A method and system for obsessive-compulsive disorder recognition and intervention based on deep learning

Abstract

The present invention relates to an obsessive-compulsive disorder recognition and intervention method and intervention system based on deep learning. When the target object appears within the field of view, the video image of the target object collected is segmented and trained to predict the relevance of obsessive-compulsive disorder. When the target object is not within the field of view, multiple physiological index data of obsessive-compulsive disorder patients are collected through a bracelet, and a fuzzy neural network is used for prediction. The relevance of obsessive-compulsive disorder detected in the two scenarios can achieve data fusion. Finally, by judging the target control function, when obsessive-compulsive disorder behaviors occur, a remote control instruction is sent to the bracelet in a timely manner to play soothing music, and at the same time, the bracelet display screen prompts the user "relax your mind", thereby temporarily relieving the body and mind of obsessive-compulsive disorder patients and getting out of the obsessive-compulsive disorder mode. Therefore, this technical solution accurately and efficiently predicts the obsessive-compulsive disorder behaviors of people, and on the basis of the prediction, timely intervenes and guides the obsessive-compulsive disorder behaviors through audio-visual means.

Description

Technical Field

[0001] The present invention relates to the field of healthcare information, and in particular, to an obsessive-compulsive disorder recognition and intervention method and intervention system based on deep learning. Background Art

[0002] Obsessive-compulsive disorder (i.e., Obsessive-compulsive disorder, abbreviated as OCD) is characterized by the repeated intrusion of obsessive thoughts in the mind regardless of personal will, and the performance of compulsive actions to eliminate anxiety or prevent threatening consequences, and then continuous performance of compulsive behaviors. The clinical manifestations of obsessive-compulsive disorder mainly include obsessive thoughts and compulsive behaviors. Obsessive thoughts are certain unwanted or intrusive thoughts, doubts or impulses that repeatedly appear in the mind. People with obsessive-compulsive disorder clearly know that these are meaningless, but they cannot get rid of them, thus causing distress and anxiety. Compulsive behaviors refer to repeated behaviors or mental activities that may lead to loss of social function. For example, people with obsessive-compulsive disorder often worry that "what if the food is not washed clean and makes them sick" and repeatedly wash the food, or worry that "what if they are not careful when crossing the road and get hit by a car" and are absent-minded when crossing the road, looking around repeatedly to confirm. Obsessive-compulsive disorder is not an infectious disease and has no infectivity. Obsessive-compulsive disorder is a cognitive disorder that occurs during the process of seeking peace of mind. This cognitive disorder is a mental illness, but it will affect an individual's normal life and work. In severe cases, people will fall into extreme pain, anxiety and even self-harm.

[0003] For traditional obsessive-compulsive disorder intervention methods, generally, the method of consulting a psychologist is adopted, and there are also methods of predicting and evaluating the risk of obsessive-compulsive disorder by using machine learning. For example, Chinese invention patents CN112037914B - A method, system and device for constructing an obsessive-compulsive disorder risk assessment model, CN119361124A - An obsessive-compulsive disorder diagnosis and symptom degree assessment system, CN109360112B - A method for authenticating mental illness based on data processing and related devices. However, the existing methods mostly predict and judge obsessive-compulsive disorder based on image acquisition, existing statistical data or brain wave analysis. However, in many scenarios, people with obsessive-compulsive disorder are not always within the visual area and cannot upload data at any time. Therefore, these methods are not only difficult to accurately identify obsessive-compulsive disorder behaviors at the practical level, but also cannot intervene and guide in a timely and effective manner. Therefore, the existing machine learning methods cannot adapt to different scenarios of such people. Summary of the Invention

[0004] In view of the defects and problems in the above background art, the present invention proposes an obsessive-compulsive disorder recognition and intervention method and intervention system based on deep learning. The specific technical solutions are as follows:

[0005] An obsessive-compulsive disorder recognition and intervention method based on deep learning, which determines whether the target object is within the field of view. If so, it enters the image data acquisition and processing steps: acquires image data of the target object, uses a CNN network for image segmentation, classification, and prediction, and finally outputs the obsessive-compulsive disorder relevance S. If not, it enters the physiological index parameter acquisition and processing steps: acquires physiological index parameters of the target object, preprocesses, trains, and predicts the physiological index parameters, and inputs the obsessive-compulsive disorder relevance S into the target control function Ψ. Based on the obsessive-compulsive disorder relevance S and the target control function Ψ, it predicts whether the target object has obsessive-compulsive disorder behaviors. If so, it intervenes in the obsessive-compulsive disorder behaviors of the target object.

[0006] The target object is a person with obsessive-compulsive disorder, and the intervention uses video reminders and / or music playback.

[0007] The image is acquired through a camera, and the physiological index parameters are acquired through a bracelet worn by the target object.

[0008] The image data acquisition and processing steps include the following steps S11 - S14:

[0009] Step S11: When the target object is within the field of view, acquire the video frame image D of the target object within the acquisition area through the camera n,t , where t is the time and n represents a certain frame;

[0010] Step S12: Import the video frame image D n,t into the CNN network, perform image segmentation on the video frame image D n,t to obtain multiple feature maps D i,j,t , use a weighted bidirectional feature pyramid structure to assign different weights to the feature layers to transmit more effective feature information, where the subscript "i,j" represents the coordinate position of the feature map D i,j,t in the original video frame image D n,t ;

[0011] Step S13: In the output result of the CNN network, set the classification types, which include face, action, and associated objects, and preset the weights W i,j corresponding to each classification type in the softmax layer. Through the CNN network combined with the bidirectional feature pyramid and Yolo head, output each classification type, and use the classification type as the classification prediction value of the video frame image to determine the event type R t at the current moment, where R is the event type and t is the time;

[0012] The Yolo head includes, but is not limited to, the head module units of the Yolo (short for You Only Look Once) series of algorithms, such as Yolov5 head, Yolov6 head, Yolov7 head, and Yolov8 head.

[0013] Step S14: Traverse the event type R at a certain preset time before the current moment t-1 , R t-2 , R t-3 , R t-n , and determine whether R t-n is the same as R t . If they are the same, it is determined as the same event type, and the obsessive-compulsive disorder relevance S is judged according to the number of occurrences of the same event;

[0014] The levels of the obsessive-compulsive disorder relevance S are defined as: not relevant (the same event appears only once or does not appear); moderately relevant (the same event appears 2 times); highly relevant (the same event appears 3 times or more).

[0015] The steps for collecting and processing physiological index parameters include the following steps S21 - S25:

[0016] Step S21: When the target object is in the blind area or out of sight, through the bracelet worn on the hand of the obsessive-compulsive disorder person, a number of physiological index data of the obsessive-compulsive disorder person are collected in real time. The physiological index data includes the geographical location Dxy at the current moment, heart rate TH, respiratory rate ζ1, pulse frequency ζ2, blood pressure P, and movement posture M m and the obsessive-compulsive disorder relevance S, and the value of S is obtained from the step S14;

[0017] Step S22: Data preprocessing: Normalize a number of physiological index data to obtain a data set {X i，t}, specifically, the normalization parameter X i is:

[0018] (1)

[0019] In the formula, Y represents a certain physiological index value before normalization, X represents the physiological index after normalization, the subscript "t" represents a certain time point, the subscript "i" represents the i-th item, and Y i, 0 represents the i-th physiological index value at the initial moment when t = 0;

[0020] Step S23: Use the preprocessed data set {X i，t} in step S22 as the input parameter, set the number of neurons and the number of training times, and import them into the fuzzy neural network for training;

[0021] The T-S prediction model is used for prediction, and the fuzzy function of non-linear prediction is as follows:

[0022] (2)

[0023] In the above formula, both f( ) and g( ) are smooth non-linear functions, and S t represents the obsessive-compulsive disorder relevance of the person at time t, and u is the actual output value of fuzzy prediction. The format of g( ) is:

[0024] (3)

[0025] Step S24: Train and learn the normalized parameter X in the dataset i and the obsessive-compulsive disorder relevance S at the current moment, and update the parameters in the fuzzy network;

[0026] Step S25: On the basis of step S24, calculate the target control function Ψ, and judge whether the obsessive-compulsive disorder relevance S is medium or above. If so, further judge whether it reaches the preset threshold Ψ set If so, the MCU control unit sends a remote control instruction to the bracelet to intervene in the obsessive-compulsive disorder behavior of the target object through audio-visual prompts.

[0027] (4)

[0028] In the formula, the coefficients a1~a6 are obtained by repeated iterative calculation of the above fuzzy neural network under the given initial parameters, and Ψ(t) is used as the control parameter in the neural network training process.

[0029] On the basis of the above-mentioned obsessive-compulsive disorder intervention method based on image recognition, an obsessive-compulsive disorder recognition and intervention system based on deep learning is further proposed. The intervention system includes a camera, a bracelet and a controller. The camera has an image data acquisition unit, and the bracelet is provided with an index acquisition unit. The index acquisition unit includes a geographical location sensor, a heart rate sensor, a respiratory rate sensor, a pulse sensor, a blood pressure sensor and a motion posture sensor. The camera and the bracelet are both communicatively connected to the controller.

[0030] The controller includes an image processing unit, a fuzzy prediction unit and an MCU control unit. Among them, the video frame image data collected by the image data acquisition unit is transmitted to the image processing unit for image segmentation, denoising and classification prediction. The multiple physiological index data collected by the index acquisition unit is transmitted to the fuzzy prediction unit for preprocessing, fuzzy training and output of the target control function Ψ. When the target control function Ψ reaches the preset threshold Ψ setWhen the relevance S of obsessive-compulsive disorder is medium or above, the MCU control unit sends a remote control instruction to the bracelet.

[0031] In summary, compared with the prior art, the method and system for identifying and intervening in obsessive-compulsive disorder based on deep learning have the following beneficial effects:

[0032] When the target object appears within the field of view, the video images of the target object collected are segmented and trained to predict the relevance of obsessive-compulsive disorder. When the target object is not within the field of view, multiple physiological index data of the obsessive-compulsive disorder person are collected through the bracelet, and a fuzzy neural network is used for prediction. The relevance of obsessive-compulsive disorder detected in the two scenarios can achieve data fusion. Finally, by judging the target control function, when obsessive-compulsive disorder behaviors occur, a remote control instruction is sent to the bracelet in a timely manner to play soothing music, and at the same time, the bracelet display screen prompts the user "relax your mind", so that the body and mind of the obsessive-compulsive disorder person are temporarily relieved and they can get out of the obsessive-compulsive disorder mode. Therefore, the technical solution accurately and efficiently predicts the obsessive-compulsive disorder behaviors of people, and on the basis of the prediction, timely intervenes and guides the obsessive-compulsive disorder behaviors through audiovisual means. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flowchart of a method for identifying and intervening in obsessive-compulsive disorder based on deep learning according to the present invention;

[0034] Figure 2 is a schematic diagram of the SAME unit according to the present invention;

[0035] Figure 3 is a structural diagram of a system for identifying and intervening in obsessive-compulsive disorder based on deep learning according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] The present invention will be described in detail below in conjunction with the specific embodiments. The following embodiments will help those of ordinary skill in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

[0037] Please examine Figure 1 A method for identifying and intervening in obsessive-compulsive disorder based on deep learning includes two basic steps S11-S14 (image data acquisition and processing steps) and steps S21-S25 (physiological index parameter acquisition and processing steps):

[0038] Step S11: When the target object is within the field of view, the video frame image D of the target object in the area is collected through the camera n,t , where t is time and n represents a certain frame, and the target object is a person with obsessive-compulsive disorder;

[0039] Step S12: Import the video frame image D n,t into a CNN network (Convolutional Neural Network), and perform image segmentation on the video frame image D n,t to obtain multiple feature maps D i,j,t , and use a weighted bidirectional feature pyramid (BiFPN) structure to assign different weights to the feature layers to transmit more effective feature information; add a detection layer scale to improve the detection accuracy of the network for small targets, solve the problem of many targets and inaccurate recognition of small targets within the field of view, and also facilitate the rapid convergence of the results;

[0040] In the above formula, the subscripts "i,j" represent the coordinate positions of the feature map D i,j,t in the original video frame image D n,t ;

[0041] Step S13: In the output result of the CNN network, set the classification types, which include faces, actions, and associated objects. Preset the weights corresponding to each classification type in the softmax layer, and then output each classification type through this CNN network combined with the bidirectional feature pyramid and Yolo head. Use the output classification type as the classification prediction value of the video frame image to determine the event type R t at the current moment, where R is the event type and t is the time;

[0042] Step S14: Traverse the event types R t-1 , R t-2 , R t-3 , R t-n at a certain preset time before the current moment, and judge whether R t-n is the same as R t . For example, if the event type recognized in the image at time t is that the target object is cleaning the living room, and if the event type recognized at time t - 2 is that the target object is cleaning the living room and this action has been completed at time t - 1, then judge that R t and R t-2 are the same event type, and judge the obsessive-compulsive disorder relevance S according to the number of occurrences of the same event;

[0043] The levels of the obsessive-compulsive disorder relevance S are defined as: not relevant (the same event appears only once or does not appear); moderately relevant (the same event appears 2 times); highly relevant (the same event appears 3 times or more).

[0044] The above steps S11 - S14 are for segmenting and training the video images of the target object when the target object is within the field of view, and predicting the obsessive-compulsive disorder relevance. When the target object is not within the field of view, in order to predict and timely intervene in the obsessive-compulsive disorder behavior of the target object, the following steps S21 - S25 are also included:

[0045] Step S21: The target object is in the blind spot or out of view, data collection and information entry: through the wristband worn by the obsessive-compulsive person, real-time data collection of multiple physiological indicators of the obsessive-compulsive person is performed, including the current geographical location Dxy, heart rate TH, respiratory rate ζ1, pulse rate ζ2, blood pressure P, movement posture M m and the obsessive-compulsive disorder correlation S, where the S value is obtained by the step S14;

[0046] Based on a large number of obsessive-compulsive disorder intervention and treatment cases, the physiological indicator parameters selected in this embodiment can determine the degree of repetitiveness of a person's actions in daily life through geographic location and movement posture. When the obsessive-compulsive disorder is strong, due to neurological factors, endocrine disorders, etc., it is easy to cause significant changes in physiological parameters such as heart rate, respiratory rate, pulse frequency, and blood pressure. The changes in these physiological parameters can reflect the psychological state of the obsessive-compulsive disorder patient to a certain extent. For example, the detection of a person's accelerated respiratory rate, arrhythmia, sudden changes in pulse frequency, and increased blood pressure reflects the anxious psychological state of the obsessive-compulsive disorder patient who knows that it is a compulsive behavior but cannot control it autonomously. On this basis, combined with the degree of repetition of the person's actions, the obsessive-compulsive disorder relevance S of the person in the blind spot / non-video state can be inferred.

[0047] Step S22: Data preprocessing: normalize multiple physiological index data to obtain a data set {X i，t}, specifically, the normalization parameter X i for:

[0048] (1)

[0049] Where Y represents a physiological index value before normalization (including geographic location Dxy, heart rate TH, respiratory rate ζ1, pulse rate ζ2, blood pressure P, exercise posture M m ), X represents the normalized physiological index, the subscript "t" represents a certain time point, the subscript "i" represents the i-th item, and Y i, 0 represents the value of the ith physiological index at the initial moment t=0;

[0050] Step S23: The preprocessed data set {X i，t} as input parameters, set the number of neurons and the number of training times, and import them into the fuzzy neural network for training;

[0051] Since the change of obsessive-compulsive disorder over time is nonlinear, in the above step S23, in order to achieve a more accurate prediction, the present invention adopts the TS prediction model, and the fuzzy function of the nonlinear prediction is as follows:

[0052] (2)

[0053] In the above formula, both f( ) and g( ) are smooth non-linear functions, u is the actual output value of fuzzy prediction, and S t represents the obsessive-compulsive disorder relevance of a person at time t, where the format of g( ) is:

[0054] (3)

[0055] Step S24: Train and learn the normalization parameter X in the dataset i and the obsessive-compulsive disorder relevance S at the current moment, and update the parameters in the fuzzy network;

[0056] Step S25: Based on Step S24, calculate the target control function Ψ, and determine whether it reaches the preset threshold Ψ set , if so, further judge the obsessive-compulsive disorder relevance S according to the result. If the obsessive-compulsive disorder relevance S is medium or above, the MCU control unit sends a remote control instruction to the bracelet to turn on the music playback. The music played is soothing music, and at the same time, the bracelet display screen prompts the user "Relax your mind".

[0057] (4)

[0058] In the formula, the coefficients a1~a6 are obtained by repeated iterative calculation of the above fuzzy neural network under the given initial parameters, and Ψ(t) is used as the control parameter in the neural network training process.

[0059] In the above Step S23, the basic steps of the fuzzy prediction algorithm are as follows:

[0060] 1) Define the input eX of the system state i,t , eS t and the universes of discourse, membership functions and fuzzy prediction rules of the actual output value u of fuzzy prediction. Among them, the input eX i,t , eS t correspond to the dataset {X i，t} at time t and the obsessive-compulsive disorder relevance S t respectively;

[0061] 2) Map the real inputs (eX i,t , eS t ) and the actual output value u of fuzzy prediction to the inputs (eX i,t *, eS t *) and the output in the fuzzy inference universe of discourse u* ;

[0062] 3) Determine the active fuzzy subsets A i,t * and B i of (eX t *) and Bj and its action fuzzy prediction rule R k ;

[0063] 4) Calculate the membership degree values of each action rule μ k , μ k used to describe the degree to which an element eX i,t or eS t belongs to a fuzzy subset A i and B j , and its calculation is shown in formula (5):

[0064] (5)

[0065] 5) Based on formula (5), use the centroid method for defuzzification to obtain the fuzzy output quantity u* , as shown in formula (6):

[0066] (6)

[0067] 6) Map the output obtained by fuzzy prediction u* to the actual output value u, and the actual output value u is used as the adjustment value of the difference ΔS of the obsessive-compulsive disorder relevance S. At time t + 1, the difference ΔS t+1 is inferred and calculated according to the difference ΔS at time t t and the actual output value of fuzzy prediction, and the calculation method is specifically listed in formula (7) and formula (8):

[0068] (7)

[0069] (8)

[0070] Among them, select the variable values of ΔS, (S * dS / dt), and u. Positive large is PB, positive medium is PM, positive small is PS, negative small is NS, negative medium is NM, negative large is NB. Both ΔS and (S * dS / dt) correspond to six fuzzy rules of PB, PM, PS, NS, NM, and NB. Fill the six fuzzy rules of ΔS and (S * dS / dt) in rows and columns, and the assignment of u can be queried according to the fuzzy rules of any ΔS and (S * dS / dt). The assignment of u is shown in Table 1:

[0071] Table 1 Fuzzy rule table

[0072] u PB PM PS NS NM NB PB 1 2 1 2 1 -1 PM 2 0 1 1 1 -1 PS 2 1 0 1 0 1 NS 1 1 -1 -1 0 -1 NM 1 0 -1 0 0 -1 NB 2 1 0 -1 -2 -2

[0073] To improve the recognition effect of images and reduce the recognition loss, a soft attention mask is embedded between step S13 and step S14. The masked image is fed into the Transformer encoder for encoding, and the soft attention weights of the high-level features are embedded into the mask, thereby supplementing the mask regions that are not segmented due to the limited local receptive field. To enhance the recognition effect of multi-scale text sequences, the masked features and text features of different scales are hierarchically embedded, and then a soft attention weight map is obtained to assist text recognition. The recognition loss can be backpropagated to the Transformer multi-head encoder, and the detection can be guided by optimizing the form of the attention weight map. Specifically, the soft attention mask embedding module includes 1 Transformer encoder and 4 upsampling structures. The input of SAME is the masked image Mask ( Mask After encoding, adding weights, and activation, it becomes Mask 1 , Mask 2 , Mask 3 ) and 3 levels of soft attention feature maps (including , , . The soft attention feature map is a feature map in text format, and the text type is determined by the event type R t ). After the above-mentioned masked image Mask is encoded by the Transformer encoder, the soft attention weights of the high-level features d 1 , d 2 are embedded into the above-mentioned masked image. The masked image is activated using the common Sigmoid function. On this basis, the original masked image Mask is transformed into Mask 1 , Mask 2 , Mask 3 , and is successively fused and / or cross-multiplied with the 3 levels of soft attention feature maps , , to respectively form soft attention mask fusion maps SM 1 , SM 2 , SM 3 . On the basis of SM 1 , through feature upsampling U up , we get SM 2 . OnSM 2 Based on this, through feature upsampling U up to obtain SM 3 , and finally multiply the original image with Mask 3 and then input it into the recognizer. At the same time, input Mask 3 into the recognizer. The two are compared in the recognizer, and then a soft attention weight map is obtained to assist in the recognition of classified text. The recognition loss L rog ( L rog Calculated using the gradient descent method. Gradient descent is used to find the corresponding value of the independent variable when minimizing a certain function) can be backpropagated (i.e., the recognition gradient flows back) to the Transformer multi-head encoder. By optimizing the attention weights d 1 , d 2 adjust the recognition accuracy. Through the above method, the recognition effect of face or human body movement images in the area is improved, the event type is accurately recognized, and the recognition loss is reduced. The specific structure of the SAME unit is as shown in the appendix Figure 2 as shown.

[0074] In the above-mentioned embodiments, when the target object appears in the field of view, the video image of the collected target object is segmented and trained to predict the relevance of obsessive-compulsive disorder. When the target object is not in the field of view, in order to predict and timely intervene in the obsessive-compulsive disorder behavior of the target object, multiple physiological index data of obsessive-compulsive disorder patients are collected through a bracelet, and a non-linear fuzzy prediction function is used for prediction. The relevance of obsessive-compulsive disorder detected in both scenarios can achieve data fusion. Finally, through the judgment of the target control function, when obsessive-compulsive disorder behavior appears, a remote control instruction is sent to the bracelet in a timely manner to play soothing music, and at the same time, the bracelet display screen prompts the user "relax your mind", so that the obsessive-compulsive disorder patients can temporarily relieve their body and mind and jump out of the obsessive-compulsive disorder mode. Therefore, this technical solution actually combines image segmentation and non-linear fuzzy prediction function to accurately and efficiently predict the obsessive-compulsive disorder behavior of personnel, and on the basis of the prediction, timely intervene and guide the obsessive-compulsive disorder behavior through audio-visual means.

[0075] Please examine Figure 3, based on the obsessive-compulsive disorder intervention method based on image recognition in the foregoing embodiments, a deep learning-based obsessive-compulsive disorder recognition and intervention system is further proposed. The system includes a camera, a bracelet, and a controller. The camera has an image data acquisition unit, and the bracelet is provided with an index acquisition unit. The index acquisition unit includes a geographical location sensor, a heart rate sensor, a respiratory rate sensor, a pulse sensor, a blood pressure sensor, and a motion posture sensor. The geographical location sensor is used to detect the geographical location Dxy at the current moment, the heart rate sensor is used to detect the heart rate TH at the current moment, the respiratory rate sensor is used to detect the respiratory rate ζ1 at the current moment, the pulse sensor is used to detect the pulse frequency ζ2 at the current moment, the blood pressure sensor is used to detect the blood pressure P at the current moment, and the motion posture sensor is used to detect the motion posture M of the person at the current moment m The camera and the bracelet are both communicatively connected to the controller.

[0076] The controller includes an image processing unit, a fuzzy prediction unit, and an MCU control unit. Among them, the video frame image data collected by the image data acquisition unit is transmitted to the image processing unit for image segmentation, denoising, and classification prediction. The multiple physiological index data collected by the index acquisition unit is transmitted to the fuzzy prediction unit for preprocessing, fuzzy training, and outputting the target control function Ψ. When the target control function Ψ reaches the preset threshold Ψ set , and when the obsessive-compulsive disorder relevance S is medium relevant or above, the MCU control unit sends a remote control instruction to the bracelet to turn on music playback. The music played is soothing music, and at the same time, the user is prompted "relax your mind" through the bracelet display screen.

[0077] This embodiment also relates to a computer-readable storage medium. The readable storage medium is provided in the MCU control unit. The readable storage medium includes a stored computer program; when the computer program is processed by a processor, it controls the device where the computer-readable storage medium is located to execute the above embodiments.

[0078] Those skilled in the art should understand that the above embodiments only describe the optimal implementation methods of the present invention and are not limited to the above technical solutions. Without departing from the basic concept of the above invention, other technical solutions formed by any combination of the above technical features or their equivalent features, such as the above technical features being mutually replaced with the technical features having similar functions in the present invention, the protection scope of the present invention is determined by the technical solutions covered by the claims.

Claims

1. An obsessive-compulsive disorder identification and intervention system based on deep learning, characterized by: The intervention system includes a camera, a bracelet and a controller. The camera has an image data acquisition unit. The bracelet is provided with an index acquisition unit. The index acquisition unit includes a geographic location sensor, a heart rate sensor, a respiratory rate sensor, a pulse sensor, a blood pressure sensor and a motion posture sensor. The camera and the bracelet are both connected to the controller for communication. The intervention system uses the following method to identify obsessive-compulsive behavior and determine whether the target object is within the field of view. If so, the system enters the image data acquisition and processing step: image data of the target object is acquired, and image segmentation, classification, and prediction are performed using the CNN network, and finally the obsessive-compulsive disorder relevance S is output; if not, the system enters the physiological indicator parameter acquisition and processing step: physiological indicator parameters of the target object are acquired, and physiological indicator parameters are pre-processed, trained, and predicted, and the obsessive-compulsive disorder relevance S is input into the target control function Ψ; based on the obsessive-compulsive disorder relevance S and the target control function Ψ, it is predicted whether the target object has obsessive-compulsive behavior. If so, intervention is performed on the obsessive-compulsive behavior of the target object; The image data acquisition and processing steps include the following steps S11-S14: Step S11: The target object is within the field of view, and the camera collects a video frame image D of the target object in the area. n,t , where t is time and n represents a frame; Step S12: Convert the video frame image D n,t Import the CNN network and process the video frame image D n,t Perform image segmentation to obtain multiple feature maps D i,j,t , using the weighted bidirectional feature pyramid structure, different weights are assigned to the feature layers, where the subscript "i,j" represents the feature map D i,j,t In the original video frame image D n,t The coordinate position in ; Step S13: In the output result of the CNN network, set the classification type, which includes face, action, and associated object, and preset the weight W corresponding to each classification type in the softmax layer i,j , through the CNN network combined with the bidirectional feature pyramid and Yolo head, each classification type is output, and the output classification type is used as the classification prediction value of the video frame image to determine the event type R at the current moment t , where R is the event type and t is the time; Step S14: traverse the event type R of a preset time before the current time t-1 , R t-2 , R t-3 , R t-n , judge R t-n With R t Are they the same? If they are the same, they are judged to be the same event type, and the obsessive-compulsive disorder relevance S is determined based on the number of times the same event occurs; The physiological index parameters include the current geographical location Dxy, heart rate TH, respiratory rate ζ1, pulse rate ζ2, blood pressure P, movement posture M m and the obsessive-compulsive disorder correlation S, the S value is obtained by step S14; The physiological index parameter collection and processing step includes the following steps S21-S25: Step S21: when the target object is in the blind spot or out of sight, multiple physiological index data of the obsessive-compulsive disorder person are collected in real time through the wristband worn by the obsessive-compulsive disorder person; Step S22: Data preprocessing: normalize multiple physiological index data to obtain a data set {X i，t }, specifically, the normalization parameter X i for: X i，t+1 (And i，t+1 -AND i，t ) / AND i,0 (1) In the formula, Y represents the value of a physiological index before normalization, X represents the physiological index after normalization, the subscript "t" represents a time point, the subscript "i" represents the i-th item, and Y i,0 represents the value of the ith physiological index at the initial moment t=0; Step S23: The preprocessed data set {X i，t } as input parameters, set the number of neurons and the number of training times, and import them into the fuzzy neural network for training; The TS prediction model is used for prediction, and the fuzzy function of nonlinear prediction is as follows: X i，t+1 =f({X i，t+1 },S t )+g({X i，t },S t-1 )u (2) In the above formula, S t represents the relevance of the person's obsessive-compulsive disorder at time t, u is the actual output value of the fuzzy prediction, and the format of g() is: Step S24: Normalize the parameters X in the data set i And the obsessive-compulsive disorder correlation S at the current moment is trained and learned, and the parameters in the fuzzy network are updated; Step S25: Based on step S24, the target control function Ψ is calculated, and it is determined whether the obsessive-compulsive disorder correlation S is medium correlation or above. If so, it is further determined whether it reaches the preset threshold Ψ set ,If so, the MCU control unit sends remote control instructions to the bracelet to intervene in the obsessive-compulsive ,behavior of the target object through audio-visual prompts; Ψ(t)=a1*X 1，t +a2*X 2，t +a3*X 3，t +a4*X 4，t +a5*X 5，t +a6*X 6，t +S(t) (4) Wherein, coefficients a1~a6 are obtained by repeatedly iteratively solving the above fuzzy neural network when the initial parameters are given.

2. The deep learning-based obsessive-compulsive disorder identification and intervention system according to claim 1, characterized in that: The target subjects are people with obsessive-compulsive disorder, and the intervention uses video reminders and / or music playing.

3. The deep learning-based obsessive-compulsive disorder identification and intervention system according to claim 1, characterized in that: The image data is collected through a camera, and the physiological index parameters are collected through a bracelet worn by the target object.

4. The deep learning-based obsessive-compulsive disorder identification and intervention system according to claim 2, characterized in that: The intervention uses video reminders and / or music playback, specifically: the music playback is soothing music, and the user is prompted to "relax" through the bracelet display screen.

5. The deep learning-based obsessive-compulsive disorder identification and intervention system according to claim 1, characterized in that: The level of the obsessive-compulsive disorder correlation S is defined as: if the same event occurs only once or does not occur, it is irrelevant; if the same event occurs twice, it is moderately correlated; if the same event occurs three times or more, it is highly correlated.

6. The deep learning-based obsessive-compulsive disorder identification and intervention system according to claim 1, characterized in that: The controller includes an image processing unit, a fuzzy prediction unit and an MCU control unit, wherein the video frame image data collected by the image data acquisition unit is transmitted to the image processing unit for image segmentation, denoising and classification prediction, and the multiple physiological index data collected by the index acquisition unit are transmitted to the fuzzy prediction unit for preprocessing, fuzzy training and output of the target control function Ψ. When the target control function Ψ reaches a preset threshold Ψ set , and when the obsessive-compulsive disorder relevance S is medium or above, the MCU control unit sends a remote control command to the bracelet.

Images