normalizing the attention scores of each input; this invention uses the softmax function. Using the above method, the final predicted speed of the vehicle ahead at time t can be obtained.

[0032] Steps two and three are the overall structure. Figure 1 The offline learning module has the following structure: Figure 2 As shown.

[0033] The specific method for step four is as follows:

[0034] In the process of predicting the speed of the vehicle in front, in order to solve the problem of training efficiency and adaptive capability of deep learning models under large-scale data and multiple parameters, this invention designs a parameter parallel optimization mode.

[0035] Based on four parameters—acceleration of the vehicle ahead, rate of change of acceleration, traffic density, and road grade—a multi-condition fusion trigger condition is designed. The acceleration trigger condition is expressed as follows:

[0036] Among them, a th It sets a threshold. The acceleration rate of change trigger condition is expressed as:

[0037] Where, j th It sets a threshold. The trigger condition for the traffic density change rate is expressed as:

[0038] Where, ρ th This sets a threshold. The road level trigger condition is expressed as:

[0039] Among them, Rperv This refers to the road grade at the previous moment. The above single-parameter triggers are then fused together to design a multi-condition fusion trigger strategy, as shown below:

[0040] Where, ω i (i = a, j, ρ, R) represent the weight coefficients of the single triggering conditions in Formulas 11-14, respectively. The weight λ represents the dynamic threshold, which is designed in this invention as λ = 0.6-0.1 sigmoid(ρ norm -0.2), this design adjusts the trigger threshold in real time, lowering the trigger threshold in high traffic density scenarios and maintaining a conservative triggering strategy in low density scenarios, thereby balancing the sensitivity of model updates and computational efficiency. Wherein, ρ norm It is a normalized traffic density, obtained based on the connected environment. Through a multi-condition fusion triggering mechanism, it fully considers changes in the speed of the vehicle ahead, traffic density, and road type, making it more adaptable to dynamic traffic scenarios.

[0041] This invention employs an asynchronous update rule, with an online prediction module handling real-time prediction requests and using the current stable parameter θ. t Once the trigger mechanism is activated, the background executes an incremental update mechanism to generate new parameters θ. t+1 The system switches offline-obtained parameters to the online prediction module via atomic pointer swapping, a process that takes very little time, approximately 1 μs. In practical applications, a hardware FPGA environment can be used. This further improves the system's real-time performance, achieving <10 ms, which meets the real-time requirements of autonomous driving control closed loops.

[0042] In summary, this invention first designs a scrolling window to segment and normalize the sensor-acquired data. Each preprocessed window of data is then input into a deep belief network, where different levels of higher-order representations are learned layer by layer. These higher-order representations are then concatenated according to feature dimensions and input into a GRU network to learn the temporal dependencies of these features. During the learning process, a parallel parameter optimization structure is designed. When multiple event triggering conditions are met, the prediction model parameters are updated in real time through pointer swapping, thereby improving the adaptive capability of the prediction method. A comparison of test results is attached. Figure 3 and attached Figure 4 As shown.

[0043] As can be seen from the comparison chart, compared to BPNN and LSTM predictions, the method of this invention can more smoothly and accurately follow the actual vehicle speed, while the other two methods produce larger fluctuations, especially the BPNN network, which generates a large error when performing multi-step predictions. (Appendix) Figure 4The comparison of mean absolute error and root mean square error under test conditions shows that the method of the present invention not only achieves higher overall prediction accuracy, but also effectively avoids large prediction errors, thereby obtaining more stable prediction results that are closer to the actual vehicle speed trajectory.

[0044] Regarding computation time, the comparison results of the computation time of the three prediction algorithms are shown in the appendix. Figure 5 As shown, the maximum time consumed during the rolling update and prediction processes are 0.58 seconds, 0.14 seconds, and 0.18 seconds, respectively. However, the average time consumed by the three algorithms is approximately 0.52 seconds, 0.12 seconds, and 0.08 seconds, respectively. In the initial setup phase of the test, the data sampling interval of the vehicle sensor was set to 0.5 seconds. Based on the time consumption data of the prediction algorithm, the method of this invention can meet the real-time sampling requirements of vehicle applications.

Claims

1. A parallel prediction method for the speed of the preceding vehicle based on deep belief and gated recurrent units, characterized in that, Includes the following steps: Step 1: Collect the speed and acceleration of the vehicle in front to construct the original dataset, set a rolling window to segment the time series and perform normalization preprocessing, which will serve as the training and validation dataset for the prediction model. Step 2: Construct and train a deep belief network with a convolutional restricted Boltzmann machine to extract multi-level features from the normalized traffic data and obtain high-order semantic feature representations. Step 3: The extracted high-order features are concatenated and input into the gated recurrent unit to learn temporal dependencies and achieve velocity prediction; Step 4: Design a multi-condition fusion event triggering mechanism. When the data reaches the trigger threshold, the parameter learning module is activated and the online module parameters are updated synchronously. The online prediction is based on these parameters to make real-time speed predictions.

2. The parallel prediction method for the speed of the preceding vehicle based on deep belief and gated recurrent units according to claim 1, characterized in that, The process described in step one involves collecting the speed and acceleration of the vehicle in front to construct the original dataset, setting a rolling window to segment the time series and performing normalization preprocessing, which serves as the training and validation dataset for the prediction model. The process is as follows: Historical data of the preceding vehicle, including speed and acceleration, is obtained through network connectivity and sensors. A rolling window length of n is set, and the historical preceding vehicle speed v at time t is segmented within the rolling window. h and acceleration a h Data sample X t It is expressed as follows: Where T represents the sampling time interval; after selecting the dataset, this invention uses the maximum-minimum method to normalize the data to ensure that the scale of each feature is consistent, as expressed in the following formula: Among them, X max X represents the maximum value of the input data. min This represents the minimum value of the input data.

3. The parallel prediction method for the speed of the preceding vehicle based on deep belief and gated recurrent units according to claim 1, characterized in that, In step two, a deep belief network (DBN) with a convolutional restricted Boltzmann machine is constructed and trained to perform multi-level feature extraction on the normalized traffic data to obtain high-order semantic feature representations. The process is as follows: To address the nonlinearity of vehicle speed and multi-factor coupling, this invention designs a feature extraction method based on a deep belief network, where the core of the deep belief network is a restricted Boltzmann machine (RBM). To reduce computational complexity and improve the ability to express temporal local nonlinearity and multi-source coupling, a convolutional restricted Boltzmann machine is designed in each RBM layer, as shown below: In the formula, v = X' = [v h ,a h ] T The input data to the network is a two-dimensional matrix with dimensions (2, T); T is the sampling time interval; h represents the entire hidden layer information matrix; t and k are the time index and feature index, respectively; b is the bias of the hidden layer; h is the neuron state of the hidden layer; c c This 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); v c,t This represents the elements in the input data (c=1 is the velocity element, c=2 is the acceleration element); (W) k *v) t W represents the local matching degree of the k-th filter at time point t of the input sequence, where * denotes the cross-correlation operation. k This is the weight matrix; the learning process of the visible and hidden layers is represented as follows: Where σ is the sigmoid activation function; the training objective is to optimize the network parameters Θ = W by maximizing the log-likelihood of the data. k ,b k ,c c The log-likelihood function of RBM is expressed as: in, This is the assignment function, used to ensure the normalization of the probability distribution; to maximize the log-likelihood, we need to calculate its gradient with respect to the parameter Θ, and update the parameter using the contrastive divergence method. The overall expression is as follows: v (0) and v (1) These are the expected values ​​calculated from the data and the expected values ​​calculated after one model reconstruction, respectively. Approximate values ​​of the parameter gradients are calculated using a divergence algorithm. Then, the model parameters are updated using gradient descent rules, allowing the DBN to gradually learn the probability distribution of the input data. The hidden layer output after the first RBM training is used as the visible layer input for the second RBM. The pre-training parameters are iterated layer by layer. After completing the pre-training of multiple RBMs, the high-dimensional feature representation of the DBN network output is denoted as X. DBN This feature contains deep abstract information about the input data; after training, the DBN network waits to receive data from the next time window before starting a new learning process, and outputs a response X for each time window. DBN ; well-trained features X DBN The data is then fed into step three to achieve modeling and prediction of the time series data.

4. The parallel prediction method for the speed of the preceding vehicle based on deep belief and gated recurrent units according to claim 1, characterized in that, Step 3 involves concatenating the extracted high-order features and inputting them into a gated recurrent unit (GRU) to learn temporal dependencies and achieve velocity prediction. The specific process is as follows: After extracting high-order features from a deep belief network and concatenating the data, the data is fed into a GRU network to learn the temporal dependencies of these features. Gated recurrent unit networks are a type of recurrent neural network with a simple structure, few parameters, and high computational efficiency. They can effectively capture long-term dependencies in sequences and effectively avoid the gradient vanishing and exploding problems in traditional recurrent neural networks (RNNs). They are good at capturing dependencies in long-term sequences. Compared with the methods of Long Short-Term Memory (LSTM) networks and transformer models, the lightweight structure of GRU networks effectively reduces the amount of computation, thereby improving the real-time performance of the system. The specific GRU network design is as follows: Among them, X DBN These are high-dimensional features output by deep belief networks; z t For updating the gate; r t It is a reset door; ξ t This represents the system's output state; a fully connected layer is used to map the final state to the predicted value, as shown below: Where, ω y and θ t These are the weights of the fully connected layer; ξ t-1 It is the hidden state from the previous moment; ω z U z ,θ z ,ω r U r ,θ r ,ω h U h ,θ h , and ω y ,θ t These are the weight parameters that need to be trained; This represents element-wise multiplication; The goal of a GRU network is to predict, for example, the speed of the car in front, using time-series data. Therefore, the entire process can be optimized through end-to-end training, with the objective of minimizing the following loss function: Where n is the length of the time window; V t It is the actual vehicle speed collected at time t; This represents the predicted vehicle speed of the GRU network. After the GRU network learns, this invention uses a self-attention mechanism to generate the final predicted value of the vehicle ahead. The attention mechanism allocates sufficient attention to key information through weight distribution, highlighting the influence of important information, thereby improving the accuracy of the neural network model. The attention mechanism is represented as follows: Where e t The attention weights at time t are represented by the output vector ξ from the GRU network layer. t Determined; u is the weighting coefficient matrix; ψ is the bias parameter; e j It is the attention score, obtained by calculating the relationship between the input and the query; α t The attention weights are obtained by normalizing the attention scores of each input. This invention uses the softmax function, and through the above method, the final predicted speed of the vehicle ahead at time t can be obtained.

5. The parallel prediction method for the speed of the preceding vehicle based on deep belief and gated recurrent units according to claim 1, characterized in that, Step four involves designing a multi-condition fusion event triggering mechanism. When the data reaches the trigger threshold, the parameter learning module is activated and the online module parameters are updated synchronously. The online prediction then performs real-time speed prediction based on this. The specific process is as follows: Design a multi-condition fusion event triggering mechanism. When the data reaches the trigger threshold, start the parameter learning module and update the online module parameters synchronously. The online prediction is based on this to make real-time speed prediction. In the process of predicting the speed of the vehicle in front, in order to solve the problem of training efficiency and adaptive capability of deep learning models under large-scale data and multiple parameters, this invention designs a parameter parallel optimization mode. Based on four parameters—acceleration of the vehicle ahead, rate of change of acceleration, traffic density, and road grade—a multi-condition fusion trigger condition is designed. The acceleration trigger condition is expressed as follows: Among them, a th It sets a threshold; the acceleration rate of change trigger condition is expressed as: Where, j th It sets a threshold; the traffic density change rate trigger condition is expressed as: Where, ρ th It sets a threshold; the road level trigger condition is expressed as: Among them, R perv It represents the road level at the previous moment; by fusing the above single-parameter triggers, a multi-condition fusion trigger strategy is designed, as shown below: Where, ω i (i = a, j, ρ, R) represent the weight coefficients of the single triggering conditions in Formulas 11-14, respectively; λ weight represents the dynamic threshold, which is designed in this invention as λ = 0.6-0.1 sigmoid(ρ norm -0.2), this design adjusts the trigger threshold in real time, achieving a balance between the sensitivity of model updates and computational efficiency by lowering the trigger threshold in high traffic density scenarios and maintaining a conservative triggering strategy in low density scenarios; where ρ norm It is a normalized traffic density, obtained based on the connected environment. Through a multi-condition fusion triggering mechanism, it fully considers changes in the speed of the vehicle ahead, traffic density, and road type, making it more adaptable to dynamic traffic scenarios. Simultaneously, asynchronous update rules are designed, and an online prediction module handles real-time prediction requests. The current stable parameter θ is used. t Once the triggering mechanism is activated, the background executes an incremental update mechanism to generate new parameters θ. t+1 The parameters obtained offline are switched to the online prediction module by atomic pointer swapping. This process takes very little time, less than 1μs. In actual application, a hardware FPGA environment can be used to further improve the real-time performance of the system, which can basically reach <10ms, thus meeting the real-time requirements of the autonomous driving control closed loop.

Images