Complex task-oriented multi-level electroencephalogram decoding and dynamic instruction generation method

Abstract

The invention discloses a complex task-oriented multi-level electroencephalogram decoding and dynamic instruction generation method, which comprises the following steps of S1, strategic layer intention recognition: analyzing a task characteristic wave, namely a mu / beta rhythm, in an electroencephalogram signal of a user during imagination, and recognizing a target instruction of the user; s2, tactical layer task decomposition: performing action decomposition on the target instruction identified in the S1, automatically planning execution steps, and matching an optimal execution path by using a'task knowledge base '. According to the method, three-layer assembly line decoding of'strategic layer-tactical layer-action layer 'is adopted, multi-level electroencephalogram decoding and dynamic instruction generation oriented to complex tasks based on electroencephalogram signals are achieved, the complex brain control tasks are converted into natural interaction of'once desired', and the problems of intention expression overload and sequential logic disorder in the prior art can be effectively solved.

Description

Technical Field

[0001] This invention relates to the field of electroencephalography (EEG) systems, and more particularly to a multi-level EEG decoding and dynamic instruction generation method for complex tasks. Background Technology

[0002] Existing brain control systems suffer from the following fundamental defects in complex tasks: 1. Intent expression overload: users need to generate each sub-action intent separately (e.g., "grab-move-place"); 2. Disordered temporal logic: sub-intent decoding is performed independently, lacking task logic connection (e.g., "move" is triggered without detecting "grab success"); In view of the above, this application proposes a multi-level EEG decoding and dynamic instruction generation method for complex tasks. Summary of the Invention

[0003] Based on the technical problems existing in the background technology, this invention proposes a multi-level EEG decoding and dynamic instruction generation method for complex tasks.

[0004] The multi-level EEG decoding and dynamic instruction generation method for complex tasks proposed in this invention includes the following steps:

[0005] S1: Strategic layer intention recognition: When used for imagination, it analyzes the "task characteristic wave" in the user's EEG signal, i.e. μ / β rhythm, to identify the user's target instruction;

[0006] S2: Tactical task decomposition: Decompose the target instructions identified in S1 into action steps, automatically plan the execution steps, and use the "task knowledge base" to match the optimal execution path;

[0007] S3: Action layer instruction generation: The optimal execution path in S2 is dynamically generated into a safe instruction through the "adaptive learning engine", and the instruction that the device can directly execute is output.

[0008] S4: The device executes the instruction and determines whether the execution is successful. If successful, it automatically proceeds to the next step.

[0009] S5: If the device fails to execute, perform intelligent diagnosis to determine whether the error is due to a target error, a step error, or an action error. If the target is incorrect, re-enter the target; if the step is incorrect, switch to an alternative solution; if the action is incorrect, adjust the parameters.

[0010] S6: Adopt an error emergency mechanism to monitor the strategic, tactical, and action layers in real time, monitor their confidence levels, and trigger a cross-level error correction mechanism when the confidence level of a layer is lower than the threshold.

[0011] Preferably, the specific logical steps of S1 are as follows:

[0012] S101: Extract "task characteristic waves" from EEG, namely 8-25Hz rhythms;

[0013] S102: Focus on key decision-making moments through an "attention filtering mechanism";

[0014] S103: Decodes key decision-making moments and outputs strategic objective encoding vectors. The formula used is: S = ∑ t α t ×h t ;

[0015] in α t For attention weights, h t These are characteristics of electroencephalography (EEG).

[0016] Preferably, the specific logical steps of S2 are as follows:

[0017] S201: Search the "Task Knowledge Base" to find all possible action paths;

[0018] S202: Select the solution that best matches the target, has the fewest steps, and is feasible for the equipment. The matching process uses a tactical-level path scoring formula, which is as follows:

[0019] Where λ is the path length penalty coefficient, typically 0.2, p is the number of nodes in the path, i.e., the number of steps, μ is the feasibility reward coefficient, typically 0.3; the feasibility indicator function is 0 or 1; For semantic matching, IIS feasible For feasibility;

[0020] S203: Output the optimal task sequence.

[0021] Preferably, in step S3, when dynamically generating safety instructions through the "adaptive learning engine," a reinforcement learning reward mechanism that integrates EEG matching is adopted. The tactical nodes are mapped to device control instructions through a reinforcement learning policy network, and the reward function formula used is: r t =+2e - ||V EEG -V expected ||;

[0022] in Generate instructions:

[0023] Preferably, the specific logical steps of S3 are as follows:

[0024] S301: Input: Current step + device status;

[0025] S302: Dynamic calculation: Rewards completed actions and severely punishes dangerous operations. The higher the EEG matching degree, the closer the instructions are to the user's intention.

[0026] S303: Output: Instructions that the device can execute directly.

[0027] Preferably, in step S6, the error emergency mechanism employs a three-tiered security network, monitors the confidence level of each tier in real time, and automatically triggers error correction, as detailed in the table below:

[0028]

[0029]

[0030] Compared with existing technologies, the beneficial effects of this invention are:

[0031] This invention employs a three-layer pipeline decoding approach: "strategic layer → tactical layer → action layer." It enables multi-level EEG decoding and dynamic instruction generation for complex tasks based on EEG signals, transforming complex brain-controlled tasks into natural interactions that are "instantaneous." This effectively solves the problems of overloaded intent expression and disordered temporal logic in existing technologies. Attached Figure Description

[0032] Figure 1 This is a flowchart of the multi-level EEG decoding and dynamic instruction generation method for complex tasks proposed in this invention;

[0033] Figure 2 This diagram illustrates the core steps of the multi-level EEG decoding and dynamic instruction generation method for complex tasks proposed in this invention. Detailed Implementation

[0034] The present invention will be further explained below with reference to specific embodiments.

[0035] Example

[0036] Reference Figure 1-2 This embodiment proposes a multi-level EEG decoding and dynamic instruction generation method for complex tasks, including the following steps:

[0037] S1: Strategic layer intention recognition: When used for imagination, it analyzes the "task characteristic wave" in the user's EEG signal, i.e. μ / β rhythm, to identify the user's target instruction. Its strategic layer is to analyze the "ultimate goal" in the user's brain, such as "assembling a gearbox".

[0038] The specific logical steps are as follows:

[0039] S101: Extract "task characteristic waves" from EEG, namely 8-25Hz rhythms;

[0040] S102: Focus on key decision moments through an "attention filtering mechanism". For example, when a user imagines "assembling", the system automatically ignores the noise in the first 2 seconds and locks in the high confidence signal in the last 3 seconds.

[0041] S103: Decode key decision moments and output strategic objective encoding vectors, such as: Assembly = 0.92 | Disassembly = 0.05 (the larger the value, the higher the confidence level);

[0042] The formula used is: S = ∑ t α t ×h t ;

[0043] in α t For attention weights, h t EEG characteristics;

[0044] The above steps solve the problem that traditional EEG decoding requires users to have a long period of stable imagination (≥5 seconds) and is easily affected by noise interference. Users only need to imagine naturally for 3 seconds, and the system will automatically capture the most relevant EEG segments, effectively reducing cognitive load.

[0045] S2: Tactical task decomposition: Decompose the target instructions identified in S1 into action steps, automatically plan the execution steps, and use the "task knowledge base" to match the optimal execution path. For example, GPS route planning: select the route with the shortest time, the fewest traffic lights, and avoidance of construction among all roads.

[0046] The specific logical steps are as follows:

[0047] S201: Search the "Task Knowledge Base" to find all possible action paths;

[0048] S202: Select the solution that best matches the target, has the fewest steps, and is feasible for the equipment. The matching process uses a tactical-level path scoring formula, which is as follows:

[0049] Where λ is the path length penalty coefficient, typically 0.2, p is the number of nodes in the path, i.e., the number of steps, μ is the feasibility reward coefficient, typically 0.3; the feasibility indicator function is 0 or 1; For semantic matching, IIS feasible For feasibility; this formula, through a three-level design of semantic matching as the main factor, step penalty to control effectiveness, and feasibility as a safety net, can avoid execution risks caused by incorrect path selection from the source in high-risk scenarios such as industry and medical care;

[0050] S203: Output the optimal task sequence, such as "grab A → insert B → fasten";

[0051] S3: Action layer instruction generation: The optimal execution path in S2 is dynamically generated into a safe instruction through the "adaptive learning engine", and the instruction that the device can directly execute is output.

[0052] When dynamically generating safety commands using an "adaptive learning engine," a reinforcement learning reward mechanism that integrates EEG matching is employed. A reinforcement learning policy network maps tactical nodes to device control commands, and the reward function formula used is: r t =+2e - ||V EEG -V expected ||;

[0053] in Generate instructions:

[0054] The specific logical steps are as follows:

[0055] S301: Input: Current step (e.g., gripping gear A) + device status (e.g., robotic arm position);

[0056] S302: Dynamic calculation: Rewards completed actions and severely punishes dangerous operations. The higher the EEG matching degree, the closer the instructions are to the user's intention.

[0057] S303: Output: Directly executable instructions, such as "Move to coordinates (10,20) at a speed of 5cm / s";

[0058] S4: The device executes the instruction and determines whether the execution is successful. If successful, it automatically proceeds to the next step.

[0059] S5: If the device fails to execute, perform intelligent diagnosis to determine whether the error is due to a target error, a step error, or an action error. If the target is incorrect, re-enter the target; if the step is incorrect, switch to an alternative solution; if the action is incorrect, adjust the parameters.

[0060] S6: Implement an error emergency mechanism to monitor the strategic, tactical, and action layers in real time, monitor their confidence levels, and trigger a cross-level error correction mechanism when the confidence level of a layer falls below the threshold.

[0061] The error response mechanism employs a three-tiered safety net, which monitors the confidence level of each tier in real time and automatically triggers error correction, as detailed in the table below:

[0062]

[0063] In this embodiment, the present invention adopts a three-layer pipeline decoding of "strategic layer → tactical layer → action layer" to realize multi-level EEG decoding and dynamic instruction generation based on EEG signals for complex tasks, transforming complex brain control tasks into natural interactions that can be achieved "instantly", which can effectively solve the problems of intention expression overload and temporal logic disorder in the existing technology.

[0064] Case

[0065] Scenario: A brain-controlled robotic arm grasps a fragile glass tube.

[0066] 1. Strategic Input:

[0067] Users wear EEG devices and imagine "grabbing" (only one macroscopic intention is needed);

[0068] Formula output: s = [0.12, 0.88, 0.00] (88% confidence level);

[0069] 2. Tactical Breakdown:

[0070] The knowledge base is matched with the "precise crawling process";

[0071] Score=0.85-0.2×3+0.3×1==0.95;

[0072] 3. Action generation:

[0073] RL network calculation of crawling intensity (reward function constraint):

[0074] r t =+2e -匹配度 -0.1×|force-5N|, output force = 4.8N (lower than the glass's withstand limit of 5N);

[0075] 4. Error correction scenario:

[0076] If the position detection offset is 1.5mm (>threshold 1mm): Δx = k × error = 0.5 × 1.5 = 0.75mm, then the coordinates will be automatically compensated to (x + 0.75mm, y).

[0077] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-level EEG decoding and dynamic instruction generation method for complex tasks, characterized in that, Includes the following steps: S1: Strategic layer intention recognition: When used for imagination, it analyzes the "task characteristic wave" in the user's EEG signal, namely the μ / β rhythm, to identify the user's target instruction; S2: Tactical task decomposition: Decompose the target instructions identified in S1 into action steps, automatically plan the execution steps, and use the "task knowledge base" to match the optimal execution path; S3: Action layer instruction generation: The optimal execution path in S2 is dynamically generated into a safe instruction through the "adaptive learning engine", and the instruction that the device can directly execute is output. S4: The device executes the instruction and determines whether the execution is successful. If successful, it automatically proceeds to the next step. S5: If the device fails to execute, perform intelligent diagnosis to determine whether the error is due to a target error, a step error, or an action error. If the target is incorrect, re-enter the target; if the step is incorrect, switch to an alternative solution; if the action is incorrect, adjust the parameters. S6: Adopt an error emergency mechanism to monitor the strategic, tactical, and action layers in real time, monitor their confidence levels, and trigger a cross-level error correction mechanism when the confidence level of a layer is lower than the threshold.

2. The multi-level EEG decoding and dynamic instruction generation method for complex tasks according to claim 1, characterized in that, The specific logical steps of S1 are as follows: S101: Extracts "task characteristic waves" from EEG, namely 8-25Hz rhythms; S102: Focus on key decision-making moments through an "attention filtering mechanism"; S103: Decodes key decision-making moments and outputs strategic objective encoding vectors. The formula used is: S = ∑ t α t ×h t ; in α t For attention weights, h t These are characteristics of electroencephalography (EEG).

3. The multi-level EEG decoding and dynamic instruction generation method for complex tasks according to claim 1, characterized in that, The specific logical steps of S2 are as follows: S201: Search the "Task Knowledge Base" to find all possible action paths; S202: Select the solution that best matches the target, has the fewest steps, and is feasible for the equipment. The matching process uses a tactical-level path scoring formula, which is as follows: Where λ is the path length penalty coefficient, typically 0.2, p is the number of nodes in the path, i.e., the number of steps, μ is the feasibility reward coefficient, typically 0.3; the feasibility indicator function is 0 or 1; For semantic matching, Ⅱ feasible For feasibility; S203: Output the optimal task sequence.

4. The multi-level EEG decoding and dynamic instruction generation method for complex tasks according to claim 1, characterized in that, In S3, when dynamically generating safety commands through the "adaptive learning engine," a reinforcement learning reward mechanism that integrates EEG matching is adopted. The tactical nodes are mapped to device control commands through a reinforcement learning policy network, and the reward function formula used is: r t =+2e - ||V EEG -V expected ||; in Generate instructions:

5. The multi-level EEG decoding and dynamic instruction generation method for complex tasks according to claim 1, characterized in that, The specific logical steps of S3 are as follows: S301: Input: Current step + device status; S302: Dynamic calculation: Rewards completed actions and severely punishes dangerous operations. The higher the EEG matching degree, the closer the instructions are to the user's intention. S303: Output: Instructions that the device can execute directly.

6. The multi-level EEG decoding and dynamic instruction generation method for complex tasks according to claim 1, characterized in that, In S6, the error response mechanism adopts a three-tiered security network, monitors the confidence level of each tier in real time, and automatically triggers error correction, as detailed in the table below:

Images