Search

CN-121512502-B - Method, equipment and storage medium for identifying rehabilitation actions after knee joint injury operation

CN121512502BCN 121512502 BCN121512502 BCN 121512502BCN-121512502-B

Abstract

The invention relates to the field of physiological motion detection in medical instruments, in particular to a method, equipment and storage medium for identifying rehabilitation actions after knee joint injury operation. The method comprises the steps of obtaining multi-source sensor data, preprocessing to generate a preprocessing sequence, analyzing the gesture and gait phase based on the preprocessing sequence to generate a gesture sequence and a phase sequence, then carrying out segment boundary detection and motion candidate generation, extracting time sequence characteristics, carrying out hierarchical motion map matching and sequence discrimination model reasoning to generate motion category and completion degree scoring, and finally carrying out quality evaluation and safety threshold linkage comparison and executing strategy mapping to generate a parameter updating package. The invention can realize accurate identification and quantitative evaluation of rehabilitation actions, form closed loop feedback and effectively improve the safety and individuation level of rehabilitation training.

Inventors

  • WANG WENJUAN
  • MA NING
  • LI YANAN

Assignees

  • 中国人民解放军总医院第四医学中心

Dates

Publication Date
20260508
Application Date
20251114

Claims (8)

  1. 1. The method for identifying the rehabilitation actions after the knee joint injury operation is characterized by comprising the following steps of: S100, acquiring wearable inertia, optional image or pressure data and associating time, performing acquisition channel registration, data integrity verification and time stamp standardization, performing static zero calibration, limb segment parameter estimation and binding configuration, and generating a preprocessing sequence; s200, based on the preprocessing sequence, carrying out gesture observation extraction and drift suppression processing, carrying out skeleton calculation comprising individual parameters and alignment indexes, carrying out gait phase recognition processing, and generating a gesture sequence and a phase sequence comprising a time stamp, a fragment number, a phase label and a confidence level; S310, acquiring a gesture sequence and a phase sequence, and executing segment boundary detection and action candidate generation processing containing uniform time base and phase coherent constraint to obtain an action candidate set; S320, extracting time sequence features comprising intra-segment duration, joint angle peak-valley positions, phase stay ratios, key event sequence and span connection relation from the action candidate set, performing hierarchical action pattern matching which follows a top-down sequence and checks with necessary events according to node sequence constraint based on the action candidate set and the time sequence features, and generating a time-ordered action matching sequence; Loading a hierarchical action map facing a rehabilitation stage, wherein the hierarchical action map refers to a multi-layer node structure split according to a rehabilitation plan, the upper layer is a stage node, the middle layer is an action category node, and the bottom layer is an action instance template; the matching process adopts a top-down sequence, namely screening the combination of the characteristics among candidate segments meeting the sequence constraint in the stage nodes, comparing the characteristics among the segments with the necessary events in the action category node level, returning the candidate segments which do not meet the necessary events to the original position for merging the adjacent segments; S330, based on the action matching sequence, performing sequence discrimination model reasoning comprising context window, inter-segment dependency gating and abnormal duty ratio suppression, and generating action category and completion degree score; Taking the action matching sequence as input, synchronously referencing the intra-segment features and the candidate inter-segment features reserved in the action candidate set as context supplement, and internally setting three sub-units of context window, inter-segment dependency gate control and abnormal duty ratio suppression; in the running process, firstly rearranging candidates in a context window, wherein the rearranging rule follows the replaceable sequence and the jump tolerance allowed by the stage node, then checking the necessary bearing relation between adjacent candidates in inter-section dependency gating, and triggering backtracking merging if the bearing relation is missing; the abnormal duty cycle suppresses the interval for processing the candidate segment confidence level to be low, and when the abnormal duty cycle exceeds the window threshold, the window output is set to be pending; After window level reasoning is completed, the model performs full-order check in the whole range, and local output which does not meet the full order is degraded; S400, obtaining action category and completion degree score, performing quality weight calculation and grading treatment comprising weight factor table aggregation recognition confidence, phase coherence and structure checking elements, extracting key indexes of completion degree track fluctuation amplitude, phase stay time distribution and joint angle envelope deviation, performing linkage comparison with a safety threshold, selecting strategy mapping treatment of action level and system level strategies from a strategy library, and generating a parameter updating packet.
  2. 2. The method of claim 1, wherein the process of segment boundary detection and action candidate generation processing further comprises: The method comprises the steps of carrying out preliminary screening on key turning points in the same segment according to the sequence relation and time interval of occurrence of phase labels, carrying out segment boundary detection on local extremum, change rate mutation points and phase boundary points of knee joint angles which change along with time, recording suspicious boundaries and rechecking when the adjacent key turning point interval exceeds a segment threshold or quality marking prompting interference accumulation, carrying out neighborhood consistency check on suspicious boundaries when the adjacent key turning point interval exceeds a segment threshold or quality marking prompting interference accumulation, carrying out neighbor consistency check on suspicious boundaries, registering non-passing person rollbacks as common turning points, carrying out action candidate generation processing after boundary set construction, forming candidate segments between adjacent segment boundaries, carrying start-stop time, trigger phase, segment numbers and boundary source marks on each segment, and extracting three basic attributes of angle envelopes, angle change sequences and phase residence time in the segment from the gesture sequence.
  3. 3. The method of claim 1, wherein the act candidate generation process further comprises: and if the segment session is interrupted, the generated candidate segment and the first segment of the next segment are tried to be spliced across segments, and when the splicing fails, the segments are respectively reserved and the segmentation marks are added on the candidate segment.
  4. 4. The method of claim 1, wherein extracting timing characteristics including intra-segment duration, joint angle peak-to-valley position, phase dwell ratio, critical event sequencing, and span connection relationship further comprises: extracting computable elements describing the sequence relation between the interior of the candidate segment and the candidate segment, and covering the connection relation between the duration time, the joint angle peak-valley position, the phase residence proportion, the sequence of key events and the span segment; When the segmentation mark exists, the intra-segment features are calculated only in the range of not crossing the mark, and the cross-mark relation is additionally registered as the candidate inter-segment features; loading a hierarchical action map facing to a rehabilitation stage, wherein the hierarchical action map refers to a multi-layer node structure split according to a rehabilitation plan, the upper layer is a stage node, the middle layer is an action category node, and the bottom layer is an action instance template, each node is provided with a sequential constraint, a necessary event, an optional event and an allowable deviation interval, the matching process adopts a top-down sequence, namely screening the combination of features among candidate segments meeting the sequential constraint in the stage node, comparing the features with the necessary event in the stage node layer, and returning the candidate segments which do not meet the necessary event to the original position for merging adjacent segments; After the matching is completed, generating a time-ordered action matching sequence which consists of a candidate segment identifier and an action category identifier and comprises a matching confidence level and a node path record.
  5. 5. The method of claim 1, wherein the process of hierarchical action pattern matching further comprises: And reconstructing the local sequence relation by adopting a neighboring segment supplementing strategy, if the reconstruction fails, maintaining an unassigned state in a matching result and recording a reason code.
  6. 6. The method of claim 1, wherein generating an action category and completion score further comprises: The completion degree score is derived from the coincidence degree of the features in the segments and the example templates and the fitting degree of the features between the segments and the stage constraint, and the two are synthesized according to the internal weight of the model.
  7. 7. A knee joint injury postoperative rehabilitation motion recognition device applied to the method of any one of claims 1 to 6, comprising: The acquisition and time correlation module is used for acquiring wearable inertial sensor data, image data or pressure data and correlating time, generating original multi-source data and data acquisition configuration, and providing the original multi-source data and data acquisition configuration to the static zero calibration and limb segment parameter estimation module and the binding configuration and preprocessing module; The static zero calibration and limb segment parameter estimation module is used for extracting a static standing segment from the original multi-source data, completing static zero calibration and limb segment parameter estimation, outputting individual parameters and alignment indexes and providing the individual parameters and alignment indexes to the binding configuration and preprocessing module and the skeleton resolving and joint angle generating module; the binding configuration and preprocessing module is used for carrying out binding configuration on the individual parameters and the alignment index, generating a preprocessing sequence and providing the preprocessing sequence to the gesture observation extraction and drift suppression module; the gesture observation extraction and drift suppression module is used for constructing a gesture observation sequence from the preprocessing sequence and providing the gesture observation sequence to the skeleton calculation and joint angle generation module; The skeleton resolving and joint angle generating module is used for extracting joint angle candidates from the gesture observation sequence, performing skeleton resolving based on the individual parameters and the alignment index, outputting a gesture sequence and a knee joint angle sequence and providing the gesture sequence and the knee joint angle sequence to the gait phase recognition and action candidate generating module; The gait phase recognition and motion candidate generation module is used for carrying out periodic segmentation and gait phase recognition on the knee joint angle sequence, completing segment boundary detection and motion candidate generation after acquiring the gesture sequence, outputting a motion candidate set and providing the motion candidate set to the hierarchical motion map matching and sequence discrimination model module; The hierarchical action pattern matching and sequence judging model module is used for extracting time sequence characteristics from the action candidate set, completing matching with the hierarchical action pattern facing the rehabilitation stage and carrying out sequence judging model reasoning, generating action category and completion degree scores and providing the action category and completion degree scores to the quality evaluation and threshold linkage alarm and strategy write-back module; The quality evaluation and threshold linkage alarm and strategy write-back module is used for calculating quality weight according to action category and completion degree score, grading data, performing linkage comparison with a safety threshold to generate abnormal prompt, completing strategy mapping and training suggestion generation processing, forming a parameter update package and writing back to data acquisition configuration.
  8. 8. A computer readable storage medium, wherein a computer program is stored on the storage medium, and the computer program is called and executed by a computer to implement the method for identifying a rehabilitation motion after a knee joint injury operation according to any one of claims 1 to 6.

Description

Method, equipment and storage medium for identifying rehabilitation actions after knee joint injury operation Technical Field The invention relates to the field of physiological motion detection in medical instruments, in particular to a method, equipment and storage medium for identifying rehabilitation actions after knee joint injury operation. Background In the field of physiological motion detection in medical instruments, the existing scheme of rehabilitation motion recognition after knee joint injury operation generally acquires data around wearable inertia, images or pressure channels, establishes a unified time base through time correlation, performs static zero calibration and limb segment parameter estimation, then performs skeleton calculation and gait phase recognition, and attempts segment boundary detection and motion map-based matching on the basis, so as to further provide training process records. The scheme has the limitations of Ji Suoyin on the instability of multi-source data in time sequence and source mapping, individual parameter deviation caused by unclear coupling of static zero calibration and limb segment parameter estimation, insufficient consistency of gesture sequences and knee joint angle sequences caused by skeleton calculation and gait phase identification and cleavage, and the like. In the existing method, the generation and matching of action candidates are finished by relying on single channel threshold values or experience rules, the constraint of a hierarchical action map for a rehabilitation stage is incomplete, the sequence discrimination model reasoning lacks joint processing of intra-segment features and candidate inter-segment features, the quality weight calculation and safety threshold linkage comparison lacks a write-back channel mapped with a strategy, and a parameter update package does not form a closed loop penetrating through data acquisition configuration. Under the conditions of data acquisition configuration constraint and household training, segment boundary detection is sensitive to noise and phase sequence label drift, and continuous processing of inputting a stable support action matching sequence into a sequence discrimination model reasoning is difficult to realize. Aiming at how to generate joint processing of action category and completion degree scoring under data acquisition configuration based on action matching sequences and reasoning by a sequence discrimination model, the prior art generally has link disjoint in links such as multisource fusion, time alignment, rule constraint, strategy write-back and the like, and is difficult to form a consistent flow of acquisition, alignment, judgment, control and recording in a rehabilitation training application scene, so that reusability and traceability of the action category and the completion degree scoring in cross-segment and cross-stage are insufficient. Disclosure of Invention In order to solve the technical problems, the invention provides a method for identifying rehabilitation actions after knee joint injury operation, which comprises the following steps: Acquiring wearable inertia, optional image or pressure data and associating time, executing acquisition channel registration, data integrity verification and time stamp standardization processing, performing static zero calibration, limb segment parameter estimation and binding configuration processing, and generating a preprocessing sequence; Based on the preprocessing sequence, carrying out gesture observation extraction and drift suppression processing, executing skeleton calculation containing individual parameters and alignment indexes, carrying out gait phase recognition processing, and generating a gesture sequence and a phase sequence; Extracting time sequence characteristics comprising intra-segment duration, joint angle peak-valley position, phase stay proportion, key event sequence and span connection relation, performing hierarchical action map matching which follows top-down sequence and checks with necessary events according to node sequence constraint, and sequence discrimination model reasoning comprising context window, inter-segment dependent gating and abnormal duty ratio suppression, and generating action category and completion degree score; Obtaining action category and completion degree score, executing quality weight calculation and grading treatment comprising weight factor table aggregation recognition confidence, phase coherence and structure checking elements, extracting key indexes of completion degree track fluctuation amplitude, phase stay time distribution and joint angle envelope deviation, performing linkage comparison with a safety threshold, selecting strategy mapping treatment of action level and system level strategy from a strategy library, and generating a parameter updating package. Further, the process of segment boundary detection and motion candidate generation processing further includes: The m