AI-based product ergonomics design method

Abstract

This invention discloses an AI-based product ergonomic design method, belonging to the field of product ergonomics technology. The method includes: collecting and preprocessing real-time multimodal data streams from users; using a multi-source data fusion network with an attention mechanism to perform weighted fusion and feature extraction on the preprocessed multimodal data streams to generate an assessment result of the user's current ergonomic state; when the assessment result indicates a risk, calling the user's personalized skeletal muscle digital twin model and combining it with the multimodal data stream to calculate the real-time stress distribution of the user's internal tissues, and analyzing the EEG signals in the multimodal data stream to determine the level of neural adaptability, generating intelligent collaborative intervention decision instructions. This invention, through the deep integration of biomechanical simulation and neurocognitive assessment, achieves a leap from macroscopic posture control to microscopic stress optimization, and from unidirectional intervention to adaptive learning, improving the accuracy, humanization, and continuous evolution capabilities of ergonomic design.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of product ergonomics, in particular to an AI-based product ergonomics design method. BACKGROUND

[0002] In the field of intelligent office and healthy human-computer interaction, an ergonomics automatic adjustment system based on sensors and artificial intelligence is a research hotspot in recent years. The existing technology usually monitors the user's posture through cameras, pressure pads and other sensors, and when it identifies preset bad postures such as bending over, it automatically drives actuators to intervene in a single dimension. This method realizes a basic closed loop from perception to action, and has a positive effect on relieving muscle fatigue.

[0003] However, the limitations of the existing technology are that its decision logic is relatively shallow and static, mainly relying on external posture as a macro indicator for judgment, and it cannot quantitatively assess the actual biomechanical load caused by bad posture on the internal tissues of the user's lumbar intervertebral disc, which makes the intervention may not be accurate to the deep damage risk, the intervention action is often immediate and one-way, without considering the user's current cognitive concentration, which may interrupt the work at inappropriate times and affect the experience. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an AI-based product ergonomics design method to solve the problem of lack of depth and precision in intervention decision-making and mismatch with the user's cognitive state in the prior art.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides an AI-based product ergonomics design method, which includes collecting real-time multi-modal data streams of a user and preprocessing them; using a multi-source data fusion network with an attention mechanism to perform weighted fusion and feature extraction on the preprocessed multi-modal data streams, and generating a user's current ergonomics state assessment result;

[0008] When the user's current ergonomics state assessment result indicates a risk, a user's individualized skeletal muscle digital twin model is called and combined with the multi-modal data streams to calculate the real-time stress distribution of the user's internal tissues, and the electroencephalogram signals in the multi-modal data streams are analyzed to determine the level of neural adaptability, and an intelligent collaborative intervention decision instruction is generated;

[0009] According to the multi-item device adjustment target parameter set in the intelligent collaborative intervention decision instruction, a preliminary device action sequence list is obtained;

[0010] Based on the preliminary list of equipment action sequences, a path planning algorithm is used to plan a smooth and reversible multi-device cooperative motion trajectory instruction from the current state to the target state for all equipment to be adjusted.

[0011] The controller executes multi-device coordinated motion trajectory commands, and after completion, collects new multimodal data streams and subjective feedback to generate quantitative evaluation results of the intervention effect;

[0012] The results of quantitative evaluation of intervention effects will be used to update the decision-making strategy of the multi-source data fusion network of the attention mechanism and optimize the generation of future intelligent collaborative intervention decision instructions.

[0013] As a preferred embodiment of the AI-based product ergonomic design method of the present invention, the method includes: collecting real-time multimodal data streams from users and performing preprocessing, comprising the following steps:

[0014] The system synchronously collects real-time multimodal physiological and behavioral data streams from users through a depth vision sensor, a pressure distribution sensor matrix, an inertial measurement unit, and a portable headband device.

[0015] The system extracts the three-dimensional coordinates of skeletal joints from the user's real-time multimodal physiological and behavioral data stream to generate the user's posture kinematic data stream. It also performs pressure center statistics and regional pressure distribution mapping on the user's real-time multimodal physiological and behavioral data stream to generate the user's sitting posture pressure distribution data stream.

[0016] The user's real-time multimodal physiological and behavioral data stream is subjected to attitude angle calculation and filtering denoising to generate a user desktop status data stream;

[0017] Calculate the power spectral density of EEG signals in a specific frequency band from the user's real-time multimodal physiological and behavioral data stream to generate the user's neurophysiological signal data stream;

[0018] The user posture kinematics data stream, user sitting pressure distribution data stream, user desktop status data stream, and user neurophysiological signal data stream are time-stamped and standardized in data format to generate a multimodal data stream that has been preprocessed and spatiotemporally synchronized.

[0019] As a preferred embodiment of the AI-based product ergonomic design method of the present invention, the method includes the following steps: A multi-source data fusion network with an attention mechanism is used to perform weighted fusion and feature extraction on the preprocessed multimodal data stream to generate an evaluation result of the user's current ergonomic status.

[0020] The pre-processed and spatiotemporally synchronized multimodal data stream is input into the attention mechanism multi-source data fusion network. The attention mechanism multi-source data fusion network calculates dynamic weights for different data sources and different time steps in the pre-processed and spatiotemporally synchronized multimodal data stream.

[0021] The preprocessed and spatiotemporally synchronized multimodal data streams are weighted and fused according to dynamic weights to generate weighted multimodal features. The attention-based multi-source data fusion network extracts high-level temporal features from the weighted multimodal feature representation to generate a fused deep feature vector.

[0022] The attention mechanism multi-source data fusion network maps the fused deep feature vectors to a comprehensive real-time comfort score and labels of uncomfortable body parts, generating an assessment result of the user's current ergonomic status.

[0023] As a preferred embodiment of the AI-based product ergonomic design method described in this invention, when the user's current ergonomic status assessment indicates a risk, the user's personalized musculoskeletal digital twin model is invoked and combined with multimodal data streams to calculate the real-time stress distribution of the user's internal tissues, and the EEG signals in the multimodal data streams are analyzed to determine the level of neural adaptability, generating intelligent collaborative intervention decision instructions, including the following steps:

[0024] The system continuously monitors the user's current ergonomic status assessment results. Based on the analysis of historical ergonomic data, a comfort threshold is set. When the comprehensive real-time comfort score in the user's current ergonomic status assessment results is lower than the comfort threshold, and the label of the uncomfortable body part points to the target area, the system determines that the user's current ergonomic status assessment results indicate an ergonomic risk to be analyzed.

[0025] Personalized geometric models are built based on users' medical imaging data, and combined with a general musculoskeletal biomechanical parameter library. Through scaling and parameter assignment, a personalized musculoskeletal digital twin model is constructed.

[0026] Ergonomic risks are analyzed based on the assessment results of the user's current ergonomic status, and the user's posture kinematic data stream is used as the kinematic boundary conditions required to drive the user's personalized musculoskeletal digital twin model.

[0027] By receiving user posture kinematic data stream through a user-personalized skeletal muscle digital twin model, solving the mechanical balance equation of skeletal muscles, obtaining stress values ​​of the area of ​​interest, and obtaining quantitative stress values ​​of the target internal tissues;

[0028] The user's neurophysiological signal data stream is extracted from the preprocessed and spatiotemporally synchronized multimodal data stream. The user's neurophysiological signal data stream is subjected to fast Fourier transform to calculate the power spectral density. The average power values ​​of the prefrontal cortex EEG signals in the θ band and β band are extracted. The neurophysiological fitness index is calculated according to the formula.

[0029] Based on the statistical distribution of the neural adaptability index observed in a controlled experimental environment when subjects perform familiar tasks and can receive external cues without interference, a timing judgment threshold is set. Based on experimental research data on the stress critical value at which fatigue damage occurs in the target internal tissue, a biomechanical safety threshold is set.

[0030] The quantitative stress value of the target's internal tissue is compared with the biomechanical safety threshold, and the neural adaptability index is compared with the timing judgment threshold. When the quantitative stress value of the target's internal tissue exceeds the biomechanical safety threshold and the neural adaptability index is higher than the timing judgment threshold, the logical judgment condition is met.

[0031] When the logical judgment conditions are met, a set of equipment control parameters is found based on the biomechanical optimization algorithm to reduce the quantitative stress value of the target internal tissue, and the equipment control parameters are encapsulated into an executable intelligent collaborative intervention decision instruction.

[0032] As a preferred embodiment of the AI-based product ergonomic design method of the present invention, the method includes the following steps: obtaining a preliminary list of equipment action sequences based on a set of multiple equipment adjustment target parameters in the intelligent collaborative intervention decision-making instruction:

[0033] The intelligent collaborative intervention decision command is parsed, the set of equipment adjustment target parameters is extracted, the current state parameters of the actuator are queried, the current state parameter set corresponding to the set of equipment adjustment target parameters is obtained, and the difference between the multiple sets of equipment adjustment target parameters and the current state parameter set is calculated element by element using arithmetic subtraction to generate a parameter difference set.

[0034] Based on the set of parameter differences, generate basic action instructions for each parameter;

[0035] The basic action instructions are sorted in logical order and combined into a preliminary list of equipment action sequences.

[0036] As a preferred embodiment of the AI-based product ergonomic design method of the present invention, the method includes the following steps: Based on a preliminary list of equipment action sequences, a smooth and reversible multi-device cooperative motion trajectory instruction from the current state to the target state is planned for all equipment to be adjusted using a path planning algorithm.

[0037] The initial list of equipment action sequences is mapped to a path planning start point and a path planning end point in a high-dimensional parameter space. Between the path planning start point and the path planning end point, a fast random exploration tree algorithm is used to randomly sample and generate multiple possible paths connecting the start point and the end point.

[0038] From multiple possible paths, select the optimal path that satisfies the requirements of mechanical constraints and motion smoothness, and discretize the optimal path into a series of intermediate path points arranged in time order;

[0039] The intermediate path point sequence is encoded into a multi-device collaborative motion trajectory instruction with timestamps.

[0040] As a preferred embodiment of the AI-based product ergonomic design method described in this invention, the method involves: controlling the actuator to execute multi-device collaborative motion trajectory commands, collecting new multimodal data streams and subjective feedback upon completion, and generating quantitative evaluation results of the intervention effect, including the following steps:

[0041] The multi-device collaborative motion trajectory command is sent to the actuator controller, and the actuator controller drives the actuator to move according to the trajectory and timestamp specified in the multi-device collaborative motion trajectory command;

[0042] The system collects new multimodal data streams from users through a sensor array, collects subjective feedback scores of user comfort after intervention through a human-computer interaction interface, and calculates quantitative evaluation results of intervention effect by combining the new multimodal data streams and subjective feedback scores using an evaluation function.

[0043] As a preferred embodiment of the AI-based product ergonomic design method described in this invention, the following steps are included: Quantitatively evaluating the intervention effect, updating the decision-making strategy of the multi-source data fusion network based on the attention mechanism, and optimizing the generation of future intelligent collaborative intervention decision-making instructions:

[0044] The quantitative evaluation results of the intervention effect are transformed into reward values ​​under the reinforcement learning framework, and the current user's current ergonomic status evaluation results, intelligent collaborative intervention decision instructions and reward values ​​are used to form an experience sample.

[0045] Experience samples are stored in the experience replay buffer. A batch of experience samples are sampled from the experience replay buffer. The gradient of the policy network in the multi-source data fusion network of the attention mechanism is calculated using the near-end policy optimization method based on the policy gradient theorem.

[0046] The gradient descent method is used to update the parameters of the policy network in the multi-source data fusion network of the attention mechanism, thereby optimizing the generation strategy of future intelligent collaborative intervention decision instructions.

[0047] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the AI-based product ergonomic design method as described in the first aspect of the present invention.

[0048] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the AI-based product ergonomic design method as described in the first aspect of the present invention.

[0049] The beneficial effects of this invention are as follows: Multimodal physiological and behavioral data of the user are collected and preprocessed through multi-source sensors; a multi-source data fusion network based on attention mechanisms is used to generate an assessment result of the user's current ergonomic state; when the assessment result indicates a risk, the real-time stress distribution of internal tissues is calculated by combining the user's personalized musculoskeletal digital twin model, and EEG signals are analyzed to determine the level of neural adaptability. Based on this, intelligent collaborative intervention decision instructions are generated, which are then converted into device action sequences. A smooth and reversible multi-device collaborative motion trajectory is generated through a path planning algorithm; after the actuator completes the trajectory, feedback data is collected to generate a quantitative assessment result of the intervention effect; finally, the decision strategy of the fusion network is updated using a proximal strategy optimization method to achieve closed-loop optimization. Through deep integration of biomechanical simulation and neurocognitive assessment, a leap from macroscopic posture control to microscopic stress optimization, and from unidirectional intervention to adaptive learning is achieved, improving the accuracy, humanization, and continuous evolution capability of ergonomic design. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Fig. 1 A flowchart for an AI-based product ergonomic design methodology.

[0052] Fig. 2 Generate a schematic diagram based on the user's current ergonomic status assessment results.

[0053] Fig. 3 Flowchart for generating intelligent collaborative intervention decision-making instructions.

[0054] Fig. 4 Update the flowchart for decision-making strategies. Detailed Implementation

[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0056] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0057] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0058] Reference Figs. 1-4 This is one embodiment of the present invention, which provides an AI-based product ergonomic design method, including the following steps:

[0059] S1. Collect real-time multimodal data streams from users and perform preprocessing.

[0060] S1.1 Synchronously collect real-time multimodal physiological and behavioral data streams from users through a depth vision sensor, a pressure distribution sensor matrix, an inertial measurement unit, and a portable headband device.

[0061] Furthermore, a depth vision sensor captures a sequence of depth images of the user's work scene at a specific frame rate; a pressure distribution sensor matrix continuously records the pressure distribution numerical matrix on the seat surface at the same time reference; an inertial measurement unit is fixed to the edge of the desk and measures the three-axis acceleration and three-axis angular velocity of the desktop at the same frequency; a portable headband device synchronously acquires the raw voltage signals of multi-channel EEG from the user's scalp; all the above devices are time-synchronized through a unified hardware clock or network time protocol to ensure that each data point has an accurate and consistent timestamp, together forming a real-time multimodal physiological and behavioral data stream of the user.

[0062] S1.2 Extract the three-dimensional coordinates of skeletal joints from the user's real-time multimodal physiological and behavioral data stream to generate the user's posture kinematic data stream. Perform pressure center statistics and regional pressure distribution mapping on the user's real-time multimodal physiological and behavioral data stream to generate the user's sitting posture pressure distribution data stream.

[0063] Furthermore, from the depth image sequence captured by the depth vision sensor, a pose estimation algorithm based on a convolutional neural network is used to identify and output the coordinates of the main joints of the human body (such as the shoulder, elbow, hip, and knee) in three-dimensional space, forming a user posture kinematic data stream. At the same time, the coordinates of the center point of the weighted average of the pressure values ​​of all sensing units are used as the center of pressure for the numerical matrix recorded by the pressure distribution sensor matrix. The matrix is ​​divided into several predefined regions (such as the left hip region, right hip region, and thigh region), and the average pressure value of each region is calculated to generate a user sitting posture pressure distribution data stream.

[0064] S1.3 Perform attitude angle calculation and filtering denoising on the user's real-time multimodal physiological and behavioral data stream to generate the user's desktop status data stream.

[0065] Furthermore, the raw data of triaxial acceleration and triaxial angular velocity measured by the inertial measurement unit are first denoised using a Kalman filter, and then the pitch angle, roll angle and yaw angle of the table surface are calculated by quaternion or direction cosine matrix method, which constitutes the user desktop status data stream.

[0066] S1.4 Calculate the power spectral density of EEG signals in a specific frequency band from the user's real-time multimodal physiological and behavioral data stream to generate the user's neurophysiological signal data stream.

[0067] Furthermore, the multi-channel raw EEG voltage signals acquired by the portable headband device are first bandpass filtered to remove low-frequency drift and high-frequency noise. Then, a fast Fourier transform is performed on the signal segments of each time window to calculate their power spectral density. The average power values ​​of the electrode signals in the prefrontal region in the θ and β bands are extracted to generate the user's neurophysiological signal data stream.

[0068] S1.5. The user posture kinematics data stream, user sitting pressure distribution data stream, user desktop status data stream and user neurophysiological signal data stream are time-stamped and standardized in data format to generate a multimodal data stream after preprocessing and spatiotemporal synchronization.

[0069] Furthermore, based on the timestamps carried by each data stream, all data are resampled to a unified fixed time interval to ensure that all data at the same point in time can correspond. The coordinate values ​​in the user posture kinematics data stream, the pressure values ​​in the user sitting pressure distribution data stream, the angle values ​​in the user desktop status data stream, and the power values ​​in the user neurophysiological signal data stream are respectively subjected to min-max normalization processing to map all values ​​to a unified numerical range. Finally, the data streams are packaged to generate a preprocessed and spatiotemporally synchronized multimodal data stream with a unified format and time synchronization.

[0070] S2. A multi-source data fusion network with an attention mechanism is used to perform weighted fusion and feature extraction on the preprocessed multimodal data stream to generate an evaluation result of the user's current ergonomic status.

[0071] S2.1 Input the pre-processed and spatiotemporally synchronized multimodal data stream into the attention mechanism multi-source data fusion network. The attention mechanism multi-source data fusion network calculates dynamic weights for different data sources and different time steps in the pre-processed and spatiotemporally synchronized multimodal data stream.

[0072] Furthermore, the multi-source data fusion network of the attention mechanism first maps the user posture kinematics data stream, user sitting pressure distribution data stream, user desktop state data stream, and user neurophysiological signal data stream from the preprocessed and spatiotemporally synchronized multimodal data stream into feature vectors of a unified dimension through independent embedding layers. A multi-head attention layer is applied to these feature vector sequences, and its query vector, key vector, and value vector are all obtained by linear transformation of the feature vector of the current time step. For each time step, the attention weight distribution representing the importance of different time steps is obtained by calculating the dot product of the query vector and the key vectors of all time steps and applying the Softmax function for normalization. Another parallel attention mechanism calculates the similarity between feature vectors from different data sources to generate weights representing the importance of different data sources. The time step weights and data source weights are multiplied element-wise to obtain the final dynamic weight of each data point at each time step.

[0073] S2.2. Based on the dynamic weights, the preprocessed and spatiotemporally synchronized multimodal data streams are weighted and fused to generate weighted multimodal features. The attention mechanism multi-source data fusion network extracts high-level temporal features from the weighted multimodal feature representation to generate the fused deep feature vector.

[0074] Furthermore, the dynamic weights are weighted and summed with the corresponding feature vectors in the preprocessed and spatiotemporally synchronized multimodal data stream to obtain a weighted multimodal feature representation. The weighted multimodal feature representation is then fed into an encoder composed of multiple layers of gated recurrent units. The gated recurrent units process the weighted multimodal feature representation in chronological order, capture long-term dependencies in the data through their internal gating mechanism, and extract high-level temporal features. The hidden state of the last time step of the gated recurrent unit encoder is output as the fused deep feature vector.

[0075] S2.3 The multi-source data fusion network with attention mechanism maps the fused deep feature vector to a comprehensive real-time comfort score and labels of uncomfortable body parts, generating an assessment result of the user's current ergonomic status.

[0076] Furthermore, the fused deep feature vector is fed into a fully connected layer network. This network outputs two branches. The first branch uses a Sigmoid activation function to map the fused deep feature vector to a value between 0 and 1, which serves as a comprehensive real-time comfort score (e.g., 0.85 indicates high comfort). The second branch uses a Softmax activation function to map the fused deep feature vector to a probability distribution vector. Each dimension of this vector corresponds to a predefined body part (e.g., neck, lower back, shoulder). The part with the highest probability value is labeled as the uncomfortable body part. The comprehensive real-time comfort score and the uncomfortable body part label together constitute the user's current ergonomic status assessment result.

[0077] It should be noted that the training process of the multi-source data fusion network of the attention mechanism is carried out on a dataset containing a large number of preprocessed and spatiotemporally synchronized multimodal data stream samples and corresponding manually labeled real user ergonomic state evaluation results. During training, the loss function between the network's predicted user current ergonomic state evaluation result and the real label is minimized. The gradient of the loss function with respect to the network parameters is calculated using the backpropagation algorithm, and the Adam optimizer is used to iteratively update the network parameters until the prediction accuracy on the validation set reaches the predetermined standard.

[0078] S3. When the user's current ergonomic status assessment indicates a risk, the system calls the user's personalized musculoskeletal digital twin model and combines it with multimodal data streams to calculate the real-time stress distribution of the user's internal tissues, and analyzes the EEG signals in the multimodal data streams to determine the level of neural adaptability, generating intelligent collaborative intervention decision instructions.

[0079] S3.1 Continuously monitor the user's current ergonomic status assessment results. Based on the analysis of historical ergonomic data, set a comfort threshold. When the comprehensive real-time comfort score in the user's current ergonomic status assessment results is lower than the comfort threshold, and the uncomfortable body part label points to the target area, determine that the user's current ergonomic status assessment results indicate an ergonomic risk to be analyzed.

[0080] Furthermore, the system continuously monitors the user's current ergonomic status assessment results, including the comprehensive real-time comfort score and the label of the uncomfortable body part. The comfort threshold is based on statistical analysis of the comprehensive real-time comfort scores of a large number of samples in the historical ergonomic database. For example, the lower quartile of the score distribution corresponding to mild discomfort symptoms in all samples is taken as the comfort threshold. When the comprehensive real-time comfort score is detected to be lower than the comfort threshold, and the label of the uncomfortable body part clearly points to a predefined target area, it is determined that the user's current ergonomic status assessment results indicate that there is an ergonomic risk that requires further analysis.

[0081] S3.2. Based on the user's medical imaging data, a personalized geometric model is established, and combined with a general musculoskeletal biomechanical parameter library, a personalized musculoskeletal digital twin model is constructed by scaling and parameter assignment.

[0082] Furthermore, using the user's lumbar CT or MRI scan data, a personalized geometric model of the lumbar spine and intervertebral discs is generated through image segmentation and 3D reconstruction technology; standard parameters such as the origin and insertion points, cross-sectional area, and maximum muscle strength of lumbar spine-related muscles are obtained from a general musculoskeletal biomechanical parameter library; based on the user's individual information such as height and weight, the standard muscle parameters are scaled according to size; the scaled muscle parameters are assigned and attached to the personalized geometric model, completing the construction of the user's personalized musculoskeletal digital twin model.

[0083] S3.3. Analyze ergonomic risks based on the assessment results of the user's current ergonomic status, and use the user's posture kinematic data stream as the kinematic boundary conditions required to drive the user's personalized skeletal muscle digital twin model.

[0084] Furthermore, when the user's current ergonomic status assessment results indicate an ergonomic risk, the user's posture kinematic data stream at the current moment is extracted from the preprocessed and spatiotemporally synchronized multimodal data stream. The joint kinematic parameters such as spinal curvature angle and pelvic tilt angle contained in the data stream are directly set as the kinematic boundary conditions of the corresponding segments in the user's personalized musculoskeletal digital twin model.

[0085] S3.4 Receive user posture kinematics data stream through the user's personalized skeletal muscle digital twin model, solve the mechanical balance equation of the skeletal muscle, obtain the stress value of the area of ​​interest, and obtain the quantitative stress value of the target internal tissue.

[0086] Furthermore, the user-personalized skeletal muscle digital twin model, based on the joint angles defined by the input user posture kinematic data stream, solves a muscle force optimization problem (e.g., minimizing the sum of squares of muscle activation) under gravitational load to satisfy the mechanical equilibrium condition, thus obtaining the muscle forces required to maintain the posture. Then, based on the principle of mechanical equilibrium, it calculates the resultant force acting on a specific area of ​​interest (e.g., the L4 / L5 intervertebral disc). Dividing this resultant force by the cross-sectional area calculated by the personalized geometric model of the intervertebral disc yields the pressure on the intervertebral disc, i.e., the quantitative stress value of the target internal tissue.

[0087] It should be noted that the user-personalized musculoskeletal digital twin model takes the joint angles in the user's posture kinematics data stream as input, calculates muscle force by solving a static optimization problem with the goal of minimizing the sum of squares of muscle activation, and then synthesizes the muscle force with the external force according to the principle of mechanical equilibrium to obtain the joint reaction force. The force is divided by the personalized bearing area of ​​the area of ​​interest to obtain the quantitative stress value of the target internal tissue.

[0088] S3.5 Extract the user's neurophysiological signal data stream from the preprocessed and spatiotemporally synchronized multimodal data stream, perform a fast Fourier transform on the user's neurophysiological signal data stream to calculate the power spectral density, and extract the average power values ​​of the prefrontal cortex EEG signals in the θ and β bands. Calculate the neurophysiological fitness index according to the formula.

[0089] Furthermore, by extracting user neurophysiological signal data streams from multimodal data streams and performing frequency domain analysis, this step achieves an objective and quantitative assessment of the user's cognitive load state. Compared to relying solely on external indicators such as posture or behavior, directly analyzing EEG signals can capture changes in the user's internal state, such as neural fatigue or distraction, earlier and more accurately. The core benefit of calculating the neuroadaptability index is that it provides a crucial basis for timing judgment in intervention decisions. This allows for the selection of the optimal time window for ergonomic adjustments when the user's cognitive load is low and neural efficiency is high, thereby avoiding the interference and negative experience caused by abrupt intervention when the user is highly focused. This greatly enhances the intelligence and acceptability of human-machine collaboration.

[0090] It should be noted that, from the preprocessed and spatiotemporally synchronized multimodal data stream, the user's neurophysiological signal data stream corresponding to that acquired by the portable headband device is separated, and the EEG signal channel of the prefrontal cortex region is located from it. A Fast Fourier Transform is applied to the EEG signal segment of the selected channel to convert the time-domain signal into a frequency-domain representation, thereby calculating the power spectral density of the signal. Then, the power values ​​of all frequency points in the θ band (e.g., 4-7 Hz) and β band (e.g., 13-30 Hz) are arithmetically averaged to obtain the average power values ​​of the θ band and β band, respectively. Finally, the average power values ​​of the two bands are divided according to the formula: the neural fitness index, to calculate the final neural fitness index. An increase in the index value generally indicates improved neural efficiency and reduced cognitive load.

[0091] S3.6. Based on the statistical distribution of the neural adaptability index observed in a controlled experimental environment when subjects perform familiar tasks and can accept external cues without interference, a timing judgment threshold is set. Based on experimental research data on the stress critical value of fatigue damage to the target internal tissue, a biomechanical safety threshold is set.

[0092] Furthermore, the timing judgment threshold is based on the 25th percentile of the dataset obtained by recording the neural adaptability index of a large number of subjects in a controlled laboratory environment when they were performing a highly proficient low cognitive load task and clearly indicated that they could accept the cues. The biomechanical safety threshold is directly set based on experimental research data on the stress threshold of long-term load on the annulus fibrosus of the lumbar intervertebral disc leading to fatigue injury in biomechanical literature.

[0093] S3.7. Compare the quantitative stress value of the target's internal tissue with the biomechanical safety threshold, and compare the neural adaptability index with the timing judgment threshold. When the quantitative stress value of the target's internal tissue exceeds the biomechanical safety threshold and the neural adaptability index is higher than the timing judgment threshold, the logical judgment condition is met.

[0094] Furthermore, by comparing the quantitative stress value of the target's internal tissue with the biomechanical safety threshold and the neuroadaptability index with the timing judgment threshold, intervention is triggered only when both conditions are met simultaneously. This achieves a high degree of unity between precision and humanization in intervention decision-making. It ensures that ergonomic intervention is not only based on whether it is needed, but also on when it is best. The dual judgment mechanism effectively avoids two common drawbacks: first, focusing only on mechanical risks while ignoring user experience, leading to abrupt adjustments when users are highly focused, resulting in resentment; second, considering only the timing while ignoring health risks, missing the real necessity of intervention. Thus, while ensuring user health, it improves the intelligence and acceptability of intervention, making the design control truly a considerate and efficient collaborative partner.

[0095] S3.8 When the logical judgment condition is met, a set of equipment control parameters is found based on the biomechanical optimization algorithm to reduce the quantitative stress value of the target internal tissue, and the equipment control parameters are encapsulated into an executable intelligent collaborative intervention decision instruction.

[0096] Furthermore, when the logical judgment condition is met, an optimizer is started. The design variables of the optimizer are controllable equipment parameters, and the objective function is the quantified stress value of the target internal tissue under a given equipment parameter configuration, predicted using a user-personalized skeletal muscle digital twin model. A sequential quadratic programming algorithm is used to find a set of optimal equipment control parameters that minimizes the objective function value, and the combination is encapsulated into an intelligent collaborative intervention decision instruction containing specific action instructions.

[0097] S4. Based on the set of target parameters for multiple devices in the intelligent collaborative intervention decision instruction, obtain a preliminary list of device action sequences.

[0098] S4.1 Parse the intelligent collaborative intervention decision command, extract the set of equipment adjustment target parameters, query the current state parameters of the actuator, obtain the current state parameter set corresponding to the set of equipment adjustment target parameters, use arithmetic subtraction to calculate the difference between multiple sets of equipment adjustment target parameters and the current state parameter set element by element, and generate a parameter difference set.

[0099] Furthermore, the intelligent collaborative intervention decision-making instructions are parsed to extract a clear set of target parameters for device adjustment, such as a target seat back angle of 105 degrees and a target seat height of 48 centimeters. Then, the current state parameters of the actuators corresponding to these target parameters are queried, yielding a current seat back angle of 90 degrees and a current seat height of 46 centimeters. Next, arithmetic subtraction is used to calculate the difference between the target parameter set and the corresponding parameters in the current state parameter set element by element: the seat back angle difference is 105 degrees minus 90 degrees equals +15 degrees, and the seat height difference is 48 centimeters minus 46 centimeters equals +2 centimeters. The calculated differences constitute the parameter difference set.

[0100] S4.2. Based on the parameter difference set, generate basic action instructions for each parameter.

[0101] Furthermore, based on the value and sign of each parameter in the parameter difference set, a corresponding basic action command is generated. For example, if the difference in seat back angle is positive 15 degrees, a basic action command to increase the backrest angle by 15 degrees is generated; if the difference in seat height is positive 2 centimeters, a basic action command to raise the seat height by 2 centimeters is generated.

[0102] S4.3. Sort the basic action instructions in logical order and combine them into a preliminary list of equipment action sequences.

[0103] Furthermore, all the generated basic action commands are sorted according to a preset logical order. This logical order is determined based on the priority and safety principles of equipment adjustment. Priority is given to adjusting the support height before adjusting the angle to avoid interference. After sorting, the basic action command to raise the seat height by 2 cm is placed first, and the basic action command to increase the backrest angle by 15 degrees is placed last, forming an orderly preliminary list of equipment action sequences.

[0104] S5. Based on the preliminary list of equipment action sequences, use a path planning algorithm to plan a smooth and reversible multi-device cooperative motion trajectory instruction from the current state to the target state for all equipment to be adjusted.

[0105] S5.1 Map the initial list of equipment action sequences to a path planning start point and a path planning end point in a high-dimensional parameter space. Between the path planning start point and the path planning end point, use the fast random exploration tree algorithm to perform random sampling and generate multiple possible paths connecting the start point and the end point.

[0106] Furthermore, the target parameters of the equipment included in the initial equipment action sequence list (e.g., target seat height of 48 cm and target backrest angle of 105 degrees) and the corresponding current state parameters (e.g., current seat height of 46 cm and current backrest angle of 90 degrees) together define a high-dimensional parameter space; the set of current state parameters (46 cm, 90 degrees) is mapped to the starting point of path planning in this space, and the set of equipment adjustment target parameters (48 cm, 105 degrees) is mapped to the ending point of path planning;

[0107] In the high-dimensional space between the path planning start point and the path planning end point, a fast random exploration tree algorithm is used for random sampling expansion: starting from the path planning start point, a sampling point is randomly generated in the space, and then the node closest to the sampling point is found from the existing tree structure. A new node is then expanded in the direction of the sampling point, and it is checked whether the path from the parent node to the new node satisfies the basic kinematic constraints. This process is repeated until the distance between the newly expanded node and the path planning end point is less than a preset tolerance, thus generating a possible path from the start point to the end point. Through a large number of repeated samplings, multiple possible paths connecting the path planning start point and the path planning end point are finally generated.

[0108] S5.2 Select an optimal path from multiple possible paths that meets the requirements of mechanical constraints and motion smoothness, and discretize the optimal path into a series of intermediate path points arranged in time order.

[0109] Furthermore, from the multiple possible paths generated by the fast random exploration tree algorithm, each path is scored according to a preset evaluation function. The evaluation function comprehensively considers path length, path smoothness, and whether it strictly meets the mechanical motion constraints of all devices (e.g., avoiding collisions between moving parts of each device). The path with the highest evaluation function score is selected as the optimal path that meets the requirements of mechanical constraints and motion smoothness. Continuous optimal paths are discretized and sampled at fixed time intervals to generate a series of intermediate path point sequences arranged in chronological order. Each point in the sequence represents a set of cooperative state parameters that all devices should achieve at a specific time.

[0110] S5.3 Encode the intermediate path point sequence into a multi-device collaborative motion trajectory instruction with a timestamp.

[0111] Furthermore, each point in the intermediate path point sequence, along with its corresponding timestamp, is encoded into a specific control instruction; all instructions arranged in chronological order together constitute a multi-device collaborative motion trajectory instruction, which clearly specifies the target position or angle that each actuator should reach at each specific point in time, until the final state.

[0112] S6. Control the actuator to execute multi-device collaborative motion trajectory commands. After completion, collect new multimodal data streams and subjective feedback to generate quantitative evaluation results of the intervention effect.

[0113] S6.1. Send the multi-device collaborative motion trajectory command to the actuator controller. The actuator controller drives the actuator to move according to the trajectory and timestamp specified in the multi-device collaborative motion trajectory command.

[0114] Furthermore, the multi-device collaborative motion trajectory command is sent to the corresponding actuator controller through the communication interface. The actuator controller parses the target parameters corresponding to each set of timestamps in the command and drives the motor to move smoothly to the designated position in a closed-loop control manner, ensuring that the movements of multiple actuators are precisely synchronized in time and realizing accurate reproduction of the collaborative trajectory.

[0115] S6.2. Collect new multimodal data streams from users through a sensor array, collect subjective feedback scores of user comfort after intervention through a human-computer interaction interface, and use an evaluation function to combine the new multimodal data streams and subjective feedback scores to calculate quantitative evaluation results of the intervention effect.

[0116] The quantitative evaluation result of the intervention effect is expressed as follows:

[0117] ;

[0118] in, To quantitatively evaluate the intervention effect, For normalized subjective feedback scores, The objective discomfort index after intervention. The objective discomfort index before intervention. For subjective rating weighting, Weighted by the objective improvement rate;

[0119] Furthermore, a structured evaluation function is used to quantitatively and weightedly integrate the relative improvement rates of subjective feedback and objective physiological indicators, thereby achieving a multi-dimensional and comprehensive quantitative evaluation of the intervention effect. Existing technologies typically rely on only a single-dimensional indicator, such as judging solely based on the user's subjective rating or comparing only the absolute change in a certain physiological parameter before and after the intervention. The evaluation function introduced in this invention… It not only incorporates both subjective feelings and objective data, but more importantly, it measures the effectiveness of the intervention at the physiological level through the objective improvement rate and balances the contributions of the two types of evidence through weighting coefficients, so that the evaluation results can more comprehensively and accurately reflect the true overall effect of the intervention.

[0120] S7. Quantitatively evaluate the intervention effect, update the decision-making strategy of the multi-source data fusion network of the attention mechanism, and optimize the generation of future intelligent collaborative intervention decision instructions.

[0121] S7.1 Transform the quantitative evaluation results of the intervention effect into reward values ​​under the reinforcement learning framework, and construct an experience sample by combining the current user's current ergonomic status evaluation results, intelligent collaborative intervention decision instructions and reward values.

[0122] Furthermore, the quantitative evaluation results of the intervention effect are directly used as reward values ​​under the reinforcement learning framework. For example, the quantitative evaluation results of the intervention effect combine the user's current ergonomic state evaluation result corresponding to the generation of intelligent collaborative intervention decision instructions, the generated intelligent collaborative intervention decision instructions themselves, and the reward value into an experience sample.

[0123] S7.2 Store the experience samples in the experience replay buffer, sample a batch of experience samples from the experience replay buffer, and use the near-end policy optimization method of the policy gradient theorem to calculate the gradient of the policy network in the multi-source data fusion network of the attention mechanism.

[0124] The gradient expression for the policy network is:

[0125] ;

[0126] in, For time steps gradient, For gradient operators, For time steps Expected value For time steps The probability ratio, The advantage function at time step The estimated value, For the clipping function, For the cropping range, For time steps.

[0127] Furthermore, the gradient calculation method based on Proximal Policy Optimization (PPO) differs fundamentally from traditional policy gradient methods (such as the REINFORCE algorithm) in that it introduces a pruning function to constrain the probability ratio, thereby forcing the difference between the new and old policies to remain within a preset stable range during policy updates. Existing techniques often rely on the simple application of unbiased estimation of policy gradients during policy updates, which can easily lead to severe oscillations or even policy performance collapse due to excessively large single update steps. The PPO method employed in this invention, by pruning the probability ratio and combining it with a minimum value operation, essentially constructs an alternative objective function with pessimistic estimation characteristics. This function encourages policy performance improvement while strictly constraining the update magnitude of each iteration, thus ensuring the stability and convergence reliability of the entire reinforcement learning training process. This makes training large neural network policies in complex ergonomic decision-making scenarios feasible.

[0128] It should be noted that, from a physical perspective, the parameters of a policy network... The multi-source data fusion network that determines the attention mechanism maps multimodal perceptual information into behavioral strategies for intervention decisions, with a probability ratio The physical meaning is the degree of change in the new strategy's tendency to perform historical actions relative to the old strategy. A value greater than 1 indicates that the new strategy is more inclined to take that action, while a value less than 1 indicates a decreased tendency. The advantage function estimate is... The physical meaning of "intervention" is the additional expected benefit that an intervention action performed under a specific ergonomic state can bring compared to the average action; the scope of the adjustment. This sets the maximum allowable change in behavioral pattern during a single policy update, acting similarly to a damper in a mechanical system to prevent the policy from rapidly swinging from one extreme to another, ensuring a smooth evolution of decision-making behavior. (Gradient) Therefore, the direction represents the direction in which the strategy network parameters should be fine-tuned to improve the long-term intervention effect.

[0129] S7.3. Use gradient descent to update the parameters of the policy network in the multi-source data fusion network of the attention mechanism, and optimize the generation strategy of future intelligent collaborative intervention decision instructions.

[0130] Furthermore, using an optimizer, the gradient is applied to the parameters of the policy network in the multi-source data fusion network of the attention mechanism, according to the formula... To be updated, among which, For learning rate, This is the new value after this iteration. The parameters are the old values ​​before this iteration. After the parameter update, when the policy network encounters similar user current ergonomic status evaluation results, it will generate intelligent collaborative intervention decision instructions with a higher probability that can obtain higher quantitative evaluation results of intervention effects, thereby optimizing the generation strategy of future intelligent collaborative intervention decision instructions.

[0131] This embodiment also provides a computer device applicable to the AI-based product ergonomic design method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the AI-based product ergonomic design method proposed in the above embodiment.

[0132] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0133] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the AI-based product ergonomic design method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0134] In summary, this invention collects and preprocesses multimodal physiological and behavioral data from users through multi-source sensors, and uses a multi-source data fusion network based on attention mechanisms to generate an assessment result of the user's current ergonomic state. When the assessment result indicates a risk, it combines the user's personalized skeletal muscle digital twin model to calculate the real-time stress distribution of internal tissues and analyzes EEG signals to determine the level of neural adaptability. Based on this, it generates intelligent collaborative intervention decision instructions, transforms the instructions into device action sequences, and generates smooth and reversible multi-device collaborative motion trajectories through path planning algorithms. After the actuators complete the trajectory, it collects feedback data to generate quantitative assessment results of the intervention effect. Finally, it uses proximal strategy optimization methods to update the decision strategy of the fusion network, achieving closed-loop optimization. By deeply integrating biomechanical simulation and neurocognitive assessment, it achieves a leap from macroscopic posture control to microscopic stress optimization, and from unidirectional intervention to adaptive learning, improving the accuracy, humanization, and continuous evolution capabilities of ergonomic design.

[0135] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI-based product ergonomic design method, characterized by: This includes collecting real-time multimodal data streams from users and preprocessing them; using a multi-source data fusion network with an attention mechanism to perform weighted fusion and feature extraction on the preprocessed multimodal data streams to generate an assessment result of the user's current ergonomic status; When the user's current ergonomic status assessment indicates a risk, the system calls the user's personalized musculoskeletal digital twin model and combines it with multimodal data streams to calculate the real-time stress distribution of the user's internal tissues, and analyzes the EEG signals in the multimodal data streams to determine the level of neural adaptability, generating intelligent collaborative intervention decision instructions. Based on the set of target parameters for multiple devices in the intelligent collaborative intervention decision-making instructions, a preliminary list of device action sequences is obtained. Based on the preliminary list of equipment action sequences, a path planning algorithm is used to plan a smooth and reversible multi-device cooperative motion trajectory instruction from the current state to the target state for all equipment to be adjusted. The controller executes multi-device coordinated motion trajectory commands, and after completion, collects new multimodal data streams and subjective feedback to generate quantitative evaluation results of the intervention effect; The results of quantitative evaluation of intervention effects are used to update the decision-making strategy of the multi-source data fusion network of the attention mechanism and optimize the generation of future intelligent collaborative intervention decision instructions. When the user's current ergonomic assessment indicates a risk, the system invokes the user's personalized musculoskeletal digital twin model and combines it with multimodal data streams to calculate the real-time stress distribution of the user's internal tissues. It also analyzes the electroencephalogram (EEG) signals in the multimodal data streams to determine the level of neural adaptability, generating intelligent collaborative intervention decision instructions, including the following steps: The system continuously monitors the user's current ergonomic status assessment results. Based on the analysis of historical ergonomic data, a comfort threshold is set. When the comprehensive real-time comfort score in the user's current ergonomic status assessment results is lower than the comfort threshold, and the label of the uncomfortable body part points to the target area, the system determines that the user's current ergonomic status assessment results indicate an ergonomic risk to be analyzed. Personalized geometric models are built based on users' medical imaging data, and combined with a general musculoskeletal biomechanical parameter library. Through scaling and parameter assignment, a personalized musculoskeletal digital twin model is constructed. Ergonomic risks are analyzed based on the assessment results of the user's current ergonomic status, and the user's posture kinematic data stream is used as the kinematic boundary conditions required to drive the user's personalized musculoskeletal digital twin model. By receiving user posture kinematic data stream through a user-personalized skeletal muscle digital twin model, solving the mechanical balance equation of skeletal muscles, obtaining stress values ​​of the area of ​​interest, and obtaining quantitative stress values ​​of the target internal tissues; The user's neurophysiological signal data stream is extracted from the preprocessed and spatiotemporally synchronized multimodal data stream. The user's neurophysiological signal data stream is subjected to fast Fourier transform to calculate the power spectral density. The average power values ​​of the prefrontal cortex EEG signals in the θ band and β band are extracted. The neurophysiological fitness index is calculated according to the formula. Based on the statistical distribution of the neural adaptability index observed in a controlled experimental environment when subjects perform familiar tasks and can receive external cues without interference, a timing judgment threshold is set. Based on experimental research data on the stress critical value at which fatigue damage occurs in the target internal tissue, a biomechanical safety threshold is set. The quantitative stress value of the target's internal tissue is compared with the biomechanical safety threshold, and the neural adaptability index is compared with the timing judgment threshold. When the quantitative stress value of the target's internal tissue exceeds the biomechanical safety threshold and the neural adaptability index is higher than the timing judgment threshold, the logical judgment condition is met. When the logical judgment conditions are met, a set of equipment control parameters is found based on the biomechanical optimization algorithm to reduce the quantitative stress value of the target internal tissue, and the equipment control parameters are encapsulated into an executable intelligent collaborative intervention decision instruction.

2. The AI-based product ergonomic design method as described in claim 1, characterized in that: Collect real-time multimodal data streams from users and perform preprocessing, including the following steps: The system synchronously collects real-time multimodal physiological and behavioral data streams from users through a depth vision sensor, a pressure distribution sensor matrix, an inertial measurement unit, and a portable headband device. The system extracts the three-dimensional coordinates of skeletal joints from the user's real-time multimodal physiological and behavioral data stream to generate the user's posture kinematic data stream. It also performs pressure center statistics and regional pressure distribution mapping on the user's real-time multimodal physiological and behavioral data stream to generate the user's sitting posture pressure distribution data stream. The user's real-time multimodal physiological and behavioral data stream is subjected to attitude angle calculation and filtering denoising to generate a user desktop status data stream; Calculate the power spectral density of EEG signals in a specific frequency band from the user's real-time multimodal physiological and behavioral data stream to generate the user's neurophysiological signal data stream; The user posture kinematics data stream, user sitting pressure distribution data stream, user desktop status data stream, and user neurophysiological signal data stream are time-stamped and standardized in data format to generate a multimodal data stream that has been preprocessed and spatiotemporally synchronized.

3. The AI-based product ergonomic design method as described in claim 2, characterized in that: A multi-source data fusion network utilizing an attention mechanism is used to perform weighted fusion and feature extraction on preprocessed multimodal data streams to generate an assessment result of the user's current ergonomic status. This includes the following steps: The pre-processed and spatiotemporally synchronized multimodal data stream is input into the attention mechanism multi-source data fusion network. The attention mechanism multi-source data fusion network calculates dynamic weights for different data sources and different time steps in the pre-processed and spatiotemporally synchronized multimodal data stream. The preprocessed and spatiotemporally synchronized multimodal data streams are weighted and fused according to dynamic weights to generate weighted multimodal features. The attention-based multi-source data fusion network extracts high-level temporal features from the weighted multimodal feature representation to generate a fused deep feature vector. The attention mechanism multi-source data fusion network maps the fused deep feature vectors to a comprehensive real-time comfort score and labels of uncomfortable body parts, generating an assessment result of the user's current ergonomic status.

4. The AI-based product ergonomic design method as described in claim 3, characterized in that: Based on the set of target parameters for multiple devices in the intelligent collaborative intervention decision-making instruction, a preliminary list of device action sequences is obtained, including the following steps: The intelligent collaborative intervention decision command is parsed, the set of equipment adjustment target parameters is extracted, the current state parameters of the actuator are queried, the current state parameter set corresponding to the set of equipment adjustment target parameters is obtained, and the difference between the multiple sets of equipment adjustment target parameters and the current state parameter set is calculated element by element using arithmetic subtraction to generate a parameter difference set. Based on the set of parameter differences, generate basic action instructions for each parameter; The basic action instructions are sorted in logical order and combined into a preliminary list of equipment action sequences.

5. The AI-based product ergonomic design method as described in claim 4, characterized in that: Based on a preliminary list of device action sequences, a smooth and reversible multi-device cooperative motion trajectory instruction from the current state to the target state is planned for all devices to be adjusted using a path planning algorithm, including the following steps: The initial list of equipment action sequences is mapped to a path planning start point and a path planning end point in a high-dimensional parameter space. Between the path planning start point and the path planning end point, a fast random exploration tree algorithm is used to randomly sample and generate multiple possible paths connecting the start point and the end point. From multiple possible paths, select the optimal path that satisfies the requirements of mechanical constraints and motion smoothness, and discretize the optimal path into a series of intermediate path points arranged in time order; The intermediate path point sequence is encoded into a multi-device collaborative motion trajectory instruction with timestamps.

6. The AI-based product ergonomic design method as described in claim 5, characterized in that: The actuator is controlled to execute multi-device coordinated motion trajectory commands. After completion, new multimodal data streams and subjective feedback are collected to generate quantitative evaluation results of the intervention effect, including the following steps: The multi-device collaborative motion trajectory command is sent to the actuator controller, and the actuator controller drives the actuator to move according to the trajectory and timestamp specified in the multi-device collaborative motion trajectory command; The system collects new multimodal data streams from users through a sensor array, collects subjective feedback scores of user comfort after intervention through a human-computer interaction interface, and calculates quantitative evaluation results of intervention effect by combining the new multimodal data streams and subjective feedback scores using an evaluation function.

7. The AI-based product ergonomic design method as described in claim 6, characterized in that, The quantitative evaluation of intervention effects, updating the decision-making strategy of the multi-source data fusion network based on the attention mechanism, and optimizing the generation of future intelligent collaborative intervention decision instructions include the following steps: The quantitative evaluation results of the intervention effect are transformed into reward values ​​under the reinforcement learning framework, and the current user's current ergonomic status evaluation results, intelligent collaborative intervention decision instructions and reward values ​​are used to form an experience sample. Experience samples are stored in the experience replay buffer. A batch of experience samples are sampled from the experience replay buffer. The gradient of the policy network in the multi-source data fusion network of the attention mechanism is calculated using the near-end policy optimization method based on the policy gradient theorem. The gradient descent method is used to update the parameters of the policy network in the multi-source data fusion network of the attention mechanism, thereby optimizing the generation strategy of future intelligent collaborative intervention decision instructions.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the AI-based product ergonomic design method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the AI-based product ergonomic design method according to any one of claims 1 to 7.

Images