A structured instruction constraint and task guidance method based on dynamic feature weight
By dynamically calculating feature weights and generating structured instructions, the problem of fixed feature weights in multi-stage tasks is solved, enabling the model to focus precisely on different stages and accurately execute task objectives. This improves the adaptability and stability of task analysis and is suitable for multimodal data processing and complex task scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2026-04-28
- Publication Date
- 2026-06-19
AI Technical Summary
Existing technologies lack dynamic feature weighting mechanisms in multi-stage tasks, which makes it difficult to effectively address the differences in the importance of features at different stages when task objectives change. This is especially true in fields such as rehabilitation assessment of brain injury patients and monitoring of industrial equipment, where existing methods cannot accurately reflect key features, resulting in insufficient adaptability and flexibility of the analysis results.
By calculating the standardized average change of features at each stage, the weights of each indicator are dynamically calculated, and structured instructions are generated based on these weights. This guides the model to focus on the most relevant features at each stage, forming a dynamically adjusted closed-loop process that ensures the model adapts to task objectives and stage changes in multi-stage tasks.
It improves the accuracy and efficiency of multi-stage task execution. The model can dynamically adjust the focus at different stages to adapt to changes in task objectives, thereby enhancing the adaptability and stability of task analysis. It is particularly suitable for the medical and industrial fields.
Smart Images
Figure CN122241164A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of artificial intelligence, behavior analysis, and multimodal data processing technology, and specifically relates to a structured instruction constraint and task guidance method based on dynamic feature weights. Background Technology
[0002] Currently, technologies in the fields of video understanding and behavior analysis heavily rely on deep learning models, especially architectures such as Convolutional Neural Networks (CNNs) and Transformers. These models help identify, classify, and predict behaviors and events in videos by automatically extracting visual features. Existing technologies typically treat video data as continuous temporal data, generating an understanding of the video through processing steps such as feature extraction, aggregation, and classification. However, when handling multi-stage tasks, these models often use a uniform feature processing and weighting mechanism, lacking the ability to dynamically adapt to different task stages.
[0003] Taking the rehabilitation assessment of brain injury patients as an example, existing technologies typically analyze videos of the patient's behavior to assess progress during rehabilitation. The technical process involves first capturing videos of the patient's behavior using high-precision sensors or cameras, and then extracting key features such as movement patterns and gait stability from each frame using a convolutional neural network. Next, a temporal model utilizes these features to analyze the dynamic changes in the patient during rehabilitation and identify changes in behavioral patterns.
[0004] In existing technologies, feature weighting is a fixed process where all extracted features are typically assigned the same weight and then fused and standardized according to pre-defined rules. This allows different types of features to be unified into a single evaluation result. For example, in rehabilitation assessment, the acute phase may focus more on the patient's motor ability and basic balance control, while in the chronic phase, the focus shifts to the patient's coordination and fine motor control. Existing technologies use a uniform feature weighting method in this process, failing to dynamically adjust the importance of features according to the task requirements of each stage.
[0005] In handling multi-stage tasks, existing technologies employ fixed processing flows, lacking flexible mechanisms to adjust the model's focus even when task objectives change. For example, in action recognition tasks, models typically focus on fixed features such as the size or speed of actions, while in video question answering tasks, they may need to pay more attention to specific objects or scenes in the video. Although existing deep learning models and video analysis methods have achieved significant results in many fields, they generally rely on uniform feature weighting mechanisms and task processing flows, failing to fully consider the changing task objectives and data evolution patterns in multi-stage tasks. When dealing with tasks with long-term characteristics, they often lack sufficient accuracy and adaptability.
[0006] While existing technologies can extract and process features using deep learning models for video understanding and behavior analysis tasks, the fixed nature of feature weighting leads to significant drawbacks in multi-stage tasks (such as rehabilitation assessment and equipment maintenance). This is particularly true in tasks like rehabilitation assessment for brain-injured patients, where the importance of features changes over time and throughout the rehabilitation process; existing technologies struggle to effectively address the differences in feature importance between different stages.
[0007] Existing technologies employ a fixed feature weighting strategy, assigning the same importance and weight to all features extracted from videos throughout the task execution process. The task objectives and required features differ at different stages, and existing methods fail to dynamically adjust feature weights to meet the needs of each stage. For example, in the rehabilitation of brain injury patients, motor ability is the most important assessment feature during the acute phase, while coordination and gait stability become key assessment focuses during the chronic phase. Existing technologies cannot dynamically adjust feature weights based on these changing needs at each stage, resulting in inaccurate assessments of patient rehabilitation progress reflecting key features at each stage. This fixed weighting mechanism presents problems not only in the medical field, such as rehabilitation assessment, but also in other fields, such as industrial equipment monitoring and risk assessment. As task stages change and data evolves, existing technologies cannot effectively address the dynamic changes in feature importance and focus, leading to insufficient adaptability and flexibility in multi-stage task processing.
[0008] To address the above problems, this invention proposes a structured instruction constraint and task guidance method based on dynamic feature weights. Summary of the Invention
[0009] The purpose of this invention is to provide a structured instruction constraint and task guidance method based on dynamic feature weights. The core idea is to dynamically calculate the weights of each indicator at different stages by calculating the standardized average change of features at each stage. These weight values are used to adjust the structured problem input to the model, thereby guiding task execution. Specifically, this invention dynamically generates structured instructions based on the weight values at different stages. These instructions, as input to the model, guide the model to focus on the most relevant features at each stage to better complete the task.
[0010] By introducing dynamic feature weights and structured instruction constraints, this invention not only flexibly adjusts feature weights across multiple stages but also continuously updates instructions during task execution, thereby guiding the model to focus on the most critical task features at each stage. In this process, structured instructions and task guidance form a dynamically adjusted closed-loop process, ensuring that the model can adapt to different task objectives and stage changes when handling multi-stage tasks, thus improving the accuracy and efficiency of task execution.
[0011] This invention solves the problems of fixed feature weighting, lack of task guidance, and poor stage adaptability in existing technologies through this innovative process. It enables the model to accurately focus on the key features of each stage when dealing with complex multi-stage tasks, ensuring the accurate execution of task objectives. It has broad application prospects, especially in fields such as medicine, industry, and multimodal data analysis.
[0012] The specific technical solution adopted by this invention is as follows: A structured instruction constraint and task guidance method based on dynamic feature weights is proposed. This invention guides the model to adaptively adjust the task execution focus according to changes in key information at different stages in task scenarios with phased evolution characteristics. The invention quantifies the significance of changes in different behavioral indicators at each stage, calculates feature weights reflecting stage sensitivity, and dynamically adjusts the structured instruction content input to the model based on these feature weights. This allows the structured instructions to emphasize different behavioral indicators at different stages, thereby achieving stage-aware guidance of the model's task execution process through changes in structured instructions. This invention does not rely on manual experience to set weights or change the model structure. Instead, it introduces a mapping relationship between stage-specific feature weights and structured instructions to achieve refined guidance for multi-stage tasks, improving the model's relevance, stability, and interpretability in staged task analysis. Specifically, it includes the following steps: S1: This invention employs multi-stage data acquisition and feature extraction to quantify the behavioral responses of a task object at different stages. First, it uses video acquisition equipment or other sensors to collect behavioral data of the task object at different stages. These stages can be time-based, state-based, or task progress-based. The acquired data undergoes necessary preprocessing operations, including denoising, standardization, and cropping, before being input into a feature extraction model to extract behavioral response features that characterize the behavioral state. These features are represented in statistical form and used to depict the response of different behavioral indicators at each stage, providing a foundation for subsequent stage-specific change analysis.
[0013] S2: Using standardized mean change, the behavioral indicators at different stages s The degree of change relative to the baseline stage 0 is quantified. For behavioral indicator data acquired at different stages, the mean value of each behavioral indicator in the corresponding stage is calculated, and a preset baseline stage is selected as a comparison reference. For the same behavioral indicator... i By calculating its stage s The mean difference between the current stage and the baseline stage 0 is normalized by combining it with the standard deviation of the behavioral indicator, thus obtaining a standardized mean change reflecting the degree of stage-specific change. The behavioral indicator... i In the stage sThe standardized mean change is defined as follows: ; in, Indicates the first i Individual behavioral indicators at the stage s The mean of the following, This represents the mean of the behavioral indicator at the baseline stage of 0. This indicates the standard deviation of the behavioral indicator across the entire period or within a preset reference interval. Indicators of behavior i In the stage s The change in the standardized mean relative to the baseline stage 0. Through the above calculation method, this invention achieves a unified quantitative description of the degree of change of each behavioral indicator at different stages, while eliminating the differences in the dimensions of different behavioral indicators, providing a basic input for subsequent stage-based weight calculations and structured instruction adjustments.
[0014] S3: Using behavioral indicators in stages s Based on the standardized mean change results, calculate the stage weight of each behavioral indicator in this stage. Based on the behavioral indicators obtained in step S2 in stage... s Standardized mean change This invention quantifies the relative importance of different behavioral indicators in a given stage. To ensure the comparability and stability of the weighting results across different stages, a normalization method is used to map the changes in each behavioral indicator to stage-specific weights, ensuring that the sum of the weights of all behavioral indicators in the same stage is a preset constant, thus forming a clear stage-specific weight distribution.
[0015] Specifically, behavioral indicators i In the stage s The corresponding stage weight is defined as follows: ; in, Indicates that behavioral indicator i is in stage s The corresponding stage weights, This represents the total number of behavioral indicators. Indicators of behavior i In the stage s The change in the standardized mean relative to the baseline period.
[0016] By employing the aforementioned weighting calculation method, this invention enables behavioral indicators that exhibit significant changes in the current stage to receive relatively higher weights, while those with smaller changes receive relatively lower weights, thereby reflecting the differences in the discriminative power of each behavioral indicator within that stage. These stage-specific weights serve as a crucial basis for subsequent adjustments to structured instructions, guiding the model to prioritize behavioral indicators with higher information content during task execution.
[0017] S4: Formalize the structured instructions using stage-specific weights to generate structured instruction representations for the corresponding stages. The stage-specific weights of the behavioral indicators obtained in step S3 are introduced into the construction process of the structured instructions, enabling the structured instructions to explicitly reflect the differences in importance of different behavioral indicators at the current stage. Specifically, this invention does not directly use weights for model parameter or feature calculations, but rather uses weights as constraints or guiding information to participate in the generation and organization of structured instruction content, thereby achieving task guidance based on stage-specific feature differences. In the stage... s The structured instructions generated below can be represented as: ; in, Indicates the stage s The structured instructions generated below, This represents the semantic description, analytical constraint, or task concern corresponding to the i-th behavioral indicator. This indicates that the behavioral indicator is in the stage. s The corresponding stage weights, This represents a structured instruction generation mapping function based on stage weights.
[0018] The technical effects achieved by this invention are as follows: This invention offers significant advantages in handling multi-stage tasks. Existing technologies typically employ fixed feature weights or pre-defined task descriptions, making it difficult to reflect key information changes arising from stage evolution during task execution. Models often focus on the same or similar features at different stages, resulting in insufficient sensitivity of analysis results to stage differences. This invention, by introducing quantitative results of stage-specific behavioral changes, directly reflects stage differences in the construction and adjustment of structured instructions. This allows the task constraints and focus received by the model at different stages to dynamically adjust with stage changes, thus better aligning with the actual task evolution process.
[0019] Furthermore, existing technologies for handling stage changes often rely on manual experience to set rules or design separate model structures for specific stages, resulting in poor versatility and scalability. This invention automatically determines the relative importance of behavioral indicators at different stages based on statistical changes in behavioral data. It does not rely on manual rules or specific model structures, and can achieve stage-aware task guidance without changing model parameters or network architecture, exhibiting good versatility and portability.
[0020] Meanwhile, even when task guidance mechanisms are introduced in existing technologies, they often remain at the level of static task descriptions or fixed templates, making it difficult to continuously reflect stage evolution information during task execution. This invention, by continuously mapping stage feature weights to structured instruction content, allows the structured instructions to be dynamically updated as the stage evolves, thereby forming a continuous task guidance process and avoiding the problem of sudden changes or lags in analysis focus during stage switching.
[0021] In summary, by combining phased feature changes with structured instruction adjustments, this invention achieves dynamic guidance that changes the focus of task execution with each phase while ensuring the stability of the model structure. This enhances the adaptability, relevance, and stability of multi-phase task analysis, making it particularly suitable for complex task scenarios with obvious phased evolution characteristics. Attached Figure Description
[0022] Figure 1 This is a flowchart of the operation of a structured instruction constraint and task guidance method based on dynamic feature weights according to the present invention. Detailed Implementation
[0023] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.
[0024] This invention relates to the fields of artificial intelligence, behavior analysis, and multimodal data processing, particularly to a technical method for multi-stage task analysis and feature weighting. This invention can dynamically adjust feature weights based on the evolutionary characteristics of data and optimize the model's performance in different behavior assessments through structured instruction constraints and task guidance. It has broad application prospects in multiple fields such as medical rehabilitation, industrial monitoring, and anomaly detection.
[0025] like Figure 1As shown, this invention presents a structured instruction constraint and task guidance method based on dynamic feature weights. In task scenarios with stage-evolutionary characteristics, this invention guides the model to adaptively adjust the task execution focus according to changes in key information at different stages. This invention quantifies the significance of changes in different behavioral indicators at each stage, calculates feature weights reflecting stage sensitivity, and dynamically adjusts the structured instruction content input to the model based on these feature weights. This allows the structured instructions to emphasize different behavioral indicators at different stages, thereby achieving stage-aware guidance of the model's task execution process through changes in structured instructions. This invention does not rely on manual experience to set weights, nor does it change the model structure. Instead, it introduces a mapping relationship between stage-specific feature weights and structured instructions to achieve refined guidance for multi-stage tasks, improving the model's relevance, stability, and interpretability in stage-specific task analysis. Specifically, it includes the following steps: S1: This invention employs multi-stage data acquisition and feature extraction to quantify the behavioral responses of a task object at different stages. First, it uses video acquisition equipment or other sensors to collect behavioral data of the task object at different stages. These stages can be time-based, state-based, or task progress-based. The acquired data undergoes necessary preprocessing operations, including denoising, standardization, and cropping, before being input into a feature extraction model to extract behavioral response features that characterize the behavioral state. These features are represented in statistical form and used to depict the response of different behavioral indicators at each stage, providing a foundation for subsequent stage-specific change analysis.
[0026] S2: Using standardized mean change, the behavioral indicators at different stages s The degree of change relative to the baseline stage 0 is quantified. For behavioral indicator data acquired at different stages, the mean value of each behavioral indicator in the corresponding stage is calculated, and a preset baseline stage is selected as a comparison reference. For the same behavioral indicator... i By calculating its stage s The mean difference between the current stage and the baseline stage 0 is normalized by combining it with the standard deviation of the behavioral indicator, thus obtaining a standardized mean change reflecting the degree of stage-specific change. The behavioral indicator... i In the stage s The standardized mean change is defined as follows: ; in, Indicates the first i Individual behavioral indicators at the stage s The mean of the following, This represents the mean of the behavioral indicator at the baseline stage of 0. This indicates the standard deviation of the behavioral indicator across the entire period or within a preset reference interval. Indicators of behaviori In the stage s The change in the standardized mean relative to the baseline stage 0. Through the above calculation method, this invention achieves a unified quantitative description of the degree of change of each behavioral indicator at different stages, while eliminating the differences in the dimensions of different behavioral indicators, providing a basic input for subsequent stage-based weight calculations and structured instruction adjustments.
[0027] S3: Using behavioral indicators in stages s Based on the standardized mean change results, calculate the stage weight of each behavioral indicator in this stage. Based on the behavioral indicators obtained in step S2 in stage... s Standardized mean change This invention quantifies the relative importance of different behavioral indicators in a given stage. To ensure the comparability and stability of the weighting results across different stages, a normalization method is used to map the changes in each behavioral indicator to stage-specific weights, ensuring that the sum of the weights of all behavioral indicators in the same stage is a preset constant, thus forming a clear stage-specific weight distribution.
[0028] Specifically, behavioral indicators i In the stage s The corresponding stage weight is defined as follows: ; in, Indicates that behavioral indicator i is in stage s The corresponding stage weights, This represents the total number of behavioral indicators. Indicators of behavior i In the stage s The change in the standardized mean relative to the baseline period.
[0029] By employing the aforementioned weighting calculation method, this invention enables behavioral indicators that exhibit significant changes in the current stage to receive relatively higher weights, while those with smaller changes receive relatively lower weights, thereby reflecting the differences in the discriminative power of each behavioral indicator within that stage. These stage-specific weights serve as a crucial basis for subsequent adjustments to structured instructions, guiding the model to prioritize behavioral indicators with higher information content during task execution.
[0030] S4: Formalize the structured instructions using stage-specific weights to generate structured instruction representations for the corresponding stages. The stage-specific weights of the behavioral indicators obtained in step S3 are introduced into the construction process of the structured instructions, enabling the structured instructions to explicitly reflect the differences in importance of different behavioral indicators at the current stage. Specifically, this invention does not directly use weights for model parameter or feature calculations, but rather uses weights as constraints or guiding information to participate in the generation and organization of structured instruction content, thereby achieving task guidance based on stage-specific feature differences. In the stage... s The structured instructions generated below can be represented as: ; in, Indicates the stage s The structured instructions generated below, This represents the semantic description, analytical constraint, or task concern corresponding to the i-th behavioral indicator. This indicates that the behavioral indicator is in the stage. s The corresponding stage weights, This represents a structured instruction generation mapping function based on stage weights.
[0031] like Figure 1 As shown, in this invention, the behavioral responses of the task object at different stages are considered as a process of gradual evolution with each stage, and the information contribution of different behavioral indicators varies significantly at each stage. Therefore, this invention quantifies the changes of different behavioral indicators relative to a baseline stage at different stages, characterizing the significance of changes in each behavioral indicator during the stage evolution process, and thus reflecting its relative importance at the corresponding stage. This process is based on the statistical characteristics of behavioral data and does not rely on manual experience settings, thereby avoiding the problem of insufficient adaptability of fixed weights or subjective rules in multi-stage tasks.
[0032] Furthermore, this invention does not directly apply the aforementioned stage-specific importance to adjusting model parameters or network structure, but rather incorporates it into the construction and adjustment process of structured instructions. Specifically, the changing importance of different behavioral indicators at the current stage is reflected in changes in the description focus, constraints, or order of attention of the behavioral indicators in the structured instructions, causing the structured instructions to exhibit different emphases at different stages. As the task progresses, the results of stage-specific changes are continuously updated, and the content of the structured instructions is adjusted accordingly, enabling the model to continuously focus on the most discriminative behavioral information at the current stage during task execution without altering its own structure.
[0033] Through the above methods, this invention organically combines stage-based feature changes, feature weight calculation, and structured instruction adjustment to form a stage-aware task guidance process. This process enables the model's task execution focus to dynamically evolve with stage changes, ensuring consistency in task analysis while highlighting key information at different stages. The scope of this invention includes not only the specific implementation of the above method but also its equivalent application under different task types, data modalities, and model architectures.
[0034] In summary, this invention, by introducing a structured instruction adjustment mechanism driven by phased behavioral changes, achieves dynamic guidance of the model task execution process, effectively improving the adaptability and relevance of multi-stage task analysis.
[0035] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.
Claims
1. A structured instruction constraint and task guidance method based on dynamic feature weights, characterized in that: Includes the following steps: Step S1: Use multi-stage data acquisition and feature extraction to quantify the behavioral responses of the task object at different stages; Step S2: Use standardized mean change to analyze behavioral indicators at different stages. s The degree of change relative to the baseline stage 0 is quantitatively calculated. Step S3: Use behavioral indicators in the phase s Based on the standardized mean change results, calculate the corresponding stage weight of each behavioral indicator in this stage; Step S4: Use stage weights to perform formal mapping on the structured instructions to generate the structured instruction representation for the corresponding stage.
2. The structured instruction constraint and task guidance method based on dynamic feature weights according to claim 1, characterized in that: In step S1, video acquisition equipment or other sensors are first used to collect behavioral data of the task object at different stages; the different stages are time stages, state stages, or task process stages; the collected data undergoes preprocessing operations, including noise reduction, standardization, and cropping, and is then input into a feature extraction model to extract behavioral response features that can characterize the behavioral state; the features are represented in statistical form and are used to characterize the response of different behavioral indicators at each stage.
3. The structured instruction constraint and task guidance method based on dynamic feature weights according to claim 1, characterized in that: In step S2, for the behavioral indicator data obtained at different stages, the mean value of each behavioral indicator at the corresponding stage is calculated, and a preset benchmark stage is selected as a comparison reference; for the same behavioral indicator... i By calculating its stage s The mean difference between the current stage and the baseline stage 0 is normalized by combining it with the standard deviation of the behavioral indicator, thereby obtaining a standardized mean change reflecting the degree of stage-specific change; the behavioral indicator i In the stage s The standardized mean change is defined as follows: ； in, Indicates the first i Individual behavioral indicators at the stage s The mean of the following, This represents the mean of the behavioral indicator at the baseline stage of 0. This indicates the standard deviation of the behavioral indicator across the entire period or within a preset reference interval. Indicators of behavior i In the stage s The change in the standardized mean relative to the baseline stage 0.
4. The structured instruction constraint and task guidance method based on dynamic feature weights according to claim 1, characterized in that: In step S3, based on the behavioral indicators obtained in step S2, the stage... s Standardized mean change The relative importance of different behavioral indicators in this stage is quantitatively modeled, and the change of each behavioral indicator is mapped to a stage weight by normalization, so that the sum of the weights of all behavioral indicators in the same stage is a preset constant, thereby forming a clear stage weight distribution. Specifically, behavioral indicators i In the stage s The corresponding stage weight is defined as follows: ； in, Indicates that behavioral indicator i is in stage s The corresponding stage weights, This represents the total number of behavioral indicators. Indicators of behavior i In the stage s The change in the standardized mean relative to the baseline period.
5. The structured instruction constraint and task guidance method based on dynamic feature weights according to claim 1, characterized in that: In step S4, the stage-specific weights of the behavioral indicators obtained in step S3 are introduced into the construction process of the structured instructions, enabling the structured instructions to explicitly reflect the differences in importance of different behavioral indicators at the current stage. Specifically, the weights are not directly used for model parameter or feature calculation, but rather as constraints or guiding information, participating in the generation and organization of the structured instruction content, thereby achieving task guidance based on stage-specific differences. s The generated structured instructions are represented as follows: ； in, Indicates the stage s The structured instructions generated below, This represents the semantic description, analytical constraints, or task focus corresponding to the i-th behavioral indicator. This indicates that the behavioral indicator is in the stage. s The corresponding stage weights, This represents a structured instruction generation mapping function based on stage weights.