Processing method and device

Abstract

The invention discloses a processing method and device, and the method comprises the steps: driving second agents to execute a target task based on a current first agent, and enabling the first agent to coordinate the operation of all second agents based on the target task; the second intelligent agent is used for performing fusion processing on data acquired by a plurality of data acquisition devices of the electronic equipment based on a target data fusion strategy to obtain a target feature, and adjusting the running state of the electronic equipment based on the target feature; obtaining feedback data generated when the second intelligent agent executes the target task; and adjusting the current first intelligent agent based on the feedback data, taking the adjusted first intelligent agent as a new current first intelligent agent, and continuing to execute the step of driving the second intelligent agent to execute the target task based on the current first intelligent agent.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a processing method and apparatus. Background Technology

[0002] Currently, multi-sensor data fusion technology is crucial in intelligent agent systems of electronic devices due to its ability to integrate heterogeneous data from multiple sources and provide comprehensive and accurate information analysis results. However, how to improve the performance and intelligence level of intelligent agent systems based on multi-sensor data fusion technology remains a challenge. Summary of the Invention

[0003] The technical solution provided in this application is as follows:

[0004] The first aspect of this application provides a processing method, including:

[0005] The first intelligent agent drives the second intelligent agent to execute the target task. The first intelligent agent is used to coordinate the operation of all second intelligent agents based on the target task. The second intelligent agent is used to fuse data collected by multiple data acquisition devices of the electronic device based on the target data fusion strategy to obtain target features, and adjust the operating state of the electronic device based on the target features.

[0006] Obtain the feedback data generated by the second intelligent agent in performing the target task;

[0007] Based on the feedback data, the current first agent is adjusted, and the adjusted first agent is used as the new current first agent. The step of driving the second agent to perform the target task based on the current first agent continues.

[0008] In one possible implementation, obtaining the feedback data generated by the second agent performing the target task includes at least one of the following:

[0009] Obtain first feedback data; the first feedback data includes the adjusted operating status of the electronic device;

[0010] Obtain second feedback data; the second feedback data is used to reflect the state changes of the data objects to be operated by the second intelligent agent and the state of the second intelligent agent's occupation of computing resources;

[0011] Obtain third feedback data; the third feedback data is used to reflect the performance of the first intelligent agent.

[0012] In one possible implementation, adjusting the current first agent based on the feedback data includes:

[0013] Based on the feedback data, adjust at least one of the following: the current first intelligent agent's internal configuration, behavioral strategy, and target model;

[0014] The internal configuration is used to characterize the computing environment required for the current first intelligent agent to run;

[0015] The behavioral strategy is used to define task allocation behavior and resource scheduling behavior;

[0016] The target model is used for task allocation and / or resource scheduling.

[0017] In one possible implementation, the second intelligent agent, based on a target data fusion strategy, fuses data collected by multiple data acquisition devices of the electronic device to obtain target features, and adjusts the operating state of the electronic device based on the target features, including:

[0018] The second intelligent agent, based on a target data fusion strategy, fuses data collected by multiple data acquisition devices of the electronic device to obtain target features;

[0019] The second intelligent agent acquires event rule description information and performs reasoning on the event rule description information and the target features based on the target model to obtain a reasoning result; the reasoning result represents the current usage scenario of the electronic device; the event rule description information is used to describe the logical relationship between different events;

[0020] Based on the reasoning results, the second intelligent agent adjusts the operating state of the electronic device.

[0021] In one possible implementation, the processing method further includes:

[0022] The second agent searches for a template in the template library that matches the target features; the template represents a candidate use scenario for the electronic device.

[0023] The second intelligent agent adjusts the operating state of the electronic device based on the found template;

[0024] If the second agent does not find a template matching the target feature in the template library, it obtains event rule description information and infers the event rule description information and the target feature based on the target model to obtain the inference result.

[0025] In one possible implementation, the processing method further includes:

[0026] The second intelligent agent adjusts the target data fusion strategy based on the adjusted operating state of the electronic device.

[0027] In one possible implementation, the step of fusing data collected by multiple data acquisition devices based on the target data fusion strategy to obtain target features includes:

[0028] Extract the features of the data collected by each data acquisition device to obtain the features corresponding to each data acquisition device;

[0029] The features corresponding to each of the data acquisition devices are enhanced to obtain the enhanced features corresponding to each of the data acquisition devices; the enhanced features represent information related to the data in the data acquired by the data acquisition devices and the data acquired by other data acquisition devices.

[0030] Determine the temporal and spatial relationships between the enhanced features corresponding to each of the data acquisition devices;

[0031] Based on the aforementioned correlation, the enhanced features corresponding to each of the data acquisition devices are fused to obtain the target features.

[0032] In one possible implementation, the step of fusing the enhanced features corresponding to each of the data acquisition devices based on the association relationship to obtain the target features includes:

[0033] Based on the aforementioned correlation, a first weighting factor is determined for the enhanced features corresponding to each of the data acquisition devices.

[0034] Based on the quality parameters of the enhanced features corresponding to each of the data acquisition devices, a second weighting factor is determined for each of the data acquisition devices.

[0035] Based on the second weight factor corresponding to each of the data acquisition devices, update the first weight factor of the enhanced feature corresponding to each of the data acquisition devices.

[0036] Based on the updated first weight factor of the enhanced features corresponding to each of the data acquisition devices, the enhanced features corresponding to each of the data acquisition devices are weighted to obtain the target features.

[0037] In one possible implementation, before fusing data collected by multiple data acquisition devices based on the target data fusion strategy to obtain the target features, at least one of the following is also included:

[0038] Data collected by multiple data acquisition devices is processed to obtain time-aligned and / or spatially aligned data from the multiple data acquisition devices.

[0039] Determine whether the time-aligned and / or spatially-aligned data from the plurality of data acquisition devices meet the adjustment conditions;

[0040] If so, adjust the length of the sliding window;

[0041] Using the adjusted sliding window as a unit, the time-aligned and / or spatially aligned data of the multiple data acquisition devices are adjusted in terms of time dimension and / or spatial dimension. The step of determining whether the time-aligned and / or spatially aligned data of the multiple data acquisition devices meet the adjustment conditions is continued until the target data of the multiple data acquisition devices is obtained.

[0042] If not, using the sliding window as a unit, adjust the time-aligned and / or spatial-aligned data of the multiple data acquisition devices in terms of time and / or spatial dimensions, and continue to execute the step of determining whether the time-aligned and / or spatial-aligned data of the multiple data acquisition devices meet the adjustment conditions, until the target data of the multiple data acquisition devices is obtained.

[0043] Another aspect of this application provides a processing apparatus, comprising:

[0044] The driving unit is used to drive the second intelligent agent to execute the target task based on the current first intelligent agent. The first intelligent agent is used to coordinate the operation of all second intelligent agents. The second intelligent agent is used to fuse data collected by multiple data acquisition devices of the electronic device based on the target data fusion strategy to obtain target features, and adjust the operating state of the electronic device based on the target features.

[0045] The acquisition unit is used to acquire feedback data generated by the second intelligent agent in performing the target task;

[0046] The adjustment unit is used to adjust the current first agent based on the feedback data, take the adjusted first agent as the new current first agent, and continue to execute the step of driving the second agent to perform the target task based on the current first agent. Attached Figure Description

[0047] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0048] Figure 1 A flowchart illustrating a processing method provided in Embodiment 1 of this application;

[0049] Figure 2 An interactive diagram of a first intelligent agent and a second intelligent agent provided in this application;

[0050] Figure 3This is a schematic diagram of the interaction between a first intelligent agent and a second intelligent agent provided in Embodiment 4 of this application;

[0051] Figure 4 This application provides a schematic diagram of a data processing and fusion scenario.

[0052] Figure 5 This is a schematic diagram of the interaction between a first intelligent agent and a second intelligent agent provided in Embodiment 5 of this application;

[0053] Figure 6 A schematic diagram illustrating another scenario of data processing and fusion provided for this application;

[0054] Figure 7 A schematic diagram illustrating another scenario of data processing and fusion provided for this application;

[0055] Figure 8 This is a schematic diagram illustrating another scenario of data processing and fusion provided in this application. Detailed Implementation

[0056] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0057] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0058] The terms "first," "second," etc., used in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of units is not necessarily limited to those units, but may include other units not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0059] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] Reference Figure 1 This is a flowchart illustrating a processing method provided in Embodiment 1 of this application, as shown below. Figure 1 As shown, the method may include, but is not limited to, the following steps:

[0061] Step S101: Drive the second intelligent agent to execute the target task based on the current first intelligent agent. The first intelligent agent is used to coordinate the operation of all second intelligent agents based on the target task. The second intelligent agent is used to fuse data collected by multiple data acquisition devices of the electronic device based on the target data fusion strategy to obtain target features, and adjust the operating state of the electronic device based on the target features.

[0062] In this embodiment, the first intelligent agent can decompose the target task into multiple independently executable sub-tasks. For example, the intelligent control task of video conferencing (i.e., an implementation of the target task) can be decomposed into a multi-source data acquisition and preprocessing sub-task, a multi-sensor data fusion and feature extraction sub-task, a device operating status adjustment sub-task, a system resource monitoring sub-task, and a fusion result verification sub-task.

[0063] The first intelligent agent can assign various sub-tasks to the corresponding second intelligent agents based on their functional positioning. For example, the multi-sensor data fusion and feature extraction sub-task and the equipment operating status adjustment sub-task can be assigned to the core second intelligent agent, while the multi-source data acquisition and preprocessing sub-task and the attitude resource monitoring sub-task can be assigned to the auxiliary second intelligent agents.

[0064] The first intelligent agent can allocate computing resources such as CPU (Central Processing Unit), memory, bandwidth, and sensor access permissions to each second intelligent agent to avoid resource conflicts and ensure efficient execution of subtasks.

[0065] The first intelligent agent can synchronize the task progress of each second intelligent agent, receive and integrate the execution results of each second intelligent agent, ensure that the sub-task results form effective collaboration, and support the achievement of the target task.

[0066] For example, with "intelligent control of desktop computer video conferencing" as the target task, the intelligent agent system includes four second intelligent agents: second intelligent agent A (data preprocessing class), second intelligent agent B (core fusion control class), second intelligent agent C (resource monitoring class), and second intelligent agent D (result verification class).

[0067] The first intelligent agent can break down the target task into a data preprocessing subtask, a data fusion and equipment control subtask, a resource monitoring subtask, and a result verification subtask, and assign them to the second intelligent agents A, B, C, and D respectively; sensor data access permissions are assigned to the second intelligent agent A, high-priority CPU resources are assigned to the second intelligent agent B, and system monitoring interface permissions are assigned to the second intelligent agent C.

[0068] The second intelligent agent A (data preprocessing class) can collect raw video data from a camera (i.e., one implementation of a data acquisition device), raw audio data from a microphone (i.e., one implementation of a data acquisition device), and bandwidth data from a network sensor (i.e., one implementation of a data acquisition device), and perform preprocessing such as video denoising, audio denoising, and multi-source data timestamp alignment, and output standardized video, audio, and network data to the second intelligent agent B.

[0069] The second intelligent agent B (core fusion control type) can perform fusion processing on the standardized data output by the second intelligent agent A based on the target data fusion strategy, and extract features such as image clarity, voice purity, and network stability. If the network stability feature is lower than the threshold (large bandwidth fluctuation), the operating status of the electronic device can be adjusted, such as reducing the video encoding bitrate or enabling network cache optimization.

[0070] The second agent C (resource monitoring type) can monitor the CPU utilization, network data transmission latency (15ms), and sensor acquisition stability (no packet loss from the camera) of the second agents A and B in real time, and output a resource monitoring report to the current first agent.

[0071] The second agent D (result verification class) can verify the effectiveness of the operational state adjustment of the second agent B. For example, if it is detected that the video frame rate is stable at 30fps and the network fluctuation is reduced by 50% after adjustment, the verification result of the effective adjustment is output to the current first agent.

[0072] Step S102: Obtain feedback data generated by the second intelligent agent in performing the target task.

[0073] Feedback data can be, but is not limited to, reflecting task execution effectiveness, resource utilization status, the collaborative effect of the second intelligent agent, and environmental adaptation.

[0074] Each second agent can collect its own feedback data in real time during task execution and upload it synchronously to the current first agent through the communication interface between agents.

[0075] Step S103: Adjust the current first agent based on the feedback data, take the adjusted first agent as the new current first agent, and continue to execute step S101.

[0076] The first intelligent agent can adjust its task decomposition logic, task allocation rules, and resource scheduling strategies based on feedback data and built-in iterative optimization strategies (such as reinforcement learning, meta-learning, and evolutionary algorithms). This enables dynamic optimization of the collaboration mode of the second intelligent agent, ensuring that the intelligent agent system can specifically improve the accuracy, real-time performance, and stability of data fusion, enhance the precision of device operation status adjustment and resource utilization efficiency, and strengthen the intelligent agent system's ability to autonomously adapt to dynamic scenarios, automatically optimize parameters, efficiently collaborate on multiple tasks, and achieve long-term self-iterative upgrades.

[0077] In this embodiment, the first intelligent agent decomposes the target task and assigns sub-tasks such as data preprocessing, multi-sensor data fusion, device status adjustment, resource monitoring, and result verification to the corresponding second intelligent agents. It also rationally allocates resources such as CPU, bandwidth, and sensor access permissions. This not only avoids resource conflicts between multiple intelligent agents and ensures the efficient progress of the entire process of multi-sensor data from acquisition and preprocessing to fusion, but also provides a stable collaborative environment and high-quality data input for the second intelligent agents to carry out data fusion work, thereby improving the accuracy and real-time performance of multi-sensor data fusion from the source.

[0078] More importantly, based on feedback data and combined with iterative optimization strategies, the first intelligent agent autonomously adjusts its task decomposition logic, task allocation rules, and resource scheduling strategies. This not only dynamically optimizes the standards and processes of data preprocessing to ensure that the data input to the fusion stage always adapts to the real-time fusion requirements, but also provides sufficient computing power support for data fusion tasks through dynamic resource allocation. This helps the second intelligent agent to perform data fusion more efficiently and makes the device operation status adjustment of the second intelligent agent based on fusion characteristics more in line with actual application needs.

[0079] Through continuous iterative optimization of the first intelligent agent, the intelligence level of the intelligent agent system can be improved, thereby enhancing performance in areas such as data fusion accuracy, real-time performance, stability, equipment adjustment precision, and resource utilization efficiency.

[0080] As another optional embodiment of this application, this embodiment provides a processing method for embodiment 2 of this application. This embodiment is mainly an implementation of step S102 in embodiment 1, and may specifically include, but is not limited to, at least one of the following:

[0081] Step S11: Obtain first feedback data; the first feedback data includes the adjusted operating status of the electronic device.

[0082] In this embodiment, the first intelligent agent can determine whether its current task decomposition logic, task allocation rules, and resource scheduling strategy are suitable for the actual scenario by analyzing the first feedback data, and then optimize its own configuration accordingly.

[0083] For example, if the second agent B performs adjustments such as reducing the video encoding bitrate and enabling network cache optimization based on the fusion result that the network stability characteristics are below the threshold, the first feedback data may specifically include: the adjusted video encoding bitrate value, network cache occupancy rate, video frame rate stability, network latency fluctuation amplitude, etc.

[0084] If the first agent determines, based on the initial feedback data, that the adjusted video frame rate remains unstable, it can determine whether the poor real-time performance of data fusion is due to insufficient CPU resources allocated to the second agent B, or whether unreasonable task allocation is causing delays in the connection between the data preprocessing subtask and the fusion subtask. Subsequently, the first agent can adjust its resource scheduling strategy (e.g., increasing the CPU priority of the second agent B) or task allocation rules (e.g., optimizing the execution sequence of subtasks) accordingly, completing self-iteration. By optimizing its global control capabilities, it indirectly ensures the effectiveness of the second agent's data fusion and the adjustment of the electronic device's operating status.

[0085] Step S12: Obtain second feedback data; the second feedback data is used to reflect the state changes of the data objects to be operated by the second intelligent agent and the state of the second intelligent agent's occupation of computing resources.

[0086] In this embodiment, the state changes of the data objects that the second intelligent agent needs to operate on can include, but are not limited to, the state changes of various data objects throughout the entire process from the second intelligent agent collecting data from multiple data acquisition devices, data preprocessing, to fusion processing. For example, the packet loss rate of the original video data of the second intelligent agent A, the deviation value of the timestamp alignment of multi-source data, or the dimensional changes of features during the fusion process of the second intelligent agent B.

[0087] The state of computing resource usage by the second agent may include, but is not limited to, the consumption of computing resources such as CPU, memory, and bandwidth by each second agent during the execution of subtasks.

[0088] The first intelligent agent can adjust its resource scheduling strategy, task allocation rules, and subtask collaboration logic based on the second feedback data.

[0089] For example, if the data object status of the second agent shows that the multi-source data timestamp alignment deviation is too large, the first agent can determine whether it is because the second agent A was not allocated enough preprocessing time slices, or the task connection timing of the second agent A and the second agent B was not properly coordinated, causing the second agent A to complete the preprocessing hastily; then the first agent can adjust its own task timing planning rules (such as extending the preprocessing time of the second agent A, or setting a task connection buffer period between the second agent A and the second agent B).

[0090] For example, if the second agent's resource occupancy status includes: the CPU utilization rate of the second agent B is too high for a long time, while the second agent D is idle, the first agent can adjust its own resource scheduling strategy and task splitting logic, split some non-core sub-tasks of the second agent B (such as the format conversion of the merged data) to the second agent D, and adjust its own resource allocation threshold (e.g., reserve more CPU cache for the second agent B), thereby indirectly improving the data processing and fusion efficiency of the second agent.

[0091] Step S13: Obtain third feedback data; the third feedback data is used to reflect the performance of the first intelligent agent.

[0092] The third feedback data can be collected by the built-in performance monitoring module of the first agent, or by a dedicated result verification second agent (such as the second agent D mentioned above).

[0093] Third feedback data may include, but is not limited to:

[0094] Indicators for the rationality of task decomposition (e.g., the matching degree between subtasks and the functions of the second agent, the deviation rate between the total execution time of the subtask and the expected execution time), indicators for the balance of task allocation (e.g., the difference in task load rate of each second agent, the proportion of idle second agents), and indicators for the effectiveness of collaborative coordination (e.g., the success rate of integrating subtask results, the resource conflict rate when multiple second agents execute in parallel, the average latency from the first agent issuing an instruction to the second agent responding).

[0095] It can adjust its own core logic and parameters without relying on external intervention.

[0096] The first agent can adjust its core logic and parameters based on the third feedback data. For example, if the third feedback data indicates that the difference in subtask load rates is too large, the first agent can determine that its task allocation rules are flawed and then adjust its task allocation algorithm (e.g., by introducing a load balancing weight coefficient) to ensure that the load of each second agent is balanced.

[0097] Alternatively, if the third feedback data indicates that the resource conflict rate is too high, the first agent can adjust its own resource scheduling strategy, such as establishing a resource reservation mechanism and clarifying the resource occupation priority of each sub-task.

[0098] Alternatively, if the third feedback data indicates that the instruction response delay is too long, the first intelligent agent can adjust its own instruction transmission and coordination logic, such as optimizing the instruction encoding format and setting priority transmission channels for instructions of critical tasks.

[0099] By utilizing the third feedback data, the first intelligent agent continuously optimizes its core capabilities, providing more reliable support for the second intelligent agent to efficiently carry out multi-sensor data fusion, equipment status adjustment, and other tasks.

[0100] In this embodiment, based on the first feedback data, the first intelligent agent can accurately judge the supporting effect of its own task decomposition, resource scheduling and other global control strategies on multi-sensor data fusion and electronic device operation status, and avoid blind adjustments.

[0101] The second feedback data can guide the first intelligent agent to adjust itself from the perspectives of task timing planning and dynamic resource allocation, thereby indirectly improving the data preprocessing quality and fusion efficiency of the second intelligent agent and ensuring the stability of multi-source data throughout the entire process from collection to fusion.

[0102] Based on the third feedback data, the first intelligent agent can adjust the task allocation algorithm, resource scheduling mechanism and collaborative coordination logic accordingly.

[0103] By using three types of feedback data, the intelligence level of the intelligent agent system can be improved, and collaborative optimization can be achieved in terms of data fusion accuracy, real-time performance, and resource utilization efficiency.

[0104] As another optional embodiment of this application, a processing method is provided for embodiment 3 of this application. This embodiment is mainly an implementation of adjusting the current first intelligent agent based on the feedback data in embodiment 1, and may specifically include, but is not limited to:

[0105] Step S21: Adjust at least one of the following based on the feedback data: the internal configuration, behavioral strategy, and target model of the current first intelligent agent.

[0106] The internal configuration can be used to characterize the computing environment required for the operation of the current first intelligent agent.

[0107] In this embodiment, the internal configuration may include, but is not limited to: computing resource configuration (e.g., the CPU, memory, and cache allocation ratio configured by the first intelligent agent for core functions such as task decomposition and collaborative coordination) and network communication configuration (e.g., communication protocols between the first intelligent agent and each second intelligent agent (e.g., TCP / IP, RPC), data transmission bandwidth priority, and timeout retry mechanism for instruction transmission).

[0108] For example, if the third feedback data indicates that the average latency from the first agent issuing an instruction to the second agent responding exceeds a threshold, the first agent can adjust its communication configuration, that is, switch the communication protocol with the core second agent to a more lightweight RPC protocol, while increasing the bandwidth priority of instruction transmission and shortening the instruction response latency.

[0109] By adjusting its internal configuration, the first intelligent agent can improve its basic operational capabilities, ensuring the efficient execution of core functions such as task decomposition, resource scheduling, and collaborative coordination, and providing stable underlying support for the optimization of subsequent behavioral strategies and target models.

[0110] The behavioral strategy can be used to define task allocation behavior and resource scheduling behavior.

[0111] In this embodiment, the behavioral strategy may include, but is not limited to, at least one of the following:

[0112] Task allocation strategy: matching rules between subtasks and the functions of the second agent (e.g., fusion tasks with high computing power requirements are preferentially allocated to second agents with redundant computing power), and load balancing thresholds for subtasks (e.g., the difference in CPU utilization among the second agents does not exceed 20%).

[0113] Resource scheduling strategies include: priority rules for allocating computing resources (e.g., the CPU priority of the core second agent is higher than that of the data preprocessing second agent), and triggering conditions for dynamic expansion and contraction of resources (e.g., automatic expansion when the CPU utilization rate of a certain second agent is consistently ≥80%).

[0114] Coordination strategies include: execution timing rules for subtasks (e.g., delaying the data preprocessing subtask by 50ms before starting the fusion subtask after completion) and arbitration rules for resource conflicts (e.g., when multiple second agents compete for sensor resources, priority is given to agents with higher current task priority).

[0115] For example, if the first feedback data indicates that the video frame rate is still unstable, and the second feedback data indicates that the CPU utilization rate of the core second agent is ≥85% for a long period of time, the first agent can adjust the resource scheduling strategy, such as increasing the CPU priority of the agent and setting a trigger rule to automatically migrate non-core subtasks to idle agents when the CPU utilization rate is ≥80%.

[0116] If the second feedback data represents a large deviation in the timestamp alignment of the multi-source data, the first agent can adjust the collaborative coordination strategy, such as adding a 50ms buffer period between the data preprocessing subtask and the fusion subtask to ensure that the preprocessed data is fully standardized before entering the fusion stage.

[0117] By optimizing behavioral strategies, the first intelligent agent can form global control rules that are more adapted to actual task scenarios, ensuring more reasonable task breakdown, more efficient resource allocation, and smoother collaboration of the second intelligent agent, thereby indirectly improving the data fusion quality and equipment adjustment accuracy of the second intelligent agent.

[0118] The target model can be used for task allocation and / or resource scheduling.

[0119] The target model may include, but is not limited to, at least one of the following:

[0120] Task allocation model: A model built based on algorithms such as reinforcement learning and decision trees. The input can be the subtask type, the functional attributes of the second agent, and the current resource state. The output can be the optimal subtask and agent matching scheme.

[0121] Resource scheduling model: A model built based on algorithms such as linear programming and deep learning. The input can be the resource utilization rate of each agent, task priority, and device operating status, and the output can be the optimal resource allocation ratio.

[0122] The first agent can adjust the parameters of the target model based on feedback data through iterative optimization algorithms (such as reinforcement learning and meta-learning). For example, if the task allocation model output scheme determined based on feedback data leads to excessive load on some second agents and poor device adjustment, the first agent can update the Q-value weights of the task allocation model through reinforcement learning algorithms, incorporating load balancing into the model's reward function and increasing the model's emphasis on load balancing.

[0123] Alternatively, if the feedback data determines that the resource allocation accuracy of the resource scheduling model is low in network fluctuation scenarios, the first agent can optimize the constraint coefficients of the resource scheduling model through meta-learning algorithms, introduce network fluctuation amplitude as a model input feature, and improve the decision-making accuracy of the model in dynamic scenarios.

[0124] By optimizing the target model, the first intelligent agent can improve the intelligence level of task allocation and resource scheduling at the algorithm level, enabling its decision-making ability to adapt to complex and dynamic scene changes.

[0125] like Figure 2 As shown, after adjusting at least one of the internal configuration, behavioral strategy, and target model of the current first agent based on the feedback data, the current first agent is updated to a new first agent with better global control capabilities. Subsequently, the new first agent continues to drive the second agent to execute the target task, generating new feedback data adapted to the new agent's state in the process. The first agent then restarts the adjustment process of its own internal configuration, behavioral strategy, and target model based on the new feedback data, forming a continuous iterative closed loop of adjustment-execution-feedback-re-adjustment, thereby realizing the dynamic optimization and steady improvement of the first agent's own global control capabilities.

[0126] In this embodiment, adjusting the current first agent based on the feedback data may include, but is not limited to:

[0127] Step S31: Using reinforcement learning algorithms (such as Q-learning, PPO, etc.), the behavior strategy and target model of the first agent are adjusted through trial and error based on the first and second feedback data.

[0128] Step S31 may include, but is not limited to:

[0129] Step S311: Parameter initialization.

[0130] 1.1 Define the state space S of the first intelligent agent: The state parameters may include, but are not limited to: the operating state indicators of the electronic device after adjustment, the data object state and resource occupancy state indicators of the second intelligent agent, the performance indicators of the first intelligent agent's own task decomposition rationality and task allocation balance, etc.

[0131] 1.2 Define the action space A of the first intelligent agent: Action parameters may include the internal configuration adjustment actions, behavior policy adjustment actions, and target model parameter adjustment actions of the first intelligent agent, such as communication protocol switching, task allocation rule modification, and target model weight update.

[0132] 1.3 Define the reward function R(s,a): Set reward rules that are strongly correlated with the performance of multi-sensor data fusion. If the accuracy, real-time performance, stability and other indicators of data fusion are improved after executing action a, a positive reward will be given; if the indicators decline or resource conflicts occur, a negative reward will be given.

[0133] 1.4 Initialize the policy π (such as a random policy) and value function V(s) or action value function Q(s,a) of the first agent.

[0134] Step S312, Interaction and Learning Iteration.

[0135] 2.1 The first agent, based on the current policy π and combined with the current state s, t ∈S Select action a t ∈A executes, specifically by adjusting its internal configuration, behavioral strategies, or target model parameters.

[0136] 2.2 The first agent drives the second agent to execute the target task, and collects the new state s after the action is executed. t+1 With the corresponding reward value r t =R(s t ,a t ).

[0137] 2.3 Employing reinforcement learning algorithms to update the strategy and value function:

[0138] If the Q-learning algorithm is used, then Q(s,a) is updated to represent the long-term benefit of performing action a in state s.

[0139] PPO (Proximal Policy Optimization): Ensures the stability of policy updates by optimizing policy gradients.

[0140] If the PPO (Proximal Policy Optimization) algorithm is used, a policy loss function is constructed, and the policy network parameters are optimized through gradient descent to limit the policy update magnitude and ensure the stability of policy iteration.

[0141] Step S313: Strategy optimization convergence.

[0142] 3.1 Repeat steps 2.1-2.3 to continuously iterate and update the strategy and action value functions.

[0143] 3.2 When the policy π iteration satisfies the convergence condition (such as the reward value tending to stabilize and the change in policy parameters being less than the threshold), the iteration stops and the optimal policy π' is output.

[0144] 3.3 The first intelligent agent determines the behavioral strategy and target model adjustment scheme based on the optimal strategy π', and completes autonomous iterative optimization.

[0145] In this embodiment, adjusting the current first agent based on the feedback data may include, but is not limited to:

[0146] Step S32: Utilize meta-learning techniques to learn how to quickly adapt to new tasks from historical tasks, and optimize the parameters, internal configuration, and behavioral strategies of the target model of the current first agent.

[0147] Step S32 may include, but is not limited to:

[0148] Step S321, Meta-training phase.

[0149] 1.1 Constructing the task distribution T: Collecting historical task sets in multi-sensor data fusion scenarios, including different types of target tasks (such as intelligent control of video conferencing, industrial equipment status monitoring, etc.) and their corresponding task characteristics, feedback data, and optimization schemes.

[0150] 1.2 Divide the meta-learning hierarchy: Set up a meta-learner (responsible for learning general optimization capabilities across tasks) and a basic learner (i.e., the target model of the first agent, responsible for handling specific tasks).

[0151] 1.3 Meta-training is performed using the Model Independent Meta-Learning (MAML) algorithm:

[0152] 1.3.1 Randomly sample a batch of tasks from the constructed task distribution as the training sample set;

[0153] 1.3.2 For each task in the sample set, first analyze the parameter deviation based on the initial parameters of the base learner using the training data (including historical feedback data) of the task, perform gradient descent calculation with a small step size, and obtain temporary parameters adapted to the specific task.

[0154] 1.3.3 Use the test data of this task to verify the adaptation effect of the temporary parameters and calculate the adaptation error (i.e., meta-loss).

[0155] 1.3.4 Based on the meta-loss of all sampling tasks, the core parameters of the meta-learner are adjusted in reverse so that the meta-learner can master the general rule of "how to achieve task adaptation by adjusting a small number of parameters";

[0156] 1.4 Repeat steps 1.3.1 to 1.3.4 to continuously iterate and optimize the parameters of the meta-learner and the base learner until the meta-loss tends to stabilize (i.e., the adaptability of the meta-learner no longer improves significantly). At this point, output the initial parameters of the base learner with fast adaptability.

[0157] Step S322, Meta-testing and rapid adaptation phase.

[0158] 2.1 New Task Trigger Adaptation Process: When the first agent faces a brand-new multi-sensor data fusion task (such as the newly added "multi-sensor fusion control of vehicle equipment" task), there is no need to retrain the model; the initial parameters of the basic learner obtained in the meta-training stage can be directly called.

[0159] 2.2 Rapid parameter tuning with limited data: Utilizing a small amount of feedback data generated during the startup phase of a new task (without waiting for full-process data), gradient adjustments with a small step size are performed based on the initial parameters to quickly obtain the target model parameters adapted to the new task.

[0160] 2.3 Full-dimensional adaptation and optimization: Based on the adjusted target model parameters, the first intelligent agent simultaneously optimizes its internal configuration (such as communication protocol and resource reservation threshold) and behavior strategy (such as task allocation rules and resource scheduling priority) to complete the rapid adaptation to new tasks and ensure that efficient multi-sensor data fusion and global control can be achieved in the early stage of new task startup.

[0161] Step S323: Continuous iterative optimization of the model.

[0162] 3.1 After a new task has been running for a period of time, collect its complete execution data, full-process feedback data and adaptation effect data, and include it in the historical task set according to the standardized format of task distribution to enrich the coverage of task distribution.

[0163] 3.2 According to the preset cycle (e.g., once a month), based on the expanded task distribution, repeat all the steps of the meta-training stage, continuously optimize the general adaptability of the meta-learner and the initial parameters of the basic learner, so that the first agent can gradually adapt to more complex and diverse multi-sensor data fusion scenarios, and strengthen the system's dynamic adaptation and self-upgrading capabilities.

[0164] In this embodiment, adjusting the current first agent based on the feedback data may include, but is not limited to:

[0165] Step S33: Adjust the internal configuration and behavioral strategies of the first intelligent agent by simulating the core processes of selection, crossover, and mutation in biological evolution.

[0166] Step S33 may include, but is not limited to:

[0167] Step S331: Initialize the population.

[0168] Based on the actual needs of multi-sensor data fusion scenarios, a population size (i.e., the number of candidate solutions) is set, and an initial population is constructed through random generation. The initial population includes multiple candidate solutions. Each candidate solution can contain the internal configuration of the first agent (such as CPU / memory allocation ratio, communication protocol type, resource reservation threshold, etc.) and behavioral strategies (such as task allocation rules, resource scheduling priority, sub-task coordination logic, etc.) to ensure that the initial population covers a sufficiently diverse range of parameter combinations.

[0169] Step S332: Evaluation of candidate solutions.

[0170] Each candidate solution in the current population is loaded into the first agent, driving the second agent to execute the target task related to multi-sensor data fusion; the feedback data of the entire process during task execution (i.e., the three types of feedback data in Example 2) is collected, and the fitness score of each candidate solution is calculated according to the preset fitness function to complete the ranking of all candidate solutions.

[0171] Step S323: Select high-quality candidate solutions.

[0172] The selection operation is performed based on fitness scores, selecting candidate solutions with higher fitness scores from the current population to form the parent set. The selection method can be flexible, such as roulette wheel selection (allocating selection probabilities according to the proportion of fitness scores, with higher scores having a greater probability of being selected), or tournament selection (randomly selecting a portion of candidate solutions and choosing the highest-scoring one to enter the parent set), ensuring that the parent set consists of high-quality solutions.

[0173] Step S324: Crossover and mutation generate a new population:

[0174] Cross operation: Randomly select two candidate solutions from the parent set, and partially combine their core features (e.g., combine the resource configuration parameters of candidate solution A with the task allocation strategy of candidate solution B) to generate a new candidate solution (offspring), thereby achieving the inheritance and fusion of high-quality features.

[0175] Mutation operation: Randomly fine-tunes some features of the offspring candidate solutions (such as slightly modifying the CPU allocation ratio, fine-tuning the resource scheduling priority threshold, etc.), introducing new parameter combinations. The probability of the mutation operation needs to be set reasonably (usually 0.01-0.1) to ensure population diversity while avoiding excessive mutation that destroys high-quality features.

[0176] New population formation: The offspring candidate solutions generated by crossover and mutation are merged with the candidate solutions with the highest fitness scores in the parent set (retaining the core high-quality solutions) to form a new generation population, and the population size remains consistent with the initial setting.

[0177] Step S325: Iterative optimization and termination.

[0178] Repeat steps S322 (candidate solution evaluation), S323 (selection), and S324 (crossover and mutation) to continuously advance the population iteration. Stop the iteration when the preset termination condition is met.

[0179] Termination conditions may include, but are not limited to: reaching the set maximum number of iterations; or, the fitness score of the optimal candidate solution in the population for several consecutive generations tends to stabilize (the change is less than a preset threshold).

[0180] From the population after the final iteration, the candidate solution with the highest fitness score is selected as the optimal solution, and its corresponding internal configuration and behavior strategy are loaded into the first agent to complete the iterative optimization of the first agent.

[0181] The optimized first agent can work with the second agent to efficiently carry out multi-sensor data fusion.

[0182] As another optional embodiment of this application, refer to Figure 3 This is a schematic diagram of the interaction between a first intelligent agent and a second intelligent agent provided in Embodiment 4 of this application. This embodiment mainly describes an implementation method whereby the second intelligent agent in Embodiment 1 fuses data collected by multiple data acquisition devices of an electronic device based on a target data fusion strategy to obtain target features, and adjusts the operating state of the electronic device based on these target features. Figure 3 As shown, the specific steps may include, but are not limited to, the following:

[0183] Step S41: The second intelligent agent performs fusion processing on the data collected by multiple data acquisition devices of the electronic device based on the target data fusion strategy to obtain target features.

[0184] In this embodiment, a unified data interface can be used to perform preprocessing such as format conversion, syntax verification, and basic noise reduction on multi-source heterogeneous data from different data acquisition devices (e.g., multiple sensors such as cameras, microphones, and depth sensors, which can be represented as sensor 1, sensor 2, ..., sensor N), and output standardized data in a unified format. This breaks down data isolation and enables unified data access.

[0185] In this embodiment, the target data fusion strategy may include, but is not limited to, at least one of the following: data preprocessing strategy, data fusion model, and fusion parameters.

[0186] Data preprocessing strategies can be used to optimize data collected by multiple data acquisition devices.

[0187] Data fusion models can be used to integrate multimodal features.

[0188] Fusion parameters can be used to define the specific operating logic of data preprocessing strategies or data fusion models.

[0189] The target data fusion strategy can be determined in the following ways:

[0190] Step S411: Determine the data quality assessment results and environmental status information based on the data collected by multiple data acquisition devices of the electronic equipment.

[0191] For example, multiple data acquisition devices may include: visual sensors (e.g., cameras, image acquisition modules integrated into the graphics card output link), audio sensors (e.g., microphones), and depth sensors. Visual sensors can acquire visual data such as screen display data and environmental image data. Audio sensors can acquire audio data such as environmental audio data and human voice data. Depth sensors can acquire point cloud data. Data quality assessment results may include: image sharpness, audio signal-to-noise ratio, multi-source data consistency deviation, and data packet loss rate.

[0192] Environmental status information may include: CPU utilization, memory load, ambient light intensity, network fluctuation, etc.

[0193] Step S412: Make adaptive strategy decisions based on the data quality assessment results and environmental status information.

[0194] Adaptive strategy decisions may include, but are not limited to, selecting a target data fusion strategy from a strategy library.

[0195] Based on the target data fusion strategy, a scheduling strategy for data processing tasks is determined, such as adjusting task priority and parallelism. Specific scheduling strategies may include:

[0196] Dynamic computing resource allocation strategy: Based on modal activation intensity, it achieves precise allocation of GPU computing power.

[0197] The modal activation intensity can be determined by both the target data fusion strategy and the acquired data. If the target data fusion strategy focuses on visual feature extraction (such as using an improved spatiotemporal YOLOv8 network to extract spatial target features and optical flow motion features), and visual modal features account for the highest proportion in real-time data (i.e., visual-dominated scenarios), then 70% of GPU resources should be allocated to the improved spatiotemporal YOLOv8 model to ensure the efficiency and accuracy of visual feature extraction. If the target data fusion strategy emphasizes audio feature processing (such as needing to extract key human voice semantic features or voiceprint features), and audio modal activation intensity is the highest (i.e., speech-dominated scenarios), then the batch size of the ASR model should be increased to 32 to improve the audio data processing throughput through batch processing.

[0198] Memory optimization strategy: A modal feature cache replacement algorithm (MFCA) is adopted, which dynamically manages memory resources based on the weight ratio of each modality feature in the target data fusion strategy. Memory space is prioritized for allocation to high-weight modal features (such as visual features output by the improved spatiotemporal YOLOv8 network in vision-dominated scenarios, and audio-text features output by the ASR model in speech-dominated scenarios). A cache replacement mechanism is established so that when memory is insufficient, the cache space of low-weight modal features is released first, ensuring real-time access to features required for the core fusion task and avoiding fusion interruptions or increased latency due to memory overflow.

[0199] In this embodiment, the scheduling strategy for data processing tasks can be sent to the resource scheduling engine. The resource scheduling engine can dynamically allocate computing resources based on the real-time resource status of the system (CPU, memory, GPU utilization). For example, more CPU cores can be allocated to high-priority fusion tasks, sufficient GPU computing power can be reserved for complex fusion models, and a reasonable degree of parallelism can be set for parallel processing tasks to ensure efficient execution of data processing and fusion.

[0200] Selecting a target data fusion strategy from the strategy library may include, but is not limited to, at least one of the following:

[0201] Step S4121: Select a target rule from the rule base that matches at least one of the data quality assessment results and environmental status information, as the target data fusion strategy.

[0202] For example, if the data quality assessment result is "image sharpness < 60 points", a Gaussian denoising preprocessing strategy (i.e., an implementation of a data preprocessing strategy) can be selected from the rule base.

[0203] If the environment status information is "CPU utilization > 80%", a lightweight fusion model (i.e., an implementation of a data fusion model) can be selected from the rule base.

[0204] Step S4122: Input the data quality assessment results and / or environmental status information into the decision tree model to obtain the target data fusion strategy output by the decision tree model.

[0205] In this embodiment, the decision tree model can be pre-trained using historical data quality indicators (image clarity, audio signal-to-noise ratio, data packet loss rate, etc.) and historical environmental state indicators (CPU utilization, light intensity, network fluctuation amplitude, etc.) as input features and the historical best fusion strategy as the label.

[0206] For example, if the data quality assessment result is: image noise > 30dB, and the environmental status information is: light intensity < 200 lux and CPU utilization = 65%, and this is input into the decision tree model, the decision tree model can output an infrared enhancement preprocessing strategy (i.e., an implementation of a data preprocessing strategy) and a multi-feature association aggregation fusion model (i.e., an implementation of a data fusion model).

[0207] Step S4123: Based on reinforcement learning, determine the target data fusion strategy.

[0208] Step S4123 may specifically include:

[0209] 1. Constructing a reinforcement learning framework:

[0210] State space (S): A combination of data quality assessment results and environmental state information (e.g., a combination of high data quality and low environmental load; or low data quality and high environmental load).

[0211] Action space (A): The combination of all available data preprocessing strategies, data fusion models, and fusion parameters in the strategy library (i.e., all potential target data fusion strategies).

[0212] Reward function (R): Combines the scene discrimination of the fused features (the higher the reward value, the higher the reward value, representing the support capability of the features for subsequent scene reasoning) and fusion latency (the lower the reward value, the higher the reward value, representing processing efficiency). If the scene discrimination of the fused features is lower than the preset threshold or the latency exceeds the threshold, a negative reward is given.

[0213] 2. Model Training: Using Q-learning or PPO algorithms, the model is allowed to try different actions (policy combinations) under different states, and the policy selection logic is continuously optimized based on the feedback of the reward function.

[0214] 3. Strategy Output: After the model training converges, for the real-time state (i.e., data quality assessment results and environmental state information), the action with the highest reward value (i.e., the optimal strategy combination) is output as the target data fusion strategy.

[0215] After determining the target data fusion strategy and scheduling strategy, data processing and fusion can proceed. The data processing and fusion process may include, but is not limited to:

[0216] Step S413: Based on the data preprocessing strategy, the data collected by each data acquisition device can be preprocessed (such as infrared enhancement, feature dimensionality reduction, spatiotemporal calibration).

[0217] Step S414: Extract each feature from the preprocessed data of each data acquisition device.

[0218] For example, such as Figure 4 As shown, in the vision branch, an improved spatiotemporal YOLOv8 network, combined with ResNet-50 and FlowNet networks, can extract visual features such as spatial target features and optical flow motion features from visual data. ResNet-50 is used to extract spatial target features, which accurately identify the object category and location in the image. FlowNet is used to extract optical flow motion features, which can capture the motion state and trend of the target.

[0219] In the speech branch, audio features such as STT text features and speaker features can be extracted from audio data in parallel via dual-path processing. BERT embedding is a technique for extracting STT text features (based on the BERT model). The STT text features extracted using this technique can convert speech into text information and understand its semantic content, such as recognizing keywords in voice commands. Mel spectrum + Conv1D is a combination of techniques for extracting speaker features (first preprocessing the audio data using Mel spectrum, then extracting features using the Conv1D network). The speaker features extracted using this combination retain the unique frequency and waveform information of the sound, which can be used to identify the speaker's identity, emotions, etc.

[0220] In the depth perception branch, after voxelizing the point cloud data (0.1m resolution), depth features can be extracted using a 3D sparse convolutional network. Depth features provide three-dimensional spatial information of objects in the scene, accurately describing the shape, size, and spatial position of objects.

[0221] Step S415: Based on the data fusion model, fuse the various features to obtain the target features.

[0222] Target features have richer connotations and stronger representational capabilities, enabling them to describe target objects or scenarios from multiple dimensions, thereby improving the ability to identify, analyze, and process complex situations.

[0223] Step S42: The second agent acquires event rule description information and performs reasoning on the event rule description information and the target features based on the target model to obtain the reasoning result.

[0224] The event rule description information is used to describe the logical relationships between different events.

[0225] Event rule description information may include, but is not limited to:

[0226] The first event rule description information can be used to describe the causal relationship between events. For example, if a high hardware load event occurs, it will inevitably be accompanied by a high power operation event of the cooling system.

[0227] The second event rule description information can be used to describe the synchronous occurrence relationship between events. For example, audio and video playback events and network high bandwidth usage events are synchronous occurrence events.

[0228] The third event rule description information can be used to describe the relationship between multiple events that trigger a specific scenario. For example, a high-performance computing scenario is triggered when and only when the dynamic rendering event, the high-load hardware event, and the high-power operation event of the heat dissipation system are all satisfied.

[0229] The reasoning result represents the current usage scenario of the electronic device.

[0230] In this embodiment, the structure of the target model can be set as needed, and the specific structure is not limited in this application. For example, the target model may include, but is not limited to, neural network models or decision tree models.

[0231] Step S43: The second intelligent agent adjusts the operating state of the electronic device based on the reasoning result.

[0232] For example, if the inference result indicates that the computer is running a large game, the second agent can perform the following operational state adjustment actions based on this inference result:

[0233] Increase the GPU computing power allocation ratio from the default 50% to 70%, lock the game process to high CPU priority, and close unnecessary background processes (such as document editors, cloud drive synchronization tools, and redundant browser tabs).

[0234] Reserve 2GB of dedicated memory for the game process and disable memory compression mechanism to ensure memory response speed when loading game textures and switching scenes.

[0235] Network bandwidth priority is shifted towards the game process, limiting the maximum bandwidth usage of other applications to no more than 20% of the total bandwidth, and enabling a dedicated network acceleration channel for the game.

[0236] The cooling system is switched to high-performance gaming mode, with the fan speed set to 5000rpm and the CPU temperature threshold set to 85℃. If the CPU temperature is detected to be continuously above 85℃, the CPU dynamic frequency adjustment is automatically activated (the maximum frequency does not exceed 110% of the rated frequency), while reducing the power supply of non-core hardware (such as USB peripherals) to prioritize game operation.

[0237] The graphics card is in overclocking mode, with the core frequency increased by 10% and the memory frequency increased by 15%. The graphics card power-saving function is turned off to ensure stable frame rate in the game.

[0238] Adjust the monitor parameters to game mode: increase brightness to 80%, adjust contrast to 75%, switch color gamut to DCI-P3 full coverage, and increase refresh rate from 60Hz to 144Hz (monitor hardware support required).

[0239] In this embodiment, after breaking down the isolation of multi-source heterogeneous data through a unified data interface, adaptive strategy decisions are made dynamically based on data quality and environmental conditions, and target data fusion strategies are dynamically matched to improve the accuracy and efficiency of data processing from multiple data acquisition devices.

[0240] Based on this, after obtaining target features using the target data fusion strategy, the discrete logic such as causality, parallelism, and conditional triggering between different events can be clarified based on the event rule description information. Reasoning is performed on the event rule description information and target features based on the target model. This provides a clear logical basis for scenario reasoning through discrete rules, while capturing subtle dynamic changes in the scenario using continuous features. This achieves joint optimization of continuous target features and discrete event rules, improving the accuracy, robustness, and generalization ability of reasoning for the current usage scenario of electronic devices. Consequently, the second intelligent agent can perform adaptive operational state adjustments based on accurate reasoning results, ensuring the accuracy, real-time performance, and stability of data fusion, and improving the precision of device operational state adjustments and resource utilization efficiency.

[0241] Meanwhile, the first intelligent agent continuously iterates and optimizes its internal configuration, behavioral strategies, and target models through feedback data. This not only provides the second intelligent agent with more reasonable task allocation, more efficient resource scheduling, and a more stable collaborative environment, ensuring the smoothness of the second intelligent agent's adaptive strategy decision-making and data fusion, but also enables the second intelligent agent to adjust the device operating status based on accurate reasoning results to better meet the needs of actual scenarios. This further improves the accuracy, real-time performance, and stability of data fusion, as well as the precision of device operating status adjustment and resource utilization efficiency.

[0242] As another optional embodiment of this application, refer to Figure 5 This is a schematic diagram of the interaction between a first intelligent agent and a second intelligent agent provided in Embodiment 5 of this application. This embodiment is mainly an implementation method of Embodiment 1, in which the second intelligent agent performs fusion processing on data collected by multiple data acquisition devices of an electronic device based on a target data fusion strategy to obtain target features, and adjusts the operating state of the electronic device based on the target features. Specifically, it may include, but is not limited to, the following steps:

[0243] Step S51: The second intelligent agent performs fusion processing on the data collected by multiple data acquisition devices of the electronic device based on the target data fusion strategy to obtain target features.

[0244] For a detailed description of step S51, please refer to the relevant description of step S41 in the above embodiments, which will not be repeated here.

[0245] Step S52: The second intelligent agent searches for a template in the template library that matches the target feature; the template represents a candidate usage scenario of the electronic device.

[0246] The template library can include multiple templates, and each template can include candidate features, candidate usage scenario labels, and device adjustment strategies.

[0247] For example, an office scenario template may include: candidate features such as dynamic intensity of the screen (0.3~0.5), CPU load (30%~50%), and fan speed (2000~3000rpm); candidate usage scenario tags such as daily office work; and device adjustment strategies such as energy-saving mode and eye-protection screen settings.

[0248] Video conferencing templates may include: candidate features such as human voice semantic dimension ≥ 0.7, network bandwidth usage 30%~60%, and dynamic range of the image 0.5~0.7; candidate usage scenario label is video conferencing; and device adjustment strategies include network priority enhancement and microphone noise reduction.

[0249] Large-scale game templates may include: candidate features such as dynamic intensity of screen ≥0.8, CPU load ≥70%, and fan speed 4000rpm; candidate usage scenario tags such as running large-scale games; and device adjustment strategies such as high-performance mode and GPU computing power enhancement.

[0250] The second agent can determine the cosine similarity between the target feature and the candidate features of all templates in the template library. If the cosine similarity of the templates meets the set similarity threshold, a successful match can be determined.

[0251] If the cosine similarity of all templates is lower than the set similarity threshold, it can be determined that the match is unsuccessful, and the process proceeds to the rule reasoning process in step S54.

[0252] Step S53: The second intelligent agent adjusts the operating state of the electronic device based on the found template.

[0253] The second intelligent agent can adjust the operating state of the electronic device based on the device adjustment strategy in the matched template. For example, if a video conferencing template is matched, the operating state of the electronic device can be adjusted as follows according to the network priority enhancement and microphone noise reduction device adjustment strategy: limit the bandwidth usage of other applications to ≤20%, enable the microphone noise reduction enhancement mode, switch the screen to conference mode (brightness 60%+, color temperature warm), and allocate high priority to the conference process for the CPU (utilization ≤60%).

[0254] If a daily office template is matched, the device adjustment strategy based on the energy-saving mode and eye-protection screen settings can be used to adjust the operating status of electronic devices as follows: switch the power plan to balanced mode, enable eye-protection mode on the screen (blue light filter ≥50%), limit the fan speed to within 3000rpm (noise ≤35dB), and automatically put non-essential background processes to sleep.

[0255] Step S54: If the second intelligent agent does not find a template that matches the target feature, it obtains event rule description information and infers the event rule description information and the target feature based on the target model to obtain an inference result; the inference result represents the current usage scenario of the electronic device; the event rule description information is used to describe the logical relationship between different events.

[0256] Step S55: The second intelligent agent adjusts the operating state of the electronic device based on the reasoning result.

[0257] For a detailed description of steps S54-S55, please refer to the relevant description of steps S42-S43 in Example 4, which will not be repeated here.

[0258] In this embodiment, by using preset templates in the template library, target features and scenes can be quickly matched, and corresponding device adjustment strategies can be invoked, improving the real-time performance of scene recognition and device adjustment after data fusion. Furthermore, in cases where a template fails to match, discrete logic is clarified through event rule description information, and joint optimization of continuous target features and discrete rules is achieved by combining the target model. This ensures the generalization ability and accuracy of scene recognition and overcomes the limitation that the template library cannot exhaust all scenes.

[0259] As another optional embodiment of this application, this embodiment provides a processing method for embodiment 6. This embodiment mainly describes an implementation method in embodiment 1 where the second intelligent agent, based on a target data fusion strategy, fuses data collected by multiple data acquisition devices of an electronic device to obtain target features, and adjusts the operating state of the electronic device based on the target features. Specifically, it may include, but is not limited to, the following steps:

[0260] Step S61: The second intelligent agent performs fusion processing on the data collected by multiple data acquisition devices of the electronic device based on the target data fusion strategy to obtain target features.

[0261] Step S62: The second intelligent agent acquires event rule description information and performs reasoning on the event rule description information and the target features based on the target model to obtain a reasoning result; the reasoning result represents the current usage scenario of the electronic device; the event rule description information is used to describe the logical relationship between different events.

[0262] Step S63: The second intelligent agent adjusts the operating state of the electronic device based on the reasoning result.

[0263] For detailed procedures of steps S61-S63, please refer to the relevant description of steps S41-S43 in Example 4, which will not be repeated here.

[0264] Step S64: The second intelligent agent adjusts the target data fusion strategy according to the adjusted operating state of the electronic device.

[0265] After the second intelligent agent completes the device adjustment in step S63, it can continuously or periodically collect the adjusted operating status data.

[0266] The adjusted operating status data may include, but is not limited to: direct performance indicators, such as adjusted CPU / GPU utilization, memory usage, network bandwidth, video frame rate, and audio signal-to-noise ratio; and indirect performance indicators, such as device stability after adjustment, user operation smoothness, and task completion efficiency.

[0267] The second intelligent agent can compare the adjusted operating status data with the preset performance target. If the preset performance target is not met, the target data fusion strategy can be adjusted.

[0268] Adjusting the target data fusion strategy may include, but is not limited to:

[0269] Adjust data preprocessing strategies: for example, change the strength of the denoising algorithm, adjust the dimension of feature dimensionality reduction, and optimize the accuracy of timestamp alignment.

[0270] Adjust the data fusion model: for example, switch to a lighter or more complex fusion model, or adjust the model's weight parameters.

[0271] Adjust fusion parameters: For example, modify the weight coefficients when performing feature weighted fusion, or adjust the threshold of the decision tree.

[0272] The adjustment method can be a deterministic adjustment based on preset rules, or an adaptive optimization based on algorithms such as reinforcement learning.

[0273] In this embodiment, by having the second intelligent agent adjust the target data fusion strategy according to the adjusted operating state of the electronic device, the device adjustment deviation caused by the solidification of the fusion strategy can be avoided, thereby reducing the waste of resources caused by invalid calculations and indirectly improving resource utilization efficiency. At the same time, the target features obtained by subsequent data fusion are more in line with the actual operating needs of the device, providing more accurate input for scene reasoning and state adjustment. Ultimately, the accuracy, real-time performance, stability and device adjustment accuracy of data fusion are synergistically improved, and the dynamic optimization of the fusion strategy can be completed without human intervention, enhancing the ability of the intelligent agent system to autonomously adapt to dynamic scenes and for long-term iterative upgrades.

[0274] As another optional embodiment of this application, this embodiment provides a processing method for embodiment 7 of this application. This embodiment is mainly an implementation of step S41 in embodiment 4, and may specifically include, but is not limited to, the following steps:

[0275] Step S71: Extract the features of the data collected by each data acquisition device to obtain the features corresponding to each data acquisition device.

[0276] In this embodiment, different feature extraction models can be used to extract features from the data collected by each data acquisition device. For example, the feature extraction methods can be found in [reference needed]. Figure 4 This will not be elaborated upon here.

[0277] Step S72: Enhance the features corresponding to each of the data acquisition devices to obtain enhanced features corresponding to each of the data acquisition devices; the enhanced features represent information related to the data in the data acquired by the data acquisition device and the data acquired by other data acquisition devices.

[0278] In this embodiment, the features corresponding to each of the data acquisition devices can be input into the attention module, and the attention module can calculate the attention weight between each pair of features.

[0279] The calculated attention weights are used to perform a weighted summation of the features to obtain the enhanced visual features and enhanced audio features.

[0280] For example, such as Figure 6As shown, the process of performing primary cross-attention processing may include:

[0281] Visual features (such as spatial target features and optical flow motion features) and audio features (such as STT text features and voiceprint features) are input into an attention module.

[0282] This module calculates the attention weights of visual features on audio features, and the attention weights of audio features on visual features. These weights reflect the correlation and importance of the two modalities.

[0283] Using the calculated weights, the original features are weighted and summed to obtain the enhanced visual features and enhanced audio features.

[0284] For depth features, similar cross-attention enhancement can be performed with visual or audio features to obtain enhanced depth features.

[0285] The enhanced features can not only contain information from their own modality, but also incorporate key information from other modalities that are relevant to them. For example, enhanced visual features may focus more on visual regions that are relevant to the current audio event.

[0286] Step S73: Determine the temporal and spatial relationships between the enhanced features corresponding to each of the data acquisition devices.

[0287] In this embodiment, the temporal and spatial relationships between the enhanced features corresponding to each data acquisition device can be determined based on the spatiotemporal graph attention network (ST-GAT network), which can be regarded as advanced spatiotemporal fusion.

[0288] Specifically, an intermodal relationship graph can be constructed, with each enhanced feature used as a node in the graph.

[0289] Edge definition: The weight of an edge represents the strength of the association between two nodes. This association strength is calculated dynamically, taking into account both temporal and spatial information.

[0290] Spatial correlation: For example, the relationship between the position of an object in visual features and the corresponding position in depth features.

[0291] Temporal correlation: For example, the relationship between the audio features of the current frame and the visual features of the previous or next frame.

[0292] Graph attention computation: Spatiotemporal graph attention networks iteratively update the feature representation of each node through multiple graph attention layers. In each layer, each node updates its own features by aggregating information based on the features and association weights of its neighboring nodes.

[0293] Step S74: Based on the association relationship, perform fusion processing on the enhanced features corresponding to each of the data acquisition devices to obtain the target features.

[0294] In this embodiment, adaptive weighted fusion can be performed based on the association weights obtained in step S73.

[0295] In this embodiment, the target features can comprehensively integrate information from multiple sensors such as vision, audio, and depth, as well as the spatiotemporal correlations between them, providing a solid foundation for subsequent scene reasoning and device state adjustment.

[0296] In this embodiment, by enhancing the features corresponding to each data acquisition device, enhanced features corresponding to each data acquisition device are obtained. The temporal and spatial relationships between the enhanced features corresponding to each data acquisition device are determined, enabling the target features to have a more comprehensive scene representation capability, directly improving the accuracy and stability of data fusion, providing more accurate input for scene reasoning, and improving the accuracy of electronic device operating state adjustment.

[0297] As another optional embodiment of this application, this embodiment provides a processing method for embodiment 8 of this application. This embodiment is mainly an implementation of step S74 in embodiment 7, and may specifically include, but is not limited to, the following steps:

[0298] Step S81: Based on the association relationship, determine the first weight factor of the enhanced features corresponding to each of the data acquisition devices.

[0299] In this embodiment, the importance score of the node features output by the last layer of the spatiotemporal graph attention network can be used directly, or the out-degree or in-degree of each node in the adjacency matrix of the intermodal relationship graph can be used as the first weight factor.

[0300] Step S82: Determine the second weighting factor for each data acquisition device based on the quality parameters of the enhanced features corresponding to each data acquisition device.

[0301] In this embodiment, the data quality of each enhanced feature can be determined to obtain quality parameters.

[0302] For example, for visual features, data quality can be determined by judging image sharpness, signal-to-noise ratio, and whether there is motion blur or occlusion.

[0303] For audio features, data quality can be determined by judging the audio signal-to-noise ratio, the activity level of human voices, and the presence of noise interference.

[0304] For depth features, data quality can be determined by assessing point cloud density, ranging error, and the presence of a large number of invalid points.

[0305] In this embodiment, the quality parameters of the enhanced features corresponding to each data acquisition device can be input into a normalization function to obtain a normalized second weighting factor. The higher the quality, the greater the weight.

[0306] Step S83: Based on the second weight factor corresponding to each of the data acquisition devices, update the first weight factor of the enhanced features corresponding to each of the data acquisition devices.

[0307] The first and second weighting factors are linearly combined or multiplied element-wise. For example, ... Figure 7 As shown, after dynamic weight allocation (i.e., updating the first weight factor of the enhanced features corresponding to each data acquisition device based on the second weight factor corresponding to each data acquisition device), we can obtain α_vis (updated first weight factor of the enhanced visual features), α_aud (updated first weight factor of the enhanced audio features), and α_depth (updated first weight factor of the enhanced depth features).

[0308] In this way, even if a certain modality is important in spatiotemporal correlation, its weight in the final fusion will be reduced if its data quality is poor, thus avoiding the negative impact of low-quality data on the fusion result.

[0309] Step S84: Based on the updated first weight factor of the enhanced features corresponding to each data acquisition device, perform weighted processing on the enhanced features corresponding to each data acquisition device to obtain the target features.

[0310] In this embodiment, each enhanced feature can be multiplied by its corresponding updated first weight factor, and then all results can be summed to obtain the target feature.

[0311] In this embodiment, based on the aforementioned correlation, a first weighting factor is determined for the enhanced features corresponding to each data acquisition device. This accurately captures the spatiotemporal correlation of each modal feature, ensuring that the fusion result conforms to the dynamic changes of the scene. Simultaneously, introducing a second weighting factor based on quality parameters and updating the first weighting factor effectively avoids interference from low-quality data on the fusion result, improving the accuracy and stability of data fusion.

[0312] As another optional embodiment of this application, this embodiment provides a processing method for embodiment 9. This embodiment mainly describes an implementation method in embodiment 1 where the second intelligent agent, based on a target data fusion strategy, fuses data collected by multiple data acquisition devices of an electronic device to obtain target features, and adjusts the operating state of the electronic device based on the target features. Specifically, it may include, but is not limited to, the following steps:

[0313] Step S91: Process the data collected by multiple data acquisition devices to obtain time-aligned and / or spatially aligned data from the multiple data acquisition devices.

[0314] In this embodiment, as Figure 8 As shown, Precision Time Protocol (PTP) or Network Time Protocol (NTP) are used to synchronize the hardware clocks of multiple data acquisition devices. PTP provides sub-microsecond clock synchronization accuracy within a local area network, suitable for scenarios with extremely high time accuracy requirements, such as industrial automation and financial transactions. NTP, on the other hand, is widely used in network applications, providing millisecond-level clock synchronization accuracy, meeting the needs of most common application scenarios. These protocols ensure that all data acquisition devices maintain a consistent time base, laying the foundation for subsequent data time alignment.

[0315] Considering that different sensors may have different sampling rates, even with hardware clock synchronization, the acquired data may still have slight temporal differences. Therefore, a software interpolation compensation method is used to interpolate the acquired data based on the differences in sensor sampling rates. For example, for sensor data with a lower sampling rate, appropriate new data points are inserted between its adjacent sampling points, so that the data from all sensors can be more evenly distributed in time, further achieving millisecond-level time alignment.

[0316] Since data acquired by different types of data acquisition devices (such as depth sensors and cameras) resides in different coordinate spaces, a unified coordinate transformation framework is needed to perform spatial coordinate transformations. Taking visual pixel coordinates and depth sensor point cloud data as an example, an affine transformation matrix is ​​obtained by establishing an affine transformation relationship between them. This matrix describes how to accurately map each point in the point cloud data acquired by the depth sensor to the visual pixel coordinate system, thereby achieving spatial alignment of data from different sensors.

[0317] Step S92: Determine whether the time-aligned and / or spatially-aligned data from the plurality of data acquisition devices meet the adjustment conditions.

[0318] In this embodiment, a dynamic sliding window mechanism can be introduced, which allows the length of the sliding window to be adjusted according to actual conditions. Step S92 may include, but is not limited to, any of the following:

[0319] Step S921: Determine whether there are drastic fluctuations (such as sudden noise or data jumps) in the time-aligned and / or spatially aligned data of the multiple data acquisition devices. If the fluctuation exceeds the threshold, the window length needs to be adjusted to smooth the data.

[0320] Step S922: Determine whether the current data granularity of the time-aligned and / or spatially aligned data from multiple data acquisition devices is suitable for the input requirements of the data fusion model. If not, the window needs to be adjusted to match the model.

[0321] If yes, proceed to step S93; otherwise, proceed to step S95.

[0322] Step S93: Adjust the length of the sliding window.

[0323] In this embodiment, when the data stability is poor (e.g., high noise) or richer contextual information is needed (e.g., the fusion model requires long-term data), the window can be increased by a preset step size (e.g., from 50ms to 100ms), but the maximum window length does not exceed the upper limit corresponding to the real-time threshold (e.g., if the real-time requirement latency is <200ms, then the maximum window length is ≤150ms, reserving 50ms processing time).

[0324] When the data is stable but lacks real-time performance (e.g., processing latency exceeds the standard), the window size can be reduced by a preset step size (e.g., from 100ms to 50ms), but the minimum window length should not be less than the minimum effective period of the data acquisition device (e.g., if the minimum acquisition period of the camera is 33ms, then the minimum window length should be ≥33ms to avoid incomplete data).

[0325] In this embodiment, the length of the sliding window can be adjusted within the range of 0.5 to 3 seconds, with the specific length selected according to the actual situation.

[0326] Step S94: Using the adjusted sliding window as a unit, adjust the time dimension and / or spatial dimension of the time-aligned and / or spatially aligned data of the multiple data acquisition devices, and continue to execute the step of determining whether the time-aligned and / or spatially aligned data of the multiple data acquisition devices meet the adjustment conditions, until the target data of the multiple data acquisition devices is obtained.

[0327] In the time dimension, data within the sliding window can be smoothed using methods such as weighted averaging and interpolation to reduce time errors.

[0328] In terms of spatial dimension, tools such as affine transformation matrices can be used to precisely adjust the spatial coordinates of the data within the sliding window, ensuring that the data from different data acquisition devices correspond accurately in space.

[0329] Step S95: Using the sliding window as a unit, adjust the time dimension and / or spatial dimension of the time-aligned and / or spatially aligned data of the multiple data acquisition devices, and continue to execute step S91 until the target data of the multiple data acquisition devices is obtained.

[0330] Step S96: The second intelligent agent performs target data fusion processing on multiple data acquisition devices of the electronic device based on the target data fusion strategy to obtain target features.

[0331] Step S97: The second intelligent agent acquires event rule description information and performs reasoning on the event rule description information and the target features based on the target model to obtain a reasoning result; the reasoning result represents the current usage scenario of the electronic device; the event rule description information is used to describe the logical relationship between different events.

[0332] Step S98: The second intelligent agent adjusts the operating state of the electronic device based on the reasoning result.

[0333] For a detailed description of steps S96-S98, please refer to the relevant description of steps S41-S43 in Example 4, which will not be repeated here.

[0334] In this embodiment, at least one step can be selected from steps S91-S95 for execution.

[0335] Next, the processing method provided in this application will be explained in the context of a specific scenario. For example, when a computer is running a large game, the second intelligent entity can capture a series of related multi-source heterogeneous data, including the sound of the fan spinning at high speed, the rise in CPU temperature, and the game screen display.

[0336] 1. Spatiotemporal alignment:

[0337] The second agent can use Precise Time Protocol (PTP) or Network Time Protocol (NTP) to ensure that data from the visual sensor (camera capturing screen images), microphone (capturing fan sounds), and temperature sensor are synchronized in time.

[0338] For example, at a certain moment, the system simultaneously records a key frame of the game screen, a peak in fan noise, and a sudden rise in CPU temperature, and these data are precisely synchronized to the same point in time.

[0339] 2. Feature fusion:

[0340] The second agent extracts features of the game screen from the data from the visual sensor, such as character actions and scene changes; extracts features of the fan sound from the microphone data, such as the frequency and volume of the sound; and extracts features of the CPU temperature from the data from the temperature sensor, such as the temperature value and its trend.

[0341] By employing a cross-modal attention mechanism, features from different sensors are fused to form target features. These target features include visual information from the game screen, auditory information from fan sounds, and tactile information from CPU temperature.

[0342] 3. Cognitive reasoning:

[0343] The second agent can first search the template library for templates that match the target features. If a template is found, the operating state of the electronic device can be adjusted based on that template.

[0344] If no matching template is found in the template library, the second agent can obtain event rule description information, such as "high CPU utilization + high fan speed + specific game screen features = running a large game", and reason about these event rule description information and target features based on the target model.

[0345] The reasoning results confirm that the current scenario is "the user is playing a game". The second intelligent agent can adjust the operating status of the electronic device, such as automatically adjusting the GPU's computing power allocation and enabling the high-performance mode of the cooling system, to ensure the smoothness and stability of the game.

[0346] For example, when a user is working, the second intelligent agent captures a series of data such as keyboard clicks, mouse movements, and low CPU usage.

[0347] 1. Spatiotemporal alignment:

[0348] The second agent can use a time synchronization protocol to ensure that data from the visual sensor (camera captures mouse movement), microphone (captures keyboard clicks), and CPU utilization sensor are aligned in time and space.

[0349] For example, at a certain moment, the second intelligence can simultaneously record a quick mouse movement, a keystroke, and a low point in CPU usage, and these data are synchronized to the same point in time.

[0350] 2. Feature fusion:

[0351] The second agent extracts mouse movement features, such as movement trajectory and speed, from data from the vision sensor; keyboard keystroke features, such as keystroke frequency and force, from data from the microphone; and low CPU utilization features, such as utilization value and trend, from data from the CPU utilization sensor.

[0352] These features are fused using a cross-modal attention mechanism to form target features. Target features include visual information from mouse movement, auditory information from keyboard clicks, and performance information from CPU usage.

[0353] 3. Cognitive reasoning:

[0354] The second agent searches the template library for templates that match the target features. If a template is found, the operating state of the electronic device is adjusted based on that template.

[0355] If no matching template is found in the template library, the second agent can obtain event rule description information, such as "low CPU usage + frequent mouse movement + continuous keyboard typing = user is working", and infer these event rule description information and comprehensive features based on the target model.

[0356] The reasoning results confirm that the current scenario is "the user is working". Based on this, the operating status of electronic devices is adjusted, such as switching the power plan to balanced mode, enabling energy-saving mode to reduce energy consumption, and optimizing background processes to improve system response speed, so as to improve the user's work experience.

[0357] In this embodiment, high-precision time alignment of multi-sensor data is achieved by combining hardware clock synchronization with software interpolation compensation. Spatial alignment is then completed through a unified coordinate transformation framework. This eliminates fusion deviations caused by spatiotemporal misalignment of multi-source heterogeneous data, laying the foundation for subsequent high-quality data fusion. Furthermore, the introduced dynamic sliding window mechanism can flexibly adjust the window length. When data stability is poor, it smooths noise and preserves rich contextual information by increasing the window size; when real-time performance is insufficient, it reduces window control processing latency. Simultaneously, it strictly limits the upper and lower limits of the window length to balance data integrity and real-time performance, effectively solving the problem that fixed windows are difficult to adapt to dynamic scenarios.

[0358] Building upon this foundation, data is finely adjusted using the modified sliding window as a unit, further enhancing data consistency and effectiveness. This allows the second agent to obtain more accurate and representative target features when fusing target data based on the target data fusion strategy. Consequently, the current usage scenario of the electronic device, obtained by combining event rule description information with target model inference, is more realistic. Finally, the second agent adjusts the device's operating state based on the accurate inference results. This not only improves the accuracy of device adjustments but also reduces resource waste caused by invalid data processing through spatiotemporal alignment and dynamic window optimization, lowering CPU and memory usage. Simultaneously, it ensures the real-time performance and stability of data fusion and scenario inference.

[0359] The processing apparatus provided in this application will be described below. The processing apparatus described below can be referred to in correspondence with the processing method described above.

[0360] The processing device may include:

[0361] The driving unit is used to drive the second intelligent agent to execute the target task based on the current first intelligent agent. The first intelligent agent is used to coordinate the operation of all second intelligent agents. The second intelligent agent is used to fuse data collected by multiple data acquisition devices of the electronic device based on the target data fusion strategy to obtain target features, and adjust the operating state of the electronic device based on the target features.

[0362] The acquisition unit is used to acquire feedback data generated by the second intelligent agent when performing the target task.

[0363] The adjustment unit is used to adjust the current first agent based on the feedback data, take the adjusted first agent as the new current first agent, and continue to execute the step of driving the second agent to perform the target task based on the current first agent.

[0364] The acquisition unit acquires feedback data generated by the second agent performing the target task, which may include at least one of the following:

[0365] Obtain first feedback data; the first feedback data includes the adjusted operating status of the electronic device;

[0366] Obtain second feedback data; the second feedback data is used to reflect the state changes of the data objects to be operated by the second intelligent agent and the state of the second intelligent agent's occupation of computing resources;

[0367] Obtain third feedback data; the third feedback data is used to reflect the performance of the first intelligent agent.

[0368] The adjustment unit adjusts the current first agent based on the feedback data, and may include:

[0369] Based on the feedback data, adjust at least one of the following: the current first intelligent agent's internal configuration, behavioral strategy, and target model;

[0370] The internal configuration is used to characterize the computing environment required for the current first intelligent agent to run;

[0371] The behavioral strategy is used to define task allocation behavior and resource scheduling behavior;

[0372] The target model is used for task allocation and / or resource scheduling.

[0373] The second intelligent agent performs fusion processing on data collected by multiple data acquisition devices of the electronic device based on the target data fusion strategy to obtain target features, and adjusts the operating state of the electronic device based on the target features. For details on this process, please refer to the relevant description in the above processing method, which will not be repeated here.

[0374] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0375] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred 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 readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0376] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0377] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A processing method, comprising: The first intelligent agent drives the second intelligent agent to execute the target task, and the first intelligent agent is used to coordinate the operation of all second intelligent agents based on the target task. The second intelligent agent is used to fuse data collected by multiple data acquisition devices of an electronic device based on a target data fusion strategy to obtain target features, and to adjust the operating state of the electronic device based on the target features. Obtain the feedback data generated by the second intelligent agent in performing the target task; Based on the feedback data, the current first agent is adjusted, and the adjusted first agent is used as the new current first agent. The step of driving the second agent to perform the target task based on the current first agent continues.

2. The processing method according to claim 1, wherein obtaining the feedback data generated by the second intelligent agent performing the target task includes at least one of the following: Obtain first feedback data; the first feedback data includes the adjusted operating status of the electronic device; Obtain second feedback data; the second feedback data is used to reflect the state changes of the data objects to be operated by the second intelligent agent and the state of the second intelligent agent's occupation of computing resources; Obtain third-party feedback data; The third feedback data is used to reflect the performance of the first intelligent agent.

3. The processing method according to claim 1, wherein adjusting the current first agent based on the feedback data includes: Based on the feedback data, adjust at least one of the following: the current first intelligent agent's internal configuration, behavioral strategy, and target model; The internal configuration is used to characterize the computing environment required for the current first intelligent agent to run; The behavioral strategy is used to define task allocation behavior and resource scheduling behavior; The target model is used for task allocation and / or resource scheduling.

4. The processing method according to claim 1, wherein the second intelligent agent, based on a target data fusion strategy, fuses data collected by multiple data acquisition devices of the electronic device to obtain target features, and adjusts the operating state of the electronic device based on the target features, including: The second intelligent agent, based on a target data fusion strategy, fuses data collected by multiple data acquisition devices of the electronic device to obtain target features; The second agent acquires event rule description information and infers the event rule description information and the target features based on the target model to obtain the inference result; The reasoning result represents the current usage scenario of the electronic device; the event rule description information is used to describe the logical relationship between different events; Based on the reasoning results, the second intelligent agent adjusts the operating state of the electronic device.

5. The processing method according to claim 4, further comprising: The second agent searches the template library for a template that matches the target features; The template represents candidate use cases for the electronic device; The second intelligent agent adjusts the operating state of the electronic device based on the found template; If the second agent does not find a template matching the target feature in the template library, it obtains event rule description information and infers the event rule description information and the target feature based on the target model to obtain the inference result.

6. The processing method according to claim 4, further comprising: The second intelligent agent adjusts the target data fusion strategy based on the adjusted operating state of the electronic device.

7. The processing method according to claim 4, wherein fusing data collected by multiple data acquisition devices based on the target data fusion strategy to obtain target features includes: Extract the features of the data collected by each data acquisition device to obtain the features corresponding to each data acquisition device; The features corresponding to each of the data acquisition devices are enhanced to obtain the enhanced features corresponding to each of the data acquisition devices. The enhanced features characterize information related to the data collected by the data acquisition device and other data collected by other data acquisition devices. Determine the temporal and spatial relationships between the enhanced features corresponding to each of the data acquisition devices; Based on the aforementioned correlation, the enhanced features corresponding to each of the data acquisition devices are fused to obtain the target features.

8. The processing method according to claim 7, wherein the step of fusing the enhanced features corresponding to each of the data acquisition devices based on the association relationship to obtain the target features includes: Based on the aforementioned correlation, a first weighting factor is determined for the enhanced features corresponding to each of the data acquisition devices. Based on the quality parameters of the enhanced features corresponding to each of the data acquisition devices, a second weighting factor is determined for each of the data acquisition devices. Based on the second weight factor corresponding to each of the data acquisition devices, update the first weight factor of the enhanced feature corresponding to each of the data acquisition devices. Based on the updated first weight factor of the enhanced features corresponding to each of the data acquisition devices, the enhanced features corresponding to each of the data acquisition devices are weighted to obtain the target features.

9. The processing method according to claim 4, before fusing the data collected by multiple data acquisition devices based on the target data fusion strategy to obtain the target features, further comprising at least one of the following: Data collected by multiple data acquisition devices is processed to obtain time-aligned and / or spatially aligned data from the multiple data acquisition devices. Determine whether the time-aligned and / or spatially-aligned data from the plurality of data acquisition devices meet the adjustment conditions; If so, adjust the length of the sliding window; Using the adjusted sliding window as a unit, the time-aligned and / or spatially aligned data of the multiple data acquisition devices are adjusted in terms of time dimension and / or spatial dimension. The step of determining whether the time-aligned and / or spatially aligned data of the multiple data acquisition devices meet the adjustment conditions is continued until the target data of the multiple data acquisition devices is obtained. If not, using the sliding window as a unit, adjust the time-aligned and / or spatial-aligned data of the multiple data acquisition devices in terms of time and / or spatial dimensions, and continue to execute the step of determining whether the time-aligned and / or spatial-aligned data of the multiple data acquisition devices meet the adjustment conditions, until the target data of the multiple data acquisition devices is obtained.

10. A processing apparatus, comprising: The driving unit is used to drive the second agent to perform the target task based on the current first agent, and the first agent is used to coordinate the operation of all second agents. The second intelligent agent is used to fuse data collected by multiple data acquisition devices of an electronic device based on a target data fusion strategy to obtain target features, and to adjust the operating state of the electronic device based on the target features. The acquisition unit is used to acquire feedback data generated by the second intelligent agent in performing the target task; The adjustment unit is used to adjust the current first agent based on the feedback data, take the adjusted first agent as the new current first agent, and continue to execute the step of driving the second agent to perform the target task based on the current first agent.

Images