Personalized load evaluation feedback method and system fusing user and task characteristics

Abstract

The invention provides a personalized load assessment feedback method and system fusing user and task features, belongs to the field of human engineering, and on the basis of generating a personalized user file according to user basic information and special requirements, supplements user body condition limitations through large language model intelligent reasoning, and improves user experience. The tasks are split according to steps and evaluated from three dimensions of body postures, action force and action time or frequency, risk scores of the tasks are calculated by adopting an RULA method, and dynamic weight adjustment is performed according to user archives; and screening improvement suggestions adapted to user features and task scenes through a retrieval enhancement generation method, and generating a personalized risk assessment feedback report by integrating the user archives, the final risk scores and the improvement suggestions. According to the method, individual differences of users and task feature diversity are fully considered, so that ergonomics risk assessment is more accurate, and improvement suggestions are more targeted and practical.

Description

Technical Field

[0001] The present invention relates to the field of ergonomics, and in particular to a personalized load assessment feedback method and system integrating user and task characteristics. Background Art

[0002] Ergonomic risk assessment is crucial in occupational health and safety management. It aims to identify risks that may lead to musculoskeletal disorders by analyzing factors such as work posture, load, and repetitive motion. Traditional assessment methods rely primarily on expert observation, such as the Rapid Upper Limb Assessment (RULA), Rapid Whole Body Assessment (REBA), and the NOISH lifting equation. While these methods provide standardized scores, they suffer from significant issues: First, the results are susceptible to evaluator subjectivity and experience differences, resulting in significant discrepancies between scores from different experts for the same task; second, the assessment is inefficient, preventing rapid coverage of a large number of work scenarios; and third, they ignore individual differences.

[0003] Recent advances in computer vision and machine learning have driven the automation of ergonomic risk assessment. Methods based on pose estimation tools such as OpenPose and MediaPipe can automatically extract joint angles and calculate scores, significantly improving assessment efficiency. However, existing technologies still suffer from core flaws such as occlusion and accuracy issues, and insufficient generalization capabilities.

[0004] For example, the invention patent with publication number CN119580934A discloses a method for assessing human work posture based on deep learning. This method uses an improved YOLOv8-pose model to detect the coordinates of key points in work images and calculates the dual reliability indicators of confidence and shooting angle to screen valid inputs. It also calculates joint angles using the cosine theorem and corrects model errors using preset empirical coefficients. It calculates ergonomic scores based on work posture assessment methods and identifies persistent poor postures through time series analysis. It ultimately outputs a risk score and generates specific optimization plans and pain relief measures for work states with scores greater than a preset value. This solution combines posture assessment with intervention recommendations, but is unable to generate personalized improvement plans based on specific user characteristics and work scenarios.

[0005] For example, the invention patent with publication number CN116704415A discloses an ergonomic risk assessment method based on posture and scene information. This method first extracts uniformly sized image sequences through a video preprocessing module and constructs a spatiotemporal human skeleton graph using the LCR-Net pose estimation network. It then employs a two-branch feature extraction architecture: the scene branch uses a VGG16 network to extract environmental spatial features from the image sequence, while the pose branch processes the spatiotemporal skeleton graph using a modified CTR-GCNs graph convolutional network to obtain dynamic pose features. Next, a multi-level fusion mechanism concatenates scene and pose features along the channel dimension at the feature level. At the decision level, the action classification results of the two features are weighted averaged. Finally, the fused features are input into an LSTM network to generate a risk assessment score. This method combines environmental context with human kinematic features to improve the accuracy of action assessment, but the resulting scores lack interpretability and cannot provide specific improvement recommendations like human experts. Summary of the Invention

[0006] The purpose of the present invention is to provide a personalized load assessment feedback method and system that integrates user and task characteristics, so as to solve the problem that the existing ergonomic risk assessment ignores the individual differences of users and the diversity of task characteristics, resulting in inaccurate human load assessment results and lack of targeted improvement suggestions in the feedback.

[0007] The embodiment provides a personalized load assessment feedback method that integrates user and task characteristics, including the following steps: Step 1: Obtain basic user information and special needs, input the user's basic information and special needs into the large language model, infer and supplement the user's physical condition restrictions, and generate a user profile; Step 2: After pre-processing the user's input operation video, it is input into the visual language model together with the force load information in text form. The user task is then divided into several operation steps according to the preset prompt word template; Step 3: Evaluate the body posture, movement force, and movement time or frequency of each operation step one by one, and calculate the risk score of each operation step based on the RULA expert observation method assisted by the visual language model; Step 4: Based on the user profile in step 1, set differentiated weights for different operation steps, calculate the weighted average risk score of all operation steps, and obtain the final risk score of the user task; Step 5: Based on the user profile and the user task scenario, the search enhancement generation method is used to query the improvement suggestions that are suitable for the user profile and the user task scenario; Step 6: Based on the user profile, final risk score, and improvement suggestions, a personalized risk assessment feedback report is generated for the user through a large language model.

[0008] In one embodiment, the basic information of the user includes the user's age, gender and height; The special needs include illness or injury that affects the user's ability to operate.

[0009] Furthermore, the disease or injury includes one or more permanent or temporary physical impairments such as muscle strain, arthritis, disability, and fracture.

[0010] Furthermore, the reasoning and supplementing of the user's physical condition limitations include: based on the user's basic information and special needs, using a large language model combined with ergonomic theory, physiological common sense and common data of peers or in the same situation, inferring and supplementing the user's operating ability limitations.

[0011] Furthermore, the large language model is GPT-4, Claude or other models with similar functions.

[0012] In one embodiment, the user profile includes a detailed description of hand grip flexibility, hand movement coordination, upper limb joint range of motion, lumbar flexion ability and whole body movement ability, and the detailed description is quantitatively evaluated according to four levels: excellent, good, average and poor.

[0013] In one embodiment, the preprocessing of the user input operation video includes: splitting the user input operation video into sequence frames according to a preset sampling rate, and using a posture estimation model to detect body and hand key points of each frame image to generate a sequence frame with a human skeleton superimposed.

[0014] Furthermore, the visual language model includes GPT-4o or Qwen2.5-VL.

[0015] In one embodiment, the body posture includes joint angles of the upper limbs, lower limbs, and torso; The action force includes the specific location where the force is applied, the quantitative measurement of the force, the direction of the force, and the duration; The action time or frequency includes the duration of a single action, the repetition frequency and the cumulative operation time.

[0016] In one embodiment, the RULA expert observation method assisted by a visual language model calculates the risk score of each operation step, including: The visual language model scores six body parts, including upper arm, forearm, wrist, neck, trunk, and leg, based on the RULA evaluation criteria; The risk score of each operation step is calculated based on the action strength level and action repetition frequency level.

[0017] Furthermore, the six body parts were scored on a scale of 1-4, including: the upper arm was scored 1-4 based on the lifting angle; The forearm is scored 1-2 points based on the bending angle; The wrist is divided into 1-3 points according to the deflection angle; The neck is scored 1-4 points based on the bending angle; The trunk is scored 1-4 based on the bending angle; Legs are scored 1-2 points based on leg posture.

[0018] Furthermore, the force levels are mild <2 kg, moderate 2-10 kg, and severe >10 kg; The levels of movement repetition frequency are static posture, high-frequency repetitive movement, and short and low-frequency movement.

[0019] Furthermore, the static posture refers to a posture maintained for more than 1 minute; The high-frequency repetitive action refers to an action repetition frequency greater than 4 times / minute.

[0020] Furthermore, the risk score of each operation step is divided into a scoring range of 1-7 points.

[0021] In one embodiment, setting differentiated weights for different operation steps includes: For users with limited hand grip dexterity, increase the weight of steps involving fine hand manipulation and grasping movements by 20%-50%; For users with poor coordination of both hands, the weight of operation steps that require both hands will be increased by 30%-60%; For users with limited full-body movement capabilities, the weight of steps involving large body movements and posture changes will be increased by 40%-70%.

[0022] In one embodiment, the use of the search enhancement generation method to screen improvement suggestions that are adapted to the user profile and the user task scenario includes: Collect data on ergonomic improvement measures and suggestions to establish a knowledge base. Then, use a large language model to extract ergonomic improvement measures and suggestions. After preprocessing, a triple dataset of task-goal-measure is formed. Based on the triple dataset, a vector database is established after vectorization processing by the text embedding model, and category labels are set; A retrieval enhancement generation method is adopted. Based on the vector database and category labels, combined with the user profile and user task scenarios, the large language model is guided to generate structured retrieval requests, and improvement suggestions that are suitable for user profiles and user task scenarios are screened.

[0023] Furthermore, the category labels include: posture and action, equipment and tools, working environment and atmosphere, work flow and workload, and training and education.

[0024] Furthermore, the search request includes three dimensions: task description, actionable measure category and specific query.

[0025] In one embodiment, the screening of improvement suggestions adapted to user profiles and user task scenarios includes: using a retrieval enhancement generation method to perform a preliminary screening of the vector database based on category labels, and then calculating the cosine similarity between the preliminarily screened text and the document, returning the top n improvement suggestions with the highest similarity, and completing the screening of improvement suggestions.

[0026] Furthermore, the user task scenarios are divided according to the type of work environment, including: industrial production scenarios, office work scenarios, home life scenarios and construction scenarios.

[0027] Furthermore, the conditions for the preliminary screening include: suggestions that match the physical condition restrictions in the user profile; and recommendations that match the working environment and equipment conditions of the user's task scenarios.

[0028] In order to clearly demonstrate the personalized load assessment feedback method integrating user and task characteristics, the present invention further provides a personalized load assessment feedback method system integrating user and task characteristics, including: The user profile generation module is used to obtain basic user information and special needs, input the user's basic information and special needs into the large language model, infer and supplement the user's physical condition restrictions, and generate a user profile; The task feature assessment module is used to pre-process the user's input operation video and input it together with the force load information in text form into the visual language model. The module then divides the user task into several operation steps based on a preset prompt word template, evaluates the body posture, movement force, and movement time or frequency of each operation step, calculates the risk score of each operation step based on the RULA expert observation method assisted by the visual language model, and sets differentiated weights for different operation steps to perform a weighted average calculation to obtain the final risk score of the user task; Improvement suggestion screening module is used to query improvement suggestions that are suitable for user profiles and user task scenarios based on the user profiles and user task scenarios using a search enhancement generation method; The report generation module is used to generate user-personalized risk assessment feedback reports based on user profiles, final risk scores, and improvement suggestions through a large language model.

[0029] Compared with the prior art, the present invention has the following beneficial effects: (1) Based on the user's basic information and special needs, a personalized user profile is generated. The user's physical condition restrictions are intelligently supplemented through a large language model, and the individual differences of users are fully considered. The task is split into steps and evaluated from three dimensions: body posture, movement force, and movement time or frequency. The RULA method is used to calculate the risk score, and the dynamic weight is adjusted according to the user profile, fully considering the diversity of task characteristics and the differences in user characteristics.

[0030] (2) Based on the diversity of task characteristics and the differences in user characteristics, the retrieval enhancement generation method is used to screen improvement suggestions that adapt to user characteristics and task scenarios to ensure the pertinence and operability of the suggestions.

[0031] (3) By integrating user profiles, final risk scores, and improvement suggestions, a personalized risk assessment feedback report is generated through a large language model, fully considering individual differences among users and the diversity of task characteristics, making ergonomic risk assessment more accurate and improvement suggestions more targeted and practical. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for describing the embodiments or the prior art.

[0033] Figure 1 This is a flow chart of the personalized load assessment feedback method that integrates user and task characteristics provided by the present invention.

[0034] Figure 2 This is a structural diagram of the personalized load assessment feedback system that integrates user and task characteristics provided by the present invention. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solutions and advantages of the present invention more clear, the following is a The present invention is further described in detail in the following examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0036] The solution of the embodiment of the present invention is as follows Figure 1 As shown, a personalized load assessment feedback method integrating user and task characteristics includes the following steps: S1. Obtain the user's basic information and special needs, input the user's basic information and special needs into the large language model, infer and supplement the user's physical condition restrictions, and generate a user profile.

[0037] In an embodiment, basic information such as the user's age, gender, height, and special needs that affect operational capabilities, such as permanent or temporary physical disabilities such as arthritis, muscle strain, disability, and fractures, are obtained through the user input interface. This information is used as basic data for generating a personalized user profile.

[0038] Next, the user's basic information and specific needs are input into the large language model. The large language model combines ergonomic theory, physiological knowledge, and common data from peers or similar situations to intelligently infer the user's operational limitations. The large language model can be GPT-4, Claude, or other models with similar capabilities; there are no restrictions here. The generated user profile includes detailed descriptions of hand grip dexterity, bimanual coordination, upper limb joint range of motion, lumbar flexion ability, and whole-body movement ability. Each ability is quantitatively assessed according to four levels: excellent, good, average, and poor, providing a basis for subsequent personalized weighting adjustments.

[0039] S2. After pre-processing the operation video input by the user, it is input into the visual language model together with the force load information in text form, and the user task is divided into several operation steps according to the preset prompt word template.

[0040] In this embodiment, users upload a video of their operation and force load information in text form. The video is first preprocessed, extracting a sequence of frames at a sampling rate of 5 frames per second. The OpenPose 25 model is then used to detect body and hand keypoints in each frame, generating a sequence of frames with a human skeleton overlay.

[0041] The preprocessed sequence frames and force load text are fed into a visual language model. Using pre-set prompt word templates, the model is guided to break down complex operation processes into several steps. The visual language model can be GPT-4o, Qwen2.5-VL, and is not limited here.

[0042] S3. Evaluate the body posture, movement force, and movement time or frequency of each operation step one by one, and calculate the risk score of each operation step based on the RULA expert observation method assisted by the visual language model.

[0043] Each step is analyzed in detail from three perspectives: posture, force, and duration / frequency. The posture dimension includes the major joint angles of the upper and lower limbs and trunk. The force dimension includes the specific location of force application, quantitative measurement of force magnitude, direction, and duration. The time / frequency dimension includes the duration of a single movement, repetition frequency, and cumulative work time.

[0044] The visual language model scores the six body parts of the upper arm, forearm, wrist, neck, trunk and leg according to the RULA evaluation criteria, and divides them into a scoring range of 1-4 points, combined with the strength level and repetition frequency level. Specifically, the upper arm is scored from 1 to 4 points based on the lifting angle; the forearm is scored from 1 to 2 points based on the bending angle; the wrist is scored from 1 to 3 points based on the deflection angle; the neck is scored from 1 to 4 points based on the bending angle; the trunk is scored from 1 to 4 points based on the bending angle; and the leg is scored from 1 to 2 points based on the leg posture. There are three strength levels: mild <2kg, moderate 2-10kg, and severe >10kg. There are also three repetition frequency levels: static posture, high-frequency repetitive movements, and short and low-frequency movements. The static posture refers to a posture maintained for more than 1 minute; the high-frequency repetitive movements refer to movements with a repetition frequency of >4 times / minute.

[0045] Based on the above-mentioned classification, the risk score of each operation step is calculated, and each operation step is classified into a risk level of 1-7 points.

[0046] S4. According to the user profile in S1, differentiated weights are set for different operation steps, and the risk scores of all operation steps are calculated by weighted average to obtain the final risk score of the user task.

[0047] In the embodiment, differentiated weights are set for different operation steps based on the ability assessment results in the user profile. Specifically, for users with limited hand grip flexibility, the weight of steps involving fine hand manipulation and grasping movements will be increased by 20%-50%; for users with poor coordination of two-handed movements, the weight of operation steps requiring two-handed coordination will be increased by 30%-60%; for users with limited whole-body movement ability, the weight of steps involving large-scale body movement and posture changes will be increased by 40%-70%. The weight adjustment range is determined based on the specific assessment results of the ability level in the user profile. The worse the ability level, the greater the increase in the weight of the relevant steps.

[0048] Next, the risk scores of all steps are weighted averaged to obtain the final risk score of the user's task as a whole, thus ensuring that the assessment results can truly reflect the user's individual risk level.

[0049] S5. Based on the user profile and the user task scenario, a search enhancement generation method is used to query improvement suggestions that are suitable for the user profile and the user task scenario.

[0050] In this example, a knowledge base of ergonomic improvement measures was pre-established. This knowledge base collected publicly available professional guidelines and research literature. After manual screening and organization, improvement measures and suggestions were extracted using a large language model. Each improvement measure and suggestion was processed according to a unified format, recording its applicable task type, improvement goal, and specific measures, forming a "task-goal-measure" triple dataset.

[0051] These triplet datasets are vectorized using a text embedding model (such as text-embedding-ada-002, which is not restricted here) and stored in a MongoDB vector database. Category labels are set for easy retrieval. The category labels include: posture and action, equipment and tools, work environment and atmosphere, workflow and workload, and training and education.

[0052] The LangChain framework implements the Retrieval Augmentation Generation (RAG) retrieval function. Based on the user profile and task context, a large language model is guided to generate structured retrieval requests, which include task descriptions, actionable action categories, and specific queries. During the retrieval process, the vector database is initially screened based on category labels. The cosine similarity between the pre-screened text and the document is then calculated, returning the top n most similar improvement suggestions, ensuring a close match between the user's characteristics and task context.

[0053] In this embodiment, n is 5; user task scenarios are divided according to the type of work environment, including: industrial production scenarios, office work scenarios, home life scenarios and construction scenarios; the preliminary screening conditions include: suggestions that match the physical condition restrictions in the user profile; and suggestions that match the work environment and equipment conditions of the user task scenario.

[0054] S6. Based on the user profile, final risk score and improvement suggestions, a user-personalized risk assessment feedback report is generated through a large language model.

[0055] In the embodiment, the user profile, final risk score and screened improvement suggestions are input into the large language model, and a prompt word template is generated through the report to generate a personalized risk assessment feedback report with standardized format and complete content, which is used to provide users with professional and feasible assessment conclusions and improvement guidance.

[0056] In order to clearly demonstrate the personalized load assessment feedback method integrating user and task characteristics, the present invention also provides a personalized load assessment feedback system integrating user and task characteristics, such as Figure 2 Shown, including: The user profile generation module is used to obtain basic user information and special needs, input the user's basic information and special needs into the large language model, infer and supplement the user's physical condition restrictions, and generate a user profile; The task feature assessment module is used to pre-process the user's input operation video and input it together with the force load information in text form into the visual language model. The module then divides the user task into several operation steps based on a preset prompt word template, evaluates the body posture, movement force, and movement time or frequency of each operation step, calculates the risk score of each operation step based on the RULA expert observation method assisted by the visual language model, and sets differentiated weights for different operation steps to perform a weighted average calculation to obtain the final risk score of the user task; Improvement suggestion screening module is used to query improvement suggestions that are suitable for user profiles and user task scenarios based on the user profiles and user task scenarios using a search enhancement generation method; The report generation module is used to generate user-personalized risk assessment feedback reports based on user profiles, final risk scores, and improvement suggestions through a large language model.

[0057] In each module, a corresponding prompt word template is set to improve the response quality of the large language model. Specifically: The prompt word template of the user profile generation module includes: based on the obtained user basic information and special needs, reasoning about the user's operating ability according to the user's common sense and reference data.

[0058] Operational Ability dimensions include: hand grip dexterity, bimanual coordination, upper limb range of motion, lumbar flexion, and overall mobility. Each ability is quantitatively assessed on a scale of excellent, good, fair, and poor. User profiles are output in JSON format. For example: {"Name":"Zhang Hua","Age":"60","Gender":"Female","Operational Ability":{"Hand Grip Dexterity":"Fair","Bimanual Coordination":"Fair","Upper Limb Range of Motion":"Poor","Lumbar Flexion":"Poor","Overall Mobility":"Fair"}}.

[0059] The prompt word template for the task feature assessment module includes: The user obtains a video frame sequence of a task at a sampling rate of x fps, the user portrait is {}, and the user's force load information is {}; please conduct an operational load and ergonomic risk assessment based on this information: First, observe the entire task process, split the task into several operational steps, and number them; The RULA method is used to evaluate each step. The evaluation steps are: 1. Select key postures to be evaluated, including those with long duration, high frequency, or far from the neutral position; 2. Obtain angular data for the upper arm, forearm, wrist, neck, torso, and leg from the image; 3. Use a scoring table to define a score for each body part; 4. Modify the score based on the input force load and movement frequency information.

[0060] Finally, the risk score of each step is output in JSON format.

[0061] Before the user outputs the final answer, please re-evaluate the user's previous answer, consider whether the posture and angle recognition in the image are accurate, and whether the score matches; if incorrect, please correct it.

[0062] Improve the prompt word template of the suggestion filtering module, including: considering the actual scenario and user characteristics of the behavior, and providing solutions that can be used to reduce the operational load in the task.

[0063] First, we select appropriate categories that align with the context and equipment conditions of the user tasks in the video: posture and action; equipment and tools; work environment and atmosphere; workflow and workload; and training and education. Next, we generate queries for searching the knowledge base, which should contain the following information: task, category, and query. Output is in JSON format.

[0064] For example: {"task":"sweeping","category":"postures and actions, equipment and tools","query":"What methods can be used to reduce the operational load in the task?"}.

[0065] The report generation module prompt word template includes: step-by-step analysis and generation of a comprehensive report based on the following structure: 1. The main risk factors of the behavior in the video; 2. Improvement suggestions to reduce the operational load and ergonomic risks of this behavior.

[0066] Output requirements: 1. Analyze the operations step by step; 2. Output in Markdown format, no more than 300 words; 3. Do not output queries and references.

[0067] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. Step 1: Obtain basic user information and special needs, input these into the large language model, infer and supplement the user's physical condition restrictions, and generate a user profile; Step 2: After pre-processing the user's input operation video, it is input into the visual language model together with the force load information in text form. The user task is then divided into several operation steps according to the preset prompt word template; Step 3: Evaluate the body posture, movement force, and movement time or frequency of each operation step one by one, and calculate the risk score of each operation step based on the RULA expert observation method assisted by the visual language model; Step 4: Based on the user profile in step 1, set differentiated weights for different operation steps, calculate the weighted average risk score of all operation steps, and obtain the final risk score of the user task; Step 5: Based on the user profile and the user task scenario, the search enhancement generation method is used to query the improvement suggestions that are suitable for the user profile and the user task scenario; Step 6: Based on the user profile, final risk score, and improvement suggestions, a personalized risk assessment feedback report is generated for the user through a large language model.

2. The personalized load assessment feedback method according to claim 1, characterized in that: In step 1, the user's basic information includes the user's age, gender and height; The special needs include illness or injury that affects the user's ability to operate.

3. The personalized load assessment feedback method according to claim 1, characterized in that: The user profile includes a detailed description of hand grip flexibility, hand movement coordination, upper limb joint range of motion, lumbar flexion ability and whole body movement ability, and the detailed description is quantitatively evaluated according to four levels: excellent, good, average and poor.

4. The personalized load assessment feedback method according to claim 1, characterized in that: The preprocessing of the user input operation video includes: splitting the user input operation video into sequence frames according to a preset sampling rate, and using a posture estimation model to detect body and hand key points of each frame image to generate a sequence frame with a human skeleton superimposed.

5. The personalized load assessment feedback method according to claim 1, characterized in that: The body posture includes the joint angles of the upper limbs, lower limbs and trunk; The action force includes the specific location where the force is applied, the quantitative measurement of the force, the direction of the force, and the duration; The action time or frequency includes the duration of a single action, the repetition frequency and the cumulative operation time.

6. The personalized load assessment feedback method according to claim 5, characterized in that: The RULA expert observation method assisted by the visual language model calculates the risk score of each operation step, including: The visual language model scores six body parts, including upper arm, forearm, wrist, neck, trunk, and leg, based on the RULA evaluation criteria; The risk score of each operation step is calculated based on the action strength level and action repetition frequency level.

7. The personalized load assessment feedback method according to claim 3, characterized in that: The setting of differentiated weights for different operation steps includes: For users with limited hand grip dexterity, increase the weight of steps involving fine hand manipulation and grasping movements by 20%-50%; For users with poor coordination of both hands, the weight of operation steps that require both hands will be increased by 30%-60%; For users with limited full-body movement capabilities, the weight of steps involving large body movements and posture changes will be increased by 40%-70%.

8. The personalized load assessment feedback method according to claim 1, characterized in that: The aforementioned method of using retrieval enhancement to generate improved suggestions adapted to user profiles and user task scenarios includes: Collect data on ergonomic improvement measures and suggestions to establish a knowledge base. Then, use a large language model to extract ergonomic improvement measures and suggestions. After preprocessing, a triple dataset of task-goal-measure is formed. Based on the triple dataset, a vector database is established after vectorization processing by the text embedding model, and category labels are set; A retrieval enhancement generation method is adopted. Based on the vector database and category labels, combined with the user profile and user task scenarios, the large language model is guided to generate structured retrieval requests, and improvement suggestions that are suitable for user profiles and user task scenarios are screened.

9. The personalized load assessment feedback method according to claim 8, characterized in that: The screening of improvement suggestions adapted to user profiles and user task scenarios includes: using a retrieval enhancement generation method to perform a preliminary screening of the vector database based on category labels, and then calculating the cosine similarity between the preliminarily screened text and the document, returning the top n improvement suggestions with the highest similarity, and completing the screening of improvement suggestions.

10. A personalized load assessment feedback system integrating user and task characteristics, characterized in that: The personalized load assessment feedback system uses the personalized load assessment feedback method integrating user and task characteristics as described in any one of claims 1 to 9, including: The user profile generation module is used to obtain basic user information and special needs, input the user's basic information and special needs into the large language model, infer and supplement the user's physical condition restrictions, and generate a user profile; The task feature assessment module is used to pre-process the user's input operation video and input it together with the force load information in text form into the visual language model. The module then divides the user task into several operation steps based on a preset prompt word template, evaluates the body posture, movement force, and movement time or frequency of each operation step, calculates the risk score of each operation step based on the RULA expert observation method assisted by the visual language model, and sets differentiated weights for different operation steps to perform a weighted average calculation to obtain the final risk score of the user task; Improvement suggestion screening module is used to query improvement suggestions that are suitable for user profiles and user task scenarios based on the user profiles and user task scenarios using a search enhancement generation method; The report generation module is used to generate user-personalized risk assessment feedback reports based on user profiles, final risk scores, and improvement suggestions through a large language model.

Images