An attention mechanism reinforced perception decision system closed loop construction method
By constructing a three-layer attention collaboration mechanism and closed-loop optimization circuit for perception, decision-making, and task, the problem of insufficient adaptability and robustness of traditional perception and decision-making systems in complex environments is solved, achieving efficient collaboration of dynamic perception and decision-making, and improving the system's adaptive capability and resource utilization efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-17
- Publication Date
- 2026-07-10
Smart Images
Figure CN122366530A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of decision-making system technology, and in particular to a method for constructing a closed-loop perception-decision system with enhanced attention mechanisms. Background Technology
[0002] Traditional perception-decision systems typically employ a sequential or partially feedback architecture of "perception-modeling-planning-execution," but they lack adaptability, efficiency, and robustness when dealing with scenarios such as sudden environmental changes, diverse objectives, or information overload.
[0003] Attention mechanisms have achieved success in machine translation, image recognition and other fields, but their application is mostly limited to feature weighting within the model. Existing related technologies cannot systematically and multi-dimensionally embed attention mechanisms into the complete closed loop from initial perception to final decision execution, making it the core hub driving the system's dynamic adaptation and autonomous evolution. Therefore, this invention proposes a method for constructing a closed loop of a perception-decision system enhanced by attention mechanisms to solve the problems existing in the prior art. Summary of the Invention
[0004] To address the aforementioned problems, the present invention aims to propose a closed-loop construction method for a perception-decision system enhanced with attention mechanisms. This method constructs a three-layer attention coordination mechanism and a closed-loop optimization loop for perception, decision-making, and task, enabling the system to dynamically focus on key information, proactively adjust perception based on decision-making needs, and optimize attention strategies online based on execution results.
[0005] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a closed-loop construction method for a perception-decision system enhanced with attention mechanisms, comprising the following steps:
[0006] Step 1: Receive the raw environmental perception signal, generate dynamic perception focus weights based on the perception attention network and the expected attention guidance signal from the decision-making process, and selectively process the raw environmental perception signal to obtain task-related perception features.
[0007] Step 2: Evaluate task-related perceptual features based on the decision attention network, generate the optimal decision action and the decision contribution distribution representing the contribution of each perceptual feature, and use it as the expected attention guidance signal for subsequent perceptual focusing;
[0008] Step 3: Execute the optimal decision action and obtain environmental feedback signals. Then, based on the task attention network and task context, analyze the correlation between environmental feedback signals and attention allocation to generate attention optimization gradients.
[0009] Step 4: Utilize attention optimization gradients to update the parameters of the perceptual attention network and the decision attention network online in order to optimize the attention allocation mechanism;
[0010] Step 5: Apply the optimized attention allocation mechanism to process the perceptual signals of the next moment. Through continuous iteration, a reinforced closed loop is formed, in which perception and decision-making are deeply coupled and the attention strategy evolves in tandem with the task and the environment.
[0011] Further improvements are made in the following ways: the perceptual attention network in step one is one or a combination of spatial attention mechanism, channel attention mechanism, and spatiotemporal attention mechanism; the original environmental perception signal is a multimodal signal, and the original environmental perception signal is received to perform feature alignment and fusion of signals of different modalities.
[0012] The further improvement lies in the following: In step one, the generation of the perception focus weight specifically involves calculating the self-attention of the original environmental perception signal to obtain preliminary saliency, and then fusing and normalizing the preliminary saliency with the expected attention-guided signal to finally generate the perception focus weight; the perception focus weight is represented in the form of a probability distribution or a continuous weight vector, and is used to perform feature-level weighting, gating, or selective activation operations on the original perception signal.
[0013] A further improvement lies in the following: the calculation of self-attention specifically involves transforming the original environmental perception signal through a linear transformation, mapping it to a query vector Q, a key vector K, and a value vector V, respectively. Then, by calculating the scaled dot product of the query vector Q and the key vector K, and normalizing it using the Softmax function, a self-attention weight matrix A is generated. self , represented as A self = Softmax(QK T / √d k ), where K T d represents the transpose of the key vector K. k Given the dimension of the key vector, the self-attention weight matrix A is finally... self Multiplying it by the value vector V yields a weighted feature representation, which serves as the initial significance.
[0014] A further improvement is that the fusion and normalization process specifically involves transforming the expected attention guidance signal from the decision-making process into a guidance weight vector G compatible with the initial saliency dimension through a learnable projection layer. Then, the initial saliency and the guidance weight vector G are fused through a fully connected layer by element-wise summation or concatenation to obtain a fusion feature F. Finally, the fusion feature F is normalized again using the Softmax function to generate the final perception focus weights used to weight the original perception signal.
[0015] The further improvement lies in the following: In step two, generating the decision contribution distribution specifically involves the decision attention network evaluating the value of task-related perceptual features or calculating the policy gradient, quantifying the value influence of each feature dimension on the current decision action, and generating a weight distribution based on the influence.
[0016] Further improvements are made in the following aspects: In step three, the task context includes the current stage task objective, the sequence of historical decisions and execution results, preset environmental constraints, and the internal state of the system; the task attention network integrates a gated recurrent unit or a memory network to maintain and update the long-term memory of the task context.
[0017] A further improvement is that the environmental feedback signal in step three includes an external explicit reward signal, a sparse reward signal, and an intrinsic reward signal or prediction error signal generated based on internal state prediction.
[0018] A further improvement is that: in step three, generating the attention optimization gradient includes evaluating the expected performance improvement that could be achieved by adjusting the historical attention allocation pattern in the current task context, and converting the expected performance improvement into the gradient direction and magnitude of the historical attention distribution parameters.
[0019] A further improvement is that the online update in step four specifically adopts a soft update strategy, that is, the optimization gradient is applied to the target network parameters through a weighted average method, and the update magnitude of the target network is less than or equal to that of the current network.
[0020] The beneficial effects of this invention are as follows: This invention constructs a three-layer attention coordination mechanism and closed-loop optimization circuit of perception, decision-making and task, enabling it to dynamically focus on key information, actively adjust perception according to decision-making needs, and optimize attention strategies online based on execution results;
[0021] The method of this invention significantly enhances the pertinence of information processing and the accuracy of decision-making in complex environments, improves the adaptive evolution speed of the overall closed loop, the robustness of situational response, and the efficiency of computing resource utilization, and achieves a fundamental improvement from static processing to dynamic perception and decision-making collaboration. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation
[0023] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.
[0024] Traditional perception-decision systems typically employ a sequential or partially feedback architecture of "perception-modeling-planning-execution," but such architectures have inherent limitations:
[0025] First, the perception module typically processes all input information in a fixed pattern, making it susceptible to interference from irrelevant information in complex or dynamic environments, leading to wasted computing resources and loss of key features. Second, the interaction between the decision-making module and the perception module is often static or unidirectional, making it difficult for decision-making needs to be reflected in the perception focus in real time and with precision. Furthermore, the system's closed-loop optimization relies heavily on slow external rewards or manual parameter tuning, lacking fine-grained, online self-adjustment capabilities based on internal attention allocation. These limitations result in insufficient adaptability, efficiency, and robustness of the system when dealing with sudden environmental changes, diverse objectives, or information overload.
[0026] In recent years, attention mechanisms have achieved success in fields such as machine translation and image recognition, but their applications are mostly limited to feature weighting within the model. Existing technologies cannot systematically and multi-dimensionally embed attention mechanisms into the complete closed loop from initial perception to final decision execution, and make it the core hub driving the system's dynamic adaptation and autonomous evolution. Existing technologies do not have a perfect solution.
[0027] according to Figure 1 As shown, this embodiment provides a closed-loop construction method for a perception-decision system enhanced with attention mechanisms, including the following steps:
[0028] Step 1: Receive the raw environmental perception signal, generate dynamic perception focus weights based on the perception attention network and the expected attention guidance signal from the decision-making process, and selectively process the raw environmental perception signal to obtain task-related perception features.
[0029] The perceptual attention network is one or more combinations of spatial attention mechanism, channel attention mechanism, and spatiotemporal attention mechanism;
[0030] The original environmental perception signal is a multimodal signal. The original environmental perception signal is received and the signals of different modes are aligned and fused.
[0031] The perception focus weight is represented in the form of a probability distribution or a continuous weight vector, and is used to perform feature-level weighting, gating or selective activation operations on the original perception signal;
[0032] The generation of perception focus weights involves calculating the self-attention of the original environmental perception signal to obtain preliminary saliency, then fusing and normalizing the preliminary saliency with the expected attention-guided signal, and finally generating the perception focus weights.
[0033] The calculation of self-attention involves linearly transforming the original environmental perception signal into a query vector Q, a key vector K, and a value vector V. Then, the self-attention weight matrix A is generated by calculating the scaled dot product of the query vector Q and the key vector K, followed by normalization using the Softmax function. self , represented as A self = Softmax(QK T / √d k ), where K T d represents the transpose of the key vector K. k Given the dimension of the key vector, the self-attention weight matrix A is finally... self Multiplying it by the value vector V yields a weighted feature representation, which serves as the initial significance.
[0034] The fusion and normalization process specifically involves transforming the expected attention guidance signal from the decision-making process into a guidance weight vector G compatible with the initial saliency dimension through a learnable projection layer. Then, the initial saliency and the guidance weight vector G are fused through a fully connected layer by element-wise summation or concatenation to obtain the fusion feature F. Finally, the fusion feature F is normalized again by applying the Softmax function to generate the final perception focus weights used to weight the original perception signal.
[0035] Step 2: Evaluate task-related perceptual features based on the decision attention network, generate the optimal decision action and the decision contribution distribution representing the contribution of each perceptual feature, and use it as the expected attention guidance signal for subsequent perception focusing. The specific decision contribution distribution is cached after generation and then aligned with the perception signal processing cycle of the next time step to serve as the expected attention guidance signal for the next time step.
[0036] Specifically, the decision contribution distribution is generated by the decision attention network evaluating the value of task-related perceptual features or calculating policy gradients, quantifying the value influence of each feature dimension on the current decision action, and generating a weight distribution based on the influence.
[0037] Step 3: Execute the optimal decision action and obtain environmental feedback signals. Then, based on the task attention network and task context, analyze the correlation between environmental feedback signals and attention allocation to generate attention optimization gradients.
[0038] The task context includes the current stage task objective, the sequence of historical decisions and execution results, preset environmental constraints, and the internal state of the system.
[0039] Task attention networks integrate gated recurrent units or memory networks to maintain and update long-term memories of the task context;
[0040] Environmental feedback signals include external explicit reward signals, sparse reward signals, and intrinsic reward signals or prediction error signals generated based on internal state predictions.
[0041] Generating attention optimization gradients involves evaluating the expected performance improvement that would result from adjusting the historical attention allocation pattern in the current task context, and then converting this expected performance improvement into the gradient direction and magnitude with respect to the historical attention distribution parameters.
[0042] Step 4: Utilize attention optimization gradients to update the parameters of the perceptual attention network and the decision attention network online in order to optimize the attention allocation mechanism;
[0043] The online update specifically adopts a soft update strategy, which means that the optimization gradient is applied to the target network parameters through a weighted average method, and the update magnitude of the target network is less than or equal to that of the current network.
[0044] Step 5: Apply the optimized attention allocation mechanism to process the perceptual signals of the next moment. Through continuous iteration, a reinforced closed loop is formed, in which perception and decision-making are deeply coupled and the attention strategy evolves in tandem with the task and the environment.
[0045] This method cyclically executes the steps from receiving perceptual signals to updating the attention network at fixed time intervals or through event triggering, forming a continuous online learning and adaptation closed loop.
[0046] "Fixed time interval trigger" is an example:
[0047] The robot control system is set to actively collect fused data from the LiDAR and vision sensors every 50 milliseconds (i.e., 20Hz) to perform a complete attention calculation and motion planning, ensuring the real-time performance and stability of the control.
[0048] The autonomous vehicle decision-making module is set to perform a complete perception modeling and trajectory prediction of the surrounding environment every 100 milliseconds (10Hz) during the high-speed cruise phase, and update the driving strategy accordingly.
[0049] The industrial process monitoring system is set to perform a joint analysis of multiple signals such as images, temperature, and pressure of products on the production line every second to determine the process status and make adjustments.
[0050] "Event triggering" is an example of:
[0051] By setting event triggers based on perceptual saliency, the security monitoring system will only trigger a high-resolution attention perception and behavior analysis process when the motion detection algorithm identifies an abnormal moving area in the continuous video stream.
[0052] An internal event trigger based on decision uncertainty is set up. When the confidence of the optimal action output by the decision attention network is lower than a preset threshold, a more refined perception scan and decision re-evaluation process is actively triggered.
[0053] By setting events based on external commands or state changes, the intelligent voice assistant will only initiate a complete closed loop of voice signal attention enhancement, semantic understanding, and task planning after detecting the wake word; or it will immediately initiate path planning and dynamic obstacle avoidance decision-making processes when the robot receives a new navigation target point command.
[0054] By setting event triggers based on sudden changes in environmental conditions, the UAV immediately triggers a high-frequency environmental perception and stability control decision loop when the IMU sensor detects severe attitude disturbances (such as encountering strong winds).
[0055] Before the process starts, this method also needs to initialize the "perceptual attention network", "decision attention network" and "task attention network" to provide them with a reasonable and non-random initial parameter state. This ensures that the core attention mechanism has basic functions when the closed loop runs for the first time, instead of starting from a completely chaotic state. This avoids the problems of extremely poor performance, slow learning speed or even failure to start in the initial stage.
[0056] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for constructing a closed-loop perceptual decision-making system enhanced with attention mechanisms, characterized in that, Includes the following steps: Step 1: Receive the raw environmental perception signal, generate dynamic perception focus weights based on the perception attention network and the expected attention guidance signal from the decision-making process, and selectively process the raw environmental perception signal to obtain task-related perception features. Step 2: Evaluate task-related perceptual features based on the decision attention network, generate the optimal decision action and the decision contribution distribution representing the contribution of each perceptual feature, and use it as the expected attention guidance signal for subsequent perceptual focusing; Step 3: Execute the optimal decision action and obtain environmental feedback signals. Then, based on the task attention network and task context, analyze the correlation between environmental feedback signals and attention allocation to generate attention optimization gradients. Step 4: Utilize attention optimization gradients to update the parameters of the perceptual attention network and the decision attention network online in order to optimize the attention allocation mechanism; Step 5: Apply the optimized attention allocation mechanism to process the perceptual signals of the next moment. Through continuous iteration, a reinforced closed loop is formed, in which perception and decision-making are deeply coupled and the attention strategy evolves in tandem with the task and the environment.
2. The method for constructing a closed-loop perception and decision-making system with enhanced attention mechanism according to claim 1, characterized in that: In step one, the perceptual attention network is one or more of the following: spatial attention mechanism, channel attention mechanism, and spatiotemporal attention mechanism; the original environmental perception signal is a multimodal signal, and the original environmental perception signal is received to perform feature alignment and fusion of signals of different modalities.
3. The method for constructing a closed-loop perception and decision-making system with enhanced attention mechanism according to claim 1, characterized in that: In step one, the generation of perception focus weights specifically involves calculating the self-attention of the original environmental perception signal to obtain preliminary saliency, then fusing and normalizing the preliminary saliency with the expected attention-guided signal, and finally generating perception focus weights. The perception focus weights are represented in the form of a probability distribution or a continuous weight vector, and are used to perform feature-level weighting, gating, or selective activation operations on the original perception signal.
4. The method for constructing a closed-loop perception and decision-making system with enhanced attention mechanism according to claim 3, characterized in that: The calculation of self-attention specifically involves linearly transforming the original environmental perception signal into a query vector Q, a key vector K, and a value vector V, respectively. Then, by calculating the scaled dot product of the query vector Q and the key vector K, and normalizing it using the Softmax function, a self-attention weight matrix A is generated. self , represented as A self = Softmax(QK T / √d k ), where K T d represents the transpose of the key vector K. k Given the dimension of the key vector, the self-attention weight matrix A is finally... self Multiplying it by the value vector V yields a weighted feature representation, which serves as the initial significance.
5. The method for constructing a closed-loop perception decision-making system with enhanced attention mechanism according to claim 3, characterized in that: The fusion and normalization process specifically involves transforming the expected attention guidance signal from the decision-making process into a guidance weight vector G compatible with the initial saliency dimension through a learnable projection layer. Then, the initial saliency and the guidance weight vector G are fused through a fully connected layer by element-wise summation or concatenation to obtain the fusion feature F. Finally, the fusion feature F is normalized again by applying the Softmax function to generate the final perception focus weights used to weight the original perception signal.
6. The method for constructing a closed-loop perception and decision-making system with enhanced attention mechanism according to claim 1, characterized in that: In step two, generating the decision contribution distribution specifically involves the decision attention network evaluating the value of task-related perceptual features or calculating the policy gradient, quantifying the value influence of each feature dimension on the current decision action, and generating a weight distribution based on the influence.
7. The method for constructing a closed-loop perception and decision-making system with enhanced attention mechanism according to claim 1, characterized in that: In step three, the task context includes the current stage task objective, the sequence of historical decisions and execution results, preset environmental constraints, and the internal state of the system; the task attention network integrates a gated recurrent unit or a memory network to maintain and update the long-term memory of the task context.
8. The method for constructing a closed-loop perception decision-making system with enhanced attention mechanism according to claim 1, characterized in that: The environmental feedback signals in step three include external explicit reward signals, sparse reward signals, and intrinsic reward signals or prediction error signals generated based on internal state prediction.
9. The method for constructing a closed-loop perception and decision-making system with enhanced attention mechanism according to claim 1, characterized in that: The third step of generating the attention optimization gradient includes evaluating the expected performance improvement that would be brought about by adjusting the historical attention allocation pattern in the current task context, and converting the expected performance improvement into the gradient direction and magnitude of the historical attention distribution parameters.
10. The method for constructing a closed-loop perception decision-making system with enhanced attention mechanism according to claim 1, characterized in that: The online update in step four specifically adopts a soft update strategy, that is, the optimization gradient is applied to the target network parameters by weighted averaging, and the update magnitude of the target network is less than or equal to that of the current network.