A psychological abnormality detection method and system based on behavior timing pattern recognition

Abstract

The application discloses a kind of psychological abnormality detection method and system based on behavior timing pattern recognition, it is related to psychological abnormality detection technical field.The method comprises: obtaining the behavior action atomic sequence of subject executing standardization behavior task, time window segmentation is carried out to sequence to obtain behavior characteristic fragment and constructs directed timing association graph, structural matching is carried out to timing association mode in graph with abnormal feature map, according to matching degree calculation behavior abnormality index and obtain psychological abnormality detection result.The application is by behavior action atomic disintegration and directed timing association graph construction, complete depiction behavior whole timing association and internal organization law, guarantee detection result real reflection subject psychological behavior state;And by the cooperation of multiscale sliding time window and timing transition probability, behavior characteristics are comprehensively covered, so as to screen stable association and eliminate random interference, effectively improve the accuracy of psychological abnormality detection.

Description

Technical Field

[0001] This invention relates to the field of psychological anomaly detection technology, and in particular to a psychological anomaly detection method and system based on behavioral temporal pattern recognition. Background Technology

[0002] In the field of behavioral-assisted detection of psychological abnormalities, existing technologies mainly collect behavioral image data of subjects when they complete standardized behavioral tasks, extract quantitative feature parameters of single behavioral actions, compare each parameter with pre-constructed benchmark data of normal populations, and comprehensively judge whether there is an abnormality in the behavior based on the degree to which the parameters deviate from the benchmark range.

[0003] In actual clinical testing scenarios, the quantitative parameters of individual actions of some subjects are within the reasonable range of normal populations, but their overall behavioral coherence, action connection mode and behavioral rhythm are significantly different from normal subjects. Existing technology only analyzes and judges the independent characteristics of individual actions, and cannot fully capture and characterize the overall coherence characteristics and internal correlation of behavior, thus the test results cannot truly reflect the psychological and behavioral state of the subjects. Summary of the Invention

[0004] This invention provides a method and system for detecting psychological anomalies based on behavioral temporal pattern recognition, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a psychological anomaly detection method based on behavioral temporal pattern recognition, comprising: S1. Obtain the atomic sequence of the subject's behavioral actions when performing a standardized behavioral task; S2. Perform time window segmentation on the atomic sequence of behavioral actions to obtain multiple behavioral feature segments, and construct a directed temporal correlation graph for each behavioral feature segment; S3. Perform graph matching of the temporal association patterns extracted from the directed temporal association graph with the abnormal feature graph, and calculate the behavioral abnormality index based on the matching degree to obtain the psychological abnormality detection results.

[0006] Preferably, the atomic sequence of behavioral actions of the subject when performing a standardized behavioral task is obtained, including: Extracting the motion trajectory of skeletal points from videos of subjects' behavior; By detecting local extreme points of motion velocity in the motion trajectory of skeletal points, the motion trajectory of skeletal points is segmented to obtain a series of indivisible behavioral action atoms. The action atoms are arranged in order of timestamps to construct a sequence of action atoms.

[0007] Preferably, the atomic sequence of behavioral actions is segmented by time windows to obtain multiple behavioral feature segments, including: On the atomic sequence of behavior actions, multiple multi-scale sliding time windows with different time spans and a preset overlap rate between adjacent windows are set; Based on the window length of the multi-scale sliding time window and the preset overlap rate, sliding acquisition is performed on the atomic sequence of behavior actions to obtain multiple behavioral feature fragments with overlapping parts.

[0008] Preferably, a directed temporal association graph is constructed for each behavioral feature segment, including: Select behavioral action atoms within behavioral feature segments as candidate nodes; Calculate the transition probability between any two action atoms within a behavioral feature segment, determined by their temporal sequence; Based on the comparison between the transition probability and the adaptive probability threshold, directed edges are established between pairs of behavioral actions to construct a directed temporal association graph.

[0009] Preferably, the method for calculating the transition probability includes: Calculate the co-occurrence probability of two action atoms appearing in the form of consecutive adjacent occurrences in all action feature segments, and use the reciprocal of the time interval between the two action atoms as the temporal proximity. The transition probability is obtained by weighting and fusing co-occurrence frequency with temporal proximity.

[0010] Preferably, based on the comparison result between the transition probability and the adaptive probability threshold, directed edges are established between pairs of action atoms, including: Based on the distribution of the transition probabilities of all behavioral action atom pairs within the behavioral feature segment, an adaptive probability threshold is dynamically calculated. When the transition probability exceeds the adaptive probability threshold, a directed edge is established between the two corresponding action atoms.

[0011] Preferably, the method for obtaining the anomaly feature map includes: Obtain a directed temporal correlation graph of the historical data of a group of subjects with psychological abnormalities; Subgraph mining is performed on the historical directed temporal association graph to extract common temporal association substructures that appear more than a preset threshold number of times, and these are used as anomaly feature graphs.

[0012] Preferably, the temporal association patterns extracted from the directed temporal association graph are subjected to structural graph matching with the anomaly feature graph, including: The directed temporal correlation graph is decomposed into multiple local temporal subgraphs, and the anomaly feature graph is decomposed into multiple anomaly subgraph templates. In the local temporal subgraph, the search finds matching instances that satisfy the isomorphic constraint in terms of both node type and temporal dependency strength attribute represented by directed edges; The matching degree is calculated by weighting the number of matching instances and the structural similarity between the matching instances and the corresponding abnormal subgraph templates.

[0013] Preferably, a behavioral abnormality index is calculated based on the matching degree to obtain the psychological abnormality detection result, including: Each matching degree is associated with the time window position of the behavioral feature segment that generated the match, resulting in an instantaneous anomaly score sequence distributed along the time axis; The behavioral anomaly index is obtained by summing the instantaneous anomaly score sequence using a time decay function with a preset time window. The contribution of earlier instantaneous anomaly scores to the current behavioral anomaly index decreases as the time distance between them and the current moment increases. The behavioral abnormality index is compared with the clinical diagnostic threshold; if it exceeds the threshold, it is considered a psychological abnormality.

[0014] To address the aforementioned problems, this invention also provides a psychological anomaly detection system based on behavioral temporal pattern recognition, the system comprising: The sequence acquisition module is used to acquire the atomic sequences of the subjects' behavioral actions when performing standardized behavioral tasks; The association graph construction module is used to perform time window segmentation on the atomic sequence of behavior actions to obtain multiple behavior feature segments, and to construct a directed temporal association graph for each behavior feature segment; The graph structure matching evaluation module is used to perform graph structure matching between the temporal association patterns extracted from the directed temporal association graph and the abnormal feature graph, and calculate the behavioral abnormality index based on the matching degree to obtain the psychological abnormality detection result.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs an atomic sequence of behavioral actions and generates a directed temporal correlation graph, which can fully characterize the temporal correlation features of behavior, accurately depict the overall temporal pattern of behavior, and ensure that the results of psychological abnormality detection truly reflect the actual psychological and behavioral state of the subject.

[0016] 2. This invention achieves a synergistic effect through multi-scale sliding time windows, temporal transition probability calculation, and adaptive threshold edge construction. The multi-scale sliding time window can comprehensively cover behavioral features across different spans and retain the transitional correlations of continuous behaviors. The temporal transition probability combined with the adaptive threshold can accurately filter stable behavioral temporal correlations, thereby eliminating random behavioral interference. The combination of the two not only fully extracts the core features of behavioral temporal sequence but also optimizes feature processing efficiency, significantly improving the accuracy of psychological anomaly detection and the operational efficiency of the detection process. Attached Figure Description

[0017] Figure 1 A flowchart of a psychological anomaly detection method based on behavioral temporal pattern recognition provided by the present invention; Figure 2 A flowchart illustrating the method for constructing a directed temporal correlation graph provided by this invention; Figure 3 The present invention provides a module structure diagram of a psychological anomaly detection system based on behavioral temporal pattern recognition. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1, referring to Figures 1-2 As shown in this embodiment, a psychological anomaly detection method based on behavioral temporal pattern recognition includes: S1. Obtain the atomic sequence of the subject's behavioral actions when performing a standardized behavioral task.

[0020] In this embodiment of the invention, obtaining the atomic sequence of behavioral actions of a subject when performing a standardized behavioral task includes: Extracting the motion trajectory of skeletal points from videos of subjects' behavior; By detecting local extreme points of motion velocity in the motion trajectory of skeletal points, the motion trajectory of skeletal points is segmented to obtain a series of indivisible behavioral action atoms. The action atoms are arranged in order of timestamps to construct a sequence of action atoms.

[0021] Specifically, for example, a standardized behavioral task is selected: continuously raising and lowering an object. Then, video footage of the subject performing the task is captured frame by frame, and the spatial positions of key skeletal points such as the shoulder, elbow, wrist, and fingers of the human body are identified in the footage.

[0022] Then, the spatial positions of the same bone point in consecutive frames are connected sequentially according to the acquisition time to obtain the bone point motion trajectory corresponding to each key bone point.

[0023] Furthermore, the positional change of the same bone point between adjacent frames is calculated, and then the positional change is divided by the time interval between adjacent frames to obtain the motion velocity of the bone point at the corresponding moment.

[0024] Then, iterate through the motion speed at all times, find the peak point where the motion speed changes from rising to falling and the trough point where the motion speed changes from falling to rising, and take these points as local extreme points of motion speed. Then, using the local extreme points of motion speed as the dividing boundary, divide the motion trajectory of the skeleton point into multiple continuous motion segments, and each motion segment corresponds to an indivisible behavior atom.

[0025] For example, we can obtain atomic actions such as raising a hand, reaching for an object, grasping an object, moving an object with a raised hand, putting down an object, and retracting an arm.

[0026] Finally, the start timestamp of each action atom in the video is recorded, and all action atoms are arranged in ascending order of their start timestamps to construct a complete sequence of action atoms. If the subject pauses during the task, the corresponding non-action segment is not included in the action atom sequence.

[0027] In summary, this solution integrates discrete behavioral data fragments by constructing a behavioral temporal correlation network, and generates a benchmark reference system adapted to an individual's own rhythm by combining long-term historical behavioral data.

[0028] The above settings effectively filter out interference from environmental noise and occasional behaviors, accurately preserve the temporal features of continuous behaviors strongly correlated with psychological states, significantly reduce the impact of irrelevant data on subsequent pattern recognition, improve the identification of abnormal signals, and adapt to the behavioral differences of different individuals through adaptive adjustment of personalized baselines, thereby enhancing the universality and robustness of the detection method and providing a high-quality feature foundation for subsequent abnormal pattern recognition. This fundamentally improves the accuracy and operational efficiency of the entire psychological anomaly detection process.

[0029] S2. Perform time window segmentation on the atomic sequence of behavior actions to obtain multiple behavior feature segments, and construct a directed temporal correlation graph for each behavior feature segment.

[0030] As a preferred implementation, the atomic sequence of behavioral actions is segmented by time windows to obtain multiple behavioral feature segments, including: On the atomic sequence of behavior actions, multiple multi-scale sliding time windows with different time spans and a preset overlap rate between adjacent windows are set; Based on the window length of the multi-scale sliding time window and the preset overlap rate, sliding acquisition is performed on the atomic sequence of behavior actions to obtain multiple behavioral feature fragments with overlapping parts.

[0031] In this embodiment of the invention, a directed temporal correlation graph is constructed for each behavioral feature fragment, including: Select behavioral action atoms within behavioral feature segments as candidate nodes; Calculate the transition probability between any two action atoms within a behavioral feature segment, determined by their temporal sequence; Based on the comparison between the transition probability and the adaptive probability threshold, directed edges are established between pairs of behavioral actions to construct a directed temporal association graph.

[0032] In this embodiment of the invention, the method for calculating the transition probability includes: Calculate the co-occurrence probability of two action atoms appearing in the form of consecutive adjacent occurrences in all action feature segments, and use the reciprocal of the time interval between the two action atoms as the temporal proximity. The transition probability is obtained by weighting and fusing co-occurrence frequency with temporal proximity.

[0033] In this embodiment of the invention, a directed edge is established between pairs of action atoms based on the comparison result of the transition probability and the adaptive probability threshold, including: Based on the distribution of the transition probabilities of all behavioral action atom pairs within the behavioral feature segment, an adaptive probability threshold is dynamically calculated. When the transition probability exceeds the adaptive probability threshold, a directed edge is established between the two corresponding action atoms.

[0034] Specifically, by combining the average completion time of a single action in a standardized behavioral task, the transition time of consecutive actions, and the behavioral correlation span that is of concern in clinical psychological assessment, three time spans corresponding to covering a single complete action, two consecutive actions, and three consecutive actions can be determined respectively.

[0035] It should be noted that the specific value of each time span can be determined by statistically analyzing the action duration distribution of a large number of normal subjects performing the same standardized behavioral task, and selecting the median value of the time interval that can cover the corresponding number of complete actions as the time span of that scale.

[0036] For example, for a standardized behavioral task of continuously raising and lowering an object, after statistically analyzing the action duration of a large number of normal subjects, we determined that the time span covering a single complete action corresponds to the average time of raising the hand, the time span covering two consecutive actions corresponds to the average time of raising the hand plus reaching for the object, and the time span covering three consecutive actions corresponds to the average time of raising the hand plus reaching for the object plus holding the object.

[0037] After setting the time span of the multi-scale sliding time window, the overlap rate between adjacent windows is preset based on the complete capture requirements of cross-window behavior transition association and the average duration of behavior action atoms.

[0038] It is important to emphasize that the default principle of the overlap rate is to ensure that the second half of the previous window can completely contain at least one complete action atom, and that the first half of the next window can completely overlap with the second half of the previous window, so as to ensure that the continuous action associations across windows are not segmented or omitted.

[0039] For example, the preset overlap rate ensures that the second half of the window covering two consecutive actions can fully contain the action of reaching for an object, and the first half of the window covering the second action can also fully contain the action of reaching for an object, thus fully capturing the continuous behavioral association from raising the hand to reaching for an object and then grasping the object.

[0040] After all parameters of the multi-scale sliding time window are set, each multi-scale sliding time window slides backward sequentially according to the preset overlap rate, starting from the beginning position of the behavior action atom sequence.

[0041] The distance of each slide is equal to the time span of the current window multiplied by the preset overlap rate. After each slide stops, all behavioral action atoms contained within the current window area are collected to form a behavioral feature fragment.

[0042] For example, the first swipe captures two action atoms: raising the hand and reaching for an object, forming the first action feature segment. The second swipe moves backward according to the preset overlap rate and captures two action atoms: reaching for an object and grasping an object, forming the second action feature segment.

[0043] The sliding acquisition process is repeated until the end of the sliding time window reaches the end of the behavior action atom sequence, resulting in multiple behavior feature fragments with overlapping parts.

[0044] Furthermore, each behavioral feature segment is processed individually, and each behavioral action atom contained within that behavioral feature segment is treated as an independent candidate node.

[0045] For example, if a certain behavioral feature segment contains three behavioral action atoms: raising hand, reaching for an object, and grasping an object, then the candidate nodes corresponding to that segment would include the raising hand node, reaching for an object node, and grasping an object node.

[0046] After the candidate nodes are determined, the total number of times two behavioral action atoms appear in the form of consecutive adjacent occurrences is counted in all behavioral feature segments.

[0047] Divide the total number of occurrences by the total number of occurrences of all adjacent pairs of behavioral action atoms in all behavioral feature segments to obtain the co-occurrence probability of the two behavioral action atoms.

[0048] For example, if the total number of times a person raises their hand and then immediately reaches out to grab an object is 100, and the total number of all adjacent action pairs is 1,000, then the co-occurrence probability of raising the hand and reaching out to grab an object is one in ten.

[0049] After calculating the co-occurrence probability, for two action atoms that appear sequentially in the current behavioral feature segment, calculate the difference between their start timestamps.

[0050] Then, the reciprocal of the difference is used as the temporal proximity between the two behavioral atoms. For example, if the start timestamp of raising the hand in the current behavioral feature segment is earlier than the start timestamp of reaching for the object, and the difference between the two start timestamps is one unit of time, then the temporal proximity from raising the hand to reaching for the object is one.

[0051] Furthermore, by combining the differences in the contribution of global behavioral pattern characteristics and local temporal proximity characteristics to the determination of psychological abnormalities in clinical psychological and behavioral analysis, corresponding weights are assigned to co-occurrence probability and temporal proximity, respectively.

[0052] It should be noted that the weights can be preset by having multiple clinical psychology experts score the importance of different features, and taking the average of all expert scores as the weight of the corresponding feature. For example, after clinical experts score the feature, the weight of co-occurrence probability is determined to be two-thirds, and the weight of temporal proximity is one-third.

[0053] Then, multiply the co-occurrence probability by its corresponding weight, multiply the temporal proximity by its corresponding weight, and add the two results together to obtain the transition probability between the two behavioral atoms determined by their temporal order.

[0054] Furthermore, after calculating the transition probabilities of all action atom pairs within the current behavioral feature segment, the transition probabilities of all action atom pairs within the current behavioral feature segment are arranged in ascending order of their numerical values.

[0055] Then, the cumulative distribution of the transition probability sequence after permutation is calculated, and the transition probability corresponding to the position where the cumulative distribution can distinguish between significant stable temporal association and accidental random association is selected as the adaptive probability threshold.

[0056] It should also be noted that the determination of this position is based on the range of transition probability distribution corresponding to the stable occurrence of behavioral associations in the normal subject population.

[0057] For example, after arranging the transition probabilities of all behavioral action pairs within a certain behavioral feature segment from smallest to largest, the cumulative distribution is calculated. When the cumulative distribution reaches a position that can cover more than 90% of stable behavioral associations in normal subjects, the transition probability corresponding to that position is selected as the adaptive probability threshold for that segment.

[0058] After the adaptive probability threshold is determined, all sequential action atomic pairs appearing in the current behavioral feature segment are traversed.

[0059] When the transition probability between a pair of action atoms exceeds the adaptive probability threshold corresponding to the action feature segment, a directed edge is established from the candidate node representing the previous action atom to the candidate node representing the next action atom.

[0060] For example, if the transition probability from raising a hand to reaching for an object in a certain behavioral feature segment exceeds the adaptive probability threshold of that segment, then a directed edge is established from the raising hand node to the reaching for object node; if the transition probability from reaching for an object to grasping an object exceeds the adaptive probability threshold, then a directed edge is established from the reaching for object node to the grasping object node.

[0061] If the transition probability of a certain action atom pair does not exceed the adaptive probability threshold, then no directed edge is established between the corresponding two candidate nodes to filter out any accidental random associations.

[0062] Finally, after all the directed edges that meet the conditions are established, all candidate nodes corresponding to the current behavioral feature segment and all the established directed edges are combined together to construct the directed temporal association graph corresponding to the behavioral feature segment.

[0063] For example, the final result is a directed temporal relationship graph containing nodes for raising hands, reaching out to pick up objects, grasping objects, and two directed edges.

[0064] If a behavioral feature segment contains only one behavioral action atom, then the directed temporal association graph corresponding to that segment contains only one candidate node, and no directed edges are established.

[0065] In summary, this solution eliminates redundant dimensions by filtering out redundant dimensions through temporal feature correlation, and dynamically calibrates matching templates based on real-time behavioral data instead of using fixed patterns. This significantly reduces the computational load caused by high-dimensional temporal features, improves the response speed of pattern recognition, effectively resolves recognition biases caused by feature coupling and individual differences, accurately captures implicit abnormal patterns in behavioral temporal sequences, thereby strengthening the correspondence between behavioral temporal patterns and abnormal psychological states, providing stable and reliable recognition results for subsequent psychological abnormality determination, and optimizing the efficiency and reliability of the overall detection process.

[0066] S3. Perform graph matching of the temporal association patterns extracted from the directed temporal association graph with the abnormal feature graph, and calculate the behavioral abnormality index based on the matching degree to obtain the psychological abnormality detection results.

[0067] In this embodiment of the invention, the method for obtaining the abnormal feature map includes: Obtain a directed temporal correlation graph of the historical data of a group of subjects with psychological abnormalities; Subgraph mining is performed on the historical directed temporal association graph to extract common temporal association substructures that appear more than a preset threshold number of times, and these are used as anomaly feature graphs.

[0068] As a preferred implementation, the temporal correlation patterns extracted from the directed temporal correlation graph are structurally matched with the anomaly feature graph, including: The directed temporal correlation graph is decomposed into multiple local temporal subgraphs, and the anomaly feature graph is decomposed into multiple anomaly subgraph templates. In the local temporal subgraph, the search finds matching instances that satisfy the isomorphic constraint in terms of both node type and temporal dependency strength attribute represented by directed edges; The matching degree is calculated by weighting the number of matching instances and the structural similarity between the matching instances and the corresponding abnormal subgraph templates.

[0069] In embodiments of the present invention, a behavioral abnormality index is calculated based on the matching degree to obtain a psychological abnormality detection result, including: Each matching degree is associated with the time window position of the behavioral feature segment that generated the match, resulting in an instantaneous anomaly score sequence distributed along the time axis; The behavioral anomaly index is obtained by summing the instantaneous anomaly score sequence using a time decay function with a preset time window. The contribution of earlier instantaneous anomaly scores to the current behavioral anomaly index decreases as the time distance between them and the current moment increases. The behavioral abnormality index is compared with the clinical diagnostic threshold; if it exceeds the threshold, it is considered a psychological abnormality.

[0070] Specifically, historical directed temporal correlation graphs generated when clinically diagnosed subjects with various types of psychological abnormalities perform the same standardized behavioral tasks can be collected to establish a temporal correlation graph library of psychological abnormal behaviors.

[0071] After collecting the historical directed temporal relationship graphs, subgraph mining is performed on all historical directed temporal relationship graphs in the graph library. All possible substructures are traversed and the occurrence frequency of each substructure in the entire historical directed temporal relationship graph is counted.

[0072] It should be noted that the preset frequency threshold is determined based on the degree of difference between the frequency of occurrence of the substructure in the psychologically abnormal group and the frequency of occurrence in the normal group. The frequency of occurrence that reaches the clinical significance level is selected as the preset frequency threshold.

[0073] In practice, if the number of occurrences of a certain common temporal correlation substructure exceeds a preset threshold, then the common temporal correlation substructure is extracted and used as an anomaly feature map.

[0074] For example, in a standardized behavioral task of repeatedly raising and lowering an object, if the substructure of immediately retracting the arm after grasping the object occurs much more frequently in the group of subjects with depression than in the normal group, then this substructure is extracted as the corresponding abnormal feature map.

[0075] Furthermore, after obtaining the abnormal feature map, each directed temporal association graph of the subject to be tested is decomposed into multiple local temporal subgraphs, each of which contains several consecutive nodes and directed edges between nodes.

[0076] At the same time, the abnormal feature map is decomposed into multiple abnormal sub-map templates, and each abnormal sub-map template corresponds to an independent abnormal behavior association unit in the abnormal feature map.

[0077] For example, an abnormal feature map containing the behavior of grasping an object and then immediately retracting the arm can be decomposed into an abnormal sub-map template where the object-grabbing node points to the arm-retracting node.

[0078] After completing the subgraph decomposition, in each local temporal subgraph of the subject to be tested, search for matching instances that satisfy the isomorphic constraint on both the node type and the temporal dependency strength attribute represented by the directed edge.

[0079] The isomorphic constraint of node type requires that the nodes in the local temporal subgraph and the nodes in the anomaly subgraph template have the same corresponding behavioral action atoms. The isomorphic constraint of the temporal dependency strength attribute of directed edges requires that the transition probabilities of the two are in the same numerical range.

[0080] For example, if a local time series subgraph contains a directed edge from a node holding an object to a node retracting its arm, and the transition probability of this directed edge is in the same range as the transition probability of the corresponding directed edge in the abnormal subgraph template, then the local time series subgraph is determined to be a matching instance.

[0081] After the search is completed, the number of all matching instances is counted, and the structural similarity between each matching instance and the corresponding abnormal subgraph template is calculated.

[0082] The structural similarity is calculated by summing the number of identical nodes and directed edges in the matching instance and the abnormal subgraph template, and dividing by the total number of nodes and directed edges in the abnormal subgraph template.

[0083] By combining the influence of the number of matching instances on the detection results and the influence of structural similarity on the detection results, corresponding weights are assigned to each. It should be noted that the weights can be preset by clinical psychological experts scoring the importance of different factors, and the average of all expert scores is taken as the weight of the corresponding factor.

[0084] Then, the number of matching instances is multiplied by its corresponding weight, the structural similarity is multiplied by its corresponding weight, and the two calculation results are added together to obtain the matching degree between the directed temporal association graph and the anomaly feature graph.

[0085] Furthermore, after obtaining the matching degree of all directed temporal correlation graphs, each matching degree can be associated with the center position of the time window corresponding to the behavioral feature segment that generated the match, thus obtaining the instantaneous anomaly score sequence distributed along the time axis.

[0086] Then, the instantaneous anomaly score sequence is accumulated and summed using a time decay function with a preset time window length to obtain the behavioral anomaly index.

[0087] It should be noted that the time decay function is designed to make the contribution of earlier instantaneous abnormal scores to the current behavioral abnormality index gradually decrease as the time distance between them and the current moment increases. The preset time window length is determined based on the duration characteristics of the abnormal psychological behavior pattern, and can be set in combination with the average duration of abnormal behavior associations observed in clinical practice.

[0088] After obtaining the behavioral abnormality index, it is compared with the clinical diagnostic threshold.

[0089] It should also be noted that the clinical diagnostic threshold can be determined by statistically analyzing the distribution of behavioral abnormality indices of a large number of normal subjects and psychologically abnormal subjects, and then selecting the behavioral abnormality index that can best distinguish between normal subjects and psychologically abnormal subjects as the clinical diagnostic threshold.

[0090] Then, based on the comparison between the abnormality index and the clinical diagnostic threshold, if the behavioral abnormality index exceeds the clinical diagnostic threshold, the subject is determined to have psychological abnormalities.

[0091] If the behavioral abnormality index does not exceed the clinical diagnostic threshold, it is determined that no obvious abnormal psychological behavior pattern was detected in the subject.

[0092] In summary, this scheme weights and integrates the identification results of temporal features of multiple time periods and types of behaviors, and dynamically adjusts the judgment threshold by combining the individual's real-time behavioral status with historical baselines instead of using a fixed threshold.

[0093] Overall, this solution effectively eliminates the interference caused by temporary behavioral fluctuations and single feature biases, improves the stability and confidence of anomaly detection, adaptively matches the behavioral rhythms and detection scenarios of different individuals, and significantly reduces the probability of misjudgment and missed judgment, thereby obtaining psychological anomaly detection results that are consistent with the actual situation.

[0094] Example 3, as Figure 3 The diagram shown is a module structure diagram of a psychological anomaly detection system based on behavioral temporal pattern recognition provided by the present invention, which includes: The sequence acquisition module is used to acquire the atomic sequences of the subjects' behavioral actions when performing standardized behavioral tasks; The association graph construction module is used to perform time window segmentation on the atomic sequence of behavior actions to obtain multiple behavior feature segments, and to construct a directed temporal association graph for each behavior feature segment; The graph structure matching evaluation module is used to perform graph structure matching between the temporal association patterns extracted from the directed temporal association graph and the abnormal feature graph, and calculate the behavioral abnormality index based on the matching degree to obtain the psychological abnormality detection result.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for detecting psychological anomalies based on behavioral temporal pattern recognition, characterized in that, The methods include: S1. Obtain the atomic sequence of the subject's behavioral actions when performing a standardized behavioral task; S2. Perform time window segmentation on the atomic sequence of behavioral actions to obtain multiple behavioral feature segments, and construct a directed temporal correlation graph for each behavioral feature segment; S3. Perform structural graph matching between the temporal association patterns extracted from the directed temporal association graph and the abnormal feature graph, and calculate the behavioral abnormality index based on the matching degree to obtain the psychological abnormality detection results.

2. The psychological anomaly detection method based on behavioral temporal pattern recognition as described in claim 1, characterized in that, Obtain the atomic sequences of the subjects' behavioral actions when performing standardized behavioral tasks, including: Extracting skeletal point motion trajectories from videos of subjects' behavior; By detecting local extreme points of motion velocity in the motion trajectory of skeletal points, the motion trajectory of skeletal points is segmented to obtain a series of indivisible behavioral action atoms. The action atoms are arranged in order of timestamps to construct a sequence of action atoms.

3. The psychological anomaly detection method based on behavioral temporal pattern recognition as described in claim 1, characterized in that, The atomic sequence of behavioral actions is segmented by time window to obtain multiple behavioral feature segments, including: On the atomic sequence of behavior actions, multiple multi-scale sliding time windows with different time spans and a preset overlap rate between adjacent windows are set; Based on the window length of the multi-scale sliding time window and the preset overlap rate, sliding acquisition is performed on the atomic sequence of behavior actions to obtain multiple behavioral feature fragments with overlapping parts.

4. The psychological anomaly detection method based on behavioral temporal pattern recognition as described in claim 3, characterized in that, Construct a directed temporal association graph for each behavioral feature fragment, including: Select behavioral action atoms within behavioral feature segments as candidate nodes; Calculate the transition probability between any two action atoms within a behavioral feature segment, determined by their temporal sequence; Based on the comparison between the transition probability and the adaptive probability threshold, directed edges are established between pairs of action atoms to construct a directed temporal association graph.

5. A psychological anomaly detection method based on behavioral temporal pattern recognition as described in claim 4, characterized in that, Methods for calculating transition probabilities include: Calculate the co-occurrence probability of two action atoms appearing in the form of consecutive adjacent occurrences in all action feature segments, and use the reciprocal of the time interval between the two action atoms as the temporal proximity. The transition probability is obtained by weighting and fusing co-occurrence frequency with temporal proximity.

6. A psychological anomaly detection method based on behavioral temporal pattern recognition as described in claim 5, characterized in that, Based on the comparison between the transition probability and the adaptive probability threshold, directed edges are established between pairs of action atoms, including: Based on the distribution of the transition probabilities of all behavioral action atom pairs within the behavioral feature segment, an adaptive probability threshold is dynamically calculated. When the transition probability exceeds the adaptive probability threshold, a directed edge is established between the two corresponding action atoms.

7. A psychological anomaly detection method based on behavioral temporal pattern recognition as described in claim 1, characterized in that, Methods for obtaining anomaly feature maps include: Obtain a directed temporal correlation graph of the historical data of a group of subjects with psychological abnormalities; Subgraph mining is performed on the historical directed temporal association graph to extract common temporal association substructures that appear more than a preset threshold number of times, and these are used as anomaly feature graphs.

8. A psychological anomaly detection method based on behavioral temporal pattern recognition as described in claim 7, characterized in that, The temporal association patterns extracted from the directed temporal association graph are structurally matched with the anomaly feature graph, including: The directed temporal correlation graph is decomposed into multiple local temporal subgraphs, and the anomaly feature graph is decomposed into multiple anomaly subgraph templates. In the local temporal subgraph, the search finds matching instances that satisfy the isomorphic constraint in terms of both node type and temporal dependency strength attribute represented by directed edges; The matching degree is calculated by weighting the number of matching instances and the structural similarity between the matching instances and the corresponding abnormal subgraph templates.

9. A psychological anomaly detection method based on behavioral temporal pattern recognition as described in claim 8, characterized in that, The behavioral abnormality index is calculated based on the matching degree to obtain the psychological abnormality detection results, including: Each matching degree is associated with the time window position of the behavioral feature segment that generated the match, resulting in an instantaneous anomaly score sequence distributed along the time axis; The behavioral anomaly index is obtained by summing the instantaneous anomaly score sequence using a time decay function with a preset time window. The contribution of earlier instantaneous anomaly scores to the current behavioral anomaly index decreases as the time distance between them and the current moment increases. The behavioral abnormality index is compared with the clinical diagnostic threshold; if it exceeds the threshold, it is considered a psychological abnormality.

10. A psychological anomaly detection system based on behavioral temporal pattern recognition, used to implement any one of the psychological anomaly detection methods based on behavioral temporal pattern recognition according to claims 1-9, characterized in that, The system includes: The sequence acquisition module is used to acquire the atomic sequences of the subjects' behavioral actions when performing standardized behavioral tasks; The association graph construction module is used to perform time window segmentation on the atomic sequence of behavior actions to obtain multiple behavior feature segments, and to construct a directed temporal association graph for each behavior feature segment; The graph structure matching evaluation module is used to perform graph structure matching between the temporal association patterns extracted from the directed temporal association graph and the abnormal feature graph, and calculate the behavioral abnormality index based on the matching degree to obtain the psychological abnormality detection result.

Images