Lifting shifting machine attitude optimization method and system based on biomechanical analysis

Abstract

The invention provides a lifting displacement machine attitude optimization method and system based on biomechanical analysis, and relates to the technical attitude optimization field, and the method comprises the following steps: obtaining operator somatosensory data, building a man-machine interaction model to calculate joint stress, and mapping the somatosensory data into a control instruction; a deep reinforcement learning network online optimization model is used, and physical exhaustion minimization and task efficiency maximization are taken as targets; and acquiring stress data in real time to evaluate the motion state and generate a correction instruction to adjust the motion posture. User experience is emphasized, and uniform stress of a patient is ensured through real-time force feedback and posture optimization; establishing a data file to monitor vital signs; a friendly interaction interface and voice prompt are provided, so that the psychological burden of the patient is relieved; scientific displacement postures can assist in evaluating the rehabilitation progress. According to the invention, the working intensity of an operator can be reduced, the safety and comfort of a patient in a displacement process are obviously improved, the win-win situation of medical staff and the patient is realized, and the occupational injury risk of the medical staff is effectively reduced.

Description

Technical Field

[0001] The present invention relates to posture optimization technology, and in particular to a lifting and transfer machine posture optimization method and system based on biomechanical analysis. Background Art

[0002] With the accelerated aging of the population and the growing demand for medical care, lifts are becoming increasingly popular in hospitals, nursing homes, and home care settings as a crucial aid for medical staff and family caregivers. These lifts are primarily used to assist patients with limited mobility or bedridden mobility, such as those moving from bed to wheelchair or wheelchair to bathroom, thereby alleviating the physical burden on caregivers and reducing the risk of work-related musculoskeletal injuries.

[0003] Traditional lifts typically utilize manual control or simple electric control systems, requiring operators to issue commands through buttons, remote controls, and other devices to control the lift's movements, such as lifting and moving. With technological advancements, some advanced lifts are integrating sensor systems that can detect the device's load state and make simple automatic adjustments based on pre-programmed procedures. Some research is also exploring the use of human-computer interaction technology, which can control the device's movements by recognizing simple gestures or voice commands from the operator.

[0004] However, the existing control technology for lifting and shifting machines has obvious shortcomings. First, the traditional control method fails to fully consider the biomechanical characteristics of the operator, resulting in caregivers needing to adopt unnatural postures during operation. Long-term use can easily lead to occupational diseases such as low back pain. Secondly, the existing system lacks an intelligent human-machine collaboration mechanism and is unable to dynamically adjust the control strategy according to the operator's physical condition and operational proficiency. This results in large differences in the experience of different caregivers using the same equipment, and the learning cost is high. In addition, most lifting and shifting machines lack real-time feedback and adaptive adjustment capabilities. If there are changes in resistance or changes in the patient's condition during the shifting process, the motion trajectory cannot be optimized in time, resulting in unsmooth operation and even safety hazards. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for optimizing the posture of a lifting and transfer machine based on biomechanical analysis, which can solve the problems in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for optimizing the posture of a lifting and transfer machine based on biomechanical analysis, comprising:

[0007] Acquiring somatosensory data of the operator, the somatosensory data including body posture data, joint angle data, and electromyographic signal data of the operator; establishing a human-computer interaction model based on the somatosensory data, and calculating joint force data of the operator during control of the lifting and shifting machine; and mapping the somatosensory data into motion control instructions for the lifting and shifting machine based on the joint force data;

[0008] The human-machine interaction model is optimized online through a deep reinforcement learning network, with the operator's physical exertion minimized and the efficiency of the shifting task maximized as the optimization goals. The mapping relationship between the somatosensory data and the motion control instructions is dynamically adjusted to achieve intelligent control of the lifting and shifting machine.

[0009] During the process of the lifting and shifting machine executing the motion control instruction, the force sensor data of the lifting and shifting machine is collected in real time, and the real-time force state of the lifting and shifting machine is calculated according to the force sensor data;

[0010] Based on the real-time force state, the human-computer interaction model is used to evaluate in real time whether the motion state of the lifting and shifting machine meets ergonomic requirements. If the requirements are not met, a correction control instruction is automatically generated, and the correction control instruction is used to adjust the motion trajectory of the lifting and shifting machine in real time;

[0011] The motion posture of the lifting and shifting machine is adjusted according to the modified control instruction, and the adjusted motion state data is fed back to the deep reinforcement learning network to optimize the human-computer interaction model.

[0012] Establishing a human-computer interaction model based on the somatosensory data and calculating the joint force data of the operator in the process of controlling the lifting and shifting machine; mapping the somatosensory data of the operator into motion control instructions of the lifting and shifting machine according to the joint force data includes:

[0013] A human-computer interaction model is used to define an allowable error range between joint forces and a preset safety threshold. Based on the human-computer interaction model, a hierarchical dynamic optimization framework is used to calculate the operator's joint force data, including: calculating the sum of the products of weight coefficients of multiple joints and the corresponding joint forces, wherein the weight coefficients are adaptively adjusted based on the importance and frequency of use of the joints to obtain overall force distribution data, and constructing a multi-joint collaborative optimization target based on the overall force distribution data;

[0014] A mapping conversion function is constructed using the human-computer interaction model in combination with a historical operation experience library, wherein the historical operation experience library contains the correspondence between somatosensory data and control instructions in typical operation scenarios, and an optimized mapping conversion function is obtained by minimizing the square of the distance between the predicted control instructions of the mapping conversion function and the multi-joint collaborative optimization target and a regularization parameter;

[0015] A pre-control instruction is generated based on the joint force data and the optimized mapping conversion function, and it is determined whether the pre-control instruction meets the constraint conditions of the human-computer interaction model within a preset time interval; when the pre-control instruction meets the constraint conditions, the pre-control instruction is smoothed to obtain a motion control instruction to ensure the continuity and smoothness of the control instruction.

[0016] The human-computer interaction model is combined with the historical operation experience library to construct a mapping conversion function, and the optimized mapping conversion function is obtained by minimizing the square of the distance between the predictive control instruction of the mapping conversion function and the multi-joint collaborative optimization target and the regularization parameter, including:

[0017] Building a historical operation experience library, combining the human-computer interaction model with the historical operation experience library, and constructing a mapping conversion function through a deep neural network, the deep neural network including multiple computing layers, the mapping conversion function being used to convert somatosensory data features into predictive control instructions;

[0018] Establishing a distance square loss function, the distance square loss function is used to calculate the deviation between the predictive control instruction and the multi-joint collaborative optimization target, and constructing the optimization target in combination with a regularization parameter;

[0019] The optimization objective is minimized using a gradient descent method, and the weight matrix and bias vector in the deep neural network are updated iteratively; based on the minimization result, the mapping conversion function is optimized and adjusted to obtain an optimized mapping conversion function.

[0020] The human-machine interaction model is optimized online through a deep reinforcement learning network, with minimizing the operator's physical exertion and maximizing the efficiency of the shifting task as the optimization goals. The mapping relationship between the somatosensory data and the motion control instructions is dynamically adjusted to achieve intelligent control of the lifting and shifting machine, including:

[0021] The human-computer interaction model is optimized using a deep reinforcement learning network, and an initial control strategy is generated using a bionic swarm intelligence algorithm. The bionic swarm intelligence algorithm uses the product of path pheromone strength, heuristic information, and state transition probability to perform swarm-style parameter optimization;

[0022] Verifying the initial control strategy in a digital twin environment, fusing actual execution rewards with simulation environment rewards through dynamic weight coefficients to obtain a mixed reality evaluation index, and optimizing and adjusting the initial control strategy based on the mixed reality evaluation index to obtain an initial control strategy;

[0023] The initial control strategy is optimized online, a global optimization target is constructed through the group coordination value and the target tracking value, and local strategy adjustment is achieved based on the proportional integral differential calculation of the control error;

[0024] The global optimization target and the result of the local strategy adjustment are combined to generate a comprehensive modulation value, and the mapping relationship between the somatosensory data and the motion control instruction is dynamically adjusted according to the comprehensive modulation value to realize intelligent control of the lifting and shifting machine.

[0025] A hierarchical collaborative algorithm is used to perform online optimization of the control strategy with emotional adaptability. The global optimization target is constructed through the group collaboration value and the target tracking value, and the local strategy adjustment is achieved based on the proportional integral differential calculation of the control error. The following steps are involved:

[0026] Constructing a state vector of a hierarchical collaborative algorithm, the state vector comprising a global layer state vector and a local layer state vector; performing feature mapping on the global layer state vector based on the hierarchical collaborative algorithm, mapping the global layer state vector to a high-dimensional feature space using an adaptive Gaussian kernel function, and obtaining a group collaborative value;

[0027] Performing state estimation and prediction on the global layer state vector based on the Kalman filter algorithm, wherein a priori state estimation value is calculated through a state prediction equation, and the observation data is fused with the prior estimation value through a measurement update equation to obtain a state estimation result, which is input into a target tracker, and the target state is tracked through a dynamic weight adaptation mechanism to obtain a target tracking value;

[0028] Constructing a global optimization target based on the group collaboration value, the target tracking value, and the emotional adaptability reward value in combination with a dynamic weight allocation strategy, wherein the dynamic weight allocation strategy adaptively adjusts the weight coefficient based on the importance and time-varying characteristics of each optimization target;

[0029] Based on the local layer state vector, a variable structure adaptive controller is used to design a local adjustment strategy. The variable structure adaptive controller includes an adaptive law and a switching function. The control parameters are dynamically adjusted through the adaptive law, and the online switching of the control strategy is achieved through the switching function. The adaptive law adjusts the control parameters according to the changing trend of the control error, and the switching function achieves smooth switching between different control strategies based on the system state.

[0030] The global layer state vector is estimated and predicted based on the Kalman filter algorithm, and the state estimation result is input into the target tracker to obtain the target tracking value including:

[0031] The global layer state vector is estimated and predicted based on the Kalman filter algorithm, wherein the optimal state estimate at the previous moment is multiplied by the system state transfer matrix through the state prediction equation and the process noise is added to obtain the prior state estimate. The observation data at the current moment is fused with the prior state estimate through the measurement update equation to obtain the state estimation result;

[0032] The state estimation result is input into the target tracker, and the target state is tracked by multiplying the deviation between the target state and the current state estimation value by the adaptive gain coefficient using a dynamic weight adaptive mechanism to obtain a target tracking value.

[0033] Based on the real-time force state, the human-machine interaction model is used to evaluate in real time whether the motion state of the lifting and shifting machine meets the ergonomic requirements. When the requirements are not met, the corrective control instructions are automatically generated, including:

[0034] A comprehensive evaluation index is constructed based on the human-computer interaction model, wherein the comprehensive evaluation index includes a posture evaluation component, a mechanical load component, and a fatigue component. The posture evaluation components are fused using an adaptive weight allocation algorithm to obtain a fused evaluation value. The adaptive weight allocation algorithm preferentially calculates an importance index and an information weight of each evaluation component, and normalizes the importance index and the information weight to obtain a weight coefficient of each evaluation component.

[0035] The difference between the fusion evaluation value and the preset ergonomic safety threshold is input into a fuzzy inference algorithm as an evaluation deviation. The fuzzy inference algorithm maps the weight coefficient to a fuzzy membership space and performs an inference operation based on a preset fuzzy rule base, wherein the fuzzy rule base includes a correspondence between the evaluation deviation and the ergonomic requirements. The fuzzy rule base is optimized by online updating to obtain a satisfaction degree determination result.

[0036] According to the satisfaction degree judgment result, the evaluation deviation is multiplied by the adaptive gain coefficient to obtain a control instruction correction amount, and the control instruction correction amount is added to the original control instruction to automatically generate a corrected control instruction, wherein the adaptive gain coefficient is dynamically adjusted according to the changing trend of the evaluation deviation to ensure that the corrected control instruction meets the control constraint conditions.

[0037] A second aspect of an embodiment of the present invention provides a lifting and transfer machine posture optimization system based on biomechanical analysis, comprising:

[0038] The first unit is configured to obtain somatosensory data of the operator, the somatosensory data including body posture data, joint angle data, and electromyographic signal data of the operator; establish a human-computer interaction model based on the somatosensory data, and calculate the joint force data of the operator when controlling the lifting and shifting machine; and map the somatosensory data into motion control instructions for the lifting and shifting machine based on the joint force data;

[0039] The second unit is configured to perform online optimization of the human-machine interaction model through a deep reinforcement learning network, with minimizing the operator's physical exertion and maximizing the efficiency of completing the shifting task as optimization goals, dynamically adjusting the mapping relationship between the somatosensory data and the motion control instructions, and realizing intelligent control of the lifting and shifting machine;

[0040] The third unit is configured to collect force sensor data of the lifting and shifting machine in real time during the process of the lifting and shifting machine executing the motion control instruction, and calculate the real-time force state of the lifting and shifting machine according to the force sensor data;

[0041] a fourth unit configured to evaluate in real time, based on the real-time force state, whether the motion state of the lifting and shifting machine meets ergonomic requirements using the human-computer interaction model, and automatically generate a correction control instruction when the requirements are not met, wherein the correction control instruction is used to adjust the motion trajectory of the lifting and shifting machine in real time;

[0042] The fifth unit is configured to adjust the motion posture of the lifting and shifting machine according to the modified control instruction, and feed back the adjusted motion state data to the deep reinforcement learning network to optimize the human-computer interaction model.

[0043] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0044] processor;

[0045] a memory for storing processor-executable instructions;

[0046] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0047] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0048] The beneficial effects of this application are as follows:

[0049] The lift posture optimization method based on biomechanical analysis provided by the present invention establishes a human-computer interaction model by collecting the operator's somatosensory data, which can accurately capture the operator's intention and physical state, making the control of the lift more intuitive and natural, and significantly improving the efficiency and comfort of human-computer interaction.

[0050] This method uses a deep reinforcement learning network to dynamically optimize the human-computer interaction model, with the goal of minimizing the operator's physical exertion and maximizing task completion efficiency. It realizes intelligent adjustment of the control mapping relationship, effectively reduces the operator's fatigue during long-term operation, reduces the risk of occupational injury, and improves work safety.

[0051] By collecting the lift's force sensor data in real time and conducting ergonomic assessments, the system can automatically generate corrective control instructions and adjust the device's motion posture to ensure a smooth and safe transfer process. At the same time, it forms a closed-loop optimization mechanism that enables the system to continuously learn and adapt to the usage habits of different operators, improving the lift's adaptability and practicality in various complex medical care scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Schematic diagram of a flow chart of a method for optimizing the posture of a lifting and transfer machine based on biomechanical analysis according to an embodiment of the present invention;

[0053] Figure 2 This is a bar chart comparing the performance of the human-computer interaction model in different scenarios according to an embodiment of the present invention;

[0054] Figure 3 Schematic diagram of the correlation matrix between adaptability parameters and control parameters according to an embodiment of the present invention;

[0055] Figure 4 A flow chart for evaluating a human-computer interaction model and generating control instructions for an embodiment of the present invention;

[0056] Figure 5 This is a bar chart comparing and analyzing the evaluation performance of the human-computer interaction model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0058] The technical solution of the present invention is described in detail below with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0059] Figure 1 FIG. 1 is a flow chart of a method for optimizing the posture of a lifting and shifting machine based on biomechanical analysis according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0060] Acquiring somatosensory data of the operator, the somatosensory data including body posture data, joint angle data, and electromyographic signal data of the operator; establishing a human-computer interaction model based on the somatosensory data, and calculating joint force data of the operator during control of the lifting and shifting machine; and mapping the somatosensory data into motion control instructions for the lifting and shifting machine based on the joint force data;

[0061] The human-machine interaction model is optimized online through a deep reinforcement learning network, with the operator's physical exertion minimized and the efficiency of the shifting task maximized as the optimization goals. The mapping relationship between the somatosensory data and the motion control instructions is dynamically adjusted to achieve intelligent control of the lifting and shifting machine.

[0062] During the process of the lifting and shifting machine executing the motion control instruction, the force sensor data of the lifting and shifting machine is collected in real time, and the real-time force state of the lifting and shifting machine is calculated according to the force sensor data;

[0063] Based on the real-time force state, the human-computer interaction model is used to evaluate in real time whether the motion state of the lifting and shifting machine meets ergonomic requirements. If the requirements are not met, a correction control instruction is automatically generated, and the correction control instruction is used to adjust the motion trajectory of the lifting and shifting machine in real time;

[0064] The motion posture of the lifting and shifting machine is adjusted according to the modified control instruction, and the adjusted motion state data is fed back to the deep reinforcement learning network to optimize the human-computer interaction model.

[0065] By acquiring the operator's somatosensory data, building a human-computer interaction model to calculate joint forces, and mapping the somatosensory data into control instructions, the system uses a deep reinforcement learning network to optimize the model online, aiming to minimize physical exertion and maximize task efficiency. Force data is collected in real time to assess motion status and generate correction instructions to adjust motion posture. This invention can reduce operator workload, improve the safety and efficiency of displacement tasks, and reduce the risk of occupational injury to medical staff.

[0066] In an optional embodiment, establishing a human-computer interaction model based on the somatosensory data and calculating the joint force data of the operator in the process of controlling the lifting and shifting machine; mapping the operator's somatosensory data into motion control instructions for the lifting and shifting machine according to the joint force data includes:

[0067] A human-computer interaction model is used to define an allowable error range between joint forces and a preset safety threshold. Based on the human-computer interaction model, a hierarchical dynamic optimization framework is used to calculate the operator's joint force data, including: calculating the sum of the products of weight coefficients of multiple joints and the corresponding joint forces, wherein the weight coefficients are adaptively adjusted based on the importance and frequency of use of the joints to obtain overall force distribution data, and constructing a multi-joint collaborative optimization target based on the overall force distribution data;

[0068] A mapping conversion function is constructed using the human-computer interaction model in combination with a historical operation experience library, wherein the historical operation experience library contains the correspondence between somatosensory data and control instructions in typical operation scenarios, and an optimized mapping conversion function is obtained by minimizing the square of the distance between the predicted control instructions of the mapping conversion function and the multi-joint collaborative optimization target and a regularization parameter;

[0069] A pre-control instruction is generated based on the joint force data and the optimized mapping conversion function, and it is determined whether the pre-control instruction meets the constraint conditions of the human-computer interaction model within a preset time interval; when the pre-control instruction meets the constraint conditions, the pre-control instruction is smoothed to obtain a motion control instruction to ensure the continuity and smoothness of the control instruction.

[0070] The operator's somatosensory data is acquired, including multi-dimensional information such as body posture, joint angles, muscle activity, and pressure distribution. This somatosensory data is collected using a multimodal sensor system and acquired through wearable devices. For example, angle sensors are installed at key joints to measure the spatial posture of joints such as the wrist, elbow, and shoulder. Pressure sensors are attached to the surface of the limbs to monitor the force applied to different muscle groups in real time. The acquisition frequency is set to 60Hz to ensure real-time and continuous data.

[0071] Based on the acquired somatosensory data, a human-computer interaction model was established. This model consists of three parts: a kinematics model, a dynamics model, and a biomechanical constraint model. The kinematics model describes the geometric constraints and range of motion of human joints. For example, the flexion and extension range of the wrist joint is -70° to 70°, and the rotation range is -90° to 90°. The dynamics model describes the relationship between joint force and motion state. The biomechanical constraint model defines the safe force thresholds for each joint, such as the maximum load on the lumbar spine not exceeding 3000N and the maximum torque on the shoulder joint not exceeding 60Nm.

[0072] Based on the human-computer interaction model, a hierarchical dynamic optimization framework is used to calculate the operator's joint force data. This framework is divided into three levels: the first level calculates the force of individual joints; the second level considers the synergy between joints; and the third level performs overall optimization. In specific implementation, weight coefficients are first assigned to different joints to reflect their importance and frequency of use. For example, for lifting movements, the initial weights of the lumbar spine, knee joint, and shoulder joint are set to 0.4, 0.3, and 0.3, respectively. As the frequency of use of each joint changes during the operation, the system automatically adjusts the weight coefficients. For example, if the shoulder joint is used frequently, its weight increases to 0.5, while the weights of the lumbar spine and knee joints decrease to 0.3 and 0.2 respectively. The sum of the products of the force on each joint and the corresponding weight coefficient is calculated to obtain the overall force distribution data. For example, the force on the lumbar spine is 2000N multiplied by the weight of 0.4, which is 800; the force on the knee joint is 1500N multiplied by the weight of 0.3, which is 450; the force on the shoulder joint is 1000N multiplied by the weight of 0.3, which is 300. The total is 1550, which is used as the overall force evaluation indicator.

[0073] Based on the overall force distribution data, a multi-joint collaborative optimization objective is constructed. This objective function comprehensively considers the force balance of each joint, operating comfort, and task completion efficiency. In order to achieve accurate mapping of somatosensory data to control instructions, the system constructs a mapping conversion function in combination with the historical operation experience library. The historical operation experience library contains the correspondence between somatosensory data and control instructions in typical operation scenarios, such as the correspondence between the angle change of the lifting action and the lifting speed, the correspondence between the horizontal push-pull action and the translation speed, etc. The system extracts 100 sets of typical data from the experience library as training samples, each set containing the operator's somatosensory data and the corresponding control instruction parameters.

[0074] The optimized mapping function is obtained by minimizing the squared distance and regularization parameter between the predicted control instructions of the mapping function and the multi-joint collaborative optimization target. Euclidean distance is used for distance calculation, and the regularization parameter is set to 0.05 to prevent overfitting. The optimization process is iterative, with a maximum number of iterations set to 500 and a convergence threshold of 0.001. For example, if the initial mapping function predicts a lifting speed of 0.2 m / s, while the optimization target is 0.15 m / s, through iterative optimization, the lifting speed output by the mapping function converges to 0.16 m / s, meeting the accuracy requirements.

[0075] Based on the joint force data and the optimized mapping conversion function, pre-control instructions are generated, including parameters such as lifting speed, translation speed, and rotation angle. The system determines whether the pre-control instructions meet the constraints of the human-computer interaction model within the preset time interval (usually set to 2 seconds). It mainly checks three aspects: whether the joint force exceeds the safety threshold, whether the motion trajectory is smooth and continuous, and whether the control parameters are within the allowable range of the equipment. For example, it is detected that the pre-control instruction causes the lumbar spine to be subjected to a force of 2800N, which does not exceed the safety threshold of 3000N; the translation speed change rate is 0.05m / s 2 , less than the maximum allowable change rate of 0.1m / s 2 The lifting speed is 0.3m / s, which is within the equipment's allowable range of 0-0.5m / s. Therefore, it is determined that the pre-control instruction meets the constraint conditions.

[0076] When the pre-control instructions meet the constraints, the system smooths them to ensure continuity and stability. This smoothing process uses a low-pass filter algorithm, with a filter window size of 5 sampling points and a cutoff frequency of 10 Hz. For example, the original pre-control instruction sequence is [0.25, 0.30, 0.28, 0.35, 0.33] m / s. After smoothing, it becomes [0.26, 0.29, 0.30, 0.32, 0.33] m / s, reducing the degree of sudden change in the instructions. The smoothed motion control instructions are ultimately output and transmitted to the actuator of the lift and shift machine, achieving precise control of the equipment.

[0077] Figure 2 This is a bar chart comparing the performance of the human-computer interaction model in different scenarios according to an embodiment of the present invention:

[0078] The figure shows the performance of the human-machine collaborative system in five different mobility scenarios, evaluated based on three dimensions: prediction accuracy, operator comfort, and response timeliness. The system performed best in the straight-line motion scenario on flat ground, achieving 92.5% prediction accuracy, 85.7% operator comfort, and 89.3% response timeliness, maintaining high levels of all three indicators. In the turning scenario, prediction accuracy dropped slightly to 88.2% and operator comfort dropped to 83.9%, but response timeliness remained high at 90.6%. In the uphill and downhill scenarios, system performance declined somewhat, with prediction accuracy at 85.7%, operator comfort dropping to 79.3%, and response timeliness at 87.4%. In the narrow passage scenario, all indicators remained relatively stable, with prediction accuracy at 83.9%, operator comfort at 81.2%, and response timeliness at 86.5%. The multi-obstacle scenario presented the greatest challenge, with prediction accuracy dropping to 81.6% and operator comfort at its lowest at 76.8%, but response timeliness remained at an acceptable level of 84.9%. Data shows that the system performs well in simple scenarios, but its performance will be reduced in complex environments, especially the operating comfort, which needs further optimization.

[0079] In an optional embodiment, the human-computer interaction model is combined with the historical operation experience library to construct a mapping conversion function, and the optimized mapping conversion function is obtained by minimizing the square of the distance between the predictive control instruction of the mapping conversion function and the multi-joint collaborative optimization target and the regularization parameter, including:

[0080] Building a historical operation experience library, combining the human-computer interaction model with the historical operation experience library, and constructing a mapping conversion function through a deep neural network, the deep neural network including multiple computing layers, the mapping conversion function being used to convert somatosensory data features into predictive control instructions;

[0081] Establishing a distance square loss function, the distance square loss function is used to calculate the deviation between the predictive control instruction and the multi-joint collaborative optimization target, and constructing the optimization target in combination with a regularization parameter;

[0082] The optimization objective is minimized using a gradient descent method, and the weight matrix and bias vector in the deep neural network are updated iteratively; based on the minimization result, the mapping conversion function is optimized and adjusted to obtain an optimized mapping conversion function.

[0083] Build a historical operation experience database. The historical operation experience database contains a large number of somatosensory data features and corresponding control instruction data pairs of human operators when controlling mechanical equipment. For example, 100 hours of operation data of 50 operators performing the same task are collected. Each data pair includes the operator's somatosensory data features such as joint angle, acceleration, speed and its corresponding mechanical equipment control instructions. Specifically, the somatosensory data features can be the three-dimensional coordinates, speed and acceleration of the operator's wrists, elbows, shoulders and other joints, and the control instructions can be parameters such as the joint angle value of the robotic arm and the opening and closing degree of the gripper. These data are collected at a frequency of 120Hz using data acquisition equipment and are preprocessed to remove noise and outliers to form a standardized data set.

[0084] The human-computer interaction model is combined with the historical operation experience library, and a mapping conversion function is constructed through a deep neural network. The deep neural network includes an input layer, multiple hidden layers, and an output layer. In this embodiment, the input layer receives somatosensory data features, including the three-dimensional coordinates, velocity, and acceleration of the operator's 18 joints, for a total of 108 input features; the hidden layer is set to 4 layers, with 256, 128, 64, and 32 nodes in each layer respectively; the output layer corresponds to the predictive control instructions, including the angle values ​​of the 6 joints of the robotic arm and the opening and closing degree of the gripper, for a total of 7 output values. A full connection is used between each layer, the hidden layer uses the ReLU activation function, and the output layer uses the Tanh activation function to limit the output range.

[0085] Deep neural networks convert the input somatosensory data features into predictive control instructions through weight matrices and bias vectors. For example, from the input layer to the first hidden layer, let the input feature be X (a 108-dimensional vector), the weight matrix be W1 (a 108×256-dimensional matrix), and the bias vector be b1 (a 256-dimensional vector). Then the output H1 of the first hidden layer is calculated as H1 equal to the matrix product of the ReLU function applied to W1 and X plus the result of b1. Similarly, through layer-by-layer calculations, the final output layer result is the predictive control instruction Y (a 7-dimensional vector).

[0086] A distance square loss function is established to calculate the deviation between the predicted control instruction and the multi-joint collaborative optimization target. The multi-joint collaborative optimization target is an ideal control instruction predefined according to specific task requirements. In this embodiment, assuming that the predicted control instruction is Y (7-dimensional vector) and the multi-joint collaborative optimization target is Y* (7-dimensional vector), the distance square loss function is calculated as the sum of the squares of the differences between the corresponding elements of Y and Y*. For example, if Y is [0.3, 0.5, -0.2, 0.1, 0.4, -0.3, 0.6] and Y* is [0.32, 0.48, -0.22, 0.12, 0.38, -0.28, 0.58], the distance square loss is 0.0028.

[0087] To prevent overfitting, a regularization parameter is introduced to construct the optimization objective. In this example, L2 regularization is used, and the regularization coefficient is set to 0.001. The regularization term is calculated as the sum of the squares of all weight matrix elements multiplied by the regularization coefficient. The distance squared loss is added to the regularization term to obtain the final optimization objective function.

[0088] The optimization objective is minimized using the gradient descent method. In this embodiment, the Adam optimizer is used, the learning rate is set to 0.001, the batch size is 64, and the number of training iterations is 10,000. The gradient of the optimization objective function with respect to each weight matrix and bias vector is calculated for each iteration, and the parameters are updated accordingly. For example, for the weight matrix W1, the update formula is: the new value of W1 is equal to the old value of W1 minus the learning rate multiplied by the gradient of the optimization objective function with respect to W1. The loss function value on the validation set is monitored during training. When the loss no longer decreases after 10 consecutive iterations, the training is stopped early.

[0089] Based on the minimization results, the mapping function is optimized and adjusted. After training, the model performance is evaluated using a test set. In this example, the average squared distance loss of the model on the test set is reduced to 0.0025, an 83.7% reduction compared to the unoptimized 0.0153. For specific tasks, such as moving an object from one location to another, controlling a robotic arm using the optimized mapping function reduces completion time from an average of 42 seconds to 28 seconds, and increases operational accuracy from an average error of 1.8 cm to 0.7 cm.

[0090] The optimized mapping conversion function enables more efficient human-machine interaction. The operator simply performs intuitive body movements, and the system accurately converts sensory data into control instructions for the mechanical equipment through this function, significantly reducing the operator's learning curve and operational difficulty. For example, a novice operator can complete basic tasks using this system after 15 minutes of training, while traditional control methods require more than four hours of training.

[0091] This embodiment effectively solves the problems of complex mapping relationships and non-intuitive operations in human-computer interaction by constructing and optimizing mapping conversion functions, improves operational efficiency and accuracy, and has broad application prospects.

[0092] In an optional embodiment, the human-machine interaction model is optimized online through a deep reinforcement learning network, with minimizing the operator's physical exertion and maximizing the efficiency of the shifting task as the optimization goals, and the mapping relationship between the somatosensory data and the motion control instructions is dynamically adjusted to achieve intelligent control of the lifting and shifting machine, including:

[0093] The human-computer interaction model is optimized using a deep reinforcement learning network, and an initial control strategy is generated using a bionic swarm intelligence algorithm. The bionic swarm intelligence algorithm uses the product of path pheromone strength, heuristic information, and state transition probability to perform swarm-style parameter optimization;

[0094] Verifying the initial control strategy in a digital twin environment, fusing actual execution rewards with simulation environment rewards through dynamic weight coefficients to obtain a mixed reality evaluation index, and optimizing and adjusting the initial control strategy based on the mixed reality evaluation index to obtain an initial control strategy;

[0095] The initial control strategy is optimized online, a global optimization target is constructed through the group coordination value and the target tracking value, and local strategy adjustment is achieved based on the proportional integral differential calculation of the control error;

[0096] The global optimization target and the result of the local strategy adjustment are combined to generate a comprehensive modulation value, and the mapping relationship between the somatosensory data and the motion control instruction is dynamically adjusted according to the comprehensive modulation value to realize intelligent control of the lifting and shifting machine.

[0097] A deep reinforcement learning network was used to optimize the human-computer interaction model. This network generated an initial control strategy based on a bionic swarm intelligence algorithm. In its implementation, an ant colony optimization algorithm was used to generate policy parameters. In the algorithm, the path pheromone intensity was set to an initial value of 0.01 and gradually decayed by a proportional factor of 0.95 with increasing iterations. Heuristic information consisted of two components: the operator's physical exertion model and the task completion time, with physical exertion weighted at 0.6 and task completion time weighted at 0.4. State transition probabilities were calculated by multiplying the pheromone intensity with the heuristic information, with the pheromone influence factor α set to 1.0 and the heuristic information influence factor β set to 2.0.

[0098] In each iteration, the top 10% of parameter combinations with the highest probability are retained and randomly perturbed within a ±15% range to generate a new set of candidate parameters. After 500 iterations, the parameter combination with the highest reward value is selected as the initial control strategy, which contains 15 key mapping parameters that cover the conversion relationship between somatosensory data and motion control commands.

[0099] The effectiveness of the initial control strategy was verified in a digital twin environment. A virtual model consistent with the actual lift was constructed, with physical parameter errors controlled within ±2%. The operator's somatosensory data was input into the digital twin environment, and the response characteristics of the virtual lift, including acceleration, velocity curve, and position trajectory, were recorded. Actual execution rewards were calculated based on the energy consumption data of the actual device and the task completion time. The energy consumption baseline was set at 120 watt-hours, with a 15-point reward added for every 10% reduction in energy consumption. The task completion time baseline was set at 35 seconds, with a 20-point reward added for every 5-second improvement in time.

[0100] The simulation environment reward is calculated based on the virtual model and uses the same evaluation criteria. The two rewards are combined using a dynamic weighting coefficient. Initially, the actual execution reward weight is 0.3, while the simulation environment reward weight is 0.7. As the number of verifications increases, the actual execution reward weight gradually increases to 0.7. Based on mixed reality evaluation metrics, a sensitivity analysis of different control strategy parameters was conducted, and 15 key parameters such as the mapping curve slope, response delay, and smoothing factor were refined and adjusted to ultimately obtain an optimized initial control strategy.

[0101] In the actual application environment, the initial control strategy is optimized online. A group synergy value evaluation model is constructed, and data from 20 users of different body shapes and operating habits are collected to extract common operating features and form a synergy feature vector. The group synergy value is obtained by calculating the similarity between the current operator's characteristics and the synergy feature vector. The similarity is calculated using cosine distance, with a numerical range of 0-1, and the synergy value weight coefficient is set to 0.45. The target tracking value is calculated based on a preset optimal trajectory template, which includes three basic action modes: linear movement, steering, and precise positioning. An ideal response curve is set for each mode. The system calculates the deviation between the actual trajectory and the ideal trajectory in real time. A full score of 100 is obtained when the deviation is less than 5%. As the deviation increases, the score decreases proportionally. The target tracking value weight coefficient is set to 0.55. The group synergy value and the target tracking value are weighted and summed to obtain the global optimization target value.

[0102] For local strategy adjustments, a proportional-integral-derivative (PID) calculation is implemented based on the control error. The proportional coefficient is initially set to 0.8 and automatically increases to 1.2 when the error exceeds a threshold of 10%. The integral coefficient is fixed at 0.25 to eliminate long-term steady-state errors. The differential coefficient is initially set to 0.15 and automatically increases to 0.3 to enhance response speed when rapid operator command changes are detected. The system calculates the error value every 100 milliseconds, continuously collecting 10 error values ​​to form an error sequence. The adjustment amount is calculated using the PID algorithm, and the adjustment range is limited to within ±25% of the original parameter value to prevent operational instability caused by sudden changes in strategy.

[0103] Finally, the global optimization goal and the local strategy adjustment results are combined to generate a comprehensive modulation value. The initial value of the global target weight is set to 0.6, and the initial value of the local adjustment weight is set to 0.4. When a specific task mode is detected, the system automatically adjusts the weight distribution. For example, in a precise positioning task, the local adjustment weight is increased to 0.7. The comprehensive modulation value is obtained by weighted summation, and the numerical range is between -1 and 1. According to the comprehensive modulation value, the mapping relationship between the somatosensory data and the motion control instructions is adjusted in real time. The mapping function uses a piecewise linear interpolation method to dynamically fine-tune the key points within the range of ±20% to ensure the smoothness and response speed of the control. In this way, the lifting and shifting machine can intelligently adjust the control strategy according to the habits of different operators and the characteristics of different tasks to achieve the control goals of minimizing physical exertion and maximizing task completion efficiency.

[0104] In an optional embodiment, a hierarchical collaborative algorithm is used to perform online optimization of the control strategy with emotional adaptability, a global optimization target is constructed through group collaboration value and target tracking value, and local strategy adjustment is achieved based on proportional integral differential calculation of control error, including:

[0105] Constructing a state vector of a hierarchical collaborative algorithm, the state vector comprising a global layer state vector and a local layer state vector; performing feature mapping on the global layer state vector based on the hierarchical collaborative algorithm, mapping the global layer state vector to a high-dimensional feature space using an adaptive Gaussian kernel function, and obtaining a group collaborative value;

[0106] Performing state estimation and prediction on the global layer state vector based on the Kalman filter algorithm, wherein a priori state estimation value is calculated through a state prediction equation, and the observation data is fused with the prior estimation value through a measurement update equation to obtain a state estimation result, which is input into a target tracker, and the target state is tracked through a dynamic weight adaptation mechanism to obtain a target tracking value;

[0107] Constructing a global optimization target based on the group collaboration value, the target tracking value, and the emotional adaptability reward value in combination with a dynamic weight allocation strategy, wherein the dynamic weight allocation strategy adaptively adjusts the weight coefficient based on the importance and time-varying characteristics of each optimization target;

[0108] Based on the local layer state vector, a variable structure adaptive controller is used to design a local adjustment strategy. The variable structure adaptive controller includes an adaptive law and a switching function. The control parameters are dynamically adjusted through the adaptive law, and the online switching of the control strategy is achieved through the switching function. The adaptive law adjusts the control parameters according to the changing trend of the control error, and the switching function achieves smooth switching between different control strategies based on the system state.

[0109] The state vector construction of the layered collaborative algorithm includes the global layer state vector and the local layer state vector. The global layer state vector contains the system position coordinates (x, y, z), velocity (v x ,v y ,v z ), acceleration (a x ,a y ,a z ) and angles (θ, φ, ψ) to characterize the overall system state; the local state vector, which contains parameters such as the control error (e), the rate of change of the error (de / dt), and the control output (u), characterizes the operating state of a single control unit. For example, for a robot swarm system, the global state vector records the position and velocity information of all robots, while the local state vector records the control parameters and error state of each robot.

[0110] When performing feature mapping on the global layer state vector, the adaptive Gaussian kernel function K(x,y)=exp(-||xy|| 2 / 2σ 2 ), where σ is an adaptive bandwidth parameter that is dynamically adjusted based on the data distribution. The bandwidth parameter σ is adjusted as follows: when the data distribution is relatively concentrated, σ is reduced to improve feature resolution; when the data distribution is relatively dispersed, σ is increased to enhance feature generalization. In practice, σ ranges from [0.1 to 10], with an initial value of 1.0. Each iteration is updated based on the square root of the variance of the data distribution. This mapping transforms the low-dimensional state vector into a high-dimensional feature representation, allowing the calculation of the group synergy value, which quantifies the degree of synergy between units in the system.

[0111] The Kalman filter algorithm performs state estimation and prediction for the global layer state vector in two stages. The prediction stage calculates a priori state estimates using the state prediction equation, taking into account the system's dynamic model and control inputs. The update stage fuses observations with the prior estimates using the measurement update equation to produce a state estimate. The Kalman gain is dynamically adjusted based on the prediction error covariance and the observation noise covariance to achieve optimal state estimation. For example, for the position state variable, the prediction error covariance is initially set to the diagonal matrix diag(0.1, 0.1, 0.1), and the observation noise covariance is set to the diagonal matrix diag(0.05, 0.05, 0.05). After the state estimate is input into the target tracker, target state tracking is performed using a dynamic weight adaptation mechanism. This mechanism dynamically adjusts tracker parameters based on the tracking error, increasing the tracking gain when the tracking error is large and decreasing it when the error is low to ensure stable tracking performance. The tracking gain is adjustable within the range [0.2, 2.0], initially at 1.0, and updated based on an exponential function of the error.

[0112] The construction of the global optimization objective comprehensively considers the group coordination value C, the target tracking value T, and the emotional adaptability reward value R. A dynamic weight allocation strategy uses normalized weight coefficients, adaptively adjusting them based on the importance and time-varying characteristics of each objective. The weight coefficients w1, w2, and w3 correspond to the group coordination value, target tracking value, and emotional adaptability reward value, respectively, satisfying the constraint of w1 + w2 + w3 = 1. The initial values ​​are set to w1 = 0.4, w2 = 0.4, and w3 = 0.2, and are subsequently dynamically adjusted based on the system state. When the system needs to strengthen collaborative behavior, the value of w1 is increased; when the system needs to accurately track the target, the value of w2 is increased; and when the system needs to emphasize emotional adaptability, the value of w3 is increased. The weight adjustment step size is 0.05, and it is updated every 100ms to ensure a smooth system transition.

[0113] The design of the local adjustment strategy is based on a variable structure adaptive controller, which includes an adaptive law and a switching function. The adaptive law adjusts the control parameters according to the control error e and its rate of change de / dt. The control parameters include the proportional gain K p , integral gain K i and differential gain K d When |e|>0.1, the proportional gain K p According to K p =K p 0+α|e| is adjusted, where K p 0 is the basic gain, α is the adaptive coefficient, and its value range is [0.5, 2.0]. When |de / dt|>0.05, the differential gain K d According to K d =K d 0+β|de / dt| is adjusted, where K d 0 is the basic gain, β is the adaptive coefficient, and the value range is [0.2, 1.0]; the integral gain K i According to the cumulative effect of the error according to K i =K i 0 / (1+γ∫|e|dt) is adjusted, where K i 0 is the base gain, and γ is the suppression coefficient, set to 0.1, to prevent integral windup. The switching function S is defined as S = λe + de / dt, where λ is a positive constant, set to 2.0. When |S| < 0.05, the system adopts a continuous control strategy; when |S| ≥ 0.05, the system switches to a discrete control strategy, achieving rapid response to large disturbances.

[0114] In actual application scenarios, taking robot collaborative work as an example, initially six robots are scattered in a 20×20 meter working area, and the global layer state vector records the position coordinates and speed information of all robots. The group collaboration value calculated by feature mapping is 0.72, indicating that the system has good collaboration; the target tracking value obtained by Kalman filtering and target tracker is 0.85, indicating that the system can track the target well; the emotional adaptability reward value is 0.68, reflecting the system's ability to adapt to environmental changes. The dynamic weights are adjusted to w1=0.35, w2=0.45, and w3=0.2 to construct a global optimization target. The local layer controller adjusts the control parameters according to the error state of each robot. The control parameters of one robot are adjusted from the initial value K p =1.2, K i =0.5, K d =0.8 adjusted to K p =1.45,K i =0.42, K d =0.95, achieving precise control of the system.

[0115] Figure 3 Schematic diagram of the correlation matrix between the adaptability parameters and the control parameters according to an embodiment of the present invention:

[0116] This heat map shows the results of a correlation analysis between eight key parameters in a human-machine collaborative control system. Several significant strong correlations can be observed in the data: the group coordination coefficient exhibits strong positive correlations with target tracking deviation (0.82) and control stability (0.87). The correlation coefficient between the switching function threshold and the adaptive law gain is as high as 0.92, indicating a close coupling between these two control parameters. Response speed also exhibits strong positive correlations with control stability (0.82) and target tracking deviation (0.85). Notably, interference rejection exhibits a weak negative correlation with the switching function threshold (-0.28) and the adaptive law gain (-0.17), indicating that improving the system's interference rejection performance will, to a certain extent, affect control sensitivity. Furthermore, the positive correlations between the adaptive reward and the switching function threshold (0.79) and the adaptive law gain (0.83) indicate a significant synergistic effect between the system's adaptability and the switching mechanism of the control strategy. Overall, this correlation analysis reveals the complex coupling relationships between the system's parameters, providing an important theoretical basis for optimizing control strategies.

[0117] In an optional embodiment, the global layer state vector is estimated and predicted based on a Kalman filter algorithm, and the state estimation result is input into a target tracker to obtain a target tracking value, which includes:

[0118] The global layer state vector is estimated and predicted based on the Kalman filter algorithm, wherein the optimal state estimate at the previous moment is multiplied by the system state transfer matrix through the state prediction equation and the process noise is added to obtain the prior state estimate. The observation data at the current moment is fused with the prior state estimate through the measurement update equation to obtain the state estimation result;

[0119] The state estimation result is input into the target tracker, and the target state is tracked by multiplying the deviation between the target state and the current state estimation value by the adaptive gain coefficient using a dynamic weight adaptive mechanism to obtain a target tracking value.

[0120] In practical applications, the global state vector typically contains key information such as the target's position, velocity, and acceleration. The dimensionality of the state vector can be determined based on tracking requirements. For example, in two-dimensional plane tracking, the state vector can be set to a six-dimensional vector, including the target's x-coordinate, y-coordinate, x-speed, y-speed, x-acceleration, and y-acceleration.

[0121] The Kalman filter algorithm's state estimation and prediction process for the global layer state vector can be divided into a prediction step and an update step. In the prediction step, the system uses the state prediction equation to calculate the prior state estimate. In specific implementation, assuming the current time is k, the system multiplies the optimal state estimate at the previous time k-1 by the system state transition matrix. The system state transition matrix describes how the state changes over time. In a uniform motion model, if the sampling interval is 0.1 seconds, the relationship between position and velocity can be expressed as: the new position equals the old position plus the velocity multiplied by the time interval of 0.1 seconds.

[0122] The system also needs to consider the impact of process noise and add it to the calculation results. Process noise represents the uncertainty in the system model and can be set as Gaussian white noise with zero mean and a covariance matrix Q. In practical applications, the diagonal elements of the Q matrix can be set based on the system characteristics. For example, the noise variance of the position component can be set to 0.01, the noise variance of the velocity component can be set to 0.05, and the noise variance of the acceleration component can be set to 0.1.

[0123] During the update step, the system fuses the current observation data with the prior state estimate using the measurement update equation. First, the Kalman gain is calculated, which determines the system's confidence in the new observation. The calculation of the Kalman gain involves the prior estimation error covariance matrix, the observation matrix, and the observation noise covariance matrix. The observation noise covariance matrix R describes the uncertainty in the measurement process. In practical applications, if high-precision sensors are used, the diagonal elements of the R matrix can be set to small values, such as 0.02 for position measurement noise variance. If low-precision sensors are used, they can be set to larger values, such as 0.2.

[0124] After obtaining the Kalman gain, the system uses the weighted sum of the prior state estimate and the observation residual (the difference between the actual observation and the predicted observation) as the posterior state estimate. The observation residual multiplied by the Kalman gain represents the correction required to the prior estimate. The system also updates the posterior estimate error covariance matrix to prepare for the next calculation.

[0125] In practical applications, assume the initial state vector is [100, 150, 5, 10, 0, 0], indicating that the target's initial position is (100, 150), the initial velocity is (5, 10), and the initial acceleration is (0, 0). The initial estimation error covariance matrix P can be set to a diagonal matrix with diagonal elements of [10, 10, 1, 1, 0.1, 0.1], representing the uncertainty of the estimate of each component of the initial state.

[0126] After state estimation is complete, the system inputs the estimated state results into the target tracker. The target tracker uses a dynamic weight adaptation mechanism to track the target state. In implementation, the system first calculates the deviation between the target state and the current state estimate. This deviation vector is represented as the difference between the target's ideal state and the Kalman filter's estimated state. For example, if the target's ideal position is (120, 170) and the Kalman filter's estimated position is (118, 173), the position deviation is (2, -3).

[0127] The system dynamically adjusts the adaptive gain coefficient based on the deviation. The adaptive gain coefficient can be calculated using a nonlinear function, resulting in a smaller gain for small deviations and a larger gain for large deviations. For example, for position deviation, the gain coefficient can be set to 0.5×(1-exp(-0.1×|deviation|)). In this case, for an x-axis deviation of 2 units, the gain coefficient is approximately 0.095; for a y-axis deviation of -3 units, the gain coefficient is approximately 0.139.

[0128] The target tracking value is calculated by multiplying the deviation between the target state and the current state estimate by the corresponding adaptive gain coefficient and then adding it to the state estimate. For example, the target tracking value in the x-direction is 118 + 2 × 0.095 = 118.19, and the target tracking value in the y-direction is 173 + (-3) × 0.139 = 172.58.

[0129] Furthermore, the system can build a dynamic model based on the target's motion history and adjust the adaptive gain coefficient calculation strategy. For example, for fast-moving targets, the gain coefficients for the velocity and acceleration components can be increased; for slow-moving targets, the gain coefficients for these components can be decreased. The system can also calculate the average velocity and acceleration by storing the target's state over the past 10 frames, and adjust the gain coefficients accordingly.

[0130] Through this method, the system can effectively address various uncertainties in target motion and achieve precise target tracking. Experiments have shown that in typical scenarios, this method can control the target position tracking error to within ±1.5 units and the velocity tracking error to within ±0.8 units / second, significantly improving the accuracy and robustness of target tracking.

[0131] In an optional embodiment, based on the real-time force state, using the human-machine interaction model to evaluate in real time whether the motion state of the lifting and shifting machine meets ergonomic requirements, and when the requirements are not met, automatically generating a correction control instruction includes:

[0132] A comprehensive evaluation index is constructed based on the human-computer interaction model, wherein the comprehensive evaluation index includes a posture evaluation component, a mechanical load component, and a fatigue component. The posture evaluation components are fused using an adaptive weight allocation algorithm to obtain a fused evaluation value. The adaptive weight allocation algorithm preferentially calculates an importance index and an information weight of each evaluation component, and normalizes the importance index and the information weight to obtain a weight coefficient of each evaluation component.

[0133] The difference between the fusion evaluation value and the preset ergonomic safety threshold is input into a fuzzy inference algorithm as an evaluation deviation. The fuzzy inference algorithm maps the weight coefficient to a fuzzy membership space and performs an inference operation based on a preset fuzzy rule base, wherein the fuzzy rule base includes a correspondence between the evaluation deviation and the ergonomic requirements. The fuzzy rule base is optimized by online updating to obtain a satisfaction degree determination result.

[0134] According to the satisfaction degree judgment result, the evaluation deviation is multiplied by the adaptive gain coefficient to obtain a control instruction correction amount, and the control instruction correction amount is added to the original control instruction to automatically generate a corrected control instruction, wherein the adaptive gain coefficient is dynamically adjusted according to the changing trend of the evaluation deviation to ensure that the corrected control instruction meets the control constraint conditions.

[0135] like Figure 4 As shown, the method includes:

[0136] By collecting the real-time force status data of the lifting and shifting machine and combining it with the human-computer interaction model for evaluation, corrective control instructions are automatically generated when ergonomic requirements are not met.

[0137] The human-computer interaction model constructs comprehensive evaluation indicators, including posture assessment, mechanical load, and fatigue components. The posture assessment component uses angle sensors to collect the operator's joint angle data, such as maintaining the waist flexion angle within the range of 20-45 degrees, the shoulder flexion angle below 60 degrees, and the wrist deflection angle within 15 degrees. The mechanical load component uses pressure sensors to monitor the pressure distribution of the operator during lifting and shifting. Under normal circumstances, the force on the lumbar vertebrae L4 / L5 should not exceed 3400 Newtons, and the pressure on the spine should be kept below 2500 Newtons.

[0138] The fatigue component is calculated based on the duration and intensity of the operation. For example, the fatigue coefficient increases by 0.05 when the continuous operation time exceeds 30 minutes, and the fatigue coefficient increases by 0.08 when the operation force exceeds 60% of the maximum autonomous force.

[0139] An adaptive weight allocation algorithm fuses these three evaluation components. The algorithm first calculates the importance index and information weight of each component. The importance index is determined based on the impact of each component in a specific operational scenario. For example, when lifting a heavy object, the importance index of the mechanical load component is set to 0.5, the posture assessment component to 0.3, and the fatigue component to 0.2.

[0140] Information weights are calculated using the entropy method, determining the amount of information contained in each component based on its volatility. For example, when posture data fluctuates dramatically, its information weight reaches 0.45, while the more stable fatigue data has an information weight of only 0.15. The importance index and information weights are weighted and normalized to obtain the final weight coefficients. For example, the weight of the posture assessment component is 0.35, the weight of the mechanical load component is 0.42, and the weight of the fatigue component is 0.23.

[0141] After the weight coefficients are determined, the three components are weighted and summed to obtain the fused assessment value. Assuming the posture assessment component score is 78, the mechanical load component score is 65, and the fatigue component score is 85, the fused assessment value is 78 × 0.35 + 65 × 0.42 + 85 × 0.23 = 74.15. The preset ergonomic safety threshold is 80 points, so the assessment deviation is 80 - 74.15 = 5.85 points.

[0142] The estimated deviations are then fed into a fuzzy inference algorithm for processing. The algorithm first maps the estimated deviations into a fuzzy membership space. For example, an estimated deviation of 5.85 would have a membership of 0.7 for "medium deviation" and 0.3 for "small deviation." The weight coefficients are also mapped into the fuzzy space. For example, a posture assessment component weight of 0.35 would have a membership of 0.6 for "moderately important" and 0.4 for "very important."

[0143] Fuzzy reasoning operates based on a pre-set fuzzy rule base. This base contains the mapping between assessment deviations and ergonomic requirements, such as "If the assessment deviation is large and the posture assessment component weight is high, the degree of non-compliance with ergonomic requirements is high." In practice, the fuzzy rule base is optimized through online updates. For example, if the system detects an operator experiencing discomfort, it adjusts the strength of the corresponding rule, changing a situation originally judged as "moderate non-compliance" to "high non-compliance." Fuzzy rule reasoning and calculations yield a satisfaction level determination, such as "satisfaction level is 65%."

[0144] Based on the satisfaction level, the system multiplies the evaluation deviation by the adaptive gain coefficient to determine the control command correction. The adaptive gain coefficient is dynamically adjusted based on the changing trend of the evaluation deviation. For example, if the evaluation deviation increases three times in a row, the gain coefficient is increased from 0.8 to 1.2. When the evaluation deviation stabilizes, the gain coefficient remains at 0.9. Assuming the current gain coefficient is 1.1 and the evaluation deviation is 5.85, the control command correction is 5.85 × 1.1 = 6.435.

[0145] The control command correction is added to the original control command to generate the final corrected control command. For example, if the original control command requires a lift speed of 10 cm / s, the corrected control command adjusts the lift speed to 3.565 cm / s. The system ensures that the corrected control command meets the control constraints, such as a lift speed of no less than 2 cm / s and no more than 15 cm / s, and a lift height within the range of 40-120 cm.

[0146] Through the above-mentioned real-time evaluation and automatic correction mechanism, the lifting and transfer machine can automatically adjust its motion parameters according to the actual situation of the operator, ensuring that the operation process meets ergonomic requirements, reducing the operator's physical burden and injury risk, and improving the safety of auxiliary equipment and the efficiency of human-machine collaboration.

[0147] Figure 5 This is a bar chart comparing and analyzing the performance of the human-computer interaction model evaluation according to an embodiment of the present invention:

[0148] This figure shows the performance evaluation results of the human-robot collaborative system under four different test scenarios, including evaluation data for posture evaluation, mechanical load, and fatigue. In the standard operation scenario, the system performed most balanced, with posture evaluation reaching 87.5%, mechanical load at 79.2%, and fatigue at 82.6%. In the high-load displacement scenario, the mechanical load component significantly improved to 83.5%, while posture evaluation and fatigue decreased to 75.8% and 71.3%, respectively. In the rapid response scenario, posture evaluation performance significantly improved to 92.4%, but mechanical load and fatigue indicators were relatively low, at 77.1% and 68.9%, respectively. In the irregular path scenario, all three indicators remained relatively stable, with posture evaluation at 81.6%, mechanical load at 73.8%, and fatigue at 79.5%. The data shows that the system exhibits different performance characteristics under different operating conditions. In particular, posture control has a significant advantage in the rapid response scenario, but also reveals shortcomings in fatigue control under high-load conditions, providing important reference for system optimization.

[0149] A second aspect of an embodiment of the present invention provides a lifting and transfer machine posture optimization system based on biomechanical analysis, comprising:

[0150] The first unit is configured to obtain somatosensory data of the operator, the somatosensory data including body posture data, joint angle data, and electromyographic signal data of the operator; establish a human-computer interaction model based on the somatosensory data, and calculate the joint force data of the operator when controlling the lifting and shifting machine; and map the somatosensory data into motion control instructions for the lifting and shifting machine based on the joint force data;

[0151] The second unit is configured to perform online optimization of the human-machine interaction model through a deep reinforcement learning network, with minimizing the operator's physical exertion and maximizing the efficiency of completing the shifting task as optimization goals, dynamically adjusting the mapping relationship between the somatosensory data and the motion control instructions, and realizing intelligent control of the lifting and shifting machine;

[0152] The third unit is configured to collect force sensor data of the lifting and shifting machine in real time during the process of the lifting and shifting machine executing the motion control instruction, and calculate the real-time force state of the lifting and shifting machine according to the force sensor data;

[0153] a fourth unit configured to evaluate in real time, based on the real-time force state, whether the motion state of the lifting and shifting machine meets ergonomic requirements using the human-computer interaction model, and automatically generate a correction control instruction when the requirements are not met, wherein the correction control instruction is used to adjust the motion trajectory of the lifting and shifting machine in real time;

[0154] The fifth unit is configured to adjust the motion posture of the lifting and shifting machine according to the modified control instruction, and feed back the adjusted motion state data to the deep reinforcement learning network to optimize the human-computer interaction model.

[0155] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0156] processor;

[0157] a memory for storing processor-executable instructions;

[0158] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0159] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0160] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing the posture of a lifting and transfer machine based on biomechanical analysis, characterized in that: include: Acquiring somatosensory data of the operator, wherein the somatosensory data includes body posture data, joint angle data, and electromyographic signal data of the operator; A human-computer interaction model is established based on the somatosensory data, and joint force data of the operator in the process of controlling the lifting and shifting machine is calculated; the somatosensory data is mapped into motion control instructions for the lifting and shifting machine according to the joint force data; The human-machine interaction model is optimized online through a deep reinforcement learning network, with the operator's physical exertion minimized and the efficiency of the shifting task maximized as the optimization goals. The mapping relationship between the somatosensory data and the motion control instructions is dynamically adjusted to achieve intelligent control of the lifting and shifting machine. During the process of the lifting and shifting machine executing the motion control instruction, the force sensor data of the lifting and shifting machine is collected in real time, and the real-time force state of the lifting and shifting machine is calculated according to the force sensor data; Based on the real-time force state, the human-computer interaction model is used to evaluate in real time whether the motion state of the lifting and shifting machine meets ergonomic requirements. If the requirements are not met, a correction control instruction is automatically generated, and the correction control instruction is used to adjust the motion trajectory of the lifting and shifting machine in real time; The motion posture of the lifting and shifting machine is adjusted according to the modified control instruction, and the adjusted motion state data is fed back to the deep reinforcement learning network to optimize the human-computer interaction model.

2. The method according to claim 1, characterized in that Establishing a human-computer interaction model based on the somatosensory data and calculating the joint force data of the operator in the process of controlling the lifting and shifting machine; Mapping the operator's body sensory data into motion control instructions for the lifting and shifting machine according to the joint force data includes: A human-computer interaction model is used to define an allowable error range between joint forces and a preset safety threshold. Based on the human-computer interaction model, a hierarchical dynamic optimization framework is used to calculate the operator's joint force data, including: calculating the sum of the products of weight coefficients of multiple joints and the corresponding joint forces, wherein the weight coefficients are adaptively adjusted based on the importance and frequency of use of the joints to obtain overall force distribution data, and constructing a multi-joint collaborative optimization target based on the overall force distribution data; A mapping conversion function is constructed using the human-computer interaction model in combination with a historical operation experience library, wherein the historical operation experience library contains the correspondence between somatosensory data and control instructions in typical operation scenarios, and an optimized mapping conversion function is obtained by minimizing the square of the distance between the predicted control instructions of the mapping conversion function and the multi-joint collaborative optimization target and a regularization parameter; A pre-control instruction is generated based on the joint force data and the optimized mapping conversion function, and it is determined whether the pre-control instruction meets the constraint conditions of the human-computer interaction model within a preset time interval; when the pre-control instruction meets the constraint conditions, the pre-control instruction is smoothed to obtain a motion control instruction to ensure the continuity and smoothness of the control instruction.

3. The method according to claim 2, characterized in that The human-computer interaction model is combined with the historical operation experience library to construct a mapping conversion function, and the optimized mapping conversion function is obtained by minimizing the square of the distance between the predictive control instruction of the mapping conversion function and the multi-joint collaborative optimization target and the regularization parameter, including: Building a historical operation experience library, combining the human-computer interaction model with the historical operation experience library, and constructing a mapping conversion function through a deep neural network, the deep neural network including multiple computing layers, the mapping conversion function being used to convert somatosensory data features into predictive control instructions; Establishing a distance square loss function, the distance square loss function is used to calculate the deviation between the predictive control instruction and the multi-joint collaborative optimization target, and constructing the optimization target in combination with a regularization parameter; The optimization objective is minimized using a gradient descent method, and the weight matrix and bias vector in the deep neural network are updated iteratively; based on the minimization result, the mapping conversion function is optimized and adjusted to obtain an optimized mapping conversion function.

4. The method according to claim 1, wherein The human-machine interaction model is optimized online through a deep reinforcement learning network, with minimizing the operator's physical exertion and maximizing the efficiency of the shifting task as the optimization goals. The mapping relationship between the somatosensory data and the motion control instructions is dynamically adjusted to achieve intelligent control of the lifting and shifting machine, including: The human-computer interaction model is optimized using a deep reinforcement learning network, and an initial control strategy is generated using a bionic swarm intelligence algorithm. The bionic swarm intelligence algorithm uses the product of path pheromone strength, heuristic information, and state transition probability to perform swarm-style parameter optimization; Verifying the initial control strategy in a digital twin environment, fusing actual execution rewards with simulation environment rewards through dynamic weight coefficients to obtain a mixed reality evaluation index, and optimizing and adjusting the initial control strategy based on the mixed reality evaluation index to obtain an initial control strategy; The initial control strategy is optimized online, a global optimization target is constructed through the group coordination value and the target tracking value, and local strategy adjustment is achieved based on the proportional integral differential calculation of the control error; The global optimization target and the result of the local strategy adjustment are combined to generate a comprehensive modulation value, and the mapping relationship between the somatosensory data and the motion control instruction is dynamically adjusted according to the comprehensive modulation value to realize intelligent control of the lifting and shifting machine.

5. The method according to claim 4, characterized in that A hierarchical collaborative algorithm is used to perform online optimization of the control strategy with emotional adaptability. The global optimization target is constructed through the group collaboration value and the target tracking value, and the local strategy adjustment is achieved based on the proportional integral differential calculation of the control error. The following steps are involved: Constructing a state vector of a hierarchical collaborative algorithm, the state vector comprising a global layer state vector and a local layer state vector; performing feature mapping on the global layer state vector based on the hierarchical collaborative algorithm, mapping the global layer state vector to a high-dimensional feature space using an adaptive Gaussian kernel function, and obtaining a group collaborative value; Performing state estimation and prediction on the global layer state vector based on the Kalman filter algorithm, wherein a priori state estimation value is calculated through a state prediction equation, and the observation data is fused with the prior estimation value through a measurement update equation to obtain a state estimation result, which is input into a target tracker, and the target state is tracked through a dynamic weight adaptation mechanism to obtain a target tracking value; Constructing a global optimization target based on the group collaboration value, the target tracking value, and the emotional adaptability reward value in combination with a dynamic weight allocation strategy, wherein the dynamic weight allocation strategy adaptively adjusts the weight coefficient based on the importance and time-varying characteristics of each optimization target; Based on the local layer state vector, a variable structure adaptive controller is used to design a local adjustment strategy. The variable structure adaptive controller includes an adaptive law and a switching function. The control parameters are dynamically adjusted through the adaptive law, and the online switching of the control strategy is achieved through the switching function. The adaptive law adjusts the control parameters according to the changing trend of the control error, and the switching function achieves smooth switching between different control strategies based on the system state.

6. The method according to claim 5, characterized in that The global layer state vector is estimated and predicted based on the Kalman filter algorithm, and the state estimation result is input into the target tracker to obtain the target tracking value including: The global layer state vector is estimated and predicted based on the Kalman filter algorithm, wherein the optimal state estimate at the previous moment is multiplied by the system state transfer matrix through the state prediction equation and the process noise is added to obtain the prior state estimate. The observation data at the current moment is fused with the prior state estimate through the measurement update equation to obtain the state estimation result; The state estimation result is input into the target tracker, and the target state is tracked by multiplying the deviation between the target state and the current state estimation value by the adaptive gain coefficient using a dynamic weight adaptive mechanism to obtain a target tracking value.

7. The method according to claim 1, characterized in that Based on the real-time force state, the human-machine interaction model is used to evaluate in real time whether the motion state of the lifting and shifting machine meets the ergonomic requirements. When the requirements are not met, the corrective control instructions are automatically generated, including: A comprehensive evaluation index is constructed based on the human-computer interaction model, wherein the comprehensive evaluation index includes a posture evaluation component, a mechanical load component, and a fatigue component. The posture evaluation components are fused using an adaptive weight allocation algorithm to obtain a fused evaluation value. The adaptive weight allocation algorithm preferentially calculates an importance index and an information weight of each evaluation component, and normalizes the importance index and the information weight to obtain a weight coefficient of each evaluation component. The difference between the fusion evaluation value and the preset ergonomic safety threshold is input into a fuzzy inference algorithm as an evaluation deviation. The fuzzy inference algorithm maps the weight coefficient to a fuzzy membership space and performs an inference operation based on a preset fuzzy rule base, wherein the fuzzy rule base includes a correspondence between the evaluation deviation and the ergonomic requirements. The fuzzy rule base is optimized by online updating to obtain a satisfaction degree determination result. According to the satisfaction degree judgment result, the evaluation deviation is multiplied by the adaptive gain coefficient to obtain a control instruction correction amount, and the control instruction correction amount is added to the original control instruction to automatically generate a corrected control instruction, wherein the adaptive gain coefficient is dynamically adjusted according to the changing trend of the evaluation deviation to ensure that the corrected control instruction meets the control constraint conditions.

8. A lifting and transfer machine posture optimization system based on biomechanical analysis, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain the operator's somatosensory data, wherein the somatosensory data includes the operator's body posture data, joint angle data, and myoelectric signal data; A human-computer interaction model is established based on the somatosensory data, and joint force data of the operator in the process of controlling the lifting and shifting machine is calculated; the somatosensory data is mapped into motion control instructions for the lifting and shifting machine according to the joint force data; The second unit is configured to perform online optimization of the human-machine interaction model through a deep reinforcement learning network, with minimizing the operator's physical exertion and maximizing the efficiency of completing the shifting task as optimization goals, dynamically adjusting the mapping relationship between the somatosensory data and the motion control instructions, and realizing intelligent control of the lifting and shifting machine; The third unit is configured to collect force sensor data of the lifting and shifting machine in real time during the process of the lifting and shifting machine executing the motion control instruction, and calculate the real-time force state of the lifting and shifting machine according to the force sensor data; a fourth unit configured to evaluate in real time, based on the real-time force state, whether the motion state of the lifting and shifting machine meets ergonomic requirements using the human-computer interaction model, and automatically generate a correction control instruction when the requirements are not met, wherein the correction control instruction is used to adjust the motion trajectory of the lifting and shifting machine in real time; The fifth unit is configured to adjust the motion posture of the lifting and shifting machine according to the modified control instruction, and feed back the adjusted motion state data to the deep reinforcement learning network to optimize the human-computer interaction model.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images