Intelligent hospital ward remote video monitoring management system

Abstract

The invention discloses an intelligent hospital ward remote video monitoring management system. The system comprises a video image acquisition module, a multi-dimensional parameter sensing module, a hypergraph feature extraction module, a model training reasoning module, an intelligent decision execution module and a remote data interaction module. The video and sensing module collects data such as images and temperature and humidity and transmits the data to the edge computing node. The hypergraph feature extraction module performs mapping fusion and dimension reduction on the multi-modal data; the model training reasoning module conducts reasoning based on the hypergraph neural network and corrects a result; the intelligent decision execution module generates an instruction control device according to a result, and the state monitoring unit feeds back a state; and the remote data interaction module realizes remote data transmission. According to the system, multi-source data are fused through edge calculation and a hypergraph neural network, closed-loop feedback is constructed, and the comprehensiveness and real-time performance of ward management are improved.

Description

Technical Field

[0001] This invention relates to the field of hospital ward monitoring, and more particularly to a smart hospital ward remote video monitoring and management system. Background Technology

[0002] With the deepening of medical informatization, hospital ward management has increasingly higher technical demands for real-time, accurate, and remote collaboration capabilities. Traditional models relying on manual inspections and single-point monitoring are no longer suitable for the management scenarios of large-scale ward clusters. Against this backdrop, remote monitoring systems that integrate multimedia sensing, multi-parameter acquisition, and intelligent analysis have become a development trend. By integrating video images and multi-dimensional sensor data, they achieve comprehensive control over patient status, equipment operation, and environmental parameters, providing data support for remote medical decision-making and driving the transformation of ward management models towards intelligence and automation.

[0003] Existing technologies in the field of remote ward monitoring have two significant shortcomings. First, the multi-source data processing mechanism has limitations. Using independent analysis of single-modal data or simple splicing methods fails to effectively establish deep correlations between video image features and sensor parameters such as temperature, humidity, and oxygen concentration. This results in the insufficient exploitation of the complementarity of cross-modal data, making it difficult to form a comprehensive feature set reflecting the ward's condition. Furthermore, information fragmentation or redundant data accumulation is prone to occur during data processing. Second, the decision-making response chain suffers from lag. In centralized computing architectures, data must be transmitted to remote servers for processing before control commands are generated. This makes the overall chain from data acquisition to equipment adjustment time-consuming. Moreover, the lack of a dynamic correction mechanism for sudden situations within the ward prevents the system from dynamically adjusting decision logic based on real-time changes in patient position and equipment operating parameters. Consequently, the system lacks flexibility in responding to abnormal states within the ward, failing to meet the requirements for high-precision, low-latency management. Summary of the Invention In order to overcome the shortcomings and deficiencies of existing technologies, this invention provides a smart hospital ward remote video monitoring and management system.

[0004] The technical solution adopted in this invention is a smart hospital ward remote video monitoring and management system, comprising: The video image acquisition module establishes a two-way data interaction channel with the edge computing node through an optical fiber transmission link to capture raw image data streams of patient position changes, equipment operating status, and ambient light and shadow distribution in the ward in real time. The multi-dimensional parameter sensing module includes a distributed temperature and humidity sensor, an air pressure monitor, an oxygen concentration detector, and an electrocardiogram signal acquisition terminal. Each sensing unit is connected to the edge computing gateway via a wireless Mesh network protocol. The collected physical quantities are converted into electrical signals and then converted from analog to digital. The quantized data is then transmitted to the edge computing node in a time-division multiplexing manner. The hypergraph feature extraction module receives image feature vectors preprocessed by the edge computing node from the video image acquisition module and quantized data output by the multidimensional parameter sensing module. It constructs cross-modal data association through a feature mapping matrix based on the hypergraph topology. The internal feature fusion unit and the cache unit of the edge computing node exchange data through the high-speed PCIe bus. The model training and inference module is equipped with an optimized HyperGraph neural network computing engine. It receives the fused feature tensor output by the HyperGraph feature extraction module through an Ethernet interface, and uses the computing power provided by the edge computing nodes to perform nonlinear mapping and high-dimensional space reconstruction of feature vectors. The output inference results are sent to the remote monitoring center through an encrypted transmission protocol. The intelligent decision execution module consists of a decision instruction generation unit based on FPGA architecture and a multi-channel relay control array. The decision instruction generation unit receives the decision matrix output by the model training and inference module through a serial communication interface. After logical operation, the generated control signal drives the relay control array through an opto-isolation circuit to adjust the lighting, temperature control and medical equipment in the ward in real time. The remote data interaction module uses 5G slicing technology to build a dedicated communication link. One end is connected to the output port of the intelligent decision execution module through an optical transceiver, and the other end establishes a data transmission channel with the remote server cluster through the core network interface. The feature data, decision instructions and equipment status information generated during system operation are transmitted bidirectionally in the form of encrypted data packets.

[0005] Furthermore, the high-definition camera array in the video image acquisition module constructs an image feature weight allocation mechanism through an optimized hypergraph neural network and edge computing multidimensional heterogeneous ward data analysis model, and the feature mapping relationship satisfies: in, This represents the feature weight coefficient of the j-th pixel in the i-th frame of the image captured by the k-th camera. The temperature sensitivity coefficient is dynamically adjusted for edge computing nodes. The number of monitoring areas within the ward. This represents the light intensity attenuation coefficient of the m-th monitoring area in the i-th frame image. Let J be the color channel response value of the j-th pixel within the m-th monitoring area. These are the spectral filtering parameters for the k-th camera. This represents the total number of pixels in a single frame image; simultaneously, this module performs time synchronization correction on multiple cameras through edge computing nodes, and the synchronization error correction model is as follows: in, This represents the time synchronization error correction value between the k-th and ith cameras. This is the base clock cycle for edge computing nodes. The number of reference markers in the ward. Let be the coordinate transformation coefficient of the p-th reference marker point on the imaging plane of the k-th camera. It is the physical distance between the k-th and ith cameras. At the speed of light, Let be the reflectance coefficient of the p-th reference marker.

[0006] Furthermore, each sensing unit in the multidimensional parameter sensing module constructs a parameter spatiotemporal correlation model based on an optimized hypergraph neural network and edge computing multidimensional heterogeneous ward data analysis model, and the data fusion process satisfies: in, This represents the fused sensing value of the q-th sensing region at time t. The spatiotemporal weighting coefficients dynamically allocated to edge computing nodes. This represents the number of static sensing units in the area. Let be the spatial attenuation factor of the s-th static sensing unit in the q-th region. Let be the sampled value of the s-th static sensing unit at time t. The number of dynamic sensing units. Let r be the time response coefficient of the r-th dynamic sensing unit in the q-th region. For the r-th dynamic sensing unit at time... The sampled values, The length of the time window set for edge computing nodes; This module transmits the calculated data via a wireless mesh network. When the feature mapping unit transmitted to the Hypergraph feature extraction module performs cross-modal association with video image features, it satisfies the following: in, Let represent the value of the m-th fused feature at time t. Let be the correlation coefficient between the q-th sensing region and the m-th fused feature. This represents the total number of sensing areas. For the number of cameras, Let be the grayscale value of the j-th pixel in the i-th frame at time t.

[0007] Furthermore, the hypergraph feature extraction module utilizes an optimized hypergraph neural network to construct a feature topology, and the hypergraph node association strength calculation satisfies: in, This represents the association strength between node a and node b in the hypergraph. To integrate the dimensions of features, Let be the sensitivity coefficient of the a-th node to the c-th fused feature. Let be the sensitivity coefficient of the b-th node to the c-th fused feature. Let c be the weight factor of the fused feature in the hypergraph. The length of the time series. Let be the value of the c-th fused feature at time t; When edge computing nodes perform dimensionality reduction on hypergraph features, the following conditions must be met: in, This represents the value of the d-th eigenvector after dimensionality reduction. Let be the mapping coefficient from the a-th hypergraph node to the d-th dimension-reduced feature. This represents the total number of nodes in the hypergraph. The number of associated nodes. The time decay coefficient, The reference time set for edge computing nodes. Let be the time derivative of the fusion feature corresponding to the a-th hypergraph node at time t.

[0008] Furthermore, the model training and inference module employs an optimized hypergraph neural network for feature inference, and the inference output value satisfies: in, This represents the output value of the p-th inference result. The dimension of the reduced-dimensional features. The connection weights from the d-th dimensionality-reduced feature to the p-th output are: is the activation function of the hypergraph neural network. This represents the number of hidden layer nodes. Let be the connection weight from the d-th dimensionality-reduced feature to the e-th hidden layer node. Let be the bias term for the e-th hidden layer node. This is the bias term for the p-th output result; When edge computing nodes dynamically correct inference results, the following conditions must be met: in, This represents the p-th reasoning result after correction. For the number of correction factors, Let f be the influence coefficient of the f-th correction factor on the p-th inference result. Let f be the change in the dimensionality reduction feature. The time decay coefficient of the f-th correction factor. This refers to the current moment.

[0009] Furthermore, the decision instruction generation unit of the intelligent decision execution module is based on the output of the model training and inference module. Construct a control instruction matrix such that the instruction output values ​​satisfy: in, This represents the output value of the q-th control instruction. For the number of inference results, Let be the influence coefficient of the p-th inference result on the q-th control command. For the floor function, is the quantization coefficient of the q-th control command; The drive signals for the relay control array satisfy: in, This represents the drive signal value of the r-th relay. To control the number of instructions, Let be the driving coefficient of the q-th control command for the r-th relay. For symbolic functions, Let be the driving nonlinearity coefficient of the r-th relay.

[0010] Furthermore, the hypergraph feature extraction module includes: The feature mapping unit receives image feature vectors from the video image acquisition module and quantized data from the multidimensional parameter sensing module. By constructing the correspondence between nodes and edges in the hypergraph topology, it maps feature data of different modalities to the high-dimensional space of the hypergraph, so that image features and sensing parameters are associated under the same hypergraph framework. It also transmits data to other units through the internal data conversion interface. The feature fusion unit obtains the hypergraph node features output by the feature mapping unit, and performs weighted combination of the features of different nodes according to the weight coefficients of the edges in the hypergraph, so that the scattered feature information is integrated into a fusion feature set with correlation. The fused features are transmitted to the feature dimensionality reduction unit through the high-speed data bus, and the intermediate data in the fusion process is temporarily stored in the internal cache. The feature dimensionality reduction unit receives the high-dimensional fused feature set output by the feature fusion unit, and uses a dimensionality reduction algorithm based on a hypergraph structure to reduce the dimensionality of the feature data while retaining the calibration feature information. The dimensionality-reduced feature data is sent to the model training and inference module through the PCIe bus interface, and at the same time, it feeds back the parameter changes during the dimensionality reduction process to the feature mapping unit. The feature verification unit verifies the dimensionality reduction features output by the feature dimensionality reduction unit. By comparing them with the preset feature threshold range in the hypergraph, it determines whether the dimensionality reduction features meet the input requirements of the model training and inference module. If there are features that exceed the range, a verification signal is generated and sent to the feature fusion unit to prompt it to perform the feature fusion operation again. The feature data that passes the verification is directly transmitted to the model training and inference module.

[0011] Furthermore, the model training and inference module includes: The inference computing unit receives the dimensionality reduction feature data transmitted by the hypergraph feature extraction module, loads the optimized hypergraph neural network model parameters, and maps the dimensionality reduction features to the corresponding inference results through the computation processing of multi-layer neurons. The intermediate data generated during the calculation process is stored in the internal temporary cache area, and the inference results are sent to the result correction unit through the output interface. The result correction unit obtains the initial inference result output by the inference calculation unit, and adjusts the initial result by combining the dynamic parameters of the ward environment transmitted in real time by the edge computing node, so as to eliminate the inference deviation caused by environmental changes. The corrected result is transmitted to the instruction conversion unit through the data bus, and the correction parameters are fed back to the inference calculation unit at the same time. The model update unit periodically receives new feature data and corresponding annotation results transmitted by the remote data interaction module. It uses the computing power resources of edge computing to update the model parameters of the hypergraph neural network, so that the model can adapt to the changes in feature data in the ward. The updated model parameters are loaded into the inference computing unit through the internal parameter configuration interface. The data feedback unit collects operational data from the inference calculation unit and the result correction unit, including inference time and correction magnitude information. After processing this data, it sends it to the remote data interaction module through a dedicated data channel. At the same time, it provides model operation status data to the model update unit as a reference for model updates.

[0012] Furthermore, the intelligent decision execution module includes: The decision instruction generation unit receives the corrected inference results output by the model training inference module, and converts the inference results into specific control instructions according to the preset decision logic and control rules. These instructions are stored in the internal register in the form of binary code and sent to the instruction execution unit through the control bus. The instruction execution unit acquires the control instructions output by the decision instruction generation unit, decodes the instructions, determines the type of ward equipment to be controlled, the control method and control parameters, transmits the decoded control signal to the relay control array through the drive circuit, and monitors the status of instruction execution in real time. The relay control array receives control signals from the instruction execution unit and controls the switching and brightness adjustment of lighting fixtures in the ward, the temperature setting of air conditioning equipment, the flow control of oxygen supply devices, and the start and stop of medical instruments through the on and off state changes of the relays. Its working status is transmitted to the status monitoring unit through the status feedback circuit. The status monitoring unit collects the working status of each relay in the relay control array in real time, including on / off status, current value and temperature parameters. After converting the status data into digital signals, it feeds back to the decision command generation unit as the basis for command adjustment, and sends it to the remote data interaction module for remote transmission of status information.

[0013] Beneficial Effects: This invention proposes a smart hospital ward remote video monitoring and management system. The system utilizes a high-definition camera array of the video image acquisition module and a distributed sensing unit of the multi-dimensional parameter sensing module to collaboratively capture multi-source data on patient status, equipment operation, and environmental parameters within the ward. The image data is mapped to the high-dimensional space of the hypergraph by the feature mapping unit of the hypergraph feature extraction module. The feature fusion unit performs weighted fusion of multimodal data based on the weight coefficients of the hypergraph edges, overcoming the information fragmentation and redundancy problems caused by single-modal analysis or simple splicing in traditional technologies, enabling cross-modal data to form deep correlations within the hypergraph framework. By leveraging edge computing nodes to provide localized computing power support for the HyperGraph neural network, the inference computing unit and result correction unit of the model training and inference module complete feature inference and dynamic adjustment at the edge. The decision command generation unit and command execution unit of the intelligent decision execution module drive the relay control array in real time based on the inference results, significantly shortening the link time from data processing to equipment adjustment. Simultaneously, the feature verification unit of the HyperGraph feature extraction module verifies the dimensionality reduction features and feeds them back to the feature fusion unit, while the status monitoring unit of the intelligent decision execution module feeds back the equipment status data to the decision command generation unit, forming a closed-loop feedback mechanism. This enables the system to dynamically adjust the decision logic based on changes in ward parameters, solving the problems of response lag and insufficient decision flexibility under centralized architecture, and achieving comprehensive perception, efficient analysis, and real-time control of ward status. Attached Figure Description

[0014] Figure 1 This is a diagram showing the system module composition of the present invention; Figure 2 This is a flowchart of the system operation of the present invention. Detailed Implementation

[0015] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] like Figure 1As shown, a smart hospital ward remote video monitoring and management system includes: The video image acquisition module consists of a high-definition camera array with multispectral imaging capabilities. It establishes a two-way data interaction channel with the edge computing node through an optical fiber transmission link to capture raw image data streams of patient position changes, equipment operating status, and ambient light and shadow distribution in the ward in real time. Specifically, the core components of the video image acquisition module include a high-definition camera array with multispectral imaging capabilities. Each camera has a resolution of no less than 5 megapixels, a spectral coverage range of 400-1000nm, and a frame rate maintained above 30fps. It can simultaneously capture images in the visible and near-infrared bands. The significance of this module lies in providing the system with a foundation for visualized information within the ward. By acquiring data from multiple perspectives, it eliminates monitoring blind spots. The implementation method adopts a distributed deployment, with one camera installed at each of the four corners of the ward ceiling and above the head of the bed, forming a 360° monitoring range without blind spots. Each camera is connected to the edge computing node via single-mode fiber optic cable with a transmission rate of no less than 10Gbps, supporting bidirectional data interaction. In practice, the camera array synchronously starts image acquisition at preset time intervals, generating a frame of raw image data containing patient position, equipment status, and ambient light and shadow every 20ms. Time-division multiplexing technology is used when transmitting this data to the edge computing node via fiber optic link, with each camera occupying an independent transmission time slot to avoid data conflicts. The control signals sent by the edge computing node to the camera array include exposure time adjustment commands (range 10-100μs), white balance parameters (color temperature 3000-6500K), and spectral filter switching signals. The cameras adjust their acquisition parameters in real time based on the received signals to ensure consistency of image data under different lighting conditions. The acquired raw image data stream is temporarily stored in the camera's built-in 8GB cache, and is then transmitted in batches after the edge computing node issues a read command.

[0017] The multi-dimensional parameter sensing module includes a distributed temperature and humidity sensor, an air pressure monitor, an oxygen concentration detector, and an electrocardiogram signal acquisition terminal. Each sensing unit is connected to the edge computing gateway via a wireless Mesh network protocol. The collected physical quantities are converted into electrical signals and then converted from analog to digital. The quantized data is then transmitted to the edge computing node in a time-division multiplexing manner. Specifically, the multi-dimensional parameter sensing module includes a temperature and humidity sensor, a barometric pressure monitor, an oxygen concentration detector, and an electrocardiogram (ECG) signal acquisition terminal. The temperature and humidity sensors have measurement ranges of 0-50℃ and 10%-95%RH, with accuracies of ±0.5℃ and ±2%RH, respectively. The barometric pressure monitor has a measurement range of 80-110 kPa and an accuracy of ±0.1 kPa. The oxygen concentration detector has a measurement range of 0-30% and an accuracy of ±0.2%. The ECG signal acquisition terminal has a sampling rate of 500 Hz and a resolution of 16 bits. This module supplements physical quantities and physiological parameters that cannot be captured by video images. It is implemented using a distributed deployment, with one temperature, humidity, and barometric pressure composite sensor installed every 2 meters along the walls at a height of 1.5 meters above the ground in the ward. The oxygen concentration detector is installed 0.5 meters directly above the head of the bed. The ECG signal acquisition terminal is connected to the patient's body surface via electrode pads. All sensing units are connected to a wireless mesh network based on the IEEE 802.11s protocol.

[0018] In practice, each sensing unit collects data according to a set cycle. The temperature, humidity, and air pressure sensors collect data every 10 seconds, the oxygen concentration detector collects data every 5 seconds, and the ECG signal acquisition terminal continuously collects data and packages it every 2 seconds. The collected electrical signals are converted into digital signals by a 16-bit analog-to-digital converter and transmitted via a wireless mesh network. The transmitted data packets contain the sensor ID, acquisition timestamp, and quantized data. The data packet size is controlled between 64 and 128 bytes, and the CSMA / CA mechanism is used to avoid channel collisions. The transmission rate is maintained at 54 Mbps. After receiving the data, the edge computing gateway classifies and stores it according to sensor type, with a storage capacity of no less than 1TB and a storage period of no less than 72 hours. At the same time, it forwards real-time data to the HyperGraph feature extraction module via an Ethernet interface (1Gbps).

[0019] The hypergraph feature extraction module receives image feature vectors preprocessed by the edge computing node from the video image acquisition module and quantized data output by the multidimensional parameter sensing module. It constructs cross-modal data association through a feature mapping matrix based on the hypergraph topology. The internal feature fusion unit and the cache unit of the edge computing node exchange data through the high-speed PCIe bus. Specifically, the HyperGraph feature extraction module consists of a feature mapping unit, a feature fusion unit, a feature dimensionality reduction unit, and a feature verification unit. The feature mapping unit uses an FPGA architecture with a processing frequency of 150MHz, supporting parallel processing of 8 feature vectors. The feature fusion unit has a built-in quad-core processor with a clock speed of 2.0GHz and a cache capacity of 4MB. The feature dimensionality reduction unit uses a dedicated ASIC chip, with a processing latency of no more than 5ms. The feature verification unit is connected to the cache unit of the edge computing node via a PCIe 3.0 bus with a bandwidth of 8GB / s. The significance of this module lies in establishing cross-modal data correlations and reducing data dimensionality to adapt to subsequent processing requirements. The implementation method achieves low latency feature processing through hardware acceleration, ensuring real-time performance.

[0020] In the specific implementation process, the feature mapping unit receives the image feature vector (512-dimensional) output by the video image acquisition module and the quantized data (64-dimensional) output by the multi-dimensional parameter sensing module. Following the preset node mapping rules in the hypergraph topology, it maps the image feature vector to 128 nodes of the hypergraph and the sensing data to 32 nodes. During the mapping process, the feature value of each node is generated through weighted summation, with the weight coefficient dynamically adjusted according to the feature type (range 0.1-0.9). The feature fusion unit, based on the weight coefficients of the hypergraph edges (dynamically updated through edge computing nodes with an update cycle of 1 minute), weights and combines the features of 160 nodes to generate a 256-dimensional fused feature set. The fused features are transmitted to the feature dimensionality reduction unit via an internal bus. The dimensionality reduction unit uses a linear transformation based on the hypergraph structure to reduce the 256-dimensional features to 64 dimensions, retaining at least 95% of the feature variance during the dimensionality reduction process. The feature verification unit compares the dimensionality reduction features with a preset threshold range (distributed through edge computing nodes). If the feature exceeds the range, a 4-digit verification code is generated and fed back to the feature fusion unit, prompting it to perform the fusion operation again.

[0021] The model training and inference module is equipped with an optimized HyperGraph neural network computing engine. It receives the fused feature tensor output by the HyperGraph feature extraction module through an Ethernet interface, and uses the computing power provided by the edge computing nodes to perform nonlinear mapping and high-dimensional space reconstruction of feature vectors. The output inference results are sent to the remote monitoring center through an encrypted transmission protocol. Specifically, the model training and inference module is equipped with an optimized hypergraph neural network computing engine, employing a GPU-accelerated architecture. It includes four computing cores, each with a single-core floating-point performance of 10 TFLOPS and 16GB of dedicated video memory, supporting dynamic memory allocation. The inference computing unit and the result correction unit are connected via an internal high-speed link with a bandwidth of 16GB / s. Model parameters are stored on a 512GB solid-state drive with a read speed of 500MB / s. The significance of this module lies in achieving the mapping from features to decision results through the hypergraph neural network, outputting inference results that can be directly used for control. The implementation combines hardware acceleration with algorithm optimization to improve inference efficiency.

[0022] In the specific implementation process, the inference computing unit receives the dimensionality-reduced feature vector (64-dimensional) from the hypergraph feature extraction module, loads the hypergraph neural network model parameters (containing a weight matrix of 1024 neurons), and performs calculations according to the forward propagation path. The output of each neuron is processed through an activation function, and 8-bit quantization is used during processing to reduce the amount of computation, keeping the inference latency within 10ms. The initial inference result (8-dimensional, corresponding to 8 types of ward states) is transmitted to the result correction unit. The correction unit obtains dynamic parameters (containing 32 environmental variables) provided by the edge computing node and adjusts the initial result dimension by dimension according to the correction rules. The adjustment range is determined based on the deviation value of the dynamic parameters (the larger the deviation value, the larger the adjustment range, with a maximum adjustment range of 15%). The corrected inference result is packaged using an encrypted transmission protocol (AES-256 encryption), with a data packet size of 128 bytes, and sent to the intelligent decision execution module via an Ethernet interface (1Gbps). At the same time, the intermediate data during the inference process (generating one log file every 10 minutes, with a size of 1MB) is stored on the local hard drive.

[0023] The intelligent decision execution module consists of a decision instruction generation unit based on FPGA architecture and a multi-channel relay control array. The decision instruction generation unit receives the decision matrix output by the model training and inference module through a serial communication interface. After logical operation, the generated control signal drives the relay control array through an opto-isolation circuit to adjust the lighting, temperature control and medical equipment in the ward in real time. Specifically, the intelligent decision execution module's decision instruction generation unit adopts an FPGA architecture with 1 million logic units and an operation frequency of 200MHz, supporting the parallel generation of 16 control instructions. The relay control array contains 32 relays, each with a rated current of 10A, a pull-in time of ≤10ms, and a release time of ≤5ms. The opto-isolation circuit has an isolation voltage of 2500V to ensure electrical isolation between control signals and high-voltage circuits. The status monitoring unit contains 16 analog signal acquisition channels with a sampling accuracy of 12 bits and a sampling rate of 1kHz. The significance of this module lies in converting the inference results into specific equipment control actions, realizing the adjustment of the ward environment and equipment. The implementation method ensures control safety through hardware isolation and redundancy design.

[0024] In the specific implementation process, the decision instruction generation unit receives the corrected inference results (8 dimensions) output by the model training inference module. According to a preset instruction mapping table (stored in internal ROM, capacity 64KB), it converts the inference results of each dimension into 4 control instructions (32 in total). The instruction format is 16-bit binary code, containing the device address (4 bits), control type (4 bits), and control value (8 bits). The instruction execution unit decodes the 32 instructions, parsing out the device to be controlled (such as lighting, air conditioning, oxygen devices, etc.), the control method (continuous adjustment or on / off control), and specific parameters (such as lighting brightness 0-100%, air conditioning temperature 16-30℃). The decoded control signal is transmitted to the relay control array through an opto-isolation circuit. The relay changes its on / off state according to the control signal, and the duration of the control signal is determined according to the device type (range 100ms-5s). The status monitoring unit collects the operating status (on / off), coil current (0-500mA), and contact temperature (-20-85℃) of 32 relays in real time, converts the collected data into 64-bit status words, feeds them back to the decision instruction generation unit through the SPI interface (10Mbps), and sends them to the remote data interaction module through the Ethernet interface.

[0025] The remote data interaction module uses 5G slicing technology to build a dedicated communication link. One end is connected to the output port of the intelligent decision execution module through an optical transceiver, and the other end establishes a data transmission channel with the remote server cluster through the core network interface. The feature data, decision instructions and equipment status information generated during system operation are transmitted bidirectionally in the form of encrypted data packets.

[0026] Specifically, the remote data interaction module employs 5G slicing technology to construct a dedicated communication link, comprising a 5G modem (supporting the Sub-6GHz band, with a maximum transmission rate of 1Gbps), an encryption chip (supporting the national standard SM4 algorithm), an optical transceiver (single-mode fiber interface, transmission distance of 20km), and a data cache unit (capacity of 2TB, read / write speed of 200MB / s). The significance of this module lies in enabling data interaction between the system and the remote monitoring center, ensuring the security and dedicated nature of data transmission. The implementation method uses 5G slicing to isolate different types of data transmission, prioritizing the transmission of data with high real-time requirements.

[0027] In the specific implementation process, the remote data interaction module receives control commands (1 frame every 100ms, frame size 256 bytes), feature data (1 frame every 500ms, frame size 1KB), and inference results (1 frame every 200ms, frame size 512 bytes) output by the intelligent decision execution module, as well as inference results (1 frame every 200ms, frame size 512 bytes) output by the model training and inference module, through an optical transceiver. After the data enters the cache unit, it is sorted according to priority (control commands have the highest priority, followed by feature data, and inference results have the lowest priority). The encryption chip encrypts the data in groups, with each group being 128 bytes in size. The encryption key is dynamically obtained from the remote server via the 5G link (update cycle 30 minutes). The encrypted data packets are appended with a 5G slice identifier (3 bits), a timestamp (32 bits), and a checksum (16 bits), and are sent to the core network through a 5G modem. During transmission, a HARQ mechanism is used to ensure transmission reliability, with no more than 3 retransmissions. Meanwhile, the module receives configuration parameters (1 frame every 5 minutes, frame size 128 bytes) from the remote server, including the threshold range of the hypergraph feature extraction module, the update parameters of the model training and inference module, etc. The received data is decrypted and distributed to the corresponding module through the Ethernet interface (1Gbps).

[0028] Preferably, the high-definition camera array in the video image acquisition module constructs an image feature weight allocation mechanism through an optimized hypergraph neural network and edge computing multidimensional heterogeneous ward data analysis model, and the feature mapping relationship satisfies: in, This represents the feature weight coefficient of the j-th pixel in the i-th frame of the image captured by the k-th camera. The temperature sensitivity coefficient is dynamically adjusted for edge computing nodes. The number of monitoring areas within the ward. This represents the light intensity attenuation coefficient of the m-th monitoring area in the i-th frame image. Let J be the color channel response value of the j-th pixel within the m-th monitoring area. These are the spectral filtering parameters for the k-th camera. This represents the total number of pixels in a single frame image; simultaneously, this module performs time synchronization correction on multiple cameras through edge computing nodes, and the synchronization error correction model is as follows: in, This represents the time synchronization error correction value between the k-th and ith cameras. This is the base clock cycle for edge computing nodes. The number of reference markers in the ward. Let be the coordinate transformation coefficient of the p-th reference marker point on the imaging plane of the k-th camera. It is the physical distance between the k-th and ith cameras. At the speed of light, Let be the reflectance coefficient of the p-th reference marker.

[0029] The video image acquisition module establishes a feature weight feedback link with the feature fusion unit of the hypergraph feature extraction module through the above model, and calculates the... and The feature fusion unit is input in matrix form to correct feature misalignment deviations during cross-camera image stitching.

[0030] Specifically, an image feature weighting mechanism and a multi-camera time synchronization correction model are constructed using an optimized HyperGraph neural network and edge computing multi-dimensional heterogeneous ward data analysis model. The image feature weighting mechanism determines the feature weight coefficients of each pixel based on the light intensity attenuation in different monitoring areas within the ward, the color channel response of pixels, and the spectral filtering parameters of the cameras. This ensures that image features from different areas are weighted according to their importance during fusion. The multi-camera time synchronization correction model corrects the time synchronization error of images acquired by different cameras based on the coordinate transformation coefficients of reference markers, the physical distance between cameras, and the reflectivity coefficient. During implementation, the video image acquisition module transmits the calculated feature weight coefficients and time synchronization error correction values ​​to the feature fusion unit of the HyperGraph feature extraction module via a feature weight feedback link. The feature fusion unit uses this data to correct potential feature misalignment deviations during cross-camera image stitching, ensuring the consistency and accuracy of the stitched image features and improving the reliability of subsequent feature processing.

[0031] Preferably, each sensing unit in the multidimensional parameter sensing module constructs a parameter spatiotemporal correlation model based on an optimized hypergraph neural network and edge computing multidimensional heterogeneous ward data analysis model, and the data fusion process satisfies: in, This represents the fused sensing value of the q-th sensing region at time t. The spatiotemporal weighting coefficients dynamically allocated to edge computing nodes. This represents the number of static sensing units in the area. Let be the spatial attenuation factor of the s-th static sensing unit in the q-th region. Let be the sampled value of the s-th static sensing unit at time t. The number of dynamic sensing units. Let r be the time response coefficient of the r-th dynamic sensing unit in the q-th region. For the r-th dynamic sensing unit at time... The sampled values, The length of the time window set for edge computing nodes.

[0032] This module transmits the calculated data via a wireless mesh network. When the feature mapping unit transmitted to the Hypergraph feature extraction module performs cross-modal association with video image features, it satisfies the following: in, Let represent the value of the m-th fused feature at time t. Let be the correlation coefficient between the q-th sensing region and the m-th fused feature. This represents the total number of sensing areas. For the number of cameras, Let be the grayscale value of the j-th pixel in the i-th frame at time t.

[0033] Specifically, a parametric spatiotemporal correlation model and a cross-modal feature correlation model are constructed using an optimized hypergraph neural network and edge computing multidimensional heterogeneous ward data analysis model. The parametric spatiotemporal correlation model comprehensively considers data collected by static and dynamic sensing units, fusing sensor data from different times and locations into a fused sensor value reflecting the state of a specific area through spatiotemporal weighting coefficients, spatial attenuation factors, and time response coefficients. The cross-modal feature correlation model associates the fused sensor value with image grayscale values ​​through correlation coefficients, achieving an effective combination of sensor data and image features. During implementation, the multidimensional parametric sensing module transmits the fused sensor value to the feature mapping unit of the hypergraph feature extraction module via a wireless mesh network. The feature mapping unit uses the cross-modal feature correlation model to fuse the sensor data and image features into a unified fused feature, enabling the system to simultaneously utilize both types of data to comprehensively reflect the ward state, providing richer and more comprehensive feature inputs for subsequent model training and inference.

[0034] Preferably, the hypergraph feature extraction module utilizes an optimized hypergraph neural network to construct a feature topology, and the hypergraph node association strength calculation satisfies: in, This represents the association strength between node a and node b in the hypergraph. To integrate the dimensions of features, Let be the sensitivity coefficient of the a-th node to the c-th fused feature. Let be the sensitivity coefficient of the b-th node to the c-th fused feature. Let c be the weight factor of the fused feature in the hypergraph. The length of the time series. Let be the value of the c-th fused feature at time t.

[0035] When edge computing nodes perform dimensionality reduction on hypergraph features, the following conditions must be met: in, This represents the value of the d-th eigenvector after dimensionality reduction. Let be the mapping coefficient from the a-th hypergraph node to the d-th dimension-reduced feature. This represents the total number of nodes in the hypergraph. The number of associated nodes. The time decay coefficient, The reference time set for edge computing nodes. Let be the time derivative of the fusion feature corresponding to the a-th hypergraph node at time t.

[0036] This module uses the PCIe bus to transfer the dimensionality-reduced feature vectors. The data is transmitted to the input layer of the model training and inference module for real-time interaction of feature data.

[0037] Specifically, an optimized hypergraph neural network and edge computing multidimensional heterogeneous ward data analysis model are used to construct a hypergraph node association strength calculation model and an edge computing node feature dimensionality reduction processing model. The hypergraph node association strength calculation model calculates the association strength between different nodes in the hypergraph based on the hypergraph node's sensitivity coefficient to the fused feature, the weight factor of the fused feature in the hypergraph, and the time series length, thus reflecting the degree of mutual influence between different features. The edge computing node feature dimensionality reduction processing model uses the mapping coefficient from the hypergraph node to the dimensionality-reduced feature, the time decay coefficient, and the time derivative of the fused feature to reduce the feature dimensionality while retaining key information. In implementation, the hypergraph feature extraction module first determines the association relationship between nodes using the hypergraph node association strength calculation model, and then the feature dimensionality reduction unit uses the feature dimensionality reduction processing model to reduce the dimensionality of the fused feature. The dimensionality-reduced feature vector is transmitted to the input layer of the model training and inference module via the PCIe bus, which reduces the complexity of data transmission and processing while ensuring the effectiveness of feature information, providing high-quality feature data for model training and inference.

[0038] Preferably, the model training and inference module uses an optimized hypergraph neural network for feature inference, and the inference output value satisfies: in, This represents the output value of the p-th inference result. The dimension of the reduced-dimensional features. The connection weights from the d-th dimensionality-reduced feature to the p-th output are: is the activation function of the hypergraph neural network. This represents the number of hidden layer nodes. Let be the connection weight from the d-th dimensionality-reduced feature to the e-th hidden layer node. Let be the bias term for the e-th hidden layer node. This is the bias term for the p-th output result.

[0039] When edge computing nodes dynamically correct inference results, the following conditions must be met: in, This represents the p-th reasoning result after correction. For the number of correction factors, Let f be the influence coefficient of the f-th correction factor on the p-th inference result. Let f be the change in the dimensionality reduction feature. The time decay coefficient of the f-th correction factor. This is the current moment. The revised inference result. The decision instruction is sent to the decision instruction generation unit of the intelligent decision execution module via an encrypted transmission protocol, serving as the basis for instruction generation.

[0040] Specifically, a feature inference model and a dynamic correction model for inference results are constructed based on an optimized hypergraph neural network and edge computing multidimensional heterogeneous ward data analysis model. The feature inference model utilizes the connection weights from dimensionality-reduced features to output results, the connection weights of hidden layer nodes, and bias terms to map dimensionality-reduced feature vectors to corresponding inference results, achieving the transformation from features to results. The dynamic correction model for inference results dynamically adjusts the initial inference results based on the influence coefficient of the correction factor on the inference results, the change in dimensionality-reduced features, and the time decay coefficient, enabling the inference results to adapt to real-time changes in the ward status. During implementation, the inference computation unit of the model training inference module loads the hypergraph neural network model and calculates the initial inference results from the dimensionality-reduced feature vectors transmitted by the hypergraph feature extraction module. The result correction unit uses the dynamic correction model to adjust the initial results. The corrected inference results are then sent to the decision instruction generation unit of the intelligent decision execution module via an encrypted transmission protocol, providing an accurate basis for the generation of decision instructions.

[0041] Preferably, the decision instruction generation unit of the intelligent decision execution module is based on the output of the model training and inference module. Construct a control instruction matrix such that the instruction output values ​​satisfy: in, This represents the output value of the q-th control instruction. For the number of inference results, Let be the influence coefficient of the p-th inference result on the q-th control command. For the floor function, is the quantization coefficient of the q-th control command.

[0042] The drive signals for the relay control array satisfy: in, This represents the drive signal value of the r-th relay. To control the number of instructions, Let be the driving coefficient of the q-th control command for the r-th relay. For symbolic functions, Let be the driving nonlinearity coefficient of the r-th relay.

[0043] drive signal The control commands are transmitted to the relay control array via an opto-isolated circuit to control and adjust the ward equipment. Simultaneously, this module transmits control commands... Feedback is sent to the remote data interaction module for bidirectional data transmission.

[0044] Specifically, a control command matrix generation model and a relay drive signal model are constructed using an optimized hypergraph neural network and edge computing multidimensional heterogeneous ward data analysis model. The control command matrix generation model converts the inference results into specific control command output values ​​based on the influence coefficient and quantization coefficient of the control commands. The relay drive signal model converts the control commands into drive signal values ​​that can drive the relays, based on the drive coefficient and nonlinearity coefficient of the control commands. In implementation, the decision command generation unit of the intelligent decision execution module generates control commands through the control command matrix generation model based on the corrected inference results output by the model training inference module. The command execution unit decodes the control commands and converts them into drive signals using the relay drive signal model. These drive signals are transmitted to the relay control array through an opto-isolation circuit, causing the relays to operate according to the commands. This enables precise adjustment of lighting, temperature control, and medical equipment within the ward. Simultaneously, the control commands are fed back to the remote data interaction module, ensuring that the remote end can monitor the control status of the equipment in real time.

[0045] Preferably, the hypergraph feature extraction module includes: The feature mapping unit receives image feature vectors from the video image acquisition module and quantized data from the multidimensional parameter sensing module. By constructing the correspondence between nodes and edges in the hypergraph topology, it maps feature data of different modalities to the high-dimensional space of the hypergraph, so that image features and sensing parameters are associated under the same hypergraph framework. It also transmits data to other units through the internal data conversion interface. The feature fusion unit obtains the hypergraph node features output by the feature mapping unit, and performs weighted combination of the features of different nodes according to the weight coefficients of the edges in the hypergraph, so that the scattered feature information is integrated into a fusion feature set with correlation. The fused features are transmitted to the feature dimensionality reduction unit through the high-speed data bus, and the intermediate data in the fusion process is temporarily stored in the internal cache. The feature dimensionality reduction unit receives the high-dimensional fused feature set output by the feature fusion unit, and uses a dimensionality reduction algorithm based on a hypergraph structure to reduce the dimensionality of the feature data while retaining the calibration feature information, thereby reducing the complexity of data transmission and processing. The dimensionality-reduced feature data is sent to the model training and inference module through the PCIe bus interface, and at the same time, it feeds back the parameter changes during the dimensionality reduction process to the feature mapping unit. The feature verification unit verifies the dimensionality reduction features output by the feature dimensionality reduction unit. By comparing them with the preset feature threshold range in the hypergraph, it determines whether the dimensionality reduction features meet the input requirements of the model training and inference module. If there are features that exceed the range, a verification signal is generated and sent to the feature fusion unit to prompt it to perform the feature fusion operation again. The feature data that passes the verification is directly transmitted to the model training and inference module.

[0046] Preferably, the model training and inference module includes: The inference computing unit receives the dimensionality reduction feature data transmitted by the hypergraph feature extraction module, loads the optimized hypergraph neural network model parameters, and maps the dimensionality reduction features to the corresponding inference results through the computation processing of multi-layer neurons. The intermediate data generated during the calculation process is stored in the internal temporary cache area, and the inference results are sent to the result correction unit through the output interface. The result correction unit obtains the initial inference result output by the inference calculation unit, and adjusts the initial result by combining the dynamic parameters of the ward environment transmitted in real time by the edge computing node, so as to eliminate the inference deviation caused by environmental changes. The corrected result is transmitted to the instruction conversion unit through the data bus, and the correction parameters are fed back to the inference calculation unit at the same time. The model update unit periodically receives new feature data and corresponding annotation results transmitted by the remote data interaction module. It uses the computing power resources of edge computing to update the model parameters of the hypergraph neural network, so that the model can adapt to the changes in feature data in the ward. The updated model parameters are loaded into the inference computing unit through the internal parameter configuration interface. The data feedback unit collects operational data from the inference calculation unit and the result correction unit, including inference time and correction magnitude information. After processing this data, it sends it to the remote data interaction module through a dedicated data channel. At the same time, it provides model operation status data to the model update unit as a reference for model updates.

[0047] Preferably, the intelligent decision execution module includes: The decision instruction generation unit receives the corrected inference results output by the model training inference module, and converts the inference results into specific control instructions according to the preset decision logic and control rules. These instructions are stored in the internal register in the form of binary code and sent to the instruction execution unit through the control bus. The instruction execution unit acquires the control instructions output by the decision instruction generation unit, decodes the instructions, determines the type of ward equipment to be controlled, the control method and control parameters, transmits the decoded control signal to the relay control array through the drive circuit, and monitors the status of instruction execution in real time. The relay control array receives control signals from the instruction execution unit and controls the switching and brightness adjustment of lighting fixtures in the ward, the temperature setting of air conditioning equipment, the flow control of oxygen supply devices, and the start and stop of medical instruments through the on and off state changes of the relays. Its working status is transmitted to the status monitoring unit through the status feedback circuit. The status monitoring unit collects the working status of each relay in the relay control array in real time, including on / off status, current value and temperature parameters. After converting these status data into digital signals, they are fed back to the decision command generation unit as the basis for command adjustment, and sent to the remote data interaction module for remote transmission of status information.

[0048] like Figure 2 As shown, a smart hospital ward remote video monitoring and management system includes the following steps in its operation: Step S1: The high-definition camera array of the video image acquisition module captures image data in the ward and sends it to the edge computing node via the fiber optic transmission link. At the same time, each sensing unit of the multi-dimensional parameter sensing module converts the collected physical quantity data into electrical signals and transmits them to the edge computing node through the wireless Mesh network to form an initial data set. Step S2: The edge computing node classifies the received image data and sensor data, transmits the image data to the feature mapping unit of the hypergraph feature extraction module, transmits the sensor data to the feature fusion unit of the same module, and establishes the initial data allocation path. Step S3: The feature mapping unit of the hypergraph feature extraction module maps the image data to the hypergraph high-dimensional space. The feature fusion unit performs weighted fusion of the sensor data and the mapped image features. The feature dimensionality reduction unit reduces the dimensionality of the fused features, generates a dimensionality-reduced feature vector, and transmits it to the model training and inference module. Step S4: The inference calculation unit of the model training inference module loads the hypergraph neural network model, calculates the dimensionality reduction feature vector to obtain the initial inference result, and the result correction unit adjusts the initial result in combination with the dynamic parameters of the edge computing node and outputs the corrected inference result. Step S5: The decision instruction generation unit of the intelligent decision execution module generates control instructions based on the corrected reasoning results. After decoding the control instructions, the instruction execution unit drives the relay control array to adjust the ward equipment. The status monitoring unit feeds back the equipment status data to the system. Step S6: The remote data interaction module encrypts the system's feature data, decision commands, and device status information and transmits them to the remote server via the 5G slicing link. At the same time, it receives the configuration parameters from the remote server to complete the entire monitoring and management process.

[0049] A smart hospital ward remote video monitoring and management system is disclosed. This system achieves comprehensive capture of multi-source data within the ward through the collaborative operation of a video image acquisition module and a multi-dimensional parameter sensing module. The high-definition camera array of the video image acquisition module captures image information such as patient position and equipment status, while the distributed units of the multi-dimensional parameter sensing module collect sensor data such as temperature, humidity, and oxygen concentration. Both types of data are transmitted to edge computing nodes via fiber optic links and wireless mesh networks, respectively. The feature mapping unit of the hypergraph feature extraction module maps the image data to a high-dimensional hypergraph space, and the feature fusion unit performs weighted fusion of multimodal data based on the weight coefficients of the hypergraph edges. This breaks down the information barriers caused by single-modal analysis or simple splicing in traditional technologies, enabling cross-modal data to form deep associations within the hypergraph framework, effectively overcoming the problems of information fragmentation and redundancy.

[0050] The system provides localized computing power support for the HyperGraph neural network through edge computing nodes, significantly improving data processing efficiency. The inference computing unit of the model training and inference module loads the HyperGraph neural network model and calculates the dimensionality-reduced feature vectors transmitted from the HyperGraph feature extraction module to obtain initial inference results. The result correction unit adjusts the initial results based on the dynamic parameters of the edge computing nodes, outputting accurate corrected results. The decision instruction generation unit of the intelligent decision execution module generates control instructions based on the corrected results. After decoding, the instruction execution unit drives the relay control array to achieve real-time adjustment of ward equipment, significantly shortening the overall link time from data acquisition to equipment response and solving the problem of response lag in centralized architectures.

[0051] The system employs a robust closed-loop feedback mechanism to ensure the dynamic adaptability of its decision-making logic. The feature verification unit of the Hypergraph feature extraction module validates the dimensionality-reduced features; if anomalies are found, they are fed back to the feature fusion unit to prompt re-fusion. The status monitoring unit of the intelligent decision execution module collects real-time equipment operating status data and feeds it back to the decision command generation unit to adjust subsequent commands. This two-way feedback mechanism enables the system to dynamically adjust its decision-making logic based on real-time changes in parameters within the ward, avoiding the inflexibility issues caused by fixed decision-making patterns in traditional systems. This achieves precise control over the ward's status and improves overall management efficiency.

[0052] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart hospital ward remote video monitoring and management system, characterized in that, include: The video image acquisition module establishes a two-way data interaction channel with the edge computing node through an optical fiber transmission link to capture raw image data streams of patient position changes, equipment operating status, and ambient light and shadow distribution in the ward in real time. The multi-dimensional parameter sensing module includes a distributed temperature and humidity sensor, an air pressure monitor, an oxygen concentration detector, and an electrocardiogram signal acquisition terminal. Each sensing unit is connected to the edge computing gateway via a wireless Mesh network protocol. The collected physical quantities are converted into electrical signals and then converted from analog to digital. The quantized data is then transmitted to the edge computing node in a time-division multiplexing manner. The hypergraph feature extraction module receives image feature vectors preprocessed by the edge computing node from the video image acquisition module and quantized data output by the multidimensional parameter sensing module. It constructs cross-modal data association through a feature mapping matrix based on the hypergraph topology. The internal feature fusion unit and the cache unit of the edge computing node exchange data through the high-speed PCIe bus. The model training and inference module is equipped with an optimized HyperGraph neural network computing engine. It receives the fused feature tensor output by the HyperGraph feature extraction module through an Ethernet interface, and uses the computing power provided by the edge computing nodes to perform nonlinear mapping and high-dimensional space reconstruction of feature vectors. The output inference results are sent to the remote monitoring center through an encrypted transmission protocol. The intelligent decision execution module consists of a decision instruction generation unit based on FPGA architecture and a multi-channel relay control array. The decision instruction generation unit receives the decision matrix output by the model training and inference module through a serial communication interface. After logical operation, the generated control signal drives the relay control array through an opto-isolation circuit to adjust the lighting, temperature control and medical equipment in the ward in real time. The remote data interaction module uses 5G slicing technology to build a dedicated communication link. One end is connected to the output port of the intelligent decision execution module through an optical transceiver, and the other end establishes a data transmission channel with the remote server cluster through the core network interface. The feature data, decision instructions and equipment status information generated during system operation are transmitted bidirectionally in the form of encrypted data packets.

2. The system according to claim 1, characterized in that, The high-definition camera array in the video image acquisition module constructs an image feature weight allocation mechanism through an optimized hypergraph neural network and edge computing multidimensional heterogeneous ward data analysis model. The feature mapping relationship satisfies: in, This represents the feature weight coefficient of the j-th pixel in the i-th frame of the image captured by the k-th camera. The temperature sensitivity coefficient is dynamically adjusted for edge computing nodes. The number of monitoring areas within the ward. This represents the light intensity attenuation coefficient of the m-th monitoring area in the i-th frame image. Let J be the color channel response value of the j-th pixel within the m-th monitoring area. These are the spectral filtering parameters for the k-th camera. This represents the total number of pixels in a single frame image; simultaneously, this module performs time synchronization correction on multiple cameras through edge computing nodes, and the synchronization error correction model is as follows: in, This represents the time synchronization error correction value between the k-th and ith cameras. This is the base clock cycle for edge computing nodes. The number of reference markers in the ward. Let be the coordinate transformation coefficient of the p-th reference marker point on the imaging plane of the k-th camera. It is the physical distance between the k-th and ith cameras. At the speed of light, Let be the reflectance coefficient of the p-th reference marker.

3. The system according to claim 1, characterized in that, Each sensing unit in the multidimensional parameter sensing module constructs a parameter spatiotemporal correlation model based on an optimized hypergraph neural network and edge computing multidimensional heterogeneous ward data analysis model. The data fusion process satisfies the following: in, This represents the fused sensing value of the q-th sensing region at time t. The spatiotemporal weighting coefficients dynamically allocated to edge computing nodes. This represents the number of static sensing units in the area. Let be the spatial attenuation factor of the s-th static sensing unit in the q-th region. Let be the sampled value of the s-th static sensing unit at time t. The number of dynamic sensing units. Let r be the time response coefficient of the r-th dynamic sensing unit in the q-th region. For the r-th dynamic sensing unit at time... The sampled values, The length of the time window set for edge computing nodes; This module transmits the calculated data via a wireless mesh network. When the feature mapping unit transmitted to the Hypergraph feature extraction module performs cross-modal association with video image features, it satisfies the following: in, Let represent the value of the m-th fused feature at time t. Let be the correlation coefficient between the q-th sensing region and the m-th fused feature. This represents the total number of sensing areas. For the number of cameras, Let be the grayscale value of the j-th pixel in the i-th frame at time t.

4. The system according to claim 1, characterized in that, The hypergraph feature extraction module utilizes an optimized hypergraph neural network to construct a feature topology, and the hypergraph node association strength calculation satisfies: in, This represents the association strength between node a and node b in the hypergraph. To integrate the dimensions of features, Let be the sensitivity coefficient of the a-th node to the c-th fused feature. Let be the sensitivity coefficient of the b-th node to the c-th fused feature. Let c be the weight factor of the fused feature in the hypergraph. The length of the time series. Let be the value of the c-th fused feature at time t; When edge computing nodes perform dimensionality reduction on hypergraph features, the following conditions must be met: in, This represents the value of the d-th eigenvector after dimensionality reduction. Let be the mapping coefficient from the a-th hypergraph node to the d-th dimension-reduced feature. This represents the total number of nodes in the hypergraph. The number of associated nodes. The time decay coefficient, The reference time set for edge computing nodes. Let be the time derivative of the fusion feature corresponding to the a-th hypergraph node at time t.

5. The system according to claim 1, characterized in that, The model training and inference module uses an optimized hypergraph neural network for feature inference, and the inference output value satisfies: in, This represents the output value of the p-th inference result. The dimension of the reduced-dimensional features. The connection weights from the d-th dimensionality-reduced feature to the p-th output are: is the activation function of the hypergraph neural network. This represents the number of hidden layer nodes. Let be the connection weight from the d-th dimensionality-reduced feature to the e-th hidden layer node. Let be the bias term for the e-th hidden layer node. This is the bias term for the p-th output result; When edge computing nodes dynamically correct inference results, the following conditions must be met: in, This represents the p-th reasoning result after correction. For the number of correction factors, Let f be the influence coefficient of the f-th correction factor on the p-th inference result. Let f be the change in the dimensionality reduction feature. The time decay coefficient of the f-th correction factor. This refers to the current moment.

6. The system according to claim 1, characterized in that, The decision instruction generation unit of the intelligent decision execution module is based on the output of the model training and inference module. Construct a control instruction matrix such that the instruction output values ​​satisfy: in, This represents the output value of the q-th control instruction. For the number of inference results, Let be the influence coefficient of the p-th inference result on the q-th control command. For the floor function, is the quantization coefficient of the q-th control command; The drive signals for the relay control array satisfy: in, This represents the drive signal value of the r-th relay. To control the number of instructions, Let be the driving coefficient of the q-th control command for the r-th relay. For symbolic functions, Let be the driving nonlinearity coefficient of the r-th relay.

7. The system according to claim 1, characterized in that, The hypergraph feature extraction module includes: The feature mapping unit receives image feature vectors from the video image acquisition module and quantized data from the multidimensional parameter sensing module. By constructing the correspondence between nodes and edges in the hypergraph topology, it maps feature data of different modalities to the high-dimensional space of the hypergraph, so that image features and sensing parameters are associated under the same hypergraph framework. It also transmits data to other units through the internal data conversion interface. The feature fusion unit obtains the hypergraph node features output by the feature mapping unit, and performs weighted combination of the features of different nodes according to the weight coefficients of the edges in the hypergraph, so that the scattered feature information is integrated into a fusion feature set with correlation. The fused features are transmitted to the feature dimensionality reduction unit through the high-speed data bus, and the intermediate data in the fusion process is temporarily stored in the internal cache. The feature dimensionality reduction unit receives the high-dimensional fused feature set output by the feature fusion unit, and uses a dimensionality reduction algorithm based on a hypergraph structure to reduce the dimensionality of the feature data while retaining the calibration feature information. The dimensionality-reduced feature data is sent to the model training and inference module through the PCIe bus interface, and at the same time, it feeds back the parameter changes during the dimensionality reduction process to the feature mapping unit. The feature verification unit verifies the dimensionality reduction features output by the feature dimensionality reduction unit. By comparing them with the preset feature threshold range in the hypergraph, it determines whether the dimensionality reduction features meet the input requirements of the model training and inference module. If there are features that exceed the range, a verification signal is generated and sent to the feature fusion unit to prompt it to perform the feature fusion operation again. The feature data that passes the verification is directly transmitted to the model training and inference module.

8. The system according to claim 1, characterized in that, The model training and inference module includes: The inference computing unit receives the dimensionality reduction feature data transmitted by the hypergraph feature extraction module, loads the optimized hypergraph neural network model parameters, and maps the dimensionality reduction features to the corresponding inference results through the computation processing of multi-layer neurons. The intermediate data generated during the calculation process is stored in the internal temporary cache area, and the inference results are sent to the result correction unit through the output interface. The result correction unit obtains the initial inference result output by the inference calculation unit, and adjusts the initial result by combining the dynamic parameters of the ward environment transmitted in real time by the edge computing node, so as to eliminate the inference deviation caused by environmental changes. The corrected result is transmitted to the instruction conversion unit through the data bus, and the correction parameters are fed back to the inference calculation unit at the same time. The model update unit periodically receives new feature data and corresponding annotation results transmitted by the remote data interaction module. It uses the computing power resources of edge computing to update the model parameters of the hypergraph neural network, so that the model can adapt to the changes in feature data in the ward. The updated model parameters are loaded into the inference computing unit through the internal parameter configuration interface. The data feedback unit collects operational data from the inference calculation unit and the result correction unit, including inference time and correction magnitude information. After processing this data, it sends it to the remote data interaction module through a dedicated data channel. At the same time, it provides model operation status data to the model update unit as a reference for model updates.

9. The system according to claim 1, characterized in that, The intelligent decision execution module includes: The decision instruction generation unit receives the corrected inference results output by the model training inference module, and converts the inference results into specific control instructions according to the preset decision logic and control rules. These instructions are stored in the internal register in the form of binary code and sent to the instruction execution unit through the control bus. The instruction execution unit acquires the control instructions output by the decision instruction generation unit, decodes the instructions, determines the type of ward equipment to be controlled, the control method and control parameters, transmits the decoded control signal to the relay control array through the drive circuit, and monitors the status of instruction execution in real time. The relay control array receives control signals from the instruction execution unit and controls the switching and brightness adjustment of lighting fixtures in the ward, the temperature setting of air conditioning equipment, the flow control of oxygen supply devices, and the start and stop of medical instruments through the on and off state changes of the relays. Its working status is transmitted to the status monitoring unit through the status feedback circuit. The status monitoring unit collects the working status of each relay in the relay control array in real time, including on / off status, current value and temperature parameters. After converting the status data into digital signals, it feeds back to the decision command generation unit as the basis for command adjustment, and sends it to the remote data interaction module for remote transmission of status information.

Images