Exoskeleton control system and method for patient displacement

Abstract

The invention relates to the technical field of exoskeleton control, and discloses an exoskeleton control system and method for patient displacement. The method comprises the following steps: querying historical displacement record information according to physiological feature categories and displacement activity modes of a patient, and carrying out space division to obtain monitoring position points; neural network simulation is conducted through the joint connection structure of the exoskeleton, a control reference unit is established, the input of the control reference unit is the joint use duration, and the output of the control reference unit is the control reference value of the monitoring position point; based on the variability of the use durations of the multiple joints, fusing the output average values of the multiple control reference units, constructing a control reference model, processing the use durations of the multiple joints, and generating a reference value of a monitoring position point; and when it is detected that the displacement action does not accord with the reference value of the monitoring position point, a delay check rule is implemented for data verification. According to the method, self-adaptive control and safety verification of the exoskeleton in the auxiliary displacement process are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of exoskeleton control, in particular to an exoskeleton control system and method for patient transfer. BACKGROUND

[0002] In the field of medical care, patients with difficulty in moving, such as people with limb disabilities, patients in rehabilitation, and the elderly with limited mobility due to age, face many difficulties in their daily life transfer activities. From the bed to the wheelchair, from the wheelchair to the toilet or other locations, each transfer is a challenge for them, not only requiring a lot of physical and mental effort, but also possibly causing physical injury due to improper operation, which puts a heavy burden on the patients and their families. At the same time, the frequent and difficult transfer process also seriously affects the quality of life and mental health of the patients, who often develop negative emotions such as self-esteem and anxiety due to their inability to move.

[0003] A patient with hemiplegy after a stroke needs the help of others due to the loss of strength in one limb during transfer. In the process of transferring from the bed to the wheelchair, the patient may fall due to the imbalance of the body, causing secondary injury; for nursing staff, each time they assist the patient in transferring, they need to exert a lot of physical effort, which can easily lead to physical fatigue and injury over time. Therefore, developing a safe, convenient, and efficient patient transfer method to meet the diverse needs of patients with difficulty in moving during the transfer process has become a problem that needs to be solved in the field of medical care.

[0004] Some transfer devices on the market have great limitations in structure design and use. The wearing process of some transfer devices is very complex, for example, some transfer devices with a hip support belt that users need to wear on the hip position before use, and need to take off after use. This is extremely difficult for patients with poor physical function and difficulty in moving, and even cannot be completed independently, which seriously limits the scope of application of such devices, making many patients in need unable to benefit from them.

[0005] The functional defects of existing transfer devices also cause great distress to patients and nursing staff. One of the prominent problems is that these devices affect the patient's ability to solve the two problems during use. Due to the wearing structure of some transfer devices, the user's pants cannot be easily unfastened and removed when they need to solve the two problems. Even with the assistance of a caregiver, the entire process is very cumbersome, and both the caregiver and the user need to spend a lot of time and effort to complete the process, and the patient will feel extremely embarrassed and uncomfortable during the process, which greatly damages their self-esteem and has a negative impact on their psychology.

[0006] With the continuous progress of technology, exoskeleton technology has gradually attracted attention in the field of rehabilitation. As a wearable device, exoskeletons can provide additional support and power to the human body, helping people with mobility difficulties to achieve autonomous movement and transfer. In rehabilitation therapy, exoskeleton technology has shown certain potential, such as helping patients with lower extremity dysfunction to perform gait training, promoting neural remodeling and functional recovery. However, current exoskeleton control methods still have some problems in patient transfer. Existing exoskeleton control often lacks in-depth analysis of the physiological characteristics and transfer activity patterns of individual patients, resulting in inaccurate and inflexible control that cannot meet the needs of different patients in various transfer scenarios. In addition, the coordination between exoskeleton joint control and actual patient transfer movements needs to be improved, which may affect the safety and comfort of patient transfer. Based on the above background, this patent is committed to researching a more scientific and effective exoskeleton control method for patient transfer to solve the shortcomings of existing technology and provide a better solution for patient transfer. SUMMARY

[0007] The purpose of the present application is to provide an exoskeleton control system and method for patient transfer to solve the problems raised in the background art.

[0008] To achieve the above-mentioned purpose, the present application provides an exoskeleton control method for patient transfer, which comprises: According to the physiological characteristics and transfer activity patterns of patients, query historical transfer record information for spatial division to obtain monitoring position points; Using the joint connection structure of the exoskeleton to simulate the neural network, a control reference unit is established, wherein the input of the control reference unit is the joint usage time, and the output is the control reference value of the monitoring position point; Based on the variability of multiple joint usage times, the average value of the outputs of multiple control reference units is fused to construct a control reference model, which processes multiple joint usage times to generate monitoring position point reference values; When the transfer action is detected to be inconsistent with the monitoring position point reference value, a delay check rule is implemented for data verification.

[0009] Preferably, the implementation of the delay check rule for data verification comprises: defining a delay check rule, wherein the continuous time rule means that the duration of inconsistent state exceeds the time threshold, and the data proportion rule means that the data amount of inconsistent state accounts for more than the proportion threshold in the total data amount; when any condition in the continuous time rule or the data proportion rule is met, an abnormal prompt operation is performed, otherwise the transfer action is continuously monitored.

[0010] Preferably, the step of performing spatial partitioning on the query history shift record information to obtain the monitoring position points comprises: collecting a plurality of historical shift detection values and corresponding detection position distribution data, with physiological feature categories and exoskeleton models of patients as fixed constraints and shift activity patterns as dynamic constraints; performing cluster analysis on the detection position distribution data according to each historical shift detection value to obtain detection position partition results; and obtaining the final detection position partition by taking the intersection of all the detection position partition results; and assigning a monitoring position point to each sub-region of the final detection position partition and adding the monitoring position point to a monitoring position point set.

[0011] Preferably, the step of taking the intersection of all the detection position partition results comprises: extracting the center point of each partition and the two boundary points farthest from the center point for each detection position partition result, constructing a triplet representation of the partition to form a partition triplet set; performing pairwise similarity evaluation on the partition triplet set to obtain a difference degree set between the partition results; removing abnormal partition results based on the difference degree set, and taking the intersection of the remaining partition results to obtain the final detection position partition.

[0012] Preferably, the pairwise similarity evaluation comprises: generating a dynamic bounding box based on the partition triplet set; and obtaining a comprehensive difference measure between the partition results by calculating a weighted combination of the spatial overlap degree and the time series trend of the bounding box; and the specific implementation steps comprise: generating a minimum circumscribed rectangle as an initial bounding box according to the coordinate point set of each triplet; introducing timestamp information of historical partition data to perform sliding window expansion of the bounding box in the time dimension; calculating the three-dimensional spatial overlap rate between the expanded bounding boxes; simultaneously extracting the distribution density change gradient of the coordinate points in each bounding box; and linearly weighting and fusing the spatial overlap rate and the distribution density change gradient to generate a final partition result difference degree value.

[0013] Preferably, the step of establishing the control reference unit by simulating the neural network using the joint connection structure of the exoskeleton comprises: collecting control feature value record data sets of preset monitoring position points in accordance with the exoskeleton model and the shift activity pattern based on joint use duration record data; performing consistency analysis on the control feature value record data sets to obtain control feature value real data; and training the control reference unit using a supervised learning algorithm according to the joint use duration record data and the control feature value real data.

[0014] Preferably, the step of constructing the control reference model based on the variability of a plurality of joint use durations and fusing the output average values of a plurality of control reference units comprises: determining the number of integrated control reference units according to the size of the variability; selecting a plurality of control reference units according to the number of integrated control reference units, calculating the output average values of the plurality of control reference units, and constructing the control reference model.

[0015] Preferably, the training the control reference unit using a supervised learning algorithm comprises defining a loss function that optimizes the training process by calculating the difference between the predicted reference value and the true reference value for each monitoring location point, wherein the difference is cumulatively summed based on the magnitude of the deviation between the predicted value and the true value.

[0016] Preferably, the method further comprises a pre-processing step for the historical shift record information, the pre-processing comprising data cleaning and format standardization.

[0017] Preferably, the application further comprises an exoskeleton control system for patient shifting, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the exoskeleton control method for patient shifting.

[0018] Compared with the prior art, the application has the following beneficial effects: The application analyzes the historical shift record information in depth according to the physiological characteristics of the patient, such as age, physical function status, disease type, etc., and the shift activity mode, such as different activity modes like sitting up from bed, transferring to wheelchair, standing, etc., to divide the space and determine the precise monitoring location points. For example, for elderly patients, considering their weak body flexibility and strength, the key points such as the back and legs will be monitored during the shift activity mode of getting up from bed; for stroke patients with hemiplegia, due to the dysfunction of one side of the body, the different location points of the healthy side and the affected side will be monitored during the shift. This space division method based on physiological characteristics and activity mode can accurately monitor and control each patient's unique situation, making the control of the exoskeleton more in line with the actual needs of the patient.

[0019] This individualized and accurate control greatly reduces the risk during the shift. Precise control can make the exoskeleton better match the patient's body movements, avoiding accidents such as collisions and falls caused by improper control. At the same time, since the exoskeleton can provide appropriate support and assistance according to the patient's physiological characteristics and activity mode, the patient will feel more comfortable during the shift, reducing physical discomfort and fatigue. For example, during the process of transferring from a wheelchair to a toilet, the exoskeleton can accurately adjust the support intensity and angle according to the patient's physical condition and shift movement, allowing the patient to complete the shift smoothly and comfortably, avoiding psychological stress caused by discomfort during the shift.

[0020] The joint connection structure of the exoskeleton is used for neural network simulation to establish a control reference unit, which is a major innovation of the present application. The neural network has strong learning and simulation capabilities and can accurately simulate and analyze the complex movement and use of the exoskeleton joints. By taking the joint use duration as the input and the control reference value of the monitoring position point as the output, the control reference unit can dynamically generate the corresponding control reference value according to the actual use of the joint. For example, when the joint is working at a certain angle for a long time, the control reference unit can adjust the output control reference value in a timely manner according to the change in the use duration to adapt to the working state of the joint and ensure that the control of the exoskeleton is more flexible and intelligent.

[0021] Based on the variability of the use duration of multiple joints, the output average values of multiple control reference units are fused to construct a control reference model, so that the model has strong adaptability. Different displacement activities will cause different changes in the use duration of the joints, and the control reference model can sensitively capture these changes. When the use duration of multiple joints is significantly different in different displacement activities, the control reference model can reasonably fuse the output average values of each control reference unit according to the variability, thereby generating more accurate and actual monitoring position point reference values. For example, in a complex displacement action such as picking up an object from the ground and then standing up, the use duration and movement mode of multiple joints are relatively complex, and the control reference model can adaptively adjust to ensure that the control of the exoskeleton can meet the needs of this complex displacement action, so that the patient can smoothly complete the action. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 The working principle diagram of the exoskeleton control method for patient displacement described in the present application; Figure 2 The flowchart for obtaining the monitoring position point; Figure 3 The flowchart for obtaining the intersection of the detection position partitions; Figure 4 The diagram for identifying the clustering distribution and center point of the monitoring position point; Figure 5 The diagram for the spatial overlap rate of the dynamic boundary box changing with time steps. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] Please refer to Figure 1 The application provides an exoskeleton control system and method for patient transfer, the method comprising: querying historical transfer record information for spatial division according to physiological characteristic categories and transfer activity patterns of the patient to obtain a monitoring position point; the historical transfer record information contains past transfer data of the patient, the physiological characteristic categories involve age, weight, and health status of the patient, and the transfer activity patterns describe action types in the transfer process such as standing up, turning around, or translating. A control reference unit is established by using joint connection structures of the exoskeleton for neural network simulation; the joint connection structures include hip joint and knee joint mechanical components, the neural network simulation adopts a multilayer perceptron model, and the input of the control reference unit is joint usage duration, and the output is a control reference value of the monitoring position point. Based on variation degrees of multiple joint usage durations, an output average value of multiple control reference units is fused to construct a control reference model, the multiple joint usage durations are processed, and a monitoring position point reference value is generated; the variation degrees reflect discrete degrees of the joint usage duration data, and the control reference model optimizes the output through a weighted integration manner. When it is detected that a transfer action does not match the monitoring position point reference value, a delay check rule is implemented for data verification; the transfer action is collected in real time through a sensor, and the delay check rule ensures data reliability.

[0025] Embodiment 1: Please refer to Figure 2 The definition of the delay check rule includes two independent determination conditions of a continuous time rule and a data proportion rule, the continuous time rule refers to a time length that an inconsistent state between the transfer action data and the monitoring position point reference value continuously exists exceeding a preset time threshold, and the data proportion rule refers to a proportion of a number of data records in a set observation window in which the inconsistent state exceeds a preset proportion threshold. The time threshold is quantified according to a minimum response period of the exoskeleton control system and a requirement of the physiological characteristic categories of the patient for transfer safety, and the proportion threshold is set by comprehensively considering a sensor sampling frequency and statistical distribution characteristics of abnormal data in the historical transfer record information. When the exoskeleton control system collects continuous transfer action data streams through the joint sensor in a real-time running process, the system compares and marks a deviation state between the transfer action data at each sampling time and the monitoring position point reference value generated by the control reference model point by point; the system maintains a sliding time window for dynamically counting a duration accumulation value of the inconsistent state and a count of inconsistent data points. If the duration accumulation value of the inconsistent state is detected to exceed the time threshold or the ratio of the count of the inconsistent data points to the total data amount in the window exceeds the proportion threshold in any calculation period, the system immediately triggers an abnormal prompt operation process. The abnormal prompt operation process activates an audible and visual alarm device of an exoskeleton human-machine interface and sends a safety state notification to a nursing staff terminal; if both determination conditions are not met, the system continues to maintain a collection and comparison cycle of the transfer action data and simultaneously updates statistical indicators in the sliding time window.

[0026] The determination of the monitoring position point is based on the spatial division of the historical displacement record information, and the spatial division process takes the physiological characteristic category of the patient and the exoskeleton model as fixed constraint conditions, and takes the displacement activity mode as a dynamic constraint condition. The physiological characteristic category of the patient includes quantitative indicators such as age segmentation, body weight classification, muscle strength level, and motor function disorder degree, the exoskeleton model specifies mechanical structure parameters and drive unit specifications, and the displacement activity mode covers typical displacement scenes such as lying-to-sitting, bed-to-chair transfer, and standing support. The system retrieves a plurality of historical displacement detection values and corresponding detection position distribution data from the historical database according to the combination condition of the fixed constraint and the dynamic constraint. The historical displacement detection value is a coordinate set of key position points in the actual displacement trajectory recorded in the past, and the detection position distribution data is a point cloud representation of these coordinate points in three-dimensional space. For each historical displacement detection value obtained by retrieval, the system calls a clustering analysis algorithm to process the corresponding detection position distribution data; the clustering analysis algorithm adopts K-means clustering based on Euclidean distance, and the number of clusters K is automatically determined by elbow rule analysis of the contour coefficient of the historical data set. The detection position distribution data corresponding to each historical displacement detection value produces a detection position partition result after clustering analysis, and the detection position partition result is composed of K spatial sub-regions, and each sub-region contains a group of point cloud data with similar spatial positions.

[0027] The intersection operation of all detection location partition results is the core step to obtain the final detection location partition, and the system performs data structure conversion on each detection location partition result. The system extracts the geometric center point coordinates of the K sub-regions of each detection location partition, and calculates the Euclidean distance of all points in each sub-region to the center point and selects the two boundary points with the farthest distance; each sub-region is represented by a triple, which contains the center point coordinates and the coordinates of the two boundary points. The triple representation of all detection location partition results constitutes a partition triple set, and the system performs similarity evaluation on the elements in the partition triple set. The similarity evaluation calculates the spatial coverage and shape similarity between each two partition triples, the spatial coverage is obtained by comparing the volume intersection ratio of the sub-regions represented by the two triples, and the shape similarity is obtained by comparing the principal direction angle of the point cloud distribution of the sub-regions; the system takes the weighted sum of the spatial coverage and the shape similarity as the difference value between the partition results, and the difference values of all pairwise comparisons constitute a difference set. The system applies the outlier detection algorithm to analyze the difference set, identifies and removes abnormal partition results whose difference values are significantly deviated from the group distribution; for the remaining partition results after screening, the system performs set intersection operation: the intersection operation is performed for each spatial sub-region, and the system calculates the overlapping area of the minimum bounding cuboid of all sub-regions with the same index in the remaining partition results, and defines the overlapping area as a sub-region of the final detection location partition. The final detection location partition is composed of K overlapping sub-regions, and the system assigns a monitoring location point to each sub-region; the coordinates of the monitoring location point are obtained by calculating the weighted centroid of the point cloud of the sub-region, and the weight coefficient is determined by the time sequence correlation strength between the data points in the point cloud and the historical displacement detection value. All monitoring location points of the sub-regions are added to the monitoring location point set, and the monitoring location point set is the output target of the control reference model.

[0028] The preprocessing of historical shift record information is a necessary step before spatial division. The preprocessing operation includes data cleaning and format standardization. In the data cleaning stage, missing values, abnormal values and noise data in the historical database are processed. The missing values are filled with the moving average of historical data of the same patient under the same shift activity mode. The abnormal values are identified and removed by the Isolation Forest algorithm. Specifically, the numerical data fields in the historical shift record information, such as joint usage time record data, are taken as input data sets. The algorithm constructs multiple isolated trees by randomly selecting features and randomly selecting split points. Each isolated tree isolates data points by recursively dividing the data space. Abnormal data points are often isolated in fewer random segmentation steps due to their large difference from normal points, and thus have shorter path lengths. The noise data is smoothed by wavelet transform filtering technology. In the format standardization stage, the time stamp format, coordinate unit and sensor dimension of different source data are unified. The time stamp is converted to the world coordinated time format. The coordinate unit is converted to the metric unit. The sensor dimension is normalized according to the calibration parameters of the exoskeleton model. The historical shift record information after preprocessing is stored in a structured database for the spatial division module to call. The data verification process in the delay check rule and the spatial division process of the monitoring position points have a cooperative relationship. The accuracy of the monitoring position point set directly affects the accuracy of the inconsistent state determination. The exoskeleton control system collects the data stream of the joint angle sensor, torque sensor and inertial measurement unit in real time during operation. After coordinate transformation, the data stream is matched with the reference value of the monitoring position point. The time threshold and the proportion threshold in the delay check rule are dynamically adjusted according to the spatial distribution density of the monitoring position point set; for the region with dense distribution of monitoring position points, the system uses stricter time threshold and proportion threshold to reduce false positives, and for the region with sparse distribution of monitoring position points, the system appropriately relaxes the threshold to avoid excessive sensitivity. This dynamic adjustment mechanism is realized by querying the local point density index of the monitoring position point set. The local point density index is calculated from the K-nearest neighbor average distance of each point in the monitoring position point space. The data verification result of the delay check rule is fed back to the spatial division module. The system records the shift action data and monitoring position point deviation information when the abnormal prompt operation is triggered each time. These records are added to the historical shift record database for subsequent iterative optimization of the spatial division process.

[0029] The joint connection structure of the exoskeleton indirectly participates in the generation of the monitoring position points in the space division process, and the kinematic parameters of the joint connection structure affect the coordinate expression of the historical displacement detection values. The system converts the joint sensor data into the position coordinates of the end effector in three-dimensional space through forward kinematics calculation, and these coordinates constitute the original input of the detection position distribution data. The point cloud data processed by the clustering analysis algorithm implies the constraint of joint motion range, and the boundary of the sub-region of the final detection position partition is associated with the mechanical limit of the exoskeleton workspace. The implementation of the delay check rule depends on the stability of the joint use time length data, and the variability index of the joint use time length data is used to evaluate the reliability of the monitoring position point reference value; when the variability exceeds the safety threshold, the system automatically triggers the re-division process of the monitoring position point set, and the re-division process calls the latest historical displacement record information for space division operation. This closed-loop control mechanism enables the exoskeleton control system to adapt to changes in patient physiological characteristics and shifts in activity patterns, maintaining the safety level of the displacement assistance process.

[0030] Example 2: see Figure 3The operation of finding the intersection of all the detection location partition results starts with the data structure conversion of each independent partition result. The system iterates through each detection location partition result generated by the clustering analysis, and each detection location partition result contains K spatial sub-regions. For each spatial sub-region, the system calculates the geometric center point coordinate of all the data points inside it. The geometric center point coordinate is obtained by calculating the arithmetic mean of all the three-dimensional coordinate points inside the sub-region. The system then calculates the Euclidean distance from each data point in the spatial sub-region to the geometric center point, and selects the two points farthest from the geometric center point as the boundary points. These two boundary points together with the geometric center point form a triplet representing the spatial sub-region. A complete detection location partition result is composed of K such triplets, and the triplets of all the detection location partition results are collected to form a partition triplet set, which serves as the basic data structure for the subsequent similarity evaluation. After the generation of the partition triplet set, the system initiates the pairwise similarity evaluation process of the elements in the set to obtain a difference degree set between the partition results. The similarity evaluation is performed between each pair of detection location partition results. The system selects two different partition results from the partition triplet set in turn, and compares the corresponding index sub-regions of the two partition results one by one. For each pair of index sub-regions, the system extracts their respective triplet representations, i.e., two geometric center point coordinates and two sets of boundary point coordinates. The similarity evaluation calculation is mainly based on two metrics: spatial coverage and shape similarity. The calculation of spatial coverage first converts each triplet into a spatial polyhedron, which is defined by the minimum convex hull of the geometric center point and the two boundary points. The system calculates the ratio of the volume intersection to the volume union of the two spatial polyhedrons, and records this ratio as the spatial coverage score. The calculation of shape similarity focuses on the geometric properties of point cloud distribution. The system performs principal component analysis on all the data points in each sub-region, and extracts the first principal component vector as the main extension direction of the sub-region. The shape similarity between two sub-regions is calculated by comparing the cosine of the angle between their first principal component vectors. The closer the cosine value is to 1, the higher the shape similarity. The system assigns appropriate weight coefficients to the spatial coverage score and the shape similarity cosine value, respectively. The weight coefficients are pre-set according to the contribution of different metrics to the consistency of the partition in historical data. The weighted spatial coverage score and the weighted shape similarity cosine value are added together to obtain the local similarity score of a pair of sub-regions. The above calculation is repeated for all K sub-regions of a detection location partition result, and the arithmetic mean of the K local similarity scores is taken as the overall difference degree value between the two detection location partition results. The overall difference degree value is stored in the difference degree set. The system repeats the above evaluation process for all possible pairwise combinations in the partition triplet set until the difference degree set contains the difference degree measurement values of all pairs of partition results.

[0031] Based on the complete set of difference degrees, the system performs identification and removal of abnormal partition results. The system takes all the difference degree values in the set as a statistical sample, calculates the average and standard deviation of the sample. The system applies an outlier detection algorithm based on standard scores to convert each difference degree value into a standard score; the calculation of the standard score uses the sample mean and sample standard deviation. The system sets a standard score threshold, and any partition result whose converted standard score exceeds the threshold in absolute value is marked as an abnormal association. The system uses a graph theory model for analysis, taking each detection location partition result as a vertex in the graph and the difference degree value between each pair of partition results as the weight of the edge; the system constructs a complete graph containing all detection location partition result vertices and edges with difference degree values as weights. The system finds vertices connected to abnormal association edges in the graph, i.e., those detection location partition results that participate in forming excessively high difference degree values; the system calculates the number of abnormal association edges and the total weight of each vertex, and determines the detection location partition result corresponding to the vertex whose number of abnormal association edges exceeds a pre-set number threshold or whose total weight exceeds a pre-set weight threshold as an abnormal partition result. The system permanently removes these marked abnormal partition results from the current partition result list, and the remaining detection location partition results form a filtered, more consistent subset of partition results.

[0032] For the filtered remaining subset of partition results, the system performs a strict set intersection operation to generate the final detection location partition. The intersection operation is performed in the spatial domain and for the i-th spatial sub-region of all partition results in the remaining subset of partition results. The system converts the i-th sub-region of each remaining partition result from its triple representation to a regularized spatial volume by taking the geometric center point of the triple as the center and the maximum distance from the geometric center point to two boundary points as the semi-axis length to construct a spatial ellipsoid; this spatial ellipsoid is considered as an approximation of the spatial range of the sub-region. The system calculates the geometric intersection of the spatial ellipsoids corresponding to the i-th sub-region of all remaining partition results, and this geometric intersection region is defined as the i-th sub-region of the final detection location partition. The system uses voxel gridding method to calculate the geometric intersection, discretizing the entire three-dimensional workspace into small voxel units; for a given voxel unit, it is only included in the i-th sub-region of the final detection location partition if it is simultaneously located within the spatial ellipsoids of the i-th sub-region of all remaining partition results. The system repeats the above voxelized intersection calculation process for the K sub-regions in turn, and finally obtains a final detection location partition containing K explicit spatial ranges. Each sub-region of the final detection location partition represents a spatial location range that has been verified by multiple historical data and has high reliability, laying a geometric foundation for the accurate allocation of subsequent monitoring location points.

[0033] The construction of the partition triplet set depends on the quality of the preliminary clustering analysis, and the selection of the geometric center point and the boundary point directly affects the expression ability of the triplet to the spatial characteristics of the original sub-region. The weight coefficient in the pairwise similarity evaluation needs to be adjusted according to the specific exoskeleton application scenario, for example, in the displacement activity mode requiring high-precision positioning, the weight of the spatial coverage degree may be set higher. The analysis method of the difference degree set is not limited to outlier detection, and the system can also use a clustering algorithm to group the difference degree values, and determine the partition result belonging to a smaller cluster as abnormal. The generation method of the final detection position partition has flexibility, in addition to the voxel intersection method, the system can also use the convex hull intersection algorithm in computational geometry, specifically, for the i-th sub-region of each remaining partition result, the system first constructs a convex hull based on the triplet representation of the sub-region; the convex hull is the smallest convex polygon containing all data points in the sub-region, which is extended to a convex polyhedron in three-dimensional space. The system generates a convex hull representation for each sub-region using a computational geometry algorithm, which processes the geometric center point and boundary point coordinates in the triplet, and forms the convex hull boundary by connecting the peripheral vertices of the point set. The system calculates the geometric intersection between the convex hulls of the i-th sub-region corresponding to all remaining partition results; the intersection calculation compares the vertex and face information of the convex hulls to determine the overlapping area using the plane cutting method. The resulting intersection region is a new convex hull, representing the i-th sub-region of the final detection position partition. The convex hull intersection algorithm directly processes continuous spatial geometric bodies, avoiding the discretization error of the voxelization method. The entire process ensures that the spatial distribution of the monitoring position points effectively reflects the consensus area of the historical displacement record information, reducing the positioning deviation caused by individual abnormal historical records.

[0034] Referring to Figure 4 The figure directly presents the spatial clustering results of the exoskeleton control system monitoring position points with X, Y coordinates, which are divided into three types of monitoring points: cluster 1, cluster 2, and cluster 3. Each cluster region has a star-shaped marker representing the geometric center point of the cluster sub-region. The system retrieves historical displacement detection values and corresponding position distribution data with physiological feature categories and exoskeleton models as fixed constraints and displacement activity patterns as dynamic constraints; divides these data into multiple spatial sub-regions through the K-means clustering algorithm; and extracts the geometric center point of each sub-region to provide a basis for determining the final monitoring position points for subsequent partition intersection operations. The chart clearly shows the spatial distribution of the three types of monitoring points, and the star-shaped center points directly reflect the representative refinement of the spatial region by clustering analysis, demonstrating the scientific nature of the system in the monitoring position point determination link. Through clustering to mine spatial distribution consensus, the monitoring points cover key displacement areas and avoid redundancy, providing reliable spatial positioning support for subsequent precise control of the exoskeleton.

[0035] Example 3: Dynamic bounding box generation takes a set of partition triplets as input, each partition triplet contains a geometric center point and two boundary points' 3D coordinates. The system constructs an initial bounding box for each partition result, the construction method is to calculate the minimum circumscribed rectangle for all coordinate points contained in the partition result; the minimum circumscribed rectangle is an axis-aligned bounding box, whose range is determined by the minimum and maximum values of all coordinate points on the X, Y, Z axes. This initial bounding box represents the static range of the partition result in space, every point in the set of partition triplets participates in the definition of the bounding box. Introduce the timestamp information of the historical partition data to perform sliding window expansion on the time dimension of the bounding box, the timestamp information records the collection time of the historical displacement record data corresponding to each partition result. The system defines a fixed-length sliding time window, the window length is set according to the frequency of the exoskeleton control system data update. For the partition result being evaluated, the system searches for all adjacent time point partition results in the historical partition data whose timestamps fall within the sliding window; the system fuses the initial bounding boxes of these adjacent time point partition results with the initial bounding box of the current partition result. The fusion operation constructs an expanded bounding box sequence in the time sequence, the system connects the center point coordinates of the initial bounding box of each time point within the sliding window to form a spatial trajectory. The system takes this trajectory as the center line, takes the average size of the initial bounding box as the cross section, and generates a tubular three-dimensional volume, this tubular volume is the dynamic bounding box after the time dimension expansion. The dynamic bounding box embeds the historical spatial change trend, enhances the time sequence continuity of the similarity evaluation.

[0036] The calculation of the three-dimensional spatial overlap rate between the expanded dynamic bounding boxes is a key step in similarity evaluation, and the overlap rate calculation of two dynamic bounding boxes needs to handle their tubular geometry. The system discretizes each tubular volume into a series of dense spatial point sets, and approximates the three-dimensional spatial overlap rate by calculating the ratio of the number of intersection points to the number of union points between two point sets. Specifically, the overlap rate calculation of two dynamic bounding boxes needs to handle their tubular geometry; the system discretizes each tubular volume into a series of dense spatial point sets, and the discretization process uses a uniform grid partitioning method, and the size of the grid cell is set according to the accuracy requirement of the exoskeleton workspace, and each grid cell represents a small spatial region in the tubular volume. The point set generation is realized by spatial sampling in each grid cell, and the sampling density ensures that the point set can adequately approximate the geometric characteristics of the tubular volume; the spatial discretization uses a uniform grid partitioning method, and the size of each grid cell is set according to the accuracy requirement of the exoskeleton workspace. The numerical value of the three-dimensional spatial overlap rate is between 0 and 1, and the closer the numerical value is to 1, the higher the degree of coincidence of the two dynamic bounding boxes in space, and the calculation of the intersection ratio of the point set relies on the spatial index structure for acceleration. At the same time, the system extracts the distribution density change gradient of the coordinate points in each dynamic bounding box, and the coordinate points are derived from all coordinate points in the partition triplet set relied on to generate the dynamic bounding box. The system divides the space inside each dynamic bounding box into uniform three-dimensional grid cells, and counts the number of coordinate points contained in each grid cell to obtain the point density distribution. The calculation of the distribution density change gradient is realized by analyzing the density difference between adjacent grid cells; the system calculates the norm of the density difference value of each grid cell and all its adjacent grid cells, and takes the average value of the density change amplitude of all grid cells as the distribution density change gradient of the dynamic bounding box. The distribution density change gradient represents the uniformity of the point cloud distribution in the bounding box, and a large gradient value indicates that the point cloud distribution has obvious dense and sparse areas.

[0037] The spatial overlap rate and the distribution density change gradient are linearly weighted and fused to generate the final partition result difference value. The purpose of weighted fusion is to balance the contribution of spatial position consistency and internal structure similarity to the overall difference measure. Linear weighted fusion follows the following mathematical relationship:

[0038] Wherein: represents the final partition result difference value; represents the three-dimensional spatial overlap rate between the two dynamic bounding boxes; represents the distribution density change gradient of the coordinate points in the dynamic bounding box; represents the reference gradient value for normalization processing; represents the weight coefficient given to the spatial overlap rate factor; represents the weight coefficient of the factor of the gradient of the distribution density change. The weight coefficient and is a preset positive real number, and the numerical relationship satisfies . In the formula, the is converted into a difference amount, so that the greater the value of the , the greater the difference.

[0039] The generation process of the dynamic bounding box depends on the integrity of the partition triplet set, and the accuracy of the timestamp information directly affects the effectiveness of the sliding window expansion. The calculation accuracy of the three-dimensional space overlap rate is affected by the space discretization density. A finer discretization grid can obtain a more accurate overlap rate estimate. The calculation of the distribution density change gradient reflects the local statistical properties of the point cloud data, and has a positive effect on identifying partition results with similar spatial distribution patterns. The weight coefficient in the linear weighted fusion needs to be adjusted according to the specific application scenario and exoskeleton model. The reference gradient value is selected based on the statistical distribution characteristics of the gradient value in the historical data set. The entire pairwise similarity evaluation process provides a quantitative difference measurement basis for subsequent identification of abnormal partition results. The partition result difference value will be collected into the difference set for subsequent analysis. The method of dynamic bounding box enhances the robustness of similarity evaluation to time series changes, so that the evaluation result not only focuses on the static spatial position, but also considers the dynamic motion pattern. Each coordinate point in the partition triplet set carries spatial information in the historical displacement process. The introduction of the time dimension makes these information more fully utilized in the temporal context. Through this comprehensive evaluation of spatial overlap and internal distribution characteristics, the system can more accurately judge the consistency between different historical partition results, laying the foundation for forming high-quality final detection position partition. The calculation formula of the difference value ensures the consistency of the dimension. The setting of the weight coefficients and needs to be calibrated according to a large amount of historical data to optimize the discrimination effect of similarity evaluation. The generation method of the tubular dynamic bounding box can effectively capture the motion trajectory features in the displacement process. The granularity of three-dimensional grid division needs to balance the demand between calculation complexity and evaluation accuracy. The calculation of the distribution density change gradient uses the twenty-six-neighbor connection relationship, which can fully reflect the change of point cloud distribution in all directions. The introduction of the reference gradient value makes the historical data of different batches and different patients comparable, enhancing the adaptability and generalization ability of the algorithm. Finally, the partition result difference value As a comprehensive quantitative index, it provides a scientific basis for subsequent abnormal partition result identification and filtering, thereby improving the safety and reliability of the entire exoskeleton control system during patient displacement. The entire pairwise similarity evaluation process realizes comprehensive analysis of historical partition results in multiple dimensions and multiple features, ensuring the accuracy and reliability of the final detection position partition.

[0040] Referring to Figure 5 The figure presents the spatial overlap degree of the dynamic bounding box of the exoskeleton control system at different time steps through green column charts with time steps as the horizontal axis and bounding box overlap rate as the vertical axis. The system generates dynamic bounding boxes based on the partition triple set, introduces timestamp information of historical partition data for sliding window expansion, and then calculates the three-dimensional spatial overlap rate of the expanded bounding box, combined with the distribution density change gradient, to finally obtain the partition result difference. The change of overlap rate in the figure reflects the time sequence consistency of the space region in the displacement process. The higher the overlap rate, the stronger the spatial coincidence degree of the displacement region at the corresponding time step, and the better the consistency of the partition result; otherwise, it indicates that the region has dynamic changes. This quantitative result provides a key basis for subsequent identification of abnormal partition and optimization of the final detection position partition, ensuring that the monitoring position point of the exoskeleton control system can accurately capture the consensus space of historical displacement, and laying a data foundation for improving the spatial accuracy and time robustness of exoskeleton control.

[0041] Example 4: Establishment of control reference unit Based on joint usage time record data, which is derived from the historical running log of each joint sensor of the exoskeleton, including the cumulative working time data of the main motion units of the hip joint, knee joint and ankle joint. The system collects control characteristic value record data set of the preset monitoring position point conforming to the exoskeleton model and displacement activity mode, and the collection process carries out data screening according to the physiological characteristics of the patient, ensuring the high relevance of the data to the current application scenario. The control characteristic value record data set contains the coordinate information, joint angle data and corresponding control instruction sequence of all monitoring position points in the historical record under a specific exoskeleton model and displacement activity mode.

[0042] The consistency analysis of the control feature value record dataset is a key step to obtain reliable training data, and the consistency analysis aims to identify and process noisy data, outliers, and mutually contradictory records in the dataset. The system adopts a multi-stage consistency verification process, the first stage performs basic integrity checking, and the records with missing data are excluded. The second stage performs range rationality checking, based on the mechanical design parameters of the exoskeleton and the principles of ergonomics, the reasonable value range of the joint use time and the monitoring position point coordinates is set, and the records beyond the reasonable range are marked as abnormal. The third stage performs logical consistency checking, which verifies the logical rationality between the joint use time and the monitoring position point coordinates by using the kinematic constraint relationship of the joint, such as checking whether there are records of joint angle data and end effector position mismatch. Referring to Table 1, a simplified control feature value consistency analysis is shown, which illustrates how the system handles different types of inconsistency.

[0043] Table 1: Control feature value consistency analysis table

[0044] According to the joint use time record data and the control feature value real data, the system uses a supervised learning algorithm to train the control reference unit. The control reference unit adopts a multi-layer feedforward neural network structure, the number of input layer nodes corresponds to the number of joint use time features, and the number of output layer nodes corresponds to the coordinate dimension of the monitoring position point. The number of layers and nodes of the hidden layer is configured and adjusted according to the complexity of the dataset and the training effect, and the network activation function uses the ReLU function to enhance the nonlinear fitting ability. The training process uses the back propagation algorithm to optimize the network weights, and the optimization goal is to minimize the loss function value. The supervised learning algorithm for training the control reference unit includes defining the loss function, which optimizes the training process by calculating the difference between the predicted reference value and the real reference value of each monitoring position point. The loss function is defined as the sum of the squares of the difference between the predicted reference value and the real reference value, and the difference is accumulated and summed based on the deviation amplitude of the predicted value and the real value. For a single training sample, the loss function calculation covers the prediction error of all coordinate components of the monitoring position point, and the error contribution of each monitoring position point is weighted according to its importance in the displacement control. The training process uses the mini-batch gradient descent method, and the batch size is set according to the computing resources and the requirements for training stability, and the learning rate uses an adaptive adjustment strategy to improve the convergence efficiency.

[0045] The training data of the control reference unit needs to cover various typical shift activity patterns, including bed-to-chair transfer, sit-to-stand transition, and walking on flat ground in different scenarios. The training data set is divided into training set, validation set and test set in chronological order, the training set is used for model weight update, the validation set is used for hyperparameter tuning and preventing overfitting, and the test set is used for final evaluation of the generalization performance of the control reference unit. The system monitors the change of the loss function value on the training set and the validation set during training, when the loss function value of the validation set does not decrease for consecutive multiple training periods, the training is terminated in advance to prevent overfitting. The training quality of the control reference unit depends on the consistency level of the control feature value record data set, and the high-quality data selected in the consistency analysis stage provides a reliable basis for supervised learning. The design of the loss function directly affects the learning direction of the control reference unit, and the sum of squares error function provides a clear gradient direction for network parameter optimization. The choice of supervised learning algorithm takes into account the nonlinear characteristics of exoskeleton control problems, and multilayer feedforward neural networks can effectively learn the complex mapping relationship between joint use time and monitored position points. The hyperparameters such as learning rate and batch size need to be determined through systematic experiments to obtain the best training effect.

[0046] After the training of the control reference unit is completed, the system saves the trained neural network model parameters and structure information for subsequent real-time control reference value generation. The control reference unit receives real-time joint use time data as input during the deployment phase, and outputs predicted reference values for monitored position points through forward network calculation. These predicted reference values will be compared with actual sensor readings as the basis for shift action anomaly detection. Periodic updating of the control reference unit is necessary, and when enough new shift record data is accumulated, the system can start the retraining process to update the network parameters with new data, so that the control reference unit adapts to changes in the physiological state of the patient or the wear of the exoskeleton device. The definition of the loss function determines the sensitivity of the control reference unit to different types of prediction errors, and the sum of squares error function will impose heavier penalties on larger prediction errors. The representativeness of the training data has a decisive influence on the generalization ability of the control reference unit, and the data set needs to fully cover various possible working conditions and boundary conditions. The complexity of the network structure needs to match the size of the available training data, and an overly complex network structure is prone to overfitting on a small data set. Monitoring the training process is an important step to ensure model quality, and the trend of the loss function on the training set and the validation set can reflect the learning state and learning effect of the model. The final performance of the control reference unit needs to be objectively evaluated by an independent test set, and the evaluation indicators include average error, maximum error and error distribution characteristics.

[0047] Example 5: Variance calculation is performed on multiple joint usage time data sequences collected within a specific time window, with each joint corresponding to an independent data sequence. The system calculates the statistical variability of each joint usage time sequence, quantified by the ratio of the standard deviation to the mean. For a record containing n displacement operations, the hip joint usage time sequence is {h1, h2, ..., h...} n The knee joint usage time sequence is {k1,k2,...,k}. n The system first calculates the average value of each sequence. and standard deviation Then calculate the variability of each joint. The maximum value of all joint variability is determined as the overall variability index, reflecting the stability level of multi-joint collaborative operation. Based on the mapping relationship between the magnitude of variability and the number of integrated control reference units, the system determines the composition scale of the control reference model. The overall variability index is divided into multiple discrete levels, each corresponding to a predefined number of integrations. For example, when the overall variability is below threshold T1, the number of integrations is set to 3; when the overall variability is between thresholds T1 and T2, the number of integrations is set to 5; and when the overall variability is above threshold T2, the number of integrations is set to 7. This hierarchical mapping relationship is established by analyzing the correlation patterns between variability and model prediction errors in historical data. Increasing the number of integrations aims to reduce the uncertainty of a single model under high variability conditions through model fusion.

[0048] Based on a predetermined number of ensembles, the system selects multiple control reference units (CLUs) from the trained pool to form an ensemble. The selection process is based on the performance ranking of the CLUs on the validation set. The system evaluates the prediction accuracy of each CLU on historical shift data and selects a specified number of top-ranked CLUs. The performance evaluation of the CLUs uses the root mean square error (RMSE) metric; a lower RMSE value indicates higher prediction accuracy. The system ensures the selected CLUs are diverse, meaning they may produce slightly different outputs under the same input conditions. This diversity is achieved by using different subsets of training data or different network structures. The average output of multiple CLUs is calculated to construct a control reference model. For a given joint usage duration input vector X, each selected CLU outputs a predicted value for a monitored location point. The final output Y of the control reference model is obtained by calculating the arithmetic mean of the outputs of all selected control reference units: ,in The number of integration. This average fusion method helps to smooth the abnormal prediction of individual control reference units and improve the robustness of the overall output. The control reference model integrates the prediction ability of multiple control reference units into a unified prediction framework, providing reliable monitoring position point reference values for real-time shift operations. The preprocessing of historical shift record information is the data preparation stage of the control reference model construction, which includes two main links: data cleaning and format standardization. The data cleaning stage aims to identify and correct incomplete, incorrect or inconsistent data in the historical shift record information. The system first detects missing values, for isolated missing data points in the joint use duration sequence, linear interpolation is used to fill in the adjacent data before and after; for continuous missing fragments, mark the record as incomplete and exclude it. The statistical method is used for outlier detection, and the Z-score of each data field is calculated. Values with absolute values exceeding 3 are considered outliers. For the identified outliers, the system uses robust statistics to replace them according to the historical distribution characteristics of the field. The logical consistency check verifies the correlation between different fields, such as the matching degree of joint use duration and shift activity pattern.

[0049] The format standardization stage ensures that all historical shift record information follows a unified data specification and structure. Timestamp information is uniformly converted to the ISO 8601 standard format, including date, time, and time zone information. Joint use duration data is uniformly converted to hours, with two decimal places. Monitoring location point coordinates are uniformly represented in the same coordinate system, with the origin defined as the center point of the exoskeleton base and the coordinate axis direction following the right-hand rule. The encoding format of all data fields is uniformly UTF-8, ensuring proper handling of special characters. The field names, data types, and unit systems of the data table remain consistent throughout the dataset, facilitating subsequent analysis and processing. Take a specific shift scenario as an example to illustrate the operation mechanism of the control reference model: a 70 kg, 175 cm tall male patient needs to be shifted from a wheelchair to a bed, and the exoskeleton model is REX-02, and the shift activity mode is "sit-to-stand transfer". The system first queries the joint use duration records of the patient's last 50 sit-to-stand transfer operations, calculates the variability of the hip, knee, and ankle joint use duration, and assumes that the overall variability is 0.15. According to the preset mapping rules, the variability level corresponds to an integration number of 5. The system selects 5 control reference units from the control reference unit pool that perform best in the sit-to-stand transfer mode. These 5 control reference units are trained based on historical data from different periods and have slightly different internal parameters. When the patient starts a new shift operation, the real-time collected joint use duration data is input into these 5 control reference units, and each unit outputs a monitoring location point prediction value. The control reference model calculates the average of these 5 prediction values as the reference value of the monitoring location point at the current time. This reference value will be compared with the actual position monitored by the sensor in real time. If the deviation exceeds the safety threshold, the system will start the delay check rule for further verification.

[0050] The specific operation of the data cleaning link can be illustrated by an example: in a historical record, the hip joint usage time is recorded as-2.5 hours, which is obviously an abnormal value. The data cleaning process will identify this abnormality and replace it with the median value of 12.3 hours according to the distribution characteristics of other normal records of the patient. At the same time, the monitoring position point X coordinate in the record is missing, and the system will check the data of other sensors in the same displacement operation, and calculate the reasonable coordinate value through the kinematics model to fill in. The format standardization stage will convert all the time information in the record into the "YYYY-MM-DDHH: MM: SS+08:00" format, ensuring the correctness of the time series analysis. The performance of the control reference model depends on the quality of the preprocessed data, and the accuracy of the variability calculation directly affects the determination of the number of integrations. Each step in the preprocessing process aims to improve the completeness, consistency and standardization of the historical displacement record information, and provides a reliable data basis for the construction of the control reference model. The output average value fusion method of the control reference unit can effectively reduce the influence of random errors and improve the stability of the monitoring position point reference value. The continuous accumulation and update of the historical displacement record information enable the control reference model to adapt to the changes in the physiological state of the patient and the development of the device usage mode, and maintain the prediction accuracy. The implementation of the entire embodiment shows a complete process from data preparation to model construction, which ensures the reliability and safety of the exoskeleton control system during the patient's displacement process through systematic processing methods. The use of variability indicators enables the model structure to be dynamically adjusted according to the data characteristics, the preprocessing process guarantees the basic quality of data analysis, and the integration strategy of the control reference model improves the robustness of the prediction output.

[0051] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An exoskeleton control method for patient mobilization, characterized by, The method is realized by the following steps: According to the physiological characteristic category and the displacement activity mode of the patient, historical displacement record information is queried for spatial division to obtain a monitoring position point; A neural network simulation is performed using the joint connection structure of the exoskeleton to establish a control reference unit, wherein the input of the control reference unit is the joint use duration, and the output is a control reference value of the monitoring position point; Based on the variability of the joint use duration, the average value of the outputs of multiple control reference units is fused to construct a control reference model, which processes the multiple joint use durations to generate a monitoring position point reference value; When it is detected that the displacement action does not match the monitoring position point reference value, a delay check rule is implemented for data verification.

2. The exoskeleton control method for patient mobilization of claim 1, wherein, The implementation of the delay check rule for data verification includes: defining a delay check rule, wherein the continuous time rule indicates that the duration of the inconsistent state exceeds a time threshold, and the data proportion rule indicates that the data amount of the inconsistent state accounts for more than a proportion threshold in the total data amount; when any condition in the continuous time rule or the data proportion rule is met, an abnormal prompt operation is performed, otherwise the displacement action continues to be monitored.

3. The exoskeleton control method for patient mobilization of claim 1, wherein, The step of querying historical displacement record information for spatial division to obtain a monitoring position point includes: taking the physiological characteristic category and the exoskeleton model of the patient as fixed constraints, and taking the displacement activity mode as a dynamic constraint, collecting multiple historical displacement detection values and corresponding detection position distribution data; for each historical displacement detection value, the detection position distribution data is analyzed according to the detection value to obtain a detection position partition result; the intersection of all detection position partition results is obtained to obtain a final detection position partition; for each sub-region of the final detection position partition, a monitoring position point is assigned to the sub-region and added to a monitoring position point set.

4. The exoskeleton control method for patient mobilization of claim 3, wherein, The intersection of all detection position partition results includes: for each detection position partition result, the center point of each partition and the two boundary points farthest from the center point are extracted to construct a three-tuple representation of the partition to form a partition three-tuple set; similarity evaluation is performed on the partition three-tuple set to obtain a difference degree set between the partition results; based on the difference degree set, abnormal partition results are removed, and the intersection of the remaining partition results is obtained to obtain the final detection position partition.

5. An exoskeleton control method for patient mobilization as claimed in claim 4, wherein, The pairwise similarity evaluation includes: generating a dynamic bounding box based on the partition three-tuple set; by calculating the weighted combination of the spatial overlap degree and the time series change trend of the bounding box, a comprehensive difference measure between the partition results is obtained; the specific implementation steps include: generating a minimum circumscribed rectangle as an initial bounding box according to the coordinate point set of each three-tuple; the timestamp information of the historical partition data is introduced to perform sliding window expansion of the bounding box in the time dimension; the three-dimensional spatial overlap rate between the expanded bounding boxes is calculated; the distribution density change gradient of the coordinate points in each bounding box is extracted; the spatial overlap rate and the distribution density change gradient are linearly weighted and fused to generate a final partition result difference degree value.

6. The exoskeleton control method for patient mobilization of claim 1, wherein, The exoskeleton joint connection structure is used for neural network simulation, and the control reference unit is established, including: based on the joint use time length record data, collecting control characteristic value record data set of preset monitoring position points conforming to the exoskeleton model and the displacement activity mode; performing consistency analysis on the control characteristic value record data set to obtain control characteristic value real data; and training the control reference unit by using a supervised learning algorithm according to the joint use time length record data and the control characteristic value real data.

7. The exoskeleton control method for patient mobilization of claim 1, wherein, The control reference model is constructed based on the variability of the plurality of joint use time lengths and the output average values of the plurality of control reference units, including: determining the integrated number of control reference units according to the size of the variability; selecting a plurality of control reference units according to the integrated number, calculating the output average values of the plurality of control reference units, and constructing the control reference model.

8. The exoskeleton control method for patient mobilization of claim 6, wherein, The control reference unit is trained by using a supervised learning algorithm, including defining a loss function, the loss function optimizing the training process by calculating the difference between the predicted reference value and the real reference value of each monitoring position point, wherein the difference is accumulated by summing the deviation amplitude of the predicted value and the real value.

9. The exoskeleton control method for patient mobilization of claim 1, wherein, The method further includes a preprocessing step of historical displacement record information, and the preprocessing includes data cleaning and format standardization.

10. An exoskeleton control system for patient mobilization, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor, when executing the computer program, realizes the steps of the exoskeleton control method for patient displacement according to any one of claims 1 to 9.

Images