Computing power distribution method and device of vehicle-mounted computing unit, electronic equipment and medium

Abstract

The invention provides a computing power distribution method and device for a vehicle-mounted computing unit, electronic equipment and a medium, and belongs to the technical field of intelligent driving, and the method comprises the steps: obtaining the multi-source sensor data of a vehicle in real time, and determining a scene type according to the multi-source sensor data; based on the scene type, dynamically determining and activating a target expert network from a hybrid expert architecture, and allocating corresponding computing unit computing power resources to the target expert network; and dynamically adjusting the calculation power distribution according to the real-time monitored power consumption data of the calculation unit. By adopting the method, the vehicle-mounted calculation power utilization efficiency and the energy efficiency performance are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving technology, specifically to a method, device, electronic device, and medium for allocating computing power in an onboard computing unit. Background Technology

[0002] With the rapid development of intelligent driving technology, in-vehicle computing systems need to process massive amounts of multi-source sensor data (such as cameras, LiDAR, millimeter-wave radar, etc.) and perform real-time perception, decision-making, and planning. This places extremely high demands on the computing power and energy efficiency of computing platforms. Based on this, the Mixture of Experts (MoE) architecture has been introduced into the field of in-vehicle computing. This architecture decomposes complex neural network models into multiple specialized expert sub-networks and uses gating mechanisms to sparsely activate some experts based on different input features, thereby significantly reducing the actual computational overhead of a single inference while maintaining a large model capacity. Meanwhile, Application-Specific Units (ASUs), as hardware acceleration modules customized for intelligent driving tasks, offer significantly better energy efficiency than general-purpose processors due to their high-efficiency parallel computing capabilities and streamlined instruction sets, becoming an important hardware carrier for such complex algorithm models.

[0003] In the process of developing this invention, the inventors discovered that existing methods have at least the following drawbacks: First, existing methods, based on gating routing algorithms with fixed strategies or simple thresholds, struggle to accurately match dynamically changing driving scenarios. This easily leads to a few experts being continuously occupied by high-priority tasks, while their corresponding ASU computing units remain underloaded for extended periods, resulting in low utilization of other units and wasted valuable onboard computing resources. Second, existing ASUs typically operate in fixed voltage and frequency modes. Even in low-load scenarios where only a small number of experts need to be activated across the entire vehicle system, the entire ASU still maintains near-peak power consumption, resulting in an excessively high static power consumption ratio, leading to unnecessary energy consumption and additional thermal management burdens.

[0004] In summary, current vehicle-mounted ASU computing systems based on the MoE architecture suffer from poor performance in terms of computing power utilization efficiency, system energy efficiency, and real-time performance due to the disconnect between computing power allocation and dynamic scenarios, the mismatch between power consumption control and sparse activation characteristics, and high task scheduling latency. Therefore, designing a vehicle-mounted computing resource management method that can perceive scenarios in real time, dynamically and finely allocate computing power, and achieve closed-loop power consumption control has become an urgent technical problem to be solved in this field. Summary of the Invention

[0005] In view of this, it is necessary to provide a computing power allocation method, device, electronic device and storage medium for vehicle computing units to solve the technical problems of low computing power utilization efficiency and poor energy efficiency caused by the existing static or semi-static resource allocation strategy and fixed voltage frequency.

[0006] To address the aforementioned technical problems, in a first aspect, the present invention provides a method for allocating computing power in an on-board computing unit, comprising:

[0007] Real-time acquisition of multi-source sensor data of the vehicle, extraction of scene features based on the multi-source sensor data, and determination of scene type based on the scene features; Based on the scenario type, a target expert network to be activated is determined from multiple expert networks in a hybrid expert architecture, and corresponding computing unit computing power resources are allocated to the target expert network. The allocation of computing resources is dynamically adjusted based on the power consumption data of the monitored computing units.

[0008] In one possible implementation, the step of extracting scene features based on the multi-source sensor data and determining the scene type using the scene features includes: Feature extraction is performed on the multi-source sensor data to obtain scene features, which include at least one of road curvature, obstacle density, vehicle speed, and number of traffic lights. Based on the scene characteristics, a first coefficient is calculated to quantify the scene complexity. The scene type is determined based on the value distribution range of the first coefficient, and the scene type includes simple scene, normal scene or complex scene.

[0009] In one possible implementation, determining the target expert network to be activated from multiple expert networks in a hybrid expert architecture based on the scenario type, and allocating corresponding computing unit resources to the target expert network, includes: Based on the scenario type and the preset mapping relationship, the target expert network subset to be activated and the proportion of computing resources allocated to it are determined; wherein, the complexity corresponding to the scenario type is positively correlated with the number of expert networks contained in the mapped expert network subset and the proportion of computing resources allocated to it.

[0010] In one possible implementation, determining the subset of target expert networks to be activated and the proportion of computing resources allocated to it based on the scenario type and the preset mapping relationship includes: When the scenario is classified as a simple scenario, the first subset of expert networks is activated, and a first proportion of computing power resources is allocated to the first subset of expert networks. When the scenario is classified as a normal scenario, the second subset expert network is activated, and a second proportion of computing power resources is allocated to the second subset expert network, wherein the second subset expert network contains more expert networks than the first subset expert network, and the second proportion is greater than the first proportion. When the scenario is classified as a complex scenario, all expert networks are activated, and a third proportion of computing resources is allocated to the activated expert networks, wherein the third proportion is greater than the second proportion.

[0011] In one possible implementation, the first subset of expert networks includes an image recognition expert network and a millimeter-wave radar data processing expert network; the second subset of expert networks includes an image recognition expert network and a point cloud processing expert network.

[0012] In one possible implementation, dynamically adjusting the allocation of computing resources based on the monitored power consumption data of the computing unit includes: The power consumption data is compared with a preset power consumption threshold. If the power consumption data exceeds the preset power consumption threshold, then according to the preset expert network priority, the target expert network with the lower priority among the activated target expert networks will be frequency-reduced or put into sleep mode until the power consumption data drops to within the preset power consumption threshold.

[0013] In one possible implementation, after allocating corresponding computing unit resources to the target expert network, the computing power allocation method for the on-board computing unit includes: The target expert network is invoked to process the corresponding sensor data task, resulting in fused data and decision results; The fused data and decision results are transmitted to the vehicle's control system.

[0014] On the other hand, the present invention also provides a computing power distribution device for an on-board computing unit, comprising: The scene type determination module is used to acquire multi-source sensor data of the vehicle in real time, extract scene features based on the multi-source sensor data, and determine the scene type through the scene features; The computing power resource allocation module is used to determine the target expert network to be activated from multiple expert networks of the hybrid expert architecture based on the scenario type, and allocate corresponding computing unit computing power resources to the target expert network. The resource dynamic adjustment module is used to dynamically adjust the allocation of computing resources based on the power consumption data of the monitored computing units.

[0015] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the computing power allocation method of the vehicle computing unit described in any of the above implementations.

[0016] Fourthly, the present invention also provides a computer-readable storage medium for storing a computer-readable program or instruction, which, when executed by a processor, can implement the steps in the computing power allocation method of the vehicle-mounted computing unit described in any of the above implementations.

[0017] The beneficial effects of this invention are as follows: The computing power allocation method for the vehicle computing unit provided by this invention improves the accuracy of computing power scheduling and matching with the dynamic driving environment by acquiring multi-source sensor data in real time and determining the driving scenario type accordingly; it dynamically determines and activates the target expert network from the hybrid expert architecture based on the scenario type and allocates corresponding computing power resources to achieve precise supply of computing resources according to scenario requirements; furthermore, it dynamically adjusts the computing power allocation according to the real-time monitored computing unit power consumption data to achieve real-time coordination between computing power supply and power consumption status, thereby improving the accuracy of computing power allocation of the vehicle computing unit in dynamic driving scenarios, which is conducive to improving the utilization efficiency and energy efficiency of vehicle computing power. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic flowchart of an embodiment of the computing power allocation method for the vehicle-mounted computing unit provided by the present invention; Figure 2 For the present invention Figure 1 A schematic diagram of an embodiment of S102; Figure 3 For the present invention Figure 1 A schematic diagram of an embodiment of S103; Figure 4 A schematic flowchart of another embodiment of the computing power allocation method for the vehicle-mounted computing unit provided by the present invention; Figure 5 A schematic diagram of an embodiment of the computing power distribution device for the vehicle-mounted computing unit provided by the present invention; Figure 6 A schematic diagram of an embodiment of the electronic device provided by the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] In the description of the embodiments of the present invention, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0022] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0023] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0024] Before demonstrating the embodiments, the following terms will be explained.

[0025] The Mixture of Experts (MoE) architecture is a neural network design paradigm that decomposes a large model into multiple smaller, function-specific subnetworks. In this embodiment, when processing input data, a gating network dynamically selects and activates one or a few of the most relevant experts for computation, while other experts remain dormant. This architecture decouples the total model parameter size from the actual amount of computational activation, significantly improving computational efficiency while maintaining powerful expressive capabilities.

[0026] In a hybrid expert (MoE) architecture, an expert network refers to a specialized sub-model trained to focus on handling specific types of tasks or data. For example, a MoE system for autonomous driving might include independent experts in image recognition, point cloud processing, decision planning, etc. Each expert network is responsible for feature extraction or inference tasks within its area of ​​expertise.

[0027] An Application-Specific Unit (ASU) is a hardware acceleration module custom-designed for specific types of computing tasks, such as artificial intelligence inference and image processing. Compared to a general-purpose processor (CPU), an ASU typically employs a reduced instruction set and a highly parallelized architecture, providing higher computing performance and better energy efficiency when performing its target task.

[0028] A gating network / gating mechanism is a component responsible for routing decisions in a hybrid expert (MoE) architecture. It receives input data and outputs a set of weights or probability distributions to determine which expert networks should be activated and their respective contributions. Gating networks are trainable, and their goal is to learn to match different input patterns to the most appropriate expert.

[0029] In this embodiment, computing power resources mainly refer to the computing capabilities provided by hardware computing units (such as ASUs), which are typically measured by metrics such as the number of processing cores, operating frequency, memory bandwidth, or the number of specific operations that can be executed per second. Computing power resources are finite entities that need to be scheduled and allocated to different software tasks (such as expert networks).

[0030] In this embodiment, the power consumption threshold is a pre-set safety or performance upper limit for the power consumption of the computing unit. This threshold is typically set based on hardware thermal design power, automotive-grade safety standards, or system energy efficiency targets. Real-time monitored power consumption data is compared with this threshold to trigger corresponding power consumption regulation actions.

[0031] Controller Area Network (CAN) / CAN FD bus is a highly reliable, real-time serial communication bus protocol standard widely used in automotive electronic systems. CAN FD is an upgraded version with higher data transmission rates and larger data frame capacity. In this embodiment, it is the main onboard network for transmitting processing results (fused data and decision results) from the computing unit to the vehicle control system.

[0032] The Vehicle Control Unit (VCU) is the core component of the vehicle's electronic control system. It is responsible for aggregating and coordinating information from various subsystems (such as the powertrain, chassis, and intelligent driving system) and making the highest-level vehicle control decisions and distributing commands.

[0033] Convolutional Neural Networks (CNNs) are a class of feedforward neural networks that incorporate convolutional computations and have a deep structure. They are one of the representative algorithms in deep learning. Their key feature is their ability to automatically extract spatial hierarchical features from input data (especially raster data such as images), and they are widely used in tasks such as image recognition and object detection.

[0034] Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network capable of learning long-term dependencies. By introducing gate structures (input gate, forget gate, output gate), they effectively solve the gradient vanishing or exploding problems in traditional RNNs, making them very suitable for processing time-series data, such as the analysis of continuous vehicle speed and other time-series information in this embodiment.

[0035] This invention provides a method, apparatus, electronic device, and storage medium for allocating computing power in an on-board computing unit. The technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0036] Figure 1 This is a schematic flowchart of an embodiment of the computing power allocation method for the vehicle-mounted computing unit provided by the present invention, as shown below. Figure 1 As shown, the computing power allocation method of the on-board computing unit includes: S101. Acquire multi-source sensor data of the vehicle in real time, extract scene features based on the multi-source sensor data, and determine the scene type through the scene features; S102. Based on the scenario type, determine the target expert network to be activated from multiple expert networks of the hybrid expert architecture, and allocate corresponding computing unit computing power resources to the target expert network. S103. Based on the monitored power consumption data of the computing units, dynamically adjust the allocation of computing resources.

[0037] Specifically, this embodiment provides a method for allocating computing power in an on-board computing unit, including the following steps: First, the data streams output by the multi-source sensors deployed on the vehicle are acquired in real time.

[0038] Subsequently, the multi-source sensor data is fused and semantically analyzed to identify and determine the current driving scenario category of the vehicle, obtaining a discretized scenario type identifier. This scenario type is used to characterize the features of the current driving environment, such as urban roads, highways, parking scenarios, or intersections.

[0039] Among them, multi-source sensors are a collection of sensors from at least two sources, including but not limited to lidar, cameras and millimeter-wave radar, etc. The data from multi-source heterogeneous sensors together constitute real-time perception information of the vehicle's surrounding environment.

[0040] The methods for fusing multi-source sensor data include, but are not limited to: data-level fusion (or early fusion), which involves aligning, registering, and directly stitching raw data from different sensors to form a unified enhanced data representation; feature-level fusion, which involves extracting feature vectors or feature maps from various sensor data and then stitching, weighting, or fusing them based on attention mechanisms; and decision-level fusion, which involves each sensor or processing channel independently performing preliminary identification or judgment, and then voting, weighting, or synthesizing the output decision results based on confidence. Specific fusion architectures and techniques can include Kalman filtering, Bayesian networks, Dempster-Shafer evidence theory, convolutional neural network feature stitching, and Transformer-based cross-modal attention fusion models.

[0041] The methods for performing semantic analysis on multi-source sensor data include, but are not limited to: Deep learning-based multimodal semantic understanding models, such as using convolutional neural networks (CNNs) to extract visual features, using point cloud networks to process LiDAR data, and establishing semantic associations between different modal features through cross-modal attention mechanisms (such as the Transformer architecture), can then complete high-level tasks such as scene understanding and target intent prediction. The reasoning method based on knowledge graphs and semantic rules constructs entities, attributes and their relationships in traffic scenarios into a graph, and uses preset semantic rules or logic to perform consistency verification and contextual reasoning on multi-source data. The fusion analysis of semantic segmentation and instance segmentation involves performing pixel-level semantic segmentation on camera images and instance segmentation on LiDAR point clouds, and aligning and fusing the results of the two in space to obtain accurate 3D environmental structure information with semantic labels. End-to-end joint perception and decision-making models directly input raw or pre-processed multi-source sensor data into a unified deep neural network. This network automatically learns feature extraction, fusion, and high-level semantic parsing through internal modules, and outputs semantic information directly used for decision-making, such as drivable areas, obstacle behavior semantics, and scene classification.

[0042] After determining the scenario type, based on a pre-defined Mixture of Experts (MoE) architecture, one or more sub-networks most relevant to the current scenario type are selected as the target expert network from multiple pre-trained expert networks. The Mixture of Experts architecture contains multiple functionally specialized and structurally independent neural network sub-models, each optimized for a specific type of driving task or environmental feature.

[0043] The specific implementation of selecting one or more sub-networks most relevant to the current scene type from multiple pre-trained expert networks as the target expert network includes: using scene type identifiers obtained from a trainable gating network (or scene feature vectors extracted directly from sensor data) as input, and calculating weights within the gating network to output a corresponding weight score or activation probability for each pre-trained expert network in the architecture. Then, based on these weights (e.g., selecting the top-k experts with the highest weights, or selecting experts with weights exceeding a threshold), one or more sub-networks most relevant to the current scene are determined and activated as the target expert network.

[0044] Next, based on the selected target expert network, the computing resources in the vehicle computing unit are dynamically allocated, so that the selected expert network obtains processing power that matches its task priority and computational complexity, while the unselected expert network is not allocated or is allocated only a very low amount of computing resources, thereby realizing on-demand scheduling and efficient utilization of computing power.

[0045] Furthermore, this embodiment also introduces a power consumption closed-loop control mechanism: continuously monitor the real-time power consumption of the computing unit (such as the dedicated AI acceleration unit ASU) that performs the target expert network computing task; if the current power consumption exceeds the preset threshold or the system energy efficiency target, a computing power reallocation operation is triggered, including but not limited to reducing the operating frequency of non-critical expert networks, limiting their activation scale, or temporarily placing them in a low-power sleep state, so as to control the overall power consumption of the system within a reasonable range while ensuring the performance of core perception and decision-making tasks.

[0046] This embodiment improves the accuracy of computing power scheduling and matching with the dynamic driving environment by acquiring multi-source sensor data in real time and determining the driving scenario type accordingly. Based on the scenario type, it dynamically determines and activates the target expert network from the hybrid expert architecture and allocates corresponding computing power resources to achieve precise supply of computing resources according to scenario requirements. Furthermore, it dynamically adjusts the computing power allocation based on real-time monitored computing unit power consumption data to achieve real-time coordination between computing power supply and power consumption status. This improves the accuracy of computing power allocation of the vehicle computing unit in dynamic driving scenarios and is conducive to improving the utilization efficiency and energy efficiency of vehicle computing power.

[0047] In some embodiments of the present invention, such as Figure 2 As shown, step S101 extracts scene features based on multi-source sensor data, and determines the scene type through the scene features, including: S201. Extract features from multi-source sensor data to obtain scene features, which include at least one of road curvature, obstacle density, vehicle speed and number of traffic lights. S202. Based on scene characteristics, calculate the first coefficient used to quantify scene complexity; S203. Determine the scene type based on the value distribution range of the first coefficient. The scene type includes simple scene, normal scene, or complex scene.

[0048] First, the real-time acquired multi-source heterogeneous sensor data is preprocessed and its features are extracted.

[0049] The multi-source sensor data includes point cloud data output by LiDAR, image data acquired by cameras, and signal data returned by millimeter-wave radar. Through appropriate signal processing and computer vision algorithms, a set of scene features characterizing the current driving environment is extracted from the aforementioned raw data.

[0050] The scene features include, but are not limited to: road curvature (reflecting road geometry), obstacle density (characterizing the density of surrounding traffic participants), the vehicle's current real-time speed, and the number of identifiable traffic lights within the field of vision. These features together constitute an objective, structured digital description of the current driving scene.

[0051] After obtaining the scene features, they are input into a pre-defined or trainable computational model to generate a scalarized output value, denoted as the first coefficient. This first coefficient is used to quantify the overall complexity of the current driving scene.

[0052] The computational model can be implemented in various forms, including but not limited to: linear models based on feature weighted summation, regression networks based on multilayer perceptrons (MLP), and hybrid neural network models using CNN-LSTM.

[0053] Preferably, this embodiment employs a CNN-LSTM hybrid neural network model to calculate the first coefficient. The convolutional neural network (CNN) submodule is used to extract local and global spatial features from spatial data such as images or rasterized point clouds; the long short-term memory network (LSTM) submodule is used to model time-dependent temporal features such as vehicle speed and the dynamic distribution of obstacles. By fusing information from both spatial and temporal dimensions, this hybrid model can more accurately assess the dynamic complexity of driving scenarios.

[0054] Subsequently, based on the value of the first coefficient and referring to the pre-set mapping rules, it is mapped to the corresponding discrete scene type.

[0055] In a specific example, the range of the first coefficient (0, 1) is divided into multiple consecutive and mutually exclusive intervals, and each interval is associated with a scenario type. For example: when the first coefficient is in the interval (0, 0.3), the current scenario is determined to be a simple scenario, which includes low-interaction, low-change environments such as highway cruising; when the first coefficient is in the interval [0.3, 0.7), it is determined to be a normal scenario, corresponding to medium-complexity conditions such as driving on urban expressways or main roads; when the first coefficient is in the interval [0.7, 1), it is determined to be a complex scenario, which includes congested road sections in urban centers, unprotected left-turn intersections, or areas with dense interaction among multiple traffic participants.

[0056] This embodiment improves the robustness and accuracy of scene recognition by fusing multi-dimensional features from multiple sources of sensors, effectively avoiding misjudgments caused by the failure of a single sensor or the one-sidedness of a single indicator. At the same time, by transforming continuous complexity indicators into discrete scene type identifiers through a predefined quantization range, it provides objective and accurate decision input for subsequent scene type-based differentiated computing power scheduling strategies, thereby enhancing the adaptability and response consistency of the entire vehicle computing system to dynamic environmental changes.

[0057] In some embodiments of the present invention, based on scenario type, determining the target expert network to be activated from multiple expert networks in a hybrid expert architecture, and allocating corresponding computing unit resources to the target expert network includes: Based on the mapping relationship between scenario type and preset, the target expert network subset to be activated and the proportion of computing resources allocated to it are determined; among them, the complexity corresponding to the scenario type is positively correlated with the number of expert networks contained in the mapped expert network subset and the proportion of computing resources allocated.

[0058] In some embodiments of the present invention, determining the subset of target expert networks to be activated and the proportion of computing resources allocated to it based on the scene type and the preset mapping relationship includes: When the scene is classified as a simple scene, the first subset of expert networks is activated, and the first proportion of computing power resources is allocated to the first subset of expert networks. When the scene is classified as a normal scene, the second subset of expert networks is activated, and a second proportion of computing resources is allocated to the second subset of expert networks. The second subset of expert networks contains more expert networks than the first subset of expert networks, and the second proportion is greater than the first proportion. When the scene is classified as a complex scene, all expert networks are activated, and the activated expert networks are allocated a third proportion of computing resources, where the third proportion is greater than the second proportion.

[0059] In some embodiments of the present invention, the first subset of expert networks includes an image recognition expert network and a millimeter-wave radar data processing expert network; the second subset of expert networks includes an image recognition expert network and a point cloud processing expert network.

[0060] This embodiment is a specific implementation of step S102, which involves determining the target expert network to be activated from multiple expert networks in a hybrid expert architecture based on scenario type, and allocating corresponding computing unit resources to the target expert network. This method relies on a pre-configured mapping relationship. This mapping relationship clarifies the specific combinations of expert networks (i.e., target expert network subsets) required to be activated for different scenario types (e.g., simple, normal, complex) and the proportion of computing resources allocated to that subset relative to the total available computing power. The core principle of this mapping relationship is that the higher the complexity of the scenario type, the more expert networks are required to work simultaneously, and the higher the proportion of computing resources reserved for it; the two are positively correlated, ensuring that the intensity of computing resource investment matches the scenario processing requirements.

[0061] Specifically, the execution of the mapping relationship can be manifested as a condition-triggered strategy.

[0062] For example, when the scenario type is determined to be a simple scenario, a minimum necessary set, namely the first subset of expert networks, is activated according to the mapping relationship, and a relatively low proportion of computing power resources is allocated to it, such as 20% to 30% of the total computing power. When the scenario type is upgraded to a normal scenario, a more feature-rich second subset of expert networks is activated, which contains more expert networks than the first subset, and a medium level of computing power resources is allocated to it, such as 50% to 60%. When facing a complex scenario, the mapping relationship will instruct the activation of all or nearly all expert networks to cope with highly challenging tasks, and a high proportion or even all of the computing power resources will be allocated to it, such as 80% to 100%, while enabling computing power overclocking mode to ensure performance.

[0063] For ease of understanding, the aforementioned subset of expert networks can be illustrated by example. In one specific implementation, the first subset of expert networks may include an image recognition expert network responsible for processing camera data, and a millimeter-wave radar data processing expert network for processing millimeter-wave radar signals. This is sufficient to handle scenarios with low requirements for perception redundancy, such as high-speed cruising. The second subset of expert networks can, based on the above, replace or add a point cloud processing expert network that specifically processes lidar point cloud data, to adapt to scenarios with higher requirements for three-dimensional geometric perception, such as urban expressways.

[0064] It should be understood that the expert network in this embodiment is a pre-trained deep learning model oriented towards a specific perception or decision-making task. For example, an image recognition expert may implement object detection based on a convolutional neural network (CNN), and a point cloud processing expert may implement 3D object recognition based on architectures such as PointNet or VoxelNet.

[0065] This embodiment achieves precise matching between computing resource supply and scenario computing needs through the aforementioned refined scheduling method based on preset mapping relationships: In low-complexity scenarios, by activating only necessary experts and allocating lower computing power, static power consumption and ineffective computing overhead are significantly reduced, improving system energy efficiency; in high-complexity scenarios, by fully activating the expert network and ensuring a high supply of computing power, the accuracy and real-time performance of perception and decision-making are ensured. This fundamentally optimizes the resource utilization efficiency and task assurance capabilities of the onboard computing unit in complex and ever-changing driving environments.

[0066] In some embodiments of the present invention, such as Figure 3 As shown, step S103, which dynamically adjusts the allocation of computing resources based on the monitored power consumption data of the computing units, includes: S301. Compare the power consumption data with the preset power consumption threshold; S302. If the power consumption data exceeds the preset power consumption threshold, then according to the preset expert network priority, the target expert network with the lower priority among the activated target expert networks will be frequency reduced or put into sleep mode until the power consumption data drops to within the preset power consumption threshold.

[0067] This embodiment provides a specific implementation method for dynamically adjusting the allocation of computing resources based on the monitored power consumption data of the computing unit in step S103. By constructing an independent power consumption closed-loop control loop, it ensures that the vehicle computing system always meets the preset power consumption constraints during operation.

[0068] Specifically, real-time power consumption data is continuously collected from the power management module, performance monitoring unit, or hardware counter of the computing unit (e.g., a dedicated AI acceleration unit, ASU) that performs the computing task. This power consumption data reflects the total power consumed by all currently active target expert networks.

[0069] To achieve effective control, this embodiment presets a power consumption threshold as a benchmark for determining whether to trigger the adjustment mechanism. This preset power consumption threshold can be pre-set based on the thermal design power, heat dissipation capacity, or vehicle energy management strategy of the automotive-grade chip, for example, 150W. Real-time power consumption data is periodically compared with this preset threshold, and a corresponding comparison result signal is generated.

[0070] When the comparison results indicate that the real-time power consumption exceeds a preset power consumption threshold, a dynamic adjustment mechanism for computing resources is automatically activated. This mechanism implements fine-grained control based on a preset expert network priority strategy.

[0071] It should be noted that the expert network priority list is pre-set, and the setting rule is to assign high, medium, and low priority levels to each expert network based on its criticality to perception, decision-making, or control tasks in different driving scenarios. As a specific example of priority setting, a core expert network list is preset for each scenario type. When power consumption adjustment is triggered, expert networks outside the core list for the current scenario type are considered to have lower priority. For example, for complex scenarios, the core list may include image recognition experts and point cloud processing experts; if power consumption exceeds the limit, decision planning experts or millimeter-wave radar data processing experts will be prioritized for frequency reduction or sleep mode.

[0072] After dynamic adjustment is triggered, power consumption suppression is first performed on the lowest priority expert in the currently activated target expert network.

[0073] The power suppression operations include frequency reduction or sleep mode. Frequency reduction reduces the operating frequency and power supply voltage of the computing cores used by the expert network through dynamic voltage and frequency adjustment technology. Sleep mode suspends the inference task of the expert network and puts the computing resources it occupies into a low-power standby state.

[0074] The adjustment process is carried out iteratively. After each frequency reduction or sleep operation, the current power consumption data is collected again and compared with the preset threshold. If the power consumption still exceeds the threshold, the same type of regulation operation is performed on the expert network of the next priority until the real-time power consumption falls back to within the preset threshold.

[0075] The closed-loop control method adopted in this embodiment effectively prevents the computing unit from overheating or degrading due to continuous overload operation, and improves thermal stability and operational reliability under high load conditions. At the same time, the hierarchical control strategy based on expert network priority prioritizes the computing resources of key algorithm modules while meeting power consumption constraints, maximizes the real-time performance and accuracy of core functions, avoids the vehicle being in a high power consumption state for a long time, and helps to reduce the overall vehicle energy consumption and extend the service life of the on-board computing hardware.

[0076] In some embodiments of the present invention, such as Figure 4 As shown, after allocating corresponding computing unit resources to the target expert network, the computing power allocation method for this vehicle-mounted computing unit includes: S401. Call the target expert network to process the corresponding sensor data task and obtain fused data and decision results; S402, Transmit the fused data and decision results to the vehicle's control system.

[0077] After selecting the target expert network and allocating computing resources according to the scenario type, the data processing stage begins, invoking all activated and allocated target expert networks to execute their respective specialized processing tasks.

[0078] The system comprises three expert networks: an image recognition expert network for target detection and semantic segmentation of visual data acquired by cameras; a point cloud processing expert network for clustering, segmentation, and geometric feature extraction of 3D point cloud data output by LiDAR; and a millimeter-wave radar data processing expert network for analyzing radar echo signals to obtain target distance, velocity, and angle information. Each expert network performs targeted and in-depth processing of its corresponding type of raw sensor data based on its functional characteristics, operating in parallel or collaborative modes.

[0079] After each expert network completes its specific processing, the resulting heterogeneous intermediate results are fused and aggregated to generate a unified output with high-level semantic information. This output includes two categories: fused data and decision results.

[0080] Among them, fused data is a consistent representation of the environment surrounding the vehicle, such as a multi-source fused obstacle list containing information such as location, category, and motion state, or a local high-precision environmental map built based on multiple sensors.

[0081] The decision-making results are strategic outputs for vehicle control, including planned driving trajectories, target speeds, lane-changing intentions, or emergency braking commands. These results are generated collaboratively by an expert network, realizing the transformation from raw perception data to executable driving semantics.

[0082] Subsequently, the fused data and decision results are encapsulated into structured data frames that conform to in-vehicle communication protocols (such as the AUTOSAR standard) and transmitted to the vehicle control system via the in-vehicle internal communication network.

[0083] In a specific example, a Controller Area Network (CAN) or its high-bandwidth extension, the CAN FD bus, is used to send data frames to the vehicle controller or relevant domain controllers. This transmission mechanism ensures that the perception and decision-making results generated by intelligent computing scheduling can be received by the underlying execution units in a timely manner and used to drive vehicle control actions such as steering, braking, and acceleration, thereby forming a complete functional closed loop from environmental perception, intelligent computing, decision generation to control execution.

[0084] This embodiment improves the accuracy and efficiency of multi-source data parsing through collaborative processing of expert networks. At the same time, it achieves efficient connection between upper-layer dynamic computing power scheduling and lower-layer stable control execution through standardized data encapsulation and vehicle bus transmission mechanism, ensuring the real-time performance and consistency of the intelligent driving system in complex scenarios.

[0085] To better implement the computing power allocation method for the vehicle-mounted computing unit in this embodiment of the invention, based on the computing power allocation method for the vehicle-mounted computing unit, correspondingly, as follows: Figure 5 As shown, this embodiment of the invention also provides a computing power allocation device for an on-board computing unit. The computing power allocation device 500 for the on-board computing unit includes: The scene type determination module 501 is used to acquire multi-source sensor data of the vehicle in real time, extract scene features based on the multi-source sensor data, and determine the scene type through the scene features; The computing power resource allocation module 502 is used to determine the target expert network to be activated from multiple expert networks of the hybrid expert architecture based on the scenario type, and allocate corresponding computing unit computing power resources to the target expert network. The resource dynamic adjustment module 503 is used to dynamically adjust the allocation of computing resources based on the power consumption data of the monitored computing units.

[0086] The computing power allocation device 500 of the vehicle computing unit provided in the above embodiments can realize the technical solutions described in the above embodiments of the computing power allocation method of the vehicle computing unit. The specific implementation principles of each module or unit can be found in the corresponding content in the above embodiments of the computing power allocation method of the vehicle computing unit, and will not be repeated here.

[0087] like Figure 6 As shown, the present invention also provides an electronic device 600. The electronic device 600 includes a processor 601, a memory 602, and a display 603. Figure 6 Only some components of the electronic device 600 are shown, but it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

[0088] In some embodiments, processor 601 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 602 or process data, such as the computing power allocation method of the vehicle-mounted computing unit in this invention.

[0089] In some embodiments, processor 601 may be a single server or a group of servers. The server group may be centralized or distributed. In some embodiments, processor 601 may be local or remote. In some embodiments, processor 601 may be implemented on a cloud platform. In one embodiment, the cloud platform may include a private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, intranet, multi-cloud, etc., or any combination thereof.

[0090] In some embodiments, memory 602 may be an internal storage unit of electronic device 600, such as a hard disk or memory of electronic device 600. In other embodiments, memory 602 may also be an external storage device of electronic device 600, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on electronic device 600.

[0091] Furthermore, the memory 602 may include both internal storage units of the electronic device 600 and external storage devices. The memory 602 is used to store application software and various types of data installed on the electronic device 600.

[0092] In some embodiments, display 603 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 603 is used to display information from electronic device 600 and to display a visual user interface. Components 601-603 of electronic device 600 communicate with each other via a system bus.

[0093] In one embodiment, when the processor 601 executes the computing power allocation program of the on-board computing unit in the memory 602, the following steps can be implemented: Real-time acquisition of multi-source sensor data of the vehicle, extraction of scene features based on the multi-source sensor data, and determination of scene type based on the scene features; Based on the scenario type, a target expert network to be activated is determined from multiple expert networks in a hybrid expert architecture, and corresponding computing unit computing power resources are allocated to the target expert network. The allocation of computing resources is dynamically adjusted based on the power consumption data of the monitored computing units.

[0094] It should be understood that when the processor 601 executes the computing power allocation program of the vehicle computing unit in the memory 602, in addition to the functions mentioned above, it can also perform other functions, as can be found in the description of the corresponding method embodiments above.

[0095] Furthermore, this embodiment of the invention does not specifically limit the type of electronic device 600 mentioned. Electronic device 600 can be a mobile phone, tablet computer, personal digital assistant (PDA), wearable device, laptop computer, or other portable electronic device. Exemplary embodiments of portable electronic devices include, but are not limited to, portable electronic devices running iOS, Android, Microsoft, or other operating systems. The aforementioned portable electronic device can also be other portable electronic devices, such as a laptop computer with a touch-sensitive surface (e.g., a touch panel). It should also be understood that in some other embodiments of the invention, electronic device 600 may not be a portable electronic device, but rather a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0096] Accordingly, this application also provides a computer-readable storage medium for storing computer-readable programs or instructions. When the programs or instructions are executed by a processor, they can implement the steps or functions of the computing power allocation method of the vehicle computing unit provided in the above-described method embodiments.

[0097] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.), and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0098] The above provides a detailed description of the computing power allocation method, device, electronic equipment, and storage medium of the vehicle-mounted computing unit provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for allocating computing power in an onboard computing unit, characterized in that, include: Real-time acquisition of multi-source sensor data of the vehicle, extraction of scene features based on the multi-source sensor data, and determination of scene type based on the scene features; Based on the scenario type, a target expert network to be activated is determined from multiple expert networks in the hybrid expert architecture, and corresponding computing unit computing power resources are allocated to the target expert network. The hybrid expert architecture is a neural network architecture that dynamically selects some expert networks for computation for each input through a gating network. The allocation of computing resources is dynamically adjusted based on the power consumption data of the monitored computing units.

2. The computing power allocation method for the vehicle-mounted computing unit according to claim 1, characterized in that, The step of extracting scene features based on the multi-source sensor data and determining the scene type using the scene features includes: Feature extraction is performed on the multi-source sensor data to obtain scene features, which include at least one of road curvature, obstacle density, vehicle speed, and number of traffic lights. Based on the scene characteristics, a first coefficient is calculated to quantify the scene complexity. The scene type is determined based on the value distribution range of the first coefficient, and the scene type includes simple scene, normal scene or complex scene.

3. The computing power allocation method for the vehicle-mounted computing unit according to claim 2, characterized in that, The step of determining the target expert network to be activated from multiple expert networks in a hybrid expert architecture based on the scenario type, and allocating corresponding computing unit resources to the target expert network includes: Based on the scenario type and the preset mapping relationship, the target expert network subset to be activated and the proportion of computing resources allocated to it are determined; wherein, the complexity corresponding to the scenario type is positively correlated with the number of expert networks contained in the mapped expert network subset and the proportion of computing resources allocated to it.

4. The computing power allocation method for the vehicle-mounted computing unit according to claim 3, characterized in that, The process of determining the subset of target expert networks to be activated and the proportion of computing resources allocated to it based on the scenario type and the preset mapping relationship includes: When the scenario is classified as a simple scenario, the first subset of expert networks is activated, and a first proportion of computing power resources is allocated to the first subset of expert networks. When the scenario is classified as a normal scenario, the second subset expert network is activated, and a second proportion of computing power resources is allocated to the second subset expert network, wherein the second subset expert network contains more expert networks than the first subset expert network, and the second proportion is greater than the first proportion. When the scenario is classified as a complex scenario, all expert networks are activated, and a third proportion of computing resources is allocated to the activated expert networks, wherein the third proportion is greater than the second proportion.

5. The computing power allocation method for the vehicle-mounted computing unit according to claim 4, characterized in that, The first subset of expert networks includes an image recognition expert network and a millimeter-wave radar data processing expert network; the second subset of expert networks includes an image recognition expert network and a point cloud processing expert network.

6. The computing power allocation method for the vehicle-mounted computing unit according to any one of claims 1 to 5, characterized in that, The dynamic adjustment of the allocation of computing resources based on the monitored power consumption data of the computing unit includes: The power consumption data is compared with a preset power consumption threshold. If the power consumption data exceeds the preset power consumption threshold, then according to the preset expert network priority, the target expert network with the lower priority among the activated target expert networks will be frequency-reduced or put into sleep mode until the power consumption data drops to within the preset power consumption threshold.

7. The computing power allocation method for the vehicle-mounted computing unit according to any one of claims 1 to 5, characterized in that, After allocating corresponding computing unit resources to the target expert network, the computing power allocation method for the vehicle-mounted computing unit includes: The target expert network is invoked to process the corresponding sensor data task, resulting in fused data and decision results; The fused data and decision results are transmitted to the vehicle's control system.

8. A computing power distribution device for an on-board computing unit, characterized in that, include: The scene type determination module is used to acquire multi-source sensor data of the vehicle in real time, extract scene features based on the multi-source sensor data, and determine the scene type through the scene features; The computing power resource allocation module is used to determine the target expert network to be activated from multiple expert networks of the hybrid expert architecture based on the scenario type, and allocate corresponding computing unit computing power resources to the target expert network. The resource dynamic adjustment module is used to dynamically adjust the allocation of computing resources based on the power consumption data of the monitored computing units.

9. An electronic device, characterized in that, Including memory and processor, among which, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the computing power allocation method of the vehicle computing unit according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the steps in the computing power allocation method of the vehicle-mounted computing unit as described in any one of claims 1 to 7.

Images