Cruise control method, system, device and computer readable storage medium

By using feature mapping and transformation of a fully connected layer, encoder, and decoder, the problem of computational real-time performance and adaptability of existing cruise speed control methods under complex slope conditions is solved, achieving efficient and smooth speed planning, reducing computational load, and improving adaptability.

CN123104639APending Publication Date: 2026-11-17DONGFENG COMML VEHICLE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611304817.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-26
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing cruise speed control methods struggle to achieve real-time computation and high-quality adaptive planning under complex and variable slope conditions. Manual rule-based strategies have weak generalization capabilities and are difficult to calibrate parameters, while online optimization algorithms have high computational loads and are sensitive to initial states.

Method used

A data-driven approach is adopted, utilizing fully connected layers, encoders, and decoders for feature mapping and transformation. By processing vehicle state and slope information through self-attention mechanisms and modulation parameters, a target speed sequence is generated, replacing the traditional iterative optimization algorithm to achieve end-to-end fast response and adaptive planning.

Benefits of technology

It significantly reduces computational load, achieves efficient adaptive cruise speed control under complex operating conditions, balances computational efficiency with planning smoothness and dynamic feasibility, eliminates strategy rigidity, and achieves rapid response.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123104639A_ABST
    Figure CN123104639A_ABST
Patent Text Reader

Abstract

A cruise speed control method, system, device, and computer-readable storage medium are disclosed, relating to the field of autonomous driving technology. Specifically, the method includes: inputting vehicle state information, current vehicle speed, and reference cruise speed into a preset fully connected layer for mapping to obtain modulation parameters; inputting a target slope sequence and modulation parameters into a preset encoder for attention and modulation processing to obtain a high-order road semantic feature vector; inputting a preset query vector, the high-order road semantic feature vector, and modulation parameters into a preset decoder for attention and modulation processing to obtain a decoded feature vector; inputting the decoded feature vector into a preset speed planning output head to obtain a target speed sequence, and controlling the cruise speed based on the target speed sequence. This application balances computational efficiency with adaptive planning under complex conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, specifically to a cruise speed control method, system, device, and computer-readable storage medium. Background Technology

[0002] With the development of intelligent technology in commercial vehicles and the increasing demand for energy conservation and emission reduction, predictive cruise control has become the key to improving vehicle economy and safety. Its core lies in how to quickly generate the optimal speed trajectory based on information about the slope ahead.

[0003] In related technologies, speed planning typically employs strategies based on manually applied rules or online optimization algorithms based on dynamic models. However, strategies based on manually applied rules have weak generalization capabilities and are difficult to calibrate parameters, while algorithms based on online optimization have high computational loads and are sensitive to initial states. This makes it difficult for existing methods to achieve adaptive, high-quality planning for complex and variable slope conditions while ensuring real-time computation.

[0004] Therefore, how to provide a cruise speed control method that balances computational efficiency with adaptive planning under complex operating conditions is an urgent problem to be solved. Summary of the Invention

[0005] This application provides a cruise speed control method, system, device, and computer-readable storage medium that can balance computational efficiency with adaptive planning under complex operating conditions.

[0006] In a first aspect, embodiments of this application provide a cruise speed control method, the cruise speed control method comprising: The vehicle status information, current vehicle speed, and reference cruise speed are input into a preset fully connected layer for mapping to obtain modulation parameters. The target ramp sequence and modulation parameters are input into a preset encoder for attention and modulation processing to obtain a high-order road semantic feature vector; The preset query vector, high-order road semantic feature vector, and modulation parameters are input into the preset decoder for attention and modulation processing to obtain the decoded feature vector; The decoded feature vector is input into the preset speed planning output head to obtain the target speed sequence, and the cruise speed is controlled based on the target speed sequence.

[0007] In conjunction with the first aspect, in one embodiment, the encoder includes a first coding layer, a second coding layer, and a third coding layer. The step of inputting the target ramp sequence and modulation parameters into a preset encoder for attention and modulation processing to obtain a high-order road semantic feature vector includes: The target ramp sequence is input into the first coding layer for semantic extraction, and the extracted features are linearly modulated by the modulation parameters to obtain the basic ramp semantic features. The basic ramp semantic features are input into the second coding layer for abstract processing, and the abstracted features are linearly modulated by the modulation parameters to obtain the context ramp semantic features. The contextual ramp semantic features are input into the third encoding layer for enhancement processing, and the enhanced features are linearly modulated using the modulation parameters to obtain higher-order ramp semantic features.

[0008] In conjunction with the first aspect, in one embodiment, the decoder includes a first decoding layer and a second decoding layer. The step of inputting a preset query vector, a high-order road semantic feature vector, and modulation parameters into the preset decoder for attention and modulation processing to obtain a decoded feature vector includes: The preset query vector and the high-order road semantic feature vector are input into the first decoding layer for attention fusion processing, and the fused features are linearly modulated by the modulation parameters to obtain the basic speed adjustment intention features. The basic speed regulation intention features are input to the second decoding layer for consistency enhancement and global correction processing, and the processed features are linearly modulated using the modulation parameters to obtain the decoded feature vector.

[0009] In conjunction with the first aspect, in one embodiment, prior to the step of inputting the target ramp sequence and modulation parameters into a preset encoder for attention and modulation processing to obtain a high-order road semantic feature vector, the method further includes: The preceding ramp sequence is input into a preset mapping layer and projected to obtain the initial ramp sequence; The initial ramp sequence is positionally encoded to obtain the target ramp sequence.

[0010] In conjunction with the first aspect, in one implementation, the target speed sequence includes a relative speed sequence and a node-by-node speed increment sequence, and the control of the cruise speed based on the target speed sequence includes: The target velocity curve is determined based on the relative velocity sequence and the node-by-node velocity increment sequence; Vehicle speed control commands are generated based on the target speed curve.

[0011] In conjunction with the first aspect, in one implementation, the vehicle state information includes the current longitudinal acceleration, vehicle mass, anchor point slope, and the deviation of the current vehicle speed from the reference cruise speed. The step of inputting the vehicle state information, current vehicle speed, and reference cruise speed into a preset fully connected layer for mapping to obtain modulation parameters includes: The current longitudinal acceleration, vehicle mass, anchor point slope, current vehicle speed, reference cruise speed, and the deviation of the current vehicle speed from the reference cruise speed are normalized and spliced ​​to obtain the state vector. The state vector is input to a preset fully connected layer for mapping to obtain modulation parameters, which include scaling factors and offset factors.

[0012] Secondly, embodiments of this application provide a cruise speed control system, the cruise speed control system comprising: The first processing module is used to input vehicle status information, current vehicle speed and reference cruise speed into a preset fully connected layer for mapping to obtain modulation parameters; The second processing module is used to input the target slope sequence and modulation parameters into a preset encoder for attention and modulation processing to obtain a high-order road semantic feature vector. The third processing module is used to input the preset query vector, high-order road semantic feature vector and modulation parameters into the preset decoder for attention and modulation processing to obtain the decoded feature vector; The fourth processing module is used to input the decoded feature vector into the preset speed planning output head to obtain the target speed sequence, and control the cruise speed based on the target speed sequence.

[0013] In conjunction with the second aspect, in one implementation, the second processing module is specifically used for: The target ramp sequence is input into the first coding layer for semantic extraction, and the extracted features are linearly modulated by the modulation parameters to obtain the basic ramp semantic features. The basic ramp semantic features are input into the second coding layer for abstract processing, and the abstracted features are linearly modulated by the modulation parameters to obtain the context ramp semantic features. The contextual ramp semantic features are input into the third encoding layer for enhancement processing, and the enhanced features are linearly modulated using the modulation parameters to obtain higher-order ramp semantic features.

[0014] Thirdly, embodiments of this application provide a cruise speed control device, which includes a processor, a memory, and a cruise speed control program stored in the memory and executable by the processor. When the cruise speed control program is executed by the processor, it implements the steps of the cruise speed control method as described in any of the preceding claims.

[0015] Fourthly, embodiments of this application provide a computer-readable storage medium storing a cruise speed control program, wherein when the cruise speed control program is executed by a processor, it implements the steps of the cruise speed control method as described in any of the preceding claims.

[0016] The beneficial effects of the technical solutions provided in this application include: By inputting vehicle status information, current vehicle speed, and reference cruise speed into a pre-defined fully connected layer for mapping, modulation parameters are obtained. This data-driven approach replaces manual calibration, addressing the poor generalization problem of multi-condition strategies. The target slope sequence and modulation parameters are input into a pre-defined encoder for attention and modulation processing to obtain a high-order road semantic feature vector. A self-attention mechanism is used to avoid iterative solutions, reducing computational load. Vehicle condition information is fused into the slope features through the modulation parameters, enabling the same slope to generate differentiated semantic features under different vehicle conditions, thus supporting different planning strategies. Finally, a pre-defined query vector, high-order road semantic feature vector, and modulation parameters are input into a pre-defined decoder for attention and modulation processing. The decoded feature vector is obtained, and vehicle condition constraints are continuously injected during sequence generation using modulation parameters to enhance the state coherence between adjacent nodes, thereby ensuring the smoothness of planning and the feasibility of dynamics. The decoded feature vector is input into a preset speed planning output head to obtain the target speed sequence, and the cruising speed is controlled based on the target speed sequence. This application uses a computational path composed of an encoder, decoder, and output head to replace the iterative optimization process of the online optimization algorithm, transforming the complex optimization problem into a deterministic feature transformation calculation, significantly reducing the computational load. Furthermore, data-driven mapping replaces manual rule calibration, eliminating policy rigidity and achieving end-to-end fast response, while balancing computational efficiency and planning adaptability under complex working conditions. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an embodiment of the cruise speed control method of this application; Figure 2 This is a detailed flowchart illustrating an embodiment of the cruise speed control method of this application; Figure 3 This is a speed planning comparison diagram in the embodiments of the cruise speed control method of this application; Figure 4 This is a schematic diagram of the functional modules of an embodiment of the cruise speed control system of this application; Figure 5 This is a schematic diagram of the hardware structure of the cruise speed control device involved in the embodiments of this application. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0020] In a first aspect, embodiments of this application provide a cruise speed control method.

[0021] In one embodiment, reference is made to Figure 1 , Figure 1 This is a schematic flowchart illustrating an embodiment of the cruise speed control method of this application. Figure 1 As shown, the cruise speed control method includes: Step S10: Input the vehicle status information, current vehicle speed and reference cruise speed into the preset fully connected layer for mapping to obtain modulation parameters.

[0022] As an example, in the embodiments of this application, vehicle status information refers to a series of parameters that can reflect the vehicle's dynamic characteristics and driving environment. The specific type can be configured according to actual control requirements. The system uses the collected vehicle status information, the vehicle's current actual driving speed (current speed) and the preset expected driving speed (referencing the cruise speed, the specific value of which can be determined according to actual requirements and is not limited here) as input data.

[0023] It should be understood that the pre-defined fully connected layer is a basic neural network structure. In this embodiment, the fully connected layer is used to map vehicle state physical quantities of different dimensions (such as vehicle speed, mass, etc.) into modulation parameter vectors that can be processed by the network, realizing the transformation from physical state to feature constraints. Specifically, the fully connected layer receives the input data and processes it through linear transformation of the internal weight matrix and nonlinear activation function, mapping the input data into a series of numerical vectors, i.e., modulation parameters. These modulation parameters do not contain specific speed commands, but are used as feature adjustment coefficients (such as scaling coefficients and offset coefficients) in subsequent network processing. This step realizes the transformation of real-time vehicle operating condition information into feature constraints that the network can understand, enabling subsequent network calculations to dynamically adjust the feature expression method according to the current vehicle condition, laying the foundation for generating speed plans that adapt to different operating conditions.

[0024] Step S20: Input the target ramp sequence and modulation parameters into the preset encoder for attention and modulation processing to obtain a high-order road semantic feature vector.

[0025] In an exemplary embodiment of this application, the system constructs a target slope sequence based on the road data ahead. This sequence represents the changes in road slope over a certain distance. The target slope sequence can be preprocessed digitized sequence data, which includes road geometric features and location information. Subsequently, the target slope sequence and the modulation parameters generated in the previous step are input into a preset encoder. This encoder adopts an attention mechanism sequence processing architecture and includes one or more encoding layers. During processing, the encoder first extracts the spatial dependencies within the slope sequence through a self-attention mechanism, capturing the slope correlation between nodes at different distances. Simultaneously, it uses the modulation parameters to perform conditional modulation processing (e.g., feature linear modulation) on the intermediate features of each layer within the encoder. This processing ensures that the features output by the encoder not only include the geometric semantics of the road itself but also incorporate the influence of the current vehicle conditions on the difficulty of road passage. Finally, the encoder outputs a high-dimensional feature vector, namely a high-order road semantic feature vector. This vector is a compressed representation of the road conditions ahead and its passage strategy under the current vehicle conditions, serving as memory information for subsequent speed planning.

[0026] Step S30: Input the preset query vector, high-order road semantic feature vector and modulation parameters into the preset decoder for attention and modulation processing to obtain the decoded feature vector.

[0027] In an exemplary embodiment of this application, the decoder is responsible for converting the road semantics output by the encoder into specific speed planning features. The preset query vector is a set of learnable parameter sequences, the number of which corresponds to the length of the speed sequence to be planned (e.g., 30). This query vector serves as the initial input of the decoder, representing the time step or distance node position information within the future planning cycle. Based on this query vector, the decoder extracts the speed planning information of the corresponding position from the road semantic features output by the encoder through an attention mechanism, thereby converting the abstract position placeholder into a specific speed feature vector.

[0028] Specifically, the decoder employs an attention-based sequence generation architecture, comprising one or more decoding layers. During processing, the decoder establishes temporal dependencies within the query sequence through a masked self-attention mechanism and reads the high-order road semantic feature vector output by the encoder through a cross-attention mechanism, injecting road information into the speed planning features. Simultaneously, modulation parameters are used again to constrain and modulate the features generated by the decoder. This step ensures that the generated speed features conform to both the geometric constraints of the road ahead and the dynamic constraints of the current vehicle. After multi-layer processing by the decoder, the output is a decoded feature vector, which contains the speed planning intent and state constraint information for multiple future time points, serving as the direct basis for generating the final speed sequence.

[0029] Step S40: Input the decoded feature vector into the preset speed planning output head to obtain the target speed sequence, and control the cruise speed based on the target speed sequence.

[0030] As an example, in an embodiment of this application, the speed planning output head is a mapping network connected after the decoder. The output head receives the decoded feature vector and converts it into a physically executable target speed sequence. The target speed sequence contains the expected speed values ​​at multiple planning nodes ahead, representing the ideal driving speed trajectory of the vehicle over a period of time in the future.

[0031] Specifically, after acquiring the target speed sequence, the vehicle controller inputs it as the desired trajectory into the vehicle's longitudinal control system. The controller compares the target speed sequence with the vehicle's current actual speed and calculates the speed tracking error. Based on the speed tracking error, the controller generates corresponding vehicle control commands, such as target engine torque, target gearbox gear, or target braking system pressure, and sends them to the vehicle chassis actuators. The actuators adjust the vehicle's power output according to these commands, causing the vehicle's actual speed to follow the target speed sequence, thereby achieving automatic control of the cruising speed and resulting in energy-saving, safe, and smooth driving.

[0032] This application addresses the poor generalization problem of multi-condition strategies by inputting vehicle state information, current speed, and reference cruise speed into a pre-defined fully connected layer for mapping to obtain modulation parameters. This data-driven approach replaces manual calibration. The target slope sequence and modulation parameters are input into a pre-defined encoder for attention and modulation processing to obtain a high-order road semantic feature vector. A self-attention mechanism is used to avoid iterative solutions, reducing computational load. Vehicle condition information is fused into the slope features through the modulation parameters, enabling the same slope to generate differentiated semantic features under different vehicle conditions, thus supporting different planning strategies. Finally, a pre-defined query vector, high-order road semantic feature vector, and modulation parameters are input into a pre-defined decoder for attention and modulation processing. The decoded feature vector is obtained through modulation processing. Vehicle condition constraints are continuously injected during sequence generation using modulation parameters to enhance the state coherence between adjacent nodes, thereby ensuring the smoothness of planning and the feasibility of dynamics. The decoded feature vector is input into a preset speed planning output head to obtain the target speed sequence, and the cruising speed is controlled based on the target speed sequence. This application uses a computational path composed of an encoder, decoder, and output head to replace the iterative optimization process of the online optimization algorithm, transforming the complex optimization problem into a deterministic feature transformation calculation, which significantly reduces the computational load. Furthermore, data-driven mapping replaces manual rule calibration, eliminating policy rigidity and achieving end-to-end fast response, while balancing computational efficiency and planning adaptability under complex working conditions.

[0033] Further, in one embodiment, the encoder includes a first coding layer, a second coding layer, and a third coding layer. The step of inputting the target ramp sequence and modulation parameters into a preset encoder for attention and modulation processing to obtain a high-order road semantic feature vector includes: The target ramp sequence is input into the first coding layer for semantic extraction, and the extracted features are linearly modulated by the modulation parameters to obtain the basic ramp semantic features. The basic ramp semantic features are input into the second coding layer for abstract processing, and the abstracted features are linearly modulated by the modulation parameters to obtain the context ramp semantic features. The contextual ramp semantic features are input into the third encoding layer for enhancement processing, and the enhanced features are linearly modulated using the modulation parameters to obtain higher-order ramp semantic features.

[0034] As an example, in this embodiment, the preset encoder can specifically adopt a sequence processing network based on the Transformer architecture, whose internal structure includes a first coding layer, a second coding layer, and a third coding layer cascaded in sequence; each layer contains a multi-head self-attention sublayer and a feedforward neural network sublayer, and each layer independently applies modulation parameters for feature constraints, as detailed below: First Encoding Layer (Basic Semantic Extraction): The target slope sequence is input into the first encoding layer. This layer captures the dependencies between adjacent nodes and short-distance nodes in the sequence through a multi-head self-attention mechanism, such as identifying local small slope undulations or short slopes. After completing attention calculation and feedforward network transformation, the extracted features are obtained. At this point, the scaling factor in the modulation parameters is used... and offset coefficient The extracted features are then linearly modulated. This step makes the extraction process of local slope features affected by the current vehicle conditions (such as vehicle weight). For example, under heavy load, the feature response to local steep slopes will be stronger, thus obtaining basic slope semantic features.

[0035] The second coding layer (contextual abstraction processing): The basic slope semantic features are input into the second coding layer. Building upon the first layer, this layer expands the receptive field and aggregates mid-to-long-range node information to abstract typical slope combination patterns, such as "long uphill followed by short downhill" or "continuous slope segments." Similarly, it utilizes modulation parameters corresponding to the second layer. and The abstracted features are linearly modulated. This step ensures that the recognition of ramp patterns is based not only on geometry but also on the current vehicle speed's constraint on the ability to pass through the pattern, thus obtaining contextual ramp semantic features.

[0036] The third coding layer (global semantic enhancement): The contextual slope semantic features are input into the third coding layer. This layer integrates global information from the entire forward-looking window (e.g., 300m), identifies key speed adjustment locations (e.g., top and bottom of the slope), and utilizes modulation parameters corresponding to the third layer. and The enhanced features are linearly modulated; the final output is a high-order slope semantic feature (i.e., encoded memory M), which not only contains the geometric trend of the road, but also implies the energy management strategy for the trend under the current vehicle conditions (how to accumulate speed and where to coast).

[0037] It should be understood that by using three cascaded layers with each layer modulated independently, this embodiment achieves deep integration of vehicle condition information into the road feature extraction process, avoiding the deep forgetting problem caused by vehicle condition information being spliced ​​only at the input layer in traditional solutions.

[0038] Further, in one embodiment, the decoder includes a first decoding layer and a second decoding layer. The step of inputting a preset query vector, a high-order road semantic feature vector, and modulation parameters into the preset decoder for attention and modulation processing to obtain a decoded feature vector includes: The preset query vector and the high-order road semantic feature vector are input into the first decoding layer for attention fusion processing, and the fused features are linearly modulated by the modulation parameters to obtain the basic speed adjustment intention features. The basic speed regulation intention features are input to the second decoding layer for consistency enhancement and global correction processing, and the processed features are linearly modulated using the modulation parameters to obtain the decoded feature vector.

[0039] As an example, in this embodiment, the preset decoder also adopts the Transformer architecture, including a first decoding layer and a second decoding layer cascaded in sequence, which are responsible for converting road semantics into speed planning features. The specific implementation process is as follows: First Decoding Layer (Attention Fusion and Initial Intent Exploration): The pre-defined query vector and high-order road semantic feature vector are input to the first decoding layer. This layer first performs masked multi-head self-attention calculation to establish the sequential dependency within the query sequence; then, it reads the high-order road semantic feature vector output from the encoder through a cross-attention mechanism, injecting road information into the query vector to complete the attention fusion process. The fused features represent the initial speed adjustment direction. At this point, the modulation parameters (corresponding to the first decoding layer) are used... and The fused features are linearly modulated. This step ensures that the initially generated speed adjustment intention (such as "need to accelerate") is consistent with the current vehicle power capability, thus obtaining the basic speed adjustment intention features.

[0040] Second Decoding Layer (Consistency Enhancement and Global Correction): The basic speed regulation intention features are input to the second decoding layer. This layer continues to perform mask self-attention and cross-attention calculations, specifically including the following consistency enhancement and global correction processing: (1) Consistency enhancement: By strengthening the feature association between adjacent planning nodes through the self-attention mechanism, drastic speed changes are penalized, making the output of adjacent nodes smoother and more continuous, and avoiding physically unexecutable abrupt changes in the speed sequence.

[0041] (2) Global correction: Combining the global slope information in the encoded memory, the speed change range at key locations (such as before the top of the slope) is corrected to match the fuel-saving speed adjustment rhythm of an excellent driver; subsequently, the modulation parameters (corresponding to the second layer of decoding) are used to correct the speed change range at key locations (such as before the top of the slope). and The processed features are linearly modulated, and the resulting decoded feature vector contains road global semantics, node order constraints, and current vehicle state information, preparing for the generation of high-quality speed sequences.

[0042] Furthermore, in one embodiment, before the step of inputting the target ramp sequence and modulation parameters into a preset encoder for attention and modulation processing to obtain a high-order road semantic feature vector, the method further includes: The preceding ramp sequence is input into a preset mapping layer and projected to obtain the initial ramp sequence; The initial ramp sequence is positionally encoded to obtain the target ramp sequence.

[0043] As an example, in this embodiment of the application, a rigorous preprocessing step is performed before the data enters the encoder to ensure the normalization and interpretability of the sequence data, as detailed below: (1) Construction of the forward slope sequence: Based on high-precision map data, the system takes the current position of the vehicle as the starting point and preferably extracts the road slope data within a range of 300m ahead; preferably, sampling is performed at a fixed interval of 10m, resulting in a total of 30 forward-looking nodes. Each node contains dimensions such as slope value and elevation change, forming the forward slope sequence (dimension is 30×4).

[0044] (2) Projection processing: Input the forward slope sequence into the preset mapping layer, which is a learnable linear transformation matrix that projects the original low-dimensional slope data into a high-dimensional hidden space (e.g., 128 dimensions) to obtain the initial slope sequence (30×128 dimensions). This step converts the physical quantity into a feature vector that can be processed by the neural network.

[0045] (3) Position Encoding: Since the Transformer architecture itself does not have the ability to perceive sequence order, this embodiment superimposes position encoding on the initial slope sequence. The position encoding vector is generated according to the index position of the node in the window, explicitly representing the order and relative distance of each forward-looking node. For example, even if the slope values ​​of the 1st node (at 10m) and the 30th node (at 300m) are the same, they will get different vectors after superimposing position encoding. This allows the network to distinguish the different effects of the same slope value appearing at different distance positions on speed planning (the slope at close range needs to be responded to immediately, while the slope at a distance can be planned in advance). The processed sequence is the target slope sequence, which serves as the formal input of the encoder.

[0046] Further, in one embodiment, the target speed sequence includes a relative speed sequence and a node-by-node speed increment sequence, and the control of the cruise speed based on the target speed sequence includes: The target velocity curve is determined based on the relative velocity sequence and the node-by-node velocity increment sequence; Vehicle speed control commands are generated based on the target speed curve.

[0047] As an example, in this embodiment of the application, in order to balance the accuracy and physical smoothness of speed planning, the speed planning output head adopts a dual-branch structure, which specifically includes the following steps; I. Generation of two sequences The decoded feature vector is input to the velocity planning output head (preferably a fully connected layer), and two sequences of length 30 are output in parallel: a relative velocity sequence and a node-by-node velocity increment sequence, as follows: (1) Relative speed sequence: Characterizes the target speed adjustment result (Δv) at each forward look-ahead node relative to the current reference cruise speed.

[0048] (2) Node-by-node velocity increment sequence: characterizing the velocity change trend between adjacent planning nodes ( This dual-head design stems from the training constraints of the multi-objective loss function, enabling the model to learn both the target velocity value and the physical laws governing velocity changes.

[0049] II. Determination of Target Velocity Curve The target velocity curve is determined based on the relative velocity sequence and the node-by-node velocity increment sequence. Specifically, the relative velocity sequence can be used as the main reference, and the node-by-node velocity increment sequence is used for smoothness verification and correction. For example, if the relative velocity sequence jumps too much between two points, interpolation smoothing is performed based on the constraints of the increment sequence, and finally a continuous and trackable target velocity curve is generated.

[0050] III. Generation of vehicle speed control commands: After acquiring the target speed curve, the vehicle longitudinal controller compares it with the vehicle's current actual speed and calculates the speed tracking error. Based on this speed tracking error, it generates a vehicle speed control command using a preset control algorithm (such as PID control, model predictive control MPC, or lookup table method). This command can specifically be the engine target torque, the transmission target gear, or the braking system target pressure. The actuator adjusts the vehicle's power output according to the vehicle speed control command, so that the actual vehicle speed follows the target speed curve, thereby achieving cruise control.

[0051] It should be noted that the encoder, decoder and speed planning output head preset in this application are obtained through offline training. In order to ensure that the model maintains good executability while learning the speed regulation law of ramps, this embodiment designs a multi-objective loss function for model optimization. The multi-objective loss function is composed of absolute speed loss, speed increment loss, consistency loss and boundary constraint loss.

[0052] Among them, the absolute speed loss is used to constrain the error between the speed value predicted by the model and the actual speed value of an excellent driver; the speed increment loss is used to constrain the error between the speed increment predicted by the model and the actual speed increment. This item mainly prevents the model learning pattern from shifting towards predicting a fixed speed based on the slope rather than predicting the speed change trend, and corresponds to the node-by-node speed increment sequence; the consistency loss is used to constrain the consistency of the predicted values ​​at each node in the sequence when the model predicts the speed values ​​of a long sequence, ensuring the smoothness of the planning curve; and the boundary loss is used to constrain the model prediction results to not exceed the upper and lower limits of the speed range.

[0053] It should be noted that the total loss function is as follows:

[0054] in, , , and These are the absolute velocity loss, velocity increment loss, consistency loss, and boundary constraint loss, respectively. , , and These are the corresponding weights; the preferred loss function for each term is smooth L1 loss.

[0055] Correspondingly, the formulas for each loss function are:

[0056] in, For each forward-looking node, For the predicted absolute velocity, For true absolute speed, For the predicted speed increment, For the actual speed increment, For the predicted relative velocity, This represents the number of forward-looking nodes. As the lower limit of speed, This represents the maximum speed.

[0057] It should be understood that, through the constraints of the above loss function, the weight file output after the model training is completed can ensure that the generated target speed sequence not only fits the speed values ​​of excellent drivers, but also conforms to the trend of vehicle dynamics changes, thereby supporting the cruise speed control steps in subsequent embodiments.

[0058] Further, in one embodiment, the vehicle state information includes the current longitudinal acceleration, vehicle mass, anchor point slope, and the deviation of the current vehicle speed from the reference cruise speed. The step of inputting the vehicle state information, current vehicle speed, and reference cruise speed into a preset fully connected layer for mapping to obtain modulation parameters includes: The current longitudinal acceleration, vehicle mass, anchor point slope, current vehicle speed, reference cruise speed, and the deviation of the current vehicle speed from the reference cruise speed are normalized and spliced ​​to obtain the state vector. The state vector is input to a preset fully connected layer for mapping to obtain modulation parameters, which include scaling factors and offset factors.

[0059] In an exemplary embodiment of this application, the vehicle state information includes current longitudinal acceleration, vehicle mass, anchor point slope, and the deviation of the current vehicle speed from the reference cruise speed. The current longitudinal acceleration reflects the current dynamic response state of the vehicle; vehicle mass directly affects climbing ability and braking distance, and is the key to distinguishing between unloaded and fully loaded; anchor point slope refers to the slope value at the current location of the vehicle or at a key sampling point ahead, used to provide real-time road condition context; and the deviation of the current vehicle speed from the reference cruise speed reflects the difference between the current driving state and the desired state. It should be noted that, due to the significant differences in the dimensions and numerical ranges of the aforementioned physical quantities (e.g., mass is several tons, slope is a percentage, speed is km / h), direct input would lead to difficulties in network convergence. Therefore, this embodiment first normalizes the current longitudinal acceleration, vehicle mass, anchor point slope, current vehicle speed, reference cruise speed, and the deviation of the current vehicle speed from the reference cruise speed (e.g., mapping to the 0-1 interval or standardizing to a distribution with a mean of 0 and a variance of 1). Then, the normalized scalar data are concatenated into a one-dimensional state vector in a fixed order.

[0060] It should be understood that the state vector is input to a pre-defined fully connected layer (MLP). The output dimension of this fully connected layer is designed to be large enough to cover the total number of modulation parameters required by all layers of the encoder and decoder. The mapped modulation parameters specifically include multiple sets of scaling factors (…). ) and offset coefficient ( ), each group and These coefficients correspond to a specific feature channel in the network; during the forward propagation of the network, they are used to perform a channel-by-channel linear transformation on the feature map. This allows for dynamic adjustment of the network's internal feature distribution based on vehicle status, enabling the same set of network weights to adapt to different load and speed conditions.

[0061] It should be noted that, referring to Figure 2 As shown, the input data includes the slope sequence road_tokens [H*4] (target slope sequence), vehicle state ego_state (vehicle state information), current initial speed v0 (current vehicle speed), and reference cruise speed v_ref (reference cruise speed). The input data flows to three branches: the first branch enters the decoding query initialization stage, obtains the query sequence Q0 [H*D] (a preset query vector, representing a feature matrix containing H nodes, each node having a feature dimension of D), and then inputs it into the Transformer decoder (a preset decoder), outputting the decoded feature Q. [H*D] (decoded feature vector); the second branch enters the conditional modulation parameter stage (preset fully connected layer mapping stage), obtaining (γ,B) pairs [L,2*D] (modulation parameters, L is the network layer number), which simultaneously transmits parameters to the Transformer decoder and Transformer encoder (preset encoder); the third branch enters the ramp information embedding stage, obtaining the source input sequence [H*D] (target ramp sequence) and inputting it into the Transformer encoder (preset encoder), outputting the encoded memory M [H*D] (high-order road semantic feature vector), the output results of the Transformer decoder and Transformer encoder are jointly fed into the speed planning output head (preset speed planning output head), finally generating the absolute speed sequence, i.e. the target speed sequence (including the relative speed sequence and the node-by-node speed increment sequence), completing the entire speed planning operation process.

[0062] It should be understood that in this embodiment, the training of the deep learning model is based on a high-performance computing platform. Referring to Table 1, the specific model training configuration is as follows: Table 1 Model Training Configuration

[0063] It should be noted that, as shown in Table 1, the experimental platform uses a supercomputing platform (such as the SCNet supercomputing platform), with computing resources configured as an 8-core CPU, one Nvidia 4090 graphics card (24GB of video memory), and the basic image environment is jupyterlab-conda:conda3-py3.11-cuda11.8-ubuntu22.04-devel.

[0064] Specifically, during model training, the training epochs are set to 80, the batch size to 32, and the learning rate (lr) to 1e. -3 The optimizer chosen is Adam, with a weight decay factor of 1e. -4 After model training is complete, a file containing network weights (e.g., .pth format) is output. During model validation, the weight file is loaded and inference is performed on 4185 samples in the test set. The prediction results are output as a CSV file and compared with prepared real results for evaluation. Each of the 30 forward-looking nodes in the forward-looking window contains the actual excellent driver's driving speed and the speed predicted based on this method (i.e., the target speed sequence).

[0065] Since in most real-world inference scenarios, the predicted sequence typically uses only the first node value for immediate control, this embodiment measures the average error of the predicted speed of the first node in each forward look-ahead window, calculating it to be approximately 0.093 m / s. Furthermore, to evaluate the difference between the speed planning results generated by this method and the performance of excellent drivers, this embodiment measures the mean of the cumulative travel time error for each 300m forward look-ahead window, calculating it to be 0.0828 s, approximately 0.63% of the sample's actual average travel time of 13.2494 s. This indicates that the model has a high fitting accuracy for the overall speed planning rhythm of excellent drivers, validating the effectiveness of the encoder and decoder structure.

[0066] Further, a speed planning comparison chart was drawn based on the prediction results. Figure 3 As can be seen, the speed planning curve generated by this method is consistent with the speed curve of an excellent driver in terms of main trends and key speed adjustment positions, and the predicted data and the real data maintain a high degree of consistency. It can well reproduce the speed trend of an excellent driver when facing a hill climb scenario, indicating that this proposal has good feasibility and can meet the requirements of commercial vehicle predictive cruise control for computational efficiency and planning adaptability.

[0067] It should be noted that, referring to Figure 3As shown, the multi-subplot curves with a longitudinal distance of 0 to 10000m as the unified horizontal axis sequentially display the undulation of the slope in the range of -2% to 2% and the fluctuation of the relative elevation in the range of -40m to 0m for this long road section. At the same time, it presents two curves: the actual driving speed of an excellent driver and the speed predicted by the DL (Deep Learning) model. The overall trends of the two are highly consistent, with only slight deviations in small local segments. This intuitively verifies that the output effect of the DL speed prediction model in this slope scenario is highly matched with excellent human driving behavior.

[0068] Secondly, embodiments of this application also provide a cruise speed control system.

[0069] In one embodiment, reference is made to Figure 4 , Figure 4 This is a functional module diagram of an embodiment of the cruise speed control system of this application. Figure 4 As shown, the cruise control system includes: The first processing module is used to input vehicle status information, current vehicle speed and reference cruise speed into a preset fully connected layer for mapping to obtain modulation parameters; The second processing module is used to input the target slope sequence and modulation parameters into a preset encoder for attention and modulation processing to obtain a high-order road semantic feature vector. The third processing module is used to input the preset query vector, high-order road semantic feature vector and modulation parameters into the preset decoder for attention and modulation processing to obtain the decoded feature vector; The fourth processing module is used to input the decoded feature vector into the preset speed planning output head to obtain the target speed sequence, and control the cruise speed based on the target speed sequence.

[0070] Furthermore, in one embodiment, the second processing module is specifically used for: The target ramp sequence is input into the first coding layer for semantic extraction, and the extracted features are linearly modulated by the modulation parameters to obtain the basic ramp semantic features. The basic ramp semantic features are input into the second coding layer for abstract processing, and the abstracted features are linearly modulated by the modulation parameters to obtain the context ramp semantic features. The contextual ramp semantic features are input into the third encoding layer for enhancement processing, and the enhanced features are linearly modulated using the modulation parameters to obtain higher-order ramp semantic features.

[0071] Furthermore, in one embodiment, the decoder includes a first decoding layer and a second decoding layer, and the third processing module is specifically used for: The preset query vector and the high-order road semantic feature vector are input into the first decoding layer for attention fusion processing, and the fused features are linearly modulated by the modulation parameters to obtain the basic speed adjustment intention features. The basic speed regulation intention features are input to the second decoding layer for consistency enhancement and global correction processing, and the processed features are linearly modulated using the modulation parameters to obtain the decoded feature vector.

[0072] Furthermore, in one embodiment, the second processing module is specifically used for: The preceding ramp sequence is input into a preset mapping layer and projected to obtain the initial ramp sequence; The initial ramp sequence is positionally encoded to obtain the target ramp sequence.

[0073] Furthermore, in one embodiment, the fourth processing module is specifically used for: The target velocity curve is determined based on the relative velocity sequence and the node-by-node velocity increment sequence; Vehicle speed control commands are generated based on the target speed curve.

[0074] Further, in one embodiment, the vehicle status information includes the current longitudinal acceleration, vehicle mass, anchor point slope, and the deviation of the current vehicle speed from the reference cruise speed. The first processing module is specifically used for: The current longitudinal acceleration, vehicle mass, anchor point slope, current vehicle speed, reference cruise speed, and the deviation of the current vehicle speed from the reference cruise speed are normalized and spliced ​​to obtain the state vector. The state vector is input to a preset fully connected layer for mapping to obtain modulation parameters, which include scaling factors and offset factors.

[0075] This application addresses the poor generalization problem of multi-condition strategies by inputting vehicle state information, current speed, and reference cruise speed into a pre-defined fully connected layer for mapping to obtain modulation parameters. This data-driven approach replaces manual calibration. The target slope sequence and modulation parameters are input into a pre-defined encoder for attention and modulation processing to obtain a high-order road semantic feature vector. A self-attention mechanism is used to avoid iterative solutions, reducing computational load. Vehicle condition information is fused into the slope features through the modulation parameters, enabling the same slope to generate differentiated semantic features under different vehicle conditions, thus supporting different planning strategies. Finally, a pre-defined query vector, high-order road semantic feature vector, and modulation parameters are input into a pre-defined decoder for attention and modulation processing. The decoded feature vector is obtained through modulation processing. Vehicle condition constraints are continuously injected during sequence generation using modulation parameters to enhance the state coherence between adjacent nodes, thereby ensuring the smoothness of planning and the feasibility of dynamics. The decoded feature vector is input into a preset speed planning output head to obtain the target speed sequence, and the cruising speed is controlled based on the target speed sequence. This application uses a computational path composed of an encoder, decoder, and output head to replace the iterative optimization process of the online optimization algorithm, transforming the complex optimization problem into a deterministic feature transformation calculation, which significantly reduces the computational load. Furthermore, data-driven mapping replaces manual rule calibration, eliminating policy rigidity and achieving end-to-end fast response, while balancing computational efficiency and planning adaptability under complex working conditions.

[0076] The functions of each module in the above-mentioned cruise speed control system correspond to the steps in the above-mentioned cruise speed control method embodiment, and their functions and implementation processes will not be described in detail here.

[0077] Thirdly, embodiments of this application provide a cruise speed control device, which can be a personal computer (PC), laptop computer, server, or other device with data processing capabilities.

[0078] Reference Figure 5 , Figure 5 This is a schematic diagram of the hardware structure of the cruise speed control device involved in the embodiments of this application. In the embodiments of this application, the cruise speed control device may include a processor, a memory, a communication interface, and a communication bus.

[0079] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0080] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces used to interconnect components within the cruise speed control device, as well as interfaces used to interconnect the cruise speed control device with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0081] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0082] The processor can be a general-purpose processor, which can call the cruise speed control program stored in the memory and execute the cruise speed control method provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the cruise speed control program is called can be referred to in the various embodiments of the cruise speed control method of this application, and will not be repeated here.

[0083] Those skilled in the art will understand that Figure 5 The hardware structure shown does not constitute a limitation of this application and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0084] Fourthly, embodiments of this application also provide a readable storage medium.

[0085] The present application has a readable storage medium storing a cruise speed control program, wherein when the cruise speed control program is executed by a processor, it implements the steps of the cruise speed control method described above.

[0086] The method implemented when the cruise speed control program is executed can be referred to in various embodiments of the cruise speed control method of this application, and will not be repeated here.

[0087] The terms "comprising" and "having," and any variations thereof, in the specification and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0088] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0089] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0090] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0091] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0092] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0093] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for controlling cruising speed, characterized in that, The cruise speed control method includes: The vehicle status information, current vehicle speed, and reference cruise speed are input into a preset fully connected layer for mapping to obtain modulation parameters. The target ramp sequence and modulation parameters are input into a preset encoder for attention and modulation processing to obtain a high-order road semantic feature vector; The preset query vector, high-order road semantic feature vector, and modulation parameters are input into the preset decoder for attention and modulation processing to obtain the decoded feature vector; The decoded feature vector is input into the preset speed planning output head to obtain the target speed sequence, and the cruise speed is controlled based on the target speed sequence.

2. The cruise speed control method as described in claim 1, characterized in that, The encoder includes a first coding layer, a second coding layer, and a third coding layer. The process of inputting the target slope sequence and modulation parameters into a preset encoder for attention and modulation processing to obtain a high-order road semantic feature vector includes: The target ramp sequence is input into the first coding layer for semantic extraction, and the extracted features are linearly modulated by the modulation parameters to obtain the basic ramp semantic features. The basic ramp semantic features are input into the second coding layer for abstract processing, and the abstracted features are linearly modulated by the modulation parameters to obtain the context ramp semantic features. The contextual ramp semantic features are input into the third encoding layer for enhancement processing, and the enhanced features are linearly modulated using the modulation parameters to obtain higher-order ramp semantic features.

3. The cruise speed control method as described in claim 1, characterized in that, The decoder includes a first decoding layer and a second decoding layer. The process of inputting a preset query vector, a high-order road semantic feature vector, and modulation parameters into the preset decoder for attention and modulation processing to obtain the decoded feature vector includes: The preset query vector and the high-order road semantic feature vector are input into the first decoding layer for attention fusion processing, and the fused features are linearly modulated by the modulation parameters to obtain the basic speed adjustment intention features. The basic speed regulation intention features are input to the second decoding layer for consistency enhancement and global correction processing, and the processed features are linearly modulated using the modulation parameters to obtain the decoded feature vector.

4. The cruise speed control method as described in claim 1, characterized in that, Before the step of inputting the target ramp sequence and modulation parameters into a preset encoder for attention and modulation processing to obtain a high-order road semantic feature vector, the method further includes: The preceding ramp sequence is input into a preset mapping layer and projected to obtain the initial ramp sequence; The initial ramp sequence is positionally encoded to obtain the target ramp sequence.

5. The cruise speed control method as described in claim 1, characterized in that, The target speed sequence includes a relative speed sequence and a node-by-node speed increment sequence. The method of controlling the cruise speed based on the target speed sequence includes: The target velocity curve is determined based on the relative velocity sequence and the node-by-node velocity increment sequence; Vehicle speed control commands are generated based on the target speed curve.

6. The cruise speed control method as described in claim 1, characterized in that, The vehicle status information includes the current longitudinal acceleration, vehicle mass, anchor point slope, and the deviation of the current vehicle speed from the reference cruise speed. The process of inputting the vehicle status information, current vehicle speed, and reference cruise speed into a preset fully connected layer for mapping to obtain modulation parameters includes: The current longitudinal acceleration, vehicle mass, anchor point slope, current vehicle speed, reference cruise speed, and the deviation of the current vehicle speed from the reference cruise speed are normalized and spliced ​​to obtain the state vector. The state vector is input to a preset fully connected layer for mapping to obtain modulation parameters, which include scaling factors and offset factors.

7. A cruise speed control system, characterized in that, The cruise speed control system includes: The first processing module is used to input vehicle status information, current vehicle speed and reference cruise speed into a preset fully connected layer for mapping to obtain modulation parameters; The second processing module is used to input the target slope sequence and modulation parameters into a preset encoder for attention and modulation processing to obtain a high-order road semantic feature vector. The third processing module is used to input the preset query vector, high-order road semantic feature vector and modulation parameters into the preset decoder for attention and modulation processing to obtain the decoded feature vector; The fourth processing module is used to input the decoded feature vector into the preset speed planning output head to obtain the target speed sequence, and control the cruise speed based on the target speed sequence.

8. The cruise speed control system as described in claim 7, characterized in that, The second processing module is specifically used for: The target ramp sequence is input into the first coding layer for semantic extraction, and the extracted features are linearly modulated by the modulation parameters to obtain the basic ramp semantic features. The basic ramp semantic features are input into the second coding layer for abstract processing, and the abstracted features are linearly modulated by the modulation parameters to obtain the context ramp semantic features. The contextual ramp semantic features are input into the third encoding layer for enhancement processing, and the enhanced features are linearly modulated using the modulation parameters to obtain higher-order ramp semantic features.

9. A cruise speed control device, characterized in that, The cruise speed control device includes a processor, a memory, and a cruise speed control program stored in the memory and executable by the processor, wherein when the cruise speed control program is executed by the processor, it implements the steps of the cruise speed control method as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a cruise speed control program, wherein when the cruise speed control program is executed by a processor, it implements the steps of the cruise speed control method as described in any one of claims 1 to 6.