Mental health monitoring and nursing robot with physical-multi-dimensional intelligent cooperation

Abstract

The invention discloses a mental health monitoring and nursing robot with body-multi-dimensional intelligent cooperation. The robot comprises a mobile robot platform equipped with a differential driving chassis, a laser SLAM navigation module and the like; a multi-mode sensing network integrating an RGB-D depth camera, a thermal imager and the like processes data in real time through an edge computing unit; an intelligent decision-making engine based on a micro-service architecture covers a patrol recognition module, a vital sign monitoring module, a psychological support Transform dialogue model, a multi-modal emotion recognition module and the like; and the man-machine interaction interface is connected through an end-to-end encrypted 5G network. A medicine compliance management module, an RFID medicine box and a privacy protection scheme of medicine taking action recognition and federal learning and differential privacy are further arranged. The method has the advantages that nursing efficiency is improved by more than 40%, adverse event occurrence rate is reduced by 35%, psychological intervention coverage rate is improved, medicine taking compliance is improved, sublingual Tibetan medicine detection rate is improved to more than 95%, data privacy is safe and controllable, and intelligent upgrading of mental health nursing is promoted.

Description

Technical Field

[0001] The present invention relates to the technical field of medical robots, and in particular to an embodied-multidimensional intelligent collaborative mental health monitoring and nursing robot. Background Art

[0002] The field of mental health care is facing unprecedented challenges. According to the World Health Organization, approximately one billion people worldwide suffer from mental illness, and the shortage of psychiatric medical staff is becoming increasingly severe. This severe shortage of human resources directly impacts the quality of care and patient safety. The traditional manual patrol model has blind spots in monitoring, making it impossible for nurses to provide close and uninterrupted observation 24 hours a day. This leads to frequent adverse events such as self-harm, impulsive harm to others, and unauthorized hospital departures. Especially during night shifts, a single nurse needs to simultaneously oversee multiple wards, making it difficult to promptly detect and address emergencies. Furthermore, the unique needs of psychiatric patients require nurses to possess professional psychological intervention skills. However, less than 30% of current nurses possess psychological counseling qualifications, failing to meet patients' psychological support needs.

[0003] Existing technologies supporting psychiatric nursing care primarily include electronic monitoring systems, smart pillboxes, and telemedicine platforms. Electronic monitoring systems typically use wearable devices to monitor patients' basic vital signs, such as heart rate and blood pressure. However, these devices are often limited in functionality, collecting only a limited number of physiological parameters and lacking the ability to comprehensively analyze patients' behavioral patterns and emotional states. While smart pillboxes can remind patients to take medications on time, they cannot effectively address the widespread problem of medication hiding and refusal among psychiatric patients, and medication adherence management still relies on manual oversight. Telemedicine platforms primarily facilitate video consultations between doctors and patients and are limited in their effectiveness for inpatients requiring continuous monitoring. More importantly, these technical solutions operate independently, lacking system integration and data fusion capabilities, and thus cannot form a complete nursing solution. Regarding data security, existing systems often utilize centralized storage architectures, which pose a high risk of privacy breaches and fail to meet the unique privacy requirements of psychiatric patients.

[0004] In recent years, with the rapid development of artificial intelligence, robotics, and the Internet of Things (IoT), the application of intelligent nursing robots in the medical field has steadily increased. In general wards, service robots already perform auxiliary tasks such as medication delivery and disinfection, reducing the workload of medical staff to a certain extent. However, these general-purpose medical robots are not suitable for the specialized environment of psychiatric care. The care needs of psychiatric patients are far more complex than those of general patients. They require not only physiological monitoring but also specialized functions such as emotion recognition, psychological support, and crisis intervention. Existing medical robots lack specialized algorithms for mental illness and are unable to accurately identify risky behaviors unique to psychiatric conditions, such as manic episodes and depression and self-harm. In terms of human-robot interaction, existing robot dialogue systems are mostly based on rule matching or simple natural language processing, making it difficult to provide in-depth psychological counseling and emotional support. Therefore, developing a multidimensional intelligent robot system specifically tailored to the needs of psychiatric care has important practical significance and application value. Summary of the Invention

[0005] Based on the above objectives, the present invention provides an embodied-multi-dimensional intelligent collaborative mental health monitoring and nursing robot

[0006] include:

[0007] A mobile robot platform comprising a differential drive chassis, a laser SLAM navigation module, a UWB indoor positioning module, and an edge computing unit. The edge computing unit utilizes an AI computing platform with 32TOPS computing power.

[0008] A multimodal perception network, provided on the mobile robot platform, includes an RGB-D depth camera, a thermal imager, and a 60GHz FMCW millimeter-wave radar. The multimodal perception network performs real-time data processing through an edge computing unit;

[0009] The intelligent decision-making engine is implemented based on a microservices architecture and includes a patrol recognition module, a vital signs monitoring module, a navigation control module, a psychological support module, and a symptom warning module. Each module interacts with data through an API gateway.

[0010] The human-computer interaction interface, including the robot-side interactive screen, medical staff's mobile terminal and family-side application, realizes data transmission through an end-to-end encrypted 5G network.

[0011] Furthermore, the patrol identification module includes:

[0012] The object detection network based on YOLOv8 adopts the CSPDarknet53 backbone network and FPN-PAN feature fusion structure, and outputs the detection results through the decoupling head;

[0013] The fall detection unit extracts the coordinates of 17 key points of the human body through OpenPose and determines a fall event when the torso tilt angle exceeds 60 degrees and the center of gravity height drops by more than 30% within 0.5 seconds;

[0014] The identity verification unit uses a deep convolutional network trained based on the ArcFace loss function for face recognition, maintains a 128-dimensional feature vector database, and combines it with an RFID dual verification mechanism.

[0015] Furthermore, the vital signs monitoring module adopts a non-contact measurement solution:

[0016] The 60GHz FMCW millimeter wave radar locates the target through range-Doppler two-dimensional FFT transform and extracts the phase information of the chest area;

[0017] The 0.1-0.5 Hz respiratory signal and the 0.8-2 Hz heartbeat signal are separated by a bandpass filter;

[0018] Phase unwrapping technology is used to obtain the chest displacement curve to calculate the respiratory rate, and an adaptive noise elimination algorithm combined with empirical mode decomposition is used to extract the heart rate;

[0019] By fusing multiple measurement cycle data through the Kalman filter, the respiratory rate measurement error is less than 2 times / minute, and the heart rate measurement error is less than 3 times / minute.

[0020] Furthermore, the navigation control module includes:

[0021] A hybrid positioning unit integrates laser SLAM based on the Cartographer algorithm, visual SLAM based on the ORB-SLAM3 framework, and UWB positioning using the TWR protocol, fusing the three positioning source data through an extended Kalman filter.

[0022] The dynamic obstacle avoidance unit uses a deep reinforcement learning network trained based on the PPO algorithm. It inputs the 360-degree range scan of the lidar, the point cloud data of the depth camera and the target direction vector, and outputs the linear velocity and angular velocity control instructions.

[0023] Furthermore, the psychological support module includes:

[0024] A Transformer-based dialogue generation model was developed. Based on GPT-4, the model was fine-tuned for domain adaptation using 100,000 desensitization counseling conversations. The model uses a 12-layer decoder with a hidden layer dimension of 768.

[0025] The multimodal emotion recognition unit uses an attention mechanism to weightedly fuse facial expression, voice, and text features. Facial expression recognition is based on the ResNet-50 network, voice emotion recognition uses the CNN-LSTM network, and text emotion analysis uses the BERT model.

[0026] The dialogue policy controller adopts a hierarchical reinforcement learning framework. The high-level policy network selects the dialogue intent, and the low-level policy network generates specific responses.

[0027] Furthermore, the symptom warning module includes:

[0028] The feature extraction unit extracts 32-dimensional motion features including pace mean and variance, turn frequency, and activity area entropy; 24-dimensional physiological features including time-domain and frequency-domain indicators of heart rate variability; and 16-dimensional environmental interaction features including social distance and eye contact duration;

[0029] LSTM-GAN hybrid network uses a bidirectional LSTM encoder to process 128-dimensional time series feature vectors, with a hidden layer dimension of 256. The generator and discriminator are trained using the WGAN-GP objective function;

[0030] The risk score calculator is based on the weighted sum of the normalized deviations between the current value of the feature and the historical distribution. When the risk score exceeds the adaptive threshold, a graded warning is triggered.

[0031] Furthermore, the robot also includes a medication compliance management module:

[0032] Smart medicine box unit, integrating UHF RFID reader and anti-collision algorithm to realize simultaneous identification of multiple medicine plates;

[0033] The medication confirmation unit uses a 3D-CNN network to process 16 frames of RGB-D sequences to identify the completeness of medication actions;

[0034] The sublingual detection unit uses dual-wavelength near-infrared spectral imaging at 850nm and 940nm, and analyzes changes in reflectance spectra through an SVM classifier to detect drug residues, with a detection rate of over 95%.

[0035] Furthermore, the family-side application includes:

[0036] The data visualization module uses D3.js to implement a tree diagram of sentiment changes. The node size indicates the intensity of the sentiment, and the color encodes the sentiment category. It supports 7-day, 30-day, and 90-day historical data backtracking.

[0037] The secure dialogue system deploys a BERT-based text classification model to identify sensitive content, with an F1 score of 0.92. It mutes the conversation in real time and records the event when risky content is detected.

[0038] Furthermore, the robot uses a federated learning architecture and differential privacy technology to protect data security:

[0039] Each hospital deploys a local model and only uploads gradient updates, adding Laplace noise to the gradients with a privacy budget of ε = 1.0;

[0040] Data desensitization uses k-anonymization (k=5) and l-diversity (l=3) technologies, and sensitive attributes are stored using homomorphic encryption;

[0041] Set up a physical privacy switch and a three-level permission management system, and generate audit logs for all data access.

[0042] Preferably, a LoRa and 5G dual-channel communication architecture is adopted between the edge computing unit and the cloud platform. The edge node is responsible for real-time data processing and primary decision-making, the cloud platform performs model updates and historical data analysis, and distributed model optimization is achieved through federated learning.

[0043] Beneficial effects of the present invention:

[0044] First, it significantly improved psychiatric care efficiency and patient safety. The deployment of this system has increased psychiatric care efficiency by over 40% and reduced the incidence of major adverse events by 35%. The system utilizes multimodal real-time monitoring technology, identifying risk signs an average of 12.3 minutes before patients exhibit abnormal behavior, reducing emergency response time from 5 minutes with traditional manual inspections to 2.3 minutes. The LSTM-GAN-based risk prediction model achieves 93% accuracy in predicting dangerous behaviors such as manic episodes, effectively preventing serious incidents such as self-harm and impulsive assaults. Furthermore, the system's 24 / 7 monitoring capability mitigates staffing shortages during nighttime care, enabling a single nurse to effectively manage more patients simultaneously and optimizing the nurse-patient ratio from the current 1:8 to nearly 1:5.

[0045] Second, it significantly improves the doctor-patient experience and treatment compliance. The system uses an intelligent dialogue engine based on the Transformer architecture to provide patients with professional psychological support and emotional care, addressing the shortage of qualified nursing staff. The multimodal emotion recognition system accurately captures patients' emotional changes and provides timely psychological intervention. 76% of patients reported that the system has improved their sense of security. Regarding medication management, the smart pill box combined with sublingual testing technology has increased the detection rate of common psychiatric medications to over 95%, effectively addressing the common problem of medication hiding among psychiatric patients. The family-side application allows family members to understand the patient's condition in real time. 82% of family members reported that the system has helped them better participate in the patient's recovery process, enhancing the role of the family support system.

[0046] Third, the system fully exploits the value of data while protecting patient privacy. The system innovatively employs a federated learning architecture, whereby patient data from each hospital is retained locally and only model parameters are shared, fundamentally avoiding the risk of sensitive information leakage. Through multiple privacy protection technologies such as k-anonymization, l-diversity, and homomorphic encryption, cross-institutional collaborative research is supported while ensuring data security. The design of the physical privacy switch fully respects patient autonomy, allowing patients to control data collection at any time. This balanced design between privacy protection and data utilization not only complies with regulatory requirements for medical data management but also provides a sustainable technical path for psychiatric big data research and the continuous optimization of artificial intelligence models, driving the digital transformation of the entire mental health care industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0048] Figure 1 This is a schematic diagram of the architecture of the multi-dimensional intelligent mental health care robot system of the present invention;

[0049] Figure 2 Schematic diagram of the mobile robot platform structure of the present invention;

[0050] Figure 3 Schematic diagram of the multimodal monitoring and early warning process of the present invention;

[0051] Figure 4 Schematic diagram of the privacy protection and data security architecture of the present invention. DETAILED DESCRIPTION

[0052] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0053] It should be noted that references in the specification to "one embodiment," "an embodiment," "an exemplary embodiment," "some embodiments," etc. indicate that the described embodiments may include specific features, structures, or characteristics, but not every embodiment necessarily includes such specific features, structures, or characteristics. In addition, when specific features, structures, or characteristics are described in conjunction with an embodiment, it is within the knowledge of persons skilled in the relevant art to implement such features, structures, or characteristics in conjunction with other embodiments (whether or not explicitly described).

[0054] In general, terms can be understood, at least in part, from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending at least in part on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0055] like Figure 1 shown

[0056] System overall architecture implementation

[0057] The multi-dimensional intelligent mental health care robot system of this invention adopts a distributed microservices architecture, enabling real-time monitoring and intelligent decision-making through edge-cloud collaborative computing. The core of the system consists of four components: a mobile robot platform, a multimodal perception network, an intelligent decision-making engine, and a human-computer interaction interface.

[0058] See Figure 2

[0059] The mobile robot platform is based on a differential drive chassis and is equipped with a laser SLAM navigation system and a UWB indoor positioning module, achieving centimeter-level positioning accuracy. The robot body adopts a modular design, stands 1.2-1.5 meters tall, and is equipped with a tiltable gimbal. It is equipped with an RGB-D depth camera, thermal imager, millimeter-wave radar, and other sensor arrays. The computing unit uses the NVIDIA Jetson AGX Xavier edge computing platform, providing 32TOPS of AI computing power.

[0060] Patrol Identification Module Implementation Method

[0061] The patrol recognition function achieves omnidirectional environmental perception through multi-sensor fusion. The system deploys a YOLOv8-based object detection network, specifically optimized and trained for psychiatric scenarios. The captured image data is first denoised and enhanced through a preprocessing module before being fed into a modified CSPDarknet53 backbone network for feature extraction. A feature pyramid network and path aggregation network architecture are used for multi-scale feature fusion, ultimately outputting detection results through a decoupling head.

[0062] The fall detection algorithm uses a pose estimation method based on skeleton keypoints. Using OpenPose, it extracts the coordinates of 17 key points on the human body and calculates characteristic parameters such as torso tilt angle, center of gravity height change rate, and joint angular velocity. The system identifies a fall event when the torso tilt angle exceeds 60 degrees and the center of gravity height drops by more than 30% within 0.5 seconds. The algorithm also incorporates a time series analysis module, using an LSTM network to learn normal behavior patterns and predictively identify abnormal behavior.

[0063] Identity verification utilizes a dual authentication mechanism, facial recognition and RFID. The facial recognition module, based on a deep convolutional network trained using the ArcFace loss function, achieves 99.83% accuracy on the LFW dataset. The system maintains a 128-dimensional feature vector database and performs identity matching through cosine similarity. An RFID reader, deployed at the front of the robot, has a reading range of 3-5 meters and supports passive tags compliant with the ISO18000-6C protocol.

[0064] Vital Signs Monitoring Implementation Methods

[0065] Non-contact vital sign monitoring is achieved using 60GHz FMCW millimeter-wave radar technology. The radar transmits a linear frequency modulated continuous wave signal and extracts respiration and heartbeat information by analyzing the phase changes in the reflected signal. The signal processing pipeline includes: first, a range-Doppler two-dimensional FFT transform is performed to locate the target. Then, phase information of the chest region is extracted, and bandpass filtering is used to separate the respiration (0.1-0.5Hz) and heartbeat (0.8-2Hz) signal components.

[0066] The respiratory rate detection algorithm, based on phase unwrapping technology, calculates the cumulative phase difference between adjacent sampling points to obtain a chest displacement curve. Heart rate detection employs an adaptive noise cancellation algorithm combined with empirical mode decomposition to extract the intrinsic mode functions of the heartbeat signal. The system uses a Kalman filter to fuse data from multiple measurement cycles to improve measurement stability. Clinical validation has shown that compared to medical monitors, the respiratory rate measurement error is less than 2 beats / minute, and the heart rate measurement error is less than 3 beats / minute.

[0067] The thermal imaging temperature monitoring module uses an uncooled infrared detector with a resolution of 640×480 pixels and a temperature resolution of 0.05°C. Blackbody calibration and ambient temperature compensation algorithms achieve a temperature measurement accuracy of ±0.3°C. The system automatically identifies the forehead as a temperature measurement target and uses multi-frame averaging and outlier rejection algorithms to improve measurement reliability.

[0068] Navigation system implementation

[0069] The hybrid navigation system integrates laser SLAM, visual SLAM, and UWB positioning technologies to achieve precise navigation in complex environments. Laser SLAM utilizes the Cartographer algorithm, a graph-optimized algorithm, to construct a grid map with a resolution of 5cm using 2D lidar. The visual SLAM module, based on the ORB-SLAM3 framework, leverages depth information provided by an RGB-D camera to enhance positioning accuracy, particularly in feature-sparse corridor environments.

[0070] The UWB positioning system deploys 4-8 base stations covering the care area. It uses the TWR (Two-Way Ranging) protocol to measure distance and the least-squares method to determine position. The system integrates data from three positioning sources and uses an extended Kalman filter (EKF) for state estimation. The state vector includes position (x, y, θ), velocity, and acceleration, and is updated using the angular velocity and acceleration information provided by the IMU.

[0071] Dynamic obstacle avoidance utilizes a local path planning algorithm based on deep reinforcement learning. Inputs include 360-degree range scans from a LiDAR, point cloud data from a depth camera, and a target direction vector. The policy network is trained using the PPO algorithm, and the reward function comprehensively considers factors such as target reaching, collision avoidance, and path smoothness. The network architecture consists of three convolutional layers to process laser data and two fully connected layers to output linear and angular velocity control commands.

[0072] Implementation of the psychological support module

[0073] The core of the psychological support module is a dialogue generation model based on the Transformer architecture. This model is fine-tuned for domain adaptation based on GPT-4. The training data includes 100,000 desensitization counseling conversation transcripts and CBT / DBT treatment manuals. The model uses a 12-layer Transformer decoder with 768 hidden layer dimensions, 12 attention heads, and a vocabulary size of 50,000.

[0074] The emotion recognition system employs a multimodal fusion strategy. Facial expression recognition is based on a modified ResNet-50 network, pre-trained on the FER-2013 and AffectNet datasets, and can identify seven basic emotion categories. Voice emotion recognition extracts acoustic features such as MFCC, fundamental frequency, and formant, and performs classification using a CNN-LSTM network. Text sentiment analysis utilizes the BERT model, fine-tuned on a Chinese sentiment classification dataset.

[0075] Multimodal fusion uses the attention mechanism to weightedly fuse the feature vectors of the three modalities. Assume that the facial features are The sound characteristics are The text features are The fusion process is as follows:

[0076] First, different modal features are mapped to a unified dimensional space through linear projection:

[0077] h v =W v f v +b v ,h a =W a f a +b a ,h t =W t f t +b t

[0078] Then calculate the cross-modal attention weights:

[0079]

[0080] in is the average feature representation. The final fusion feature is:

[0081] h fusion =α v h v +α a h a +α t h t

[0082] The dialogue strategy utilizes a hierarchical reinforcement learning framework. A high-level policy network selects conversational intent, including listening, empathy, and cognitive reconstruction, while a low-level policy network generates specific responses. The reward function is designed based on user emotion changes, conversation coherence, and achievement of treatment goals.

[0083] Implementation of the symptomatic early warning system

[0084] See Figure 3

[0085] The early warning system is built on a hybrid deep learning model based on LSTM-GAN. The LSTM encoder processes time series features, including movement trajectories, vital sign changes, and behavioral patterns over a 24-hour period. The input feature vector has a dimension of 128 and contains:

[0086] Motion features: 32-dimensional features such as pace mean and variance, turning frequency, activity area entropy, and static duration distribution. The DTW algorithm is used to calculate the similarity with normal behavior patterns.

[0087] Physiological characteristics: 24-dimensional features including time-domain and frequency-domain indicators of heart rate variability, respiratory pattern entropy, and body temperature fluctuation coefficient.

[0088] Environmental interaction features: 16-dimensional features such as social distance, eye contact duration, and voice activity.

[0089] The LSTM network uses a bidirectional architecture with a hidden layer dimension of 256. It aggregates features across different time steps using an attention mechanism. The generator network learns the distribution of normal behavior, while the discriminator distinguishes between normal and abnormal patterns. Training uses the WGAN-GP objective function, with a gradient penalty added to ensure training stability.

[0090] The risk score calculation combines multiple indicators:

[0091]

[0092] where x i is the current value of the i-th feature, μ i and σ i is the historical mean and standard deviation, w i is the feature weight, learned from clinical data. When the risk score exceeds a dynamic threshold, an alert is triggered. The threshold is adaptively adjusted based on the patient's historical data and current status.

[0093] Medication compliance management implementation methods

[0094] The smart pill box system integrates RFID recognition, computer vision, and robotic arm control technologies. An ultra-high frequency RFID reader / writer is deployed within the pill box, and each pill plate is labeled with a unique tag that stores medication information, dosing schedule, and dosage. The system uses an anti-collision algorithm to simultaneously read multiple tags, achieving 99.9% recognition accuracy.

[0095] Medication confirmation utilizes a dual verification mechanism. First, a depth camera captures the medication action sequence, and a 3D-CNN network identifies key actions such as taking the medication, placing it in the mouth, and swallowing. The network input is a 16-frame RGB-D sequence, and the completeness of medication administration is determined by extracting spatiotemporal features.

[0096] The sublingual detection module utilizes near-infrared spectroscopy technology. Using dual-wavelength LED illumination at 850nm and 940nm, it analyzes changes in the reflectance spectrum of sublingual tissue to detect drug residues. The algorithm is based on a support vector machine classifier, with features including the first derivative of the spectral curve, peak position, and area ratio. The system completes testing within 2 seconds, achieving a detection rate of over 95% for common psychiatric drugs.

[0097] Family Connection System Implementation

[0098] The family-friendly app, developed using React Native, supports both iOS and Android platforms. Data transmission is end-to-end encrypted, using the AES-256 algorithm to encrypt sensitive information. Real-time video calls are based on the WebRTC protocol and integrate an AI monitoring module to analyze conversation content and patient status.

[0099] The emotional visualization interface uses D3.js to create interactive charts. The treemap's node size represents emotional intensity, color codes emotional categories, and edge thickness represents state transition probability. Data is updated hourly, with 7-day, 30-day, and 90-day historical data available. The visualization interface also includes daily activity heatmaps, medication adherence curves, and social interaction network diagrams.

[0100] The secure conversation system deploys a BERT-based text classification model to identify sensitive words and topics that may trigger negative emotions in patients. The model was trained on a dataset containing 50,000 annotated conversations, achieving an F1 score of 0.92. When risky content is detected, the system mutes the conversation in real time, prompts the family to change the topic, and records the incident for medical staff's reference.

[0101] Data security and privacy protection implementation methods

[0102] See Figure 4 shown

[0103] The system uses a federated learning architecture to protect patient privacy. Each hospital deploys a local model and uploads only gradient updates, not raw data, to the central server. Differential privacy techniques are employed, adding Laplace noise to the gradients and setting the privacy budget ε to 1.0 to achieve a balance between privacy protection and model performance.

[0104] Data desensitization utilizes k-anonymization (k=5) and l-diversity (l=3) techniques. Quasi-identifiers, such as age group, gender, and length of hospital stay, are processed using generalization and suppression techniques. Sensitive attributes, such as diagnostic information, are stored using homomorphic encryption, supporting statistical analysis on ciphertext.

[0105] A physical privacy switch is integrated into the robot's control panel, allowing patients to disable data collection at any time. The system features three levels of access: patients can only view a summary of their personal data, nurses can view real-time monitoring data and alerts, and doctors have full data access and parameter adjustment permissions. All data access is audited and logged, and retention periods comply with medical record management regulations.

[0106] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0107] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. Embodied-multi-dimensional intelligent collaborative mental health monitoring and care robot, characterized by: include: A mobile robot platform comprising a differential drive chassis, a laser SLAM navigation module, a UWB indoor positioning module, and an edge computing unit. The edge computing unit utilizes an AI computing platform with 32TOPS computing power. A multimodal perception network, provided on the mobile robot platform, includes an RGB-D depth camera, a thermal imager, and a 60GHz FMCW millimeter-wave radar. The multimodal perception network performs real-time data processing through an edge computing unit; The intelligent decision-making engine is implemented based on a microservices architecture and includes a patrol recognition module, a vital signs monitoring module, a navigation control module, a psychological support module, and a symptom warning module. Each module interacts with data through an API gateway. The human-computer interaction interface, including the robot-side interactive screen, medical staff's mobile terminal and family-side application, realizes data transmission through an end-to-end encrypted 5G network.

2. The embodied-multi-dimensional intelligent collaborative mental health monitoring and nursing robot according to claim 1 is characterized in that: The patrol identification module includes: The object detection network based on YOLOv8 adopts the CSPDarknet53 backbone network and FPN-PAN feature fusion structure, and outputs the detection results through the decoupling head; The fall detection unit extracts the coordinates of 17 key points of the human body through OpenPose and determines a fall event when the torso tilt angle exceeds 60 degrees and the center of gravity height drops by more than 30% within 0.5 seconds; The identity verification unit uses a deep convolutional network trained based on the ArcFace loss function for face recognition, maintains a 128-dimensional feature vector database, and combines it with an RFID dual verification mechanism.

3. The embodied-multi-dimensional intelligent collaborative mental health monitoring and nursing robot according to claim 1 is characterized in that: The vital signs monitoring module adopts a non-contact measurement solution: The 60GHz FMCW millimeter wave radar locates the target through range-Doppler two-dimensional FFT transform and extracts the phase information of the chest area; The 0.1-0.5 Hz respiratory signal and the 0.8-2 Hz heartbeat signal are separated by a bandpass filter; Phase unwrapping technology is used to obtain the chest displacement curve to calculate the respiratory rate, and an adaptive noise elimination algorithm combined with empirical mode decomposition is used to extract the heart rate; By fusing multiple measurement cycle data through the Kalman filter, the respiratory rate measurement error is less than 2 times / minute, and the heart rate measurement error is less than 3 times / minute.

4. The embodied-multi-dimensional intelligent collaborative mental health monitoring and nursing robot according to claim 1 is characterized in that: The navigation control module includes: A hybrid positioning unit integrates laser SLAM based on the Cartographer algorithm, visual SLAM based on the ORB-SLAM3 framework, and UWB positioning using the TWR protocol, fusing the three positioning source data through an extended Kalman filter. The dynamic obstacle avoidance unit uses a deep reinforcement learning network trained based on the PPO algorithm. It inputs the 360-degree range scan of the lidar, the point cloud data of the depth camera and the target direction vector, and outputs the linear velocity and angular velocity control instructions.

5. The embodied-multi-dimensional intelligent collaborative mental health monitoring and nursing robot according to claim 1 is characterized in that: The psychological support module includes: A Transformer-based dialogue generation model was developed. Based on GPT-4, the model was fine-tuned for domain adaptation using 100,000 desensitization counseling conversations. The model uses a 12-layer decoder with a hidden layer dimension of 768. The multimodal emotion recognition unit uses an attention mechanism to weightedly fuse facial expression, voice, and text features. Facial expression recognition is based on the ResNet-50 network, voice emotion recognition uses the CNN-LSTM network, and text emotion analysis uses the BERT model. The dialogue policy controller adopts a hierarchical reinforcement learning framework. The high-level policy network selects the dialogue intent, and the low-level policy network generates specific responses.

6. The embodied-multi-dimensional intelligent collaborative mental health monitoring and nursing robot according to claim 1 is characterized in that: The symptom warning module includes: The feature extraction unit extracts 32-dimensional motion features including pace mean and variance, turn frequency, and activity area entropy; 24-dimensional physiological features including time-domain and frequency-domain indicators of heart rate variability; and 16-dimensional environmental interaction features including social distance and eye contact duration; LSTM-GAN hybrid network uses a bidirectional LSTM encoder to process 128-dimensional time series feature vectors, with a hidden layer dimension of 256. The generator and discriminator are trained using the WGAN-GP objective function; The risk score calculator is based on the weighted sum of the normalized deviations between the current value of the feature and the historical distribution. When the risk score exceeds the adaptive threshold, a graded warning is triggered.

7. The embodied-multi-dimensional intelligent collaborative mental health monitoring and nursing robot according to claim 1 is characterized in that: Also includes a medication adherence management module: Smart medicine box unit, integrating UHF RFID reader and anti-collision algorithm to realize simultaneous identification of multiple medicine plates; The medication confirmation unit uses a 3D-CNN network to process 16 frames of RGB-D sequences to identify the completeness of medication actions; The sublingual detection unit uses dual-wavelength near-infrared spectral imaging at 850nm and 940nm, and analyzes changes in reflectance spectra through an SVM classifier to detect drug residues, with a detection rate of over 95%.

8. The embodied-multi-dimensional intelligent collaborative mental health monitoring and nursing robot according to claim 1 is characterized in that: The family-side application includes: The data visualization module uses D3.js to implement a tree diagram of sentiment changes. The node size indicates the intensity of the sentiment, and the color encodes the sentiment category. It supports 7-day, 30-day, and 90-day historical data backtracking. The secure dialogue system deploys a BERT-based text classification model to identify sensitive content, with an F1 score of 0.

92. It mutes the conversation in real time and records the event when risky content is detected.

9. The embodied-multi-dimensional intelligent collaborative mental health monitoring and nursing robot according to claim 1, characterized in that: Use federated learning architecture and differential privacy technology to protect data security: Each hospital deploys a local model and only uploads gradient updates, adding Laplace noise to the gradients with a privacy budget of ε = 1.0; Data desensitization uses k-anonymization (k=5) and l-diversity (l=3) technologies, and sensitive attributes are stored using homomorphic encryption; Set up a physical privacy switch and a three-level permission management system, and generate audit logs for all data access.

10. The embodied-multi-dimensional intelligent collaborative mental health monitoring and nursing robot according to any one of claims 1 to 9, characterized in that: The edge computing unit and the cloud platform adopt a LoRa and 5G dual-channel communication architecture. The edge node is responsible for real-time data processing and primary decision-making, and the cloud platform performs model updates and historical data analysis, and realizes distributed model optimization through federated learning.

Images