Intelligent early warning and processing decision method and system for old-age care equipment based on internet of things

Abstract

The application relates to an Internet of Things-based intelligent early warning and processing decision method and system for an elderly care device, and relates to the field of the Internet of Things. The technical scheme effectively solves the pain points of a high false alarm rate, difficulty in fusing multi-source heterogeneous data, lag in emergency response decision-making and privacy leakage risk in existing elderly care monitoring, adopts an end-edge-cloud collaborative hierarchical processing architecture, realizes the leap of elderly care services from passive monitoring to active intervention, the edge side can realize real-time physiological signal anomaly detection and feature extraction through an improved LSTM self-encoder, data privacy protection and data dimension reduction are realized, the cloud end utilizes improved D-S evidence theory, effectively fuses multi-type sensor data, solves the sensor conflict and data uncertainty problem, combines EHR data, constructs a dynamic risk assessment model through fuzzy comprehensive evaluation and MDP, realizes graded early warning, and through an NSGA-II multi-objective genetic algorithm, optimizes emergency resource scheduling and nursing task allocation under complex constraints.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) technology, and in particular to an IoT-based intelligent early warning and decision-making method and system for elderly care equipment. Background Technology

[0002] As the global population continues to age, the care needs of empty-nest elderly and disabled elderly are becoming increasingly prominent, and related care issues have become a focus of high concern for the whole society. Although existing smart elderly care systems have integrated various Internet of Things (IoT) devices such as smart bracelets, fall detection radars, and emergency call devices in an attempt to improve care efficiency through technological means, they still face many deep-seated technical bottlenecks and practical pain points in actual application, making it difficult to meet the needs of precise and intelligent elderly care.

[0003] Specifically, the technical pain points of existing smart elderly care systems are mainly reflected in five aspects: First, single-modal monitoring has limitations and a high false alarm rate. Existing devices mostly rely on single threshold logic to trigger alarms, ignoring the influence of the elderly's movement status, past medical history, and environmental factors. The false alarm rate of single sensors remains high in complex living scenarios, which can easily lead to "alarm fatigue" among caregivers, thus missing the opportunity to deal with real emergencies. Second, the fusion of multi-source heterogeneous data is difficult. The one-dimensional time-series physiological data, three-dimensional spatial behavioral data, and scalar environmental data involved in elderly care scenarios are difficult to effectively fuse in the same dimension due to differences in equipment brands, sampling rates, and communication protocols. The phenomenon of "data silos" is prominent, which makes it impossible for the system to form a panoramic understanding of the elderly's health status. Third, there is a lack of dynamic adaptive decision-making mechanisms. Traditional systems The early warning logic is fixed and static, unable to dynamically adjust risk thresholds according to the elderly's health decline trend. Moreover, the decision-making after anomaly detection relies heavily on human experience, lacking intelligent scheduling algorithms based on location, qualifications, and current load, which can easily lead to wasted golden rescue time. Fourth, privacy and data transmission delays are prominent issues. Uploading all raw physiological data to the cloud for processing not only consumes a lot of bandwidth and increases processing delays, but also poses a serious risk of privacy leakage. The lack of effective edge computing mechanisms and end-to-end high-level encryption solutions has become a common shortcoming of existing smart elderly care systems. Fifth, uneven resource allocation and complex subsidy calculations are also problems. In the scenario of government-purchased elderly care services, how to automatically calculate subsidies based on the elderly's care level, service duration, and policy standards, and achieve matching with service work orders, remains a major challenge for the current management system. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent early warning and decision-making method and system for elderly care equipment based on the Internet of Things, so as to solve the problems existing in the prior art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides an intelligent early warning and decision-making method for elderly care equipment based on the Internet of Things, the method comprising: A hierarchical processing architecture is constructed that integrates the edge computing layer, the cloud decision-making layer, and the application execution layer, which are connected in sequence. Multi-dimensional raw monitoring data is collected through the sensing access layer, and the raw monitoring data is connected to the hierarchical processing architecture through a preset communication protocol. The original monitoring data is preprocessed by the edge computing layer, and the built-in Long Short-Term Memory Network (LSTM) autoencoder is called to perform local anomaly detection on the preprocessed time series data, generating anomaly feature vectors and anomaly confidence. Only the anomaly-related data is encrypted and uploaded to the cloud decision layer. The edge computing layer and the cloud decision layer use a reflective encrypted communication mechanism for data interaction. The cloud-based decision-making layer uses an improved Dempster-Shafer evidence theory to perform spatiotemporal evidence fusion on multi-source anomalous evidence to obtain the final fusion confidence level of the current elderly person's condition; Based on the final fusion confidence level and combined with the elderly’s electronic health record (EHR) data, a dynamic risk assessment system is constructed through a fuzzy comprehensive evaluation model to determine the risk level of the current event. Based on the risk level determined, corresponding emergency response plans are triggered. For events of high risk or above, the non-dominated sorting genetic algorithm NSGA-II with elitist strategy is used to optimize multi-objective resource scheduling and generate the optimal rescue resource scheduling plan. The application execution layer executes corresponding emergency operations based on the optimal rescue resource scheduling plan and feeds the execution results back to the cloud decision layer, thus completing the closed loop of intelligent early warning and processing decision-making.

[0006] In some implementations, the preprocessing of the original monitoring data through an edge computing layer, and the invocation of a built-in Long Short-Term Memory (LSTM) autoencoder to perform local anomaly detection on the preprocessed time-series data, generating anomaly feature vectors and anomaly confidence scores, specifically includes: Kalman filtering at the edge computing layer removes sensor noise from the original monitoring data and normalizes data of different dimensions to the [0, 1] interval. The LSTM autoencoder is invoked to extract and reconstruct features from the normalized time-series data. The state update formula for the LSTM unit is as follows: Forgotten Gate: ; Input Gate: ; ; Cell status update: ; Output gate: ; ; in, is the Sigmoid activation function; tanh is the hyperbolic tangent function; This represents the Hadamard product, which is an element-wise multiplication operation. , , , These are the weight matrices corresponding to the forget gate, input gate, cell state, and output gate, respectively. , , , These are the bias terms corresponding to the forget gate, input gate, cell state, and output gate, respectively. Let be the input vector at time t, with dimension d, where d is the number of physiological features input at a single time. Let be the hidden state at time t. This represents the hidden state at time t-1; Let t represent the cell state at time t. The cell state at time t-1; The mean square error of time series data reconstruction within the sliding window is used as the outlier score. : ; Where L is the time length of the sliding window; This represents the original input data at time t; The reconstructed output data of the autoencoder at time t; Square operations for Euclidean norms; The decision threshold is dynamically generated based on Chebyshev's inequality or Gaussian distribution assumption. Defined as: ; in, This represents the average reconstruction error of a historical normal window. The standard deviation of the reconstruction error of the historical normal window; This is the sensitivity coefficient, with a default value of 3. when When an anomaly is identified, the anomaly confidence level is calculated. ; in, The function is for finding the minimum value; This is a preset scaling factor used to adjust the output range of anomaly confidence.

[0007] In some implementations, the cloud-based decision-making layer employs a modified Dempster-Shafer evidence theory to perform spatiotemporal evidence fusion on multi-source anomalous evidence to obtain the final fusion confidence level of the current elderly person's state, specifically including: Building an identification framework ,in These correspond to three mutually exclusive elderly states: normal, suspected risk, and emergency help. The value is 3 to identify the number of elements within the frame; For N sensor evidence sources, acquire the data from each sensor. The basic probability assignment (BPA) of the output is denoted as... ,satisfy: ; in, A is an empty set; A is the identification frame. Any subset of is also called a focal element; Let i be the basic probability assignment value of focus element A to evidence source i; The difference between the two pieces of evidence was measured using the Jousselme distance metric. and Jousselme distance between for: ; Where the superscript T denotes the matrix transpose operator; D denotes the matrix transpose operator. The correlation matrix, the elements within the matrix A and B are any two focal elements within the identification frame. Let be the number of elements in the intersection of focal elements A and B. Let f be the number of elements in the union of focal elements A and B; - Assign vector differences to the basic probabilities of the two pieces of evidence; Calculate the support of evidence i : ; Where N is the total number of sensor evidence sources; j is the sequence number of other evidence sources besides evidence source i; Let Jousselme distance be the distance between evidence i and evidence j; Normalization yields the credibility weights : ; Where k is the sequence number of all evidence sources; The sum of support from all sources of evidence; Let be the normalized confidence weight corresponding to evidence source i, and the sum of all weights is 1; The original BPA was corrected by weighted average to obtain the corrected evidence. : ; in, This is the corrected basic probability assignment value corresponding to focal element A; Using Dempster's orthogonal combination rule to analyze the modified evidence Perform N-1 self-fusions to obtain the final fusion result. The formula is: ; in, For Dempster orthogonal combination operators; A1 to A N To identify any focal element within the frame; The intersection of all focal elements is the target focal element A; K is the evidence conflict coefficient, denoted as: ; in, This indicates that the intersection of all focal elements is an empty set, meaning there is a complete conflict between the evidence.

[0008] In some implementations, the step of constructing a dynamic risk assessment system based on the final fusion confidence level and the elderly person's electronic health record (EHR) data, using a fuzzy comprehensive evaluation model, to determine the risk level of the current event, specifically includes: Constructing an evaluation index set ,in To integrate anomaly confidence levels in real time, For historical health records risk coefficient, For environmental risk index, In response to resource scarcity; Constructing a set of evaluation statements ,in These represent four assessment levels: low risk, medium risk, high risk, and extremely high risk. The weight vector of each evaluation index is objectively determined using the entropy weight method. ,in These are the weights corresponding to the four evaluation indicators, and the sum of all weights is 1; The membership degree of the i-th indicator to the j-th risk level is calculated using the trapezoidal distribution membership function. Construct the membership matrix Where i is the evaluation indicator number and j is the risk level number. The value range is [0, 1]; The comprehensive evaluation result is obtained by performing fuzzy linear transformation. : ; in, For fuzzy linear composition operators, These represent the overall membership degree of the current event to the four risk levels; Determine the current risk level based on the principle of maximum membership. for: ; in, To obtain the risk level operator corresponding to the maximum value, the evaluation set element corresponding to the maximum comprehensive membership degree is selected as the final risk level.

[0009] In some implementations, for events of high risk or higher, the non-dominated sorting genetic algorithm NSGA-II with an elitist strategy is used for multi-objective resource scheduling optimization to generate an optimal rescue resource scheduling scheme, specifically including: Construct three conflicting optimization objective functions: Objective 1: Minimize total response time ; Objective 2: Minimize service costs ; Objective 3: Maximize service matching ; Where K is the set of available rescue resources, k is the sequence number of the rescue resource; V is the set of task nodes, i and j are the sequence numbers of the task nodes; The path selection variable is set to 0-1. A value of 1 indicates that the rescue resource k travels from node i to node j, while a value of 0 indicates that the path is not selected. Let be the travel time from node i to node j. For the on-site service time of node i; For the fixed cost of a single service, The cost of service varies per unit distance. The total travel distance for rescue resource k; Assign variables for tasks ranging from 0 to 1. A value of 1 indicates that task i is assigned to rescue resource k, while a value of 0 indicates that it is not assigned. The skill matching degree between rescue resource k and task i, with a value range of [0, 1]; Set constraints, including: uniqueness constraint, each emergency task must be assigned to one and only one rescue resource; golden rescue window constraint, the arrival time of the rescue resource must be within the preset golden rescue time; and capacity constraint, the number of tasks that a single rescue resource can simultaneously handle cannot exceed its maximum capacity. The resource allocation scheme is encoded using an integer permutation encoding method. The population is iteratively optimized through fast non-dominated sorting, crowding distance calculation, elite retention strategy, crossover and mutation operations until the preset iteration termination condition is reached. The knee point solution is selected from the Pareto optimal solution set after iterative convergence as the final optimal rescue resource scheduling scheme.

[0010] In some implementations, a reflective encrypted communication mechanism is used between the edge computing layer and the cloud decision layer, specifically including: The edge gateway always acts as the initiator of communication, while the cloud service acts as the passive responder, thus avoiding security vulnerabilities caused by proactive cloud scanning. The edge gateway and cloud services negotiate session keys using the ECDH elliptic curve Diffie-Hellman key negotiation algorithm. All upstream physiological data are encrypted using the AES-GCM-256 algorithm. The encryption formula is as follows: ; in, For session key based The AES-GCM-256 encryption function, where Data is the original upstream data, Timestamp is the data timestamp, and Signature is the data digital signature. Cipher is the data concatenation operator, and Cipher is the encrypted ciphertext data; Control commands issued from the cloud must carry a digital signature based on a private key. The edge device can only execute the corresponding command after verifying the signature's validity.

[0011] In some implementations, after the step of constructing a dynamic risk assessment system based on the final fusion confidence level and the elderly person's electronic health record (EHR) data, using a fuzzy comprehensive evaluation model to determine the risk level of the current event, the method further includes: The comprehensive evaluation result obtained from the fuzzy linear transformation As the state inputs to a Markov decision process, an MDP model is constructed, comprising: State space: ; Action space: ; Transition probability: ; Reward function: ; in, Let t represent the risk state. The risk state at time t+1; The decision action to be performed at time t; To perform the action After that, the risk status changed from Transferred to The probability of; Cost is the health benefit value after the action is executed, and Cost is the system cost of executing the action. , These are the weighting coefficients corresponding to health benefits and system overhead, respectively. Finding the optimal strategy using the Q-Learning algorithm The iterative update formula is: ; in, Let be the action value function for performing action a in state s. The learning rate; Based on the optimal strategy after training, the hierarchical processing architecture automatically selects the decision action with the best long-term return according to the current risk status.

[0012] In some implementations, before the step of calling the LSTM autoencoder to extract and reconstruct features from the normalized time-series data, the following steps are included: The 32-bit floating-point weight matrix of the LSTM autoencoder is quantized using INT8 weight quantization. The quantization formula is as follows: ; Where q is the quantized 8-bit integer value; r is the original 32-bit floating-point true value; S is the scaling factor; and Z is the zero-point offset. This is the rounding function; The weight matrix is ​​sparsified using the amplitude pruning algorithm, as shown in the formula: ; in, W is the pruned weight matrix; W is the original weight matrix. The preset pruning threshold; For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. The pruned weight matrix is ​​stored using a sparse matrix storage format, and the quantized LSTM autoencoder is accelerated for inference using the neural network processing unit (NPU) built into the edge gateway.

[0013] In some implementations, the triggering of the corresponding emergency response plan based on the determined risk level specifically includes: When the risk level is determined to be low, the hierarchical processing architecture records the operation log and silently monitors the elderly person's condition. When the risk level is determined to be medium, a reminder message will be pushed to the guardian's APP, and the device's voice inquiry function will be activated at the same time; When the risk level is determined to be high or extremely high, an automatic dispatch order is triggered to coordinate with 120 emergency services, and the highest level alarm information is sent to the guardian at the same time. For all events that generate service work orders, the subsidy settlement module automatically matches the elderly care level and service duration according to preset policy rules, calculates the service fee and government subsidy deduction amount, and stores the service records and subsidy settlement information together.

[0014] Secondly, the present invention provides an intelligent early warning and processing decision system for elderly care equipment based on the Internet of Things, for executing the above-mentioned intelligent early warning and processing decision method for elderly care equipment based on the Internet of Things. The system includes a sensing access layer, an edge computing layer, a cloud decision layer and an application execution layer that are connected in sequence. The perception access layer is used to collect multi-dimensional raw monitoring data and connect the raw monitoring data to the system through a preset communication protocol. The perception access layer includes at least one of smart wearable devices, environmental sensing devices, medical terminals, and security devices. The edge computing layer includes a multi-protocol adaptation module, a data cleaning and normalization module, a local anomaly detection engine, and an encrypted communication module, which are used to preprocess the raw monitoring data and perform local anomaly detection, and to complete encrypted data interaction with the cloud. The cloud-based decision-making layer includes a multi-source evidence fusion module, a comprehensive risk assessment module, a resource scheduling and dispatching module, and a subsidy settlement module, which are used to complete multi-source data fusion, risk level determination, emergency resource scheduling, and service subsidy settlement. The application execution layer includes a guardian APP, a caregiver PDA, a large-screen monitoring center, and a 120 emergency response module, which are used to perform emergency response operations and feed back the execution results to the cloud decision layer.

[0015] The beneficial effects of the technical solution provided by this invention include at least the following: This technical solution effectively addresses pain points in existing elderly care monitoring, such as high false alarm rates, difficulties in fusing multi-source heterogeneous data, delayed emergency response decisions, and privacy risks. It adopts a layered processing architecture with edge-cloud collaboration, enabling a shift in elderly care services from passive monitoring to proactive intervention. At the edge, an improved LSTM autoencoder enables real-time detection of physiological signal anomalies and feature extraction, protecting data privacy while reducing data dimensionality and transmission pressure. In the cloud, an improved DS evidence theory effectively fuses data from multiple sensor types, resolving sensor conflicts and data uncertainty. Combined with EHR data, a dynamic risk assessment model is constructed using fuzzy comprehensive evaluation and MDP to achieve tiered early warning. Furthermore, the NSGA-II multi-objective genetic algorithm optimizes emergency resource scheduling and nursing task allocation under complex constraints. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0017] Figure 1 The diagram illustrates a flowchart of an intelligent early warning and decision-making method for elderly care equipment based on the Internet of Things, provided by an exemplary embodiment of the present invention.

[0018] Figure 2 The diagram illustrates a timing diagram of reflective encrypted communication between an edge gateway and a cloud service for an IoT-based intelligent early warning and processing decision-making method for elderly care equipment, provided by an exemplary embodiment of the present invention.

[0019] Figure 3 This illustration shows a schematic diagram of the logic of an improved DS evidence theory fusion algorithm based on Jousselme distance, which is provided by an exemplary embodiment of the present invention for an intelligent early warning and processing decision-making method for elderly care equipment based on the Internet of Things.

[0020] Figure 4 This illustration shows a schematic diagram of a multi-objective rescue resource scheduling process based on the NSGA-II algorithm, which is an intelligent early warning and processing decision-making method for elderly care equipment based on the Internet of Things provided by an exemplary embodiment of the present invention.

[0021] Figure 5 The diagram shows a structural block diagram of an intelligent early warning and processing decision system for elderly care equipment based on the Internet of Things, provided by an exemplary embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Figure 1 The diagram illustrates a flowchart of an IoT-based intelligent early warning and decision-making method for elderly care equipment, provided by an exemplary embodiment of the present invention. This IoT-based intelligent early warning and decision-making method for elderly care equipment includes: Step 101: Construct a hierarchical processing architecture that integrates the end, edge, and cloud. The hierarchical processing architecture includes a perception access layer, an edge computing layer, a cloud decision layer, and an application execution layer that are connected in sequence.

[0025] In this embodiment, the perception access layer is responsible for collecting multi-dimensional raw data, and the perception access layer includes, but is not limited to: Smart wearable devices are used to collect ECG (electrocardiogram), PPG (photoplethysmography), SpO2 (oxygen saturation), body temperature, and acceleration data; Environmental sensing devices include at least millimeter-wave radar (breathing / heartbeat / presence detection), infrared door magnetic sensors, smoke / gas alarms, and water immersion sensors; Medical terminals include at least Bluetooth blood pressure monitors, blood glucose meters, and sleep monitoring belts; Security equipment includes at least cameras with human detection capabilities (video stream is activated only under specific alarm conditions). The above devices access a hierarchical processing architecture via MQTT, CoAP, or HTTP protocols, and support breakpoint resume and heartbeat keep-alive mechanisms.

[0026] In this embodiment, the edge computing layer is deployed on the smart gateway in the elderly person's home or at a floor node (similar to the large screen driver board in the reference patent). The edge computing layer includes, but is not limited to: The multi-protocol adaptation module is used to parse the proprietary protocols of different vendors and convert them into the system's internal standard JSON format. The data cleaning and normalization module uses a Kalman filter to remove sensor noise and normalizes data of different dimensions (such as heart rate and acceleration values) to a range. The local anomaly detection engine has a built-in LSTM autoencoder model for real-time inference of time series data. The encrypted communication module is responsible for secure handshakes with the cloud and data encryption / decryption.

[0027] In this embodiment, the cloud-based decision layer is the brain of the hierarchical processing architecture, deployed in a private or hybrid cloud environment. The cloud-based decision layer includes, but is not limited to: The multi-source evidence fusion module is used to receive evidence (i.e., anomaly probability) uploaded from the edge layer and calculate the global state; The comprehensive risk assessment module is used to determine the risk level by combining EHR (electronic health record) with real-time status; The resource scheduling and dispatch module is used to manage nursing staff, volunteers, and medical resources, and to execute automatic dispatch. The subsidy settlement module is used to automatically calculate service fees and government subsidy deductions based on policy rules.

[0028] In this application embodiment, the application execution layer includes, but is not limited to: The guardian / children's app is used to receive tiered alarm notifications, view health reports, and make remote voice calls. The caregiver's PDA is used to receive work orders, navigate, clock in / out for services, and enter health data. The large-screen monitoring center is used to display regional heat maps, real-time alarm lists, and resource distribution maps. The 120 emergency response module is used to quickly push data such as the elderly person's identity information, real-time location, and abnormal symptoms to the 120 emergency response platform, while simultaneously notifying on-site caregivers or family members, reducing intermediate links in the transmission of emergency information and shortening emergency response time.

[0029] Step 102: Collect multi-dimensional raw monitoring data through the perception access layer, and connect the raw monitoring data to the hierarchical processing architecture through a preset communication protocol.

[0030] In this embodiment, multi-dimensional data acquisition comprehensively captures various relevant monitoring information, avoiding decision-making biases caused by data bias and ensuring coverage of needs in different scenarios. A pre-defined communication protocol resolves the compatibility issues between different types of monitoring data and the hierarchical architecture, ensuring that raw data is efficiently, stably, and without distortion transmitted to the edge computing layer and subsequent layers, avoiding data transmission interruptions, delays, or errors.

[0031] Step 103: The original monitoring data is preprocessed through the edge computing layer. The built-in Long Short-Term Memory (LSTM) autoencoder network is called to perform local anomaly detection on the preprocessed time series data, generating anomaly feature vectors and anomaly confidence scores. Only the anomaly-related data is encrypted and uploaded to the cloud decision layer. The edge computing layer and the cloud decision layer use a reflective encrypted communication mechanism for data interaction.

[0032] In some embodiments, the above-mentioned preprocessing of the original monitoring data through the edge computing layer, and the invocation of the built-in Long Short-Term Memory (LSTM) autoencoder network to perform local anomaly detection on the preprocessed time-series data, generating anomaly feature vectors and anomaly confidence scores, specifically including: Kalman filtering at the edge computing layer removes sensor noise from the original monitoring data and normalizes data of different dimensions to the [0, 1] interval. The LSTM autoencoder is used to extract and reconstruct features from the normalized time-series data. The state update formula for the LSTM unit is as follows: Forgotten Gate: ; Input Gate: ; ; Cell status update: ; Output gate: ; ; in, is the Sigmoid activation function; tanh is the hyperbolic tangent function; This represents the Hadamard product, which is an element-wise multiplication operation. , , , These are the weight matrices corresponding to the forget gate, input gate, cell state, and output gate, respectively. , , , These are the bias terms corresponding to the forget gate, input gate, cell state, and output gate, respectively. Let be the input vector at time t, with dimension d, where d is the number of physiological features input at a single time. Let be the hidden state at time t. This represents the hidden state at time t-1; Let t represent the cell state at time t. The cell state at time t-1; The mean square error of time series data reconstruction within the sliding window is used as the outlier score. : ; Where L is the time length of the sliding window; This represents the original input data at time t; The reconstructed output data of the autoencoder at time t; Square operations for Euclidean norms; The decision threshold is dynamically generated based on Chebyshev's inequality or the Gaussian distribution assumption. Defined as: ; in, This represents the average reconstruction error of a historical normal window. The standard deviation of the reconstruction error of the historical normal window; This is the sensitivity coefficient, with a default value of 3. when When an anomaly is identified, the anomaly confidence level is calculated. ; in, The function is for finding the minimum value; This is a preset scaling factor used to adjust the output range of anomaly confidence.

[0033] In this embodiment, Kalman filtering and normalization processes first denoise, purify, and standardize the original monitoring data, eliminating sensor errors and differences in data dimensions, providing a clean and unified input basis for model inference. The LSTM autoencoder captures individual patterns in physiological time-series data, achieving adaptive anomaly detection with non-fixed thresholds, accurately identifying abnormal signals deviating from normal physiological patterns. All inference is performed locally, with only abnormal features and confidence levels uploaded, achieving millisecond-level response and eliminating the risk of original privacy data being leaked. Dynamic thresholds and confidence quantification further adapt to individual physiological baseline differences, while providing standardized and quantifiable reliable input for multi-source evidence fusion in the cloud.

[0034] In some embodiments, before the step of calling the LSTM autoencoder to extract and reconstruct features from the normalized time-series data, the method further includes: The 32-bit floating-point weight matrix of the LSTM autoencoder is quantized using INT8 weight quantization. The quantization formula is as follows: ; Where q is the quantized 8-bit integer value; r is the original 32-bit floating-point true value; S is the scaling factor; and Z is the zero-point offset. This is the rounding function; The weight matrix is ​​sparsified using the amplitude pruning algorithm, as shown in the formula: ; in, W is the pruned weight matrix; W is the original weight matrix. The preset pruning threshold; For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. The pruned weight matrix is ​​stored using a sparse matrix storage format, and the quantized LSTM autoencoder is accelerated for inference using the neural network processing unit (NPU) built into the edge gateway.

[0035] In this embodiment, INT8 weight quantization compresses the bit width of model parameters through linear mapping, significantly reducing model memory usage and computational overhead with almost no loss of anomaly detection accuracy, thus lowering the computational threshold for edge hardware. Amplitude pruning removes redundant connections in the model that do not substantially contribute to feature extraction, completing weight matrix sparsity processing and further eliminating invalid computations. Combined with sparse matrix storage optimization and NPU hardware acceleration, the model ultimately achieves low-power, millisecond-level local inference.

[0036] In some embodiments, see Figure 2 The aforementioned edge computing layer and cloud decision-making layer employ a reflective encrypted communication mechanism, specifically including: The edge gateway always acts as the initiator of communication, while the cloud service acts as the passive responder, thus avoiding security vulnerabilities caused by proactive cloud scanning. The edge gateway and cloud services negotiate session keys using the ECDH elliptic curve Diffie-Hellman key negotiation algorithm. All upstream physiological data are encrypted using the AES-GCM-256 algorithm. The encryption formula is as follows: ; in, For session key based The AES-GCM-256 encryption function, where Data is the original upstream data, Timestamp is the data timestamp, and Signature is the data digital signature. Cipher is the data concatenation operator, and Cipher is the encrypted ciphertext data; Control commands issued from the cloud must carry a digital signature based on a private key. The edge device can only execute the corresponding command after verifying the signature's validity.

[0037] In this embodiment, the reflective communication architecture designates the edge gateway as the sole communication initiator, fundamentally blocking the risks of port exposure and malicious intrusion caused by proactive cloud scanning, thus adapting to the security environment of home intranets. ECDH dynamic key negotiation combined with AES-GCM-256 high-strength encryption achieves one-time pad session protection. A cloud-based mandatory signature verification mechanism eliminates the risk of malicious commands manipulating terminal devices, meeting the privacy compliance requirements of medical-grade data without affecting real-time communication.

[0038] Step 104: The cloud-based decision-making layer uses the improved Dempster-Shafer evidence theory to perform spatiotemporal evidence fusion on multi-source anomalous evidence to obtain the final fusion confidence of the current elderly person's condition.

[0039] In some embodiments, see Figure 3 The aforementioned cloud-based decision-making layer employs an improved Dempster-Shafer evidence theory to perform spatiotemporal evidence fusion on multi-source anomalous evidence, obtaining the final fusion confidence score of the current elderly person's condition, specifically including: Building an identification framework ,in These correspond to three mutually exclusive elderly states: normal, suspected risk, and emergency help. The value is 3 to identify the number of elements within the frame; For N sensor evidence sources, acquire the data from each sensor. The basic probability assignment (BPA) of the output is denoted as... ,satisfy: ; in, A is an empty set; A is the identification frame. Any subset of is also called a focal element; Let i be the basic probability assignment value of focus element A to evidence source i; The difference between the two pieces of evidence was measured using the Jousselme distance metric. and Jousselme distance between for: ; Where the superscript T denotes the matrix transpose operator; D denotes the matrix transpose operator. The correlation matrix, the elements within the matrix A and B are any two focal elements within the identification frame. Let be the number of elements in the intersection of focal elements A and B. Let f be the number of elements in the union of focal elements A and B; - Assign vector differences to the basic probabilities of the two pieces of evidence; Calculate the support of evidence i : ; Where N is the total number of sensor evidence sources; j is the sequence number of other evidence sources besides evidence source i; Let Jousselme distance be the distance between evidence i and evidence j; Normalization yields the credibility weights : ; Where k is the sequence number of all evidence sources; The sum of support from all sources of evidence; Let be the normalized confidence weight corresponding to evidence source i, and the sum of all weights is 1; The original BPA was corrected by weighted average to obtain the corrected evidence. : ; in, This is the corrected basic probability assignment value corresponding to focal element A; Using Dempster's orthogonal combination rule to analyze the modified evidence Perform N-1 self-fusions to obtain the final fusion result. The formula is: ; in, For Dempster orthogonal combination operators; A1 to A NTo identify any focal element within the frame; The intersection of all focal elements is the target focal element A; K is the evidence conflict coefficient, denoted as: ; in, This indicates that the intersection of all focal elements is an empty set, meaning there is a complete conflict between the evidence.

[0040] In this embodiment, a standardized identification framework and basic probability allocation are used to uniformly transform the abnormal outputs of sensors of different types and protocols into quantifiable and comparable probabilistic evidence, breaking down the data fusion barriers of heterogeneous devices. The introduction of Jousselme distance for evidence weight correction accurately identifies and weakens the interference of fault and conflicting evidence, fundamentally solving the paradox of high-conflict evidence fusion in traditional DS theory and avoiding counterintuitive erroneous judgments when multiple source signals contradict each other. Finally, a weighted orthogonal fusion is used to output the global confidence level of the elderly person's state, achieving multi-device cross-validation and significantly reducing the false alarm and false negative space.

[0041] Step 105: Based on the final fusion confidence level and combined with the elderly’s electronic health record (EHR) data, a dynamic risk assessment system is constructed using a fuzzy comprehensive evaluation model to determine the risk level of the current event.

[0042] In some embodiments, the above-mentioned dynamic risk assessment system, based on the final fusion confidence level and combined with the elderly's electronic health record (EHR) data, is constructed using a fuzzy comprehensive evaluation model to determine the risk level of the current event, specifically including: Constructing an evaluation index set ,in Real-time fusion of anomaly confidence (from DS fusion results) Historical health record risk factors (EHR data, such as hypertension level). Environmental risk index (such as whether it is late at night or if there is a sudden change in temperature). In response to resource scarcity (the risk is higher if there are no caregivers nearby); Constructing a set of evaluation statements ,in These represent four assessment levels: low risk, medium risk, high risk, and extremely high risk. The weight vector of each evaluation index is objectively determined using the entropy weight method. ,in These are the weights corresponding to the four evaluation indicators, and the sum of all weights is 1; The membership degree of the i-th indicator to the j-th risk level is calculated using the trapezoidal distribution membership function. Construct the membership matrix Where i is the evaluation indicator number and j is the risk level number. The value range is [0, 1]; The comprehensive evaluation result is obtained by performing fuzzy linear transformation. : ; in, For fuzzy linear composition operators, These represent the overall membership degree of the current event to the four risk levels; Determine the current risk level based on the principle of maximum membership. for: ; in, To obtain the risk level operator corresponding to the maximum value, the evaluation set element corresponding to the maximum comprehensive membership degree is selected as the final risk level.

[0043] In this embodiment, the entropy weight method objectively assigns weights, avoiding the subjective bias of manual rule setting. It can dynamically adjust the weights of indicators according to different elderly people and different scenarios, adapting to personalized care needs. Through membership quantification and fuzzy synthesis, the difficult-to-concrete risks are transformed into standardized classification results, providing a clear decision-making basis for subsequent differentiated emergency response and resource allocation.

[0044] In some embodiments, after the step of constructing a dynamic risk assessment system based on the final fusion confidence level and the elderly person's electronic health record (EHR) data, using a fuzzy comprehensive evaluation model to determine the risk level of the current event, the following steps are also included: The comprehensive evaluation results obtained from fuzzy linear transformation As the state inputs to a Markov decision process, an MDP model is constructed. The MDP model includes: State space: ; Action space: ; Transition probability: ; Reward function: ; in, Let t represent the risk state. The risk state at time t+1; The decision action to be performed at time t; To perform the action After that, the risk status changed from Transferred to The probability of; Cost is the health benefit value after the action is executed, and Cost is the system cost of executing the action. , These are the weighting coefficients corresponding to health benefits and system overhead, respectively. Finding the optimal strategy using the Q-Learning algorithm The iterative update formula is: ; in, Let be the action value function for performing action a in state s. The learning rate; Based on the optimal strategy after training, the hierarchical processing architecture automatically selects the decision action with the best long-term return according to the current risk status.

[0045] In this embodiment, the risk results output by fuzzy comprehensive evaluation are used as state inputs. The changes in the health risk of the elderly are modeled as a continuous stochastic process, rather than a single static slice, allowing for accurate prediction of the long-term impact of different interventions on health status. The reward function achieves a two-way balance between health benefits and system operating costs, safeguarding the safety of the elderly while avoiding resource waste and unnecessary disruption caused by over-response. Through iterative self-learning of the Q-Learning algorithm, intervention strategies under different risk scenarios can be continuously optimized, adapting to the individual health characteristics of different elderly individuals, achieving an upgrade from passively triggering according to rules to autonomously learning optimal decision-making.

[0046] Step 106: Based on the determined risk level, trigger the corresponding emergency response plan. For events of high risk or above, use the non-dominated sorting genetic algorithm NSGA-II with elitist strategy to optimize multi-objective resource scheduling and generate the optimal rescue resource scheduling plan.

[0047] In some embodiments, the above-mentioned emergency response plan triggered based on the determined risk level specifically includes: When the risk level is determined to be low, the hierarchical processing architecture records the operation log and silently monitors the elderly person's status. When the risk level is determined to be medium, a reminder message will be pushed to the guardian's APP, and the device's voice inquiry function will be activated at the same time; When the risk level is determined to be high or extremely high, an automatic dispatch order is triggered to coordinate with 120 emergency services, and the highest level alarm information is sent to the guardian at the same time. For all events that generate service work orders, the subsidy settlement module automatically matches the elderly care level and service duration according to preset policy rules, calculates the service fee and government subsidy deduction amount, and stores the service records and subsidy settlement information together.

[0048] In this embodiment, differentiated tiered response matches actions with risk levels. Flexible handling of low- and medium-risk situations avoids unnecessary disturbances and alarm fatigue for family members and caregivers, balancing care safety with the privacy and tranquility of the elderly at home. Multi-objective optimized scheduling in high-risk scenarios balances response timeliness, service costs, and professional matching within limited emergency resources, maximizing the utilization of the golden rescue time. The accompanying automatic subsidy settlement module removes barriers to the implementation of government-purchased elderly care services policies.

[0049] In some embodiments, see Figure 4 For events of high risk or above, the NSGA-II non-dominated sorting genetic algorithm with an elitist strategy is used for multi-objective resource scheduling optimization to generate the optimal rescue resource scheduling scheme, specifically including: Construct three conflicting optimization objective functions: Objective 1: Minimize total response time ; Objective 2: Minimize service costs (Including the self-funded portion after deducting government subsidies) Objective 3: Maximize service matching (The degree to which caregiver skills match the needs of the elderly)

[0050] Where K is the set of available rescue resources, k is the sequence number of the rescue resource; V is the set of task nodes, i and j are the sequence numbers of the task nodes; The path selection variable is set to 0-1. A value of 1 indicates that the rescue resource k travels from node i to node j, while a value of 0 indicates that the path is not selected. Let be the travel time from node i to node j. For the on-site service time of node i; For the fixed cost of a single service, The cost of service varies per unit distance. The total travel distance for rescue resource k; Assign variables for tasks ranging from 0 to 1. A value of 1 indicates that task i is assigned to rescue resource k, while a value of 0 indicates that it is not assigned. The skill matching degree between rescue resource k and task i, with a value range of [0, 1]; Set constraints, including: uniqueness constraint, each emergency task must be assigned to one and only one rescue resource; golden rescue window constraint, the arrival time of the rescue resource must be within the preset golden rescue time; and capacity constraint, the number of tasks that a single rescue resource can simultaneously handle cannot exceed its maximum capacity. The resource allocation scheme is encoded using an integer permutation encoding method. The population is iteratively optimized through fast non-dominated sorting, crowding distance calculation, elite retention strategy, crossover and mutation operations until the preset iteration termination condition is reached. The knee point solution is selected from the Pareto optimal solution set after iterative convergence as the final optimal rescue resource scheduling scheme.

[0051] In this embodiment, by constructing three optimization objectives—response timeliness, service cost, and professional matching degree—the algorithm systematically solves the problem of traditional single-objective scheduling that suffers from inconsistencies, while also considering the operational cost control of elderly care services and ensuring that the skills of rescue personnel match the needs of the elderly, thus avoiding ineffective handling. The accompanying rigid constraints define the compliance boundaries of the scheduling scheme, ensuring its feasibility. The algorithm rapidly converges to the globally optimal solution set through elite preservation and non-dominated sorting, ultimately selecting the knee point solution to achieve optimal balance among multiple objectives, and can still efficiently output a scheduling scheme even in high-concurrency emergency scenarios.

[0052] Step 107: The application execution layer executes the corresponding emergency operation according to the optimal rescue resource scheduling plan, and feeds back the execution result to the cloud decision layer to complete the closed loop of intelligent early warning and handling decision-making.

[0053] In this embodiment, the optimal scheduling scheme output from the cloud is broken down into executable instructions adapted to different entities such as guardians, caregivers, emergency medical institutions, and management centers. These instructions are accessed through multiple terminals to ensure that emergency response actions are implemented without delay or omission. Simultaneously, the progress and results data of the entire response process are transmitted back to the cloud in real time, allowing the response effectiveness data to feed back into the front-end anomaly detection, risk assessment, and resource scheduling models, achieving self-iterative and self-optimizing hierarchical processing architecture.

[0054] To verify the effectiveness of this method, a simulation environment containing 1000 virtual elderly nodes was constructed. The specific experimental settings are as follows: Dataset: The MobiAct public dataset (containing behaviors such as falling, walking, and sitting) is used as the benchmark, and Gaussian white noise is superimposed to simulate sensor error.

[0055] Comparison algorithms: Compared with benchmark 1, single threshold judgment method (traditional scheme); compared with benchmark 2, only SVM classifier is used, without time series analysis.

[0056] The experimental results are shown in the table below: method Accuracy False Alarm Rate Miss Rate Comparison with benchmark 1 78.5% 15.2% 6.3% Comparison with benchmark 2 85.1% 8.9% 6.0% This invention 96.8% 1.2% 2.0%

[0057] The results show that, thanks to the extraction of temporal features by LSTM and the correction of conflicting evidence by DS theory, the present invention achieves an order-of-magnitude reduction in false alarm rate.

[0058] Figure 5This diagram illustrates a structural block diagram of an IoT-based intelligent early warning and processing decision-making system for elderly care equipment, provided by an exemplary embodiment of the present invention. This IoT-based intelligent early warning and processing decision-making system for elderly care equipment is used to execute the aforementioned IoT-based intelligent early warning and processing decision-making method for elderly care equipment. The system includes a sensing access layer, an edge computing layer, a cloud decision-making layer, and an application execution layer, which are sequentially connected in communication. Specifically: The perception access layer is used to collect multi-dimensional raw monitoring data and connect the raw monitoring data to the system through a preset communication protocol. The perception access layer includes at least one of smart wearable devices, environmental sensing devices, medical terminals, and security devices. The edge computing layer includes a multi-protocol adaptation module, a data cleaning and normalization module, a local anomaly detection engine, and an encrypted communication module, which are used to preprocess the raw monitoring data and detect local anomalies, and to complete encrypted data interaction with the cloud. The cloud-based decision-making layer includes a multi-source evidence fusion module, a comprehensive risk assessment module, a resource scheduling and dispatch module, and a subsidy settlement module, which are used to complete multi-source data fusion, risk level determination, emergency resource scheduling, and service subsidy settlement. The application execution layer includes the guardian's APP, caregiver's PDA, large-screen monitoring center, and 120 emergency linkage module, which are used to execute emergency response operations and feed back the execution results to the cloud decision layer; The edge computing layer hardware adopts a heterogeneous multi-core SoC architecture, including an ARM Cortex-A53 quad-core main control unit, a neural network processing unit (NPU) with a computing power of 3.0 TOPS, and a Cortex-M4 real-time coprocessor. The ARM Cortex-A53 quad-core main control unit is used to run the operating system and network protocol stack, the neural network processing unit (NPU) is used to accelerate the inference operation of the LSTM autoencoder, and the Cortex-M4 real-time coprocessor is used for GPIO control, watchdog monitoring, and power management.

[0059] In this embodiment, the smart wearable device is used to collect ECG (electrocardiogram), PPG (photoplethysmography), SpO2 (oxygen saturation), body temperature, and acceleration data; the environmental sensing device includes at least millimeter-wave radar (respiratory / heartbeat / presence detection), infrared door magnet, smoke / gas alarm, and water immersion sensor; the medical terminal includes at least a Bluetooth blood pressure monitor, blood glucose meter, and sleep monitoring belt; and the security device includes at least a camera with human detection function (video stream is only activated under specific alarm conditions); wherein, the above devices are accessed through a hierarchical processing architecture via MQTT, CoAP, or HTTP protocols, supporting breakpoint resume and heartbeat keep-alive mechanisms.

[0060] In this embodiment, the multi-protocol adaptation module is used to parse the proprietary protocols of different manufacturers and convert them into the system's internal standard JSON format; the data cleaning and normalization module uses Kalman filtering to remove sensor noise and normalizes data of different dimensions (such as heart rate and acceleration values) to a range; the local anomaly detection engine has a built-in LSTM autoencoder model to perform real-time inference on time series data; and the encrypted communication module is responsible for secure handshake with the cloud and data encryption and decryption.

[0061] In this embodiment, the multi-source evidence fusion module is used to receive evidence (i.e., anomaly probability) uploaded from the edge layer and calculate the global state; the comprehensive risk assessment module is used to combine the EHR (electronic health record) and real-time state to determine the risk level; the resource scheduling and dispatch module is used to manage nursing staff, volunteers and medical resources and execute automatic dispatch; and the subsidy settlement module is used to automatically calculate service fees and government subsidy deductions according to policy rules.

[0062] In this embodiment, the guardian / children's APP is used to receive tiered alarm push notifications, view health reports, and make remote voice calls; the caregiver's PDA is used to receive work orders, navigate, check in for services, and enter health data; the large-screen monitoring center is used to display regional heat maps, real-time alarm lists, and resource distribution maps; and the 120 emergency linkage module is used to quickly push data such as the elderly person's identity information, real-time location, and abnormal symptoms to the 120 emergency platform, while simultaneously notifying on-site caregivers or family members, reducing intermediate links in emergency information transmission and shortening emergency response time.

[0063] It should be noted that the IoT-based intelligent early warning and processing decision system for elderly care equipment provided in the above embodiments is only an example of the division of the above functional units. In actual applications, the above functions can be assigned to different functional units as needed, that is, the internal structure of the equipment can be divided into different functional units to complete all or part of the functions described above.

[0064] To better understand this application, the following example of emergency rescue for a nighttime fall in an elderly person living alone further illustrates the IoT-based intelligent early warning and decision-making method for elderly care equipment provided in this application: Scenario: The monitored subject is Mr. Li (82 years old, suffering from hypertension), who lives in a smart community apartment where this system is installed. The time is 02:15 AM.

[0065] Step S1, Edge Anomaly Detection: The smart bracelet's three-axis accelerometer captured a period of drastically fluctuating data. The edge gateway then invoked an LSTM autoencoder to calculate the reconstruction error for that time window. This value is significantly higher than the dynamic threshold for stable nighttime sleep in history. Detect anomalies at the edge and generate anomaly vectors. Confidence level ; Millimeter-wave radar: Detected a stationary human target on the bedroom floor with rapid breathing. Generate anomaly vector. ; Infrared door sensor: Indicates the bedroom door is not open, ruling out the possibility of someone leaving.

[0066] Step S2, Cloud Evidence Fusion: Received in the cloud .

[0067] Calculating the Jousselme distance revealed a high degree of consistency between the two, with a low conflict coefficient. Extremely low.

[0068] Perform weighted DS fusion to calculate the final fall confidence level. .

[0069] Step S3, Risk Classification Decision: Input fuzzy comprehensive evaluation model: real-time fusion confidence (0.98), indicating "extremely high" membership degree; EHR profile (advanced age, hypertension), indicating "high" membership degree; environmental time (02:15 AM), indicating "high" membership degree.

[0070] Fuzzy synthesis results point to (Extremely high risk / red alert).

[0071] Step S4, Intelligent Resource Scheduling: The system triggers the NSGA-II algorithm to search the resource pool: Resource A, a community security guard on duty, located 200 meters away, has no first aid qualifications; Resource B, 120 ambulance, 5 kilometers away, estimated arrival time 12 minutes; Resource C, a neighbor on the same floor (a system volunteer with a CPR certificate), is 15 meters away.

[0072] The algorithm calculates the Pareto optimal solution: Option 1: Only schedule resource B (highly specialized, but slow response). Option 2: Combine scheduling resource C (fastest response) + resource B (professional backup).

[0073] The system selects option 2 and automatically dispatches the order.

[0074] Step S5, Execution and Feedback: Neighbor C received a top-level alert on the app: "An elderly person in room 102 is suspected of falling. Please assist in checking. A temporary password for the keypad lock has been issued."

[0075] The 120 emergency center received the alarm, and the large screen displayed the elderly person's location and history of hypertension.

[0076] The children received text messages and an alarm via the app, and the system displayed "Rescue resources are being dispatched".

[0077] The subsidy system automatically records this service and deducts the fee from the elderly person's government subsidy amount according to the "emergency rescue" standard.

[0078] It is understood that the specific examples in this document are only intended to help those skilled in the art better understand this disclosure, and are not intended to limit the scope of the invention.

[0079] It is understood that in the various embodiments of this specification, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of this disclosure.

[0080] It is understood that the various implementation methods described in this specification can be implemented individually or in combination, and this disclosure does not limit them.

[0081] Unless otherwise stated, all technical and scientific terms used in this disclosure have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this specification. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items. The singular forms "a," "the," and "the" as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0082] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this specification.

[0083] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the aforementioned method implementations, and will not be repeated here.

[0084] In the several embodiments provided in this specification, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments of the systems described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0085] The above description is merely a specific embodiment of this specification, but the scope of protection of this invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this specification should be included within the scope of protection of this specification. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for intelligent early warning and decision-making for elderly care equipment based on the Internet of Things, characterized in that, The method includes: A hierarchical processing architecture is constructed that integrates the edge computing layer, the cloud decision-making layer, and the application execution layer, which are connected in sequence. Multi-dimensional raw monitoring data is collected through the sensing access layer, and the raw monitoring data is connected to the hierarchical processing architecture through a preset communication protocol. The original monitoring data is preprocessed by the edge computing layer, and the built-in Long Short-Term Memory Network (LSTM) autoencoder is called to perform local anomaly detection on the preprocessed time series data, generating anomaly feature vectors and anomaly confidence. Only the anomaly-related data is encrypted and uploaded to the cloud decision layer. The edge computing layer and the cloud decision layer use a reflective encrypted communication mechanism for data interaction. The cloud-based decision-making layer uses an improved Dempster-Shafer evidence theory to perform spatiotemporal evidence fusion on multi-source anomalous evidence to obtain the final fusion confidence level of the current elderly person's condition; Based on the final fusion confidence level and combined with the elderly’s electronic health record (EHR) data, a dynamic risk assessment system is constructed through a fuzzy comprehensive evaluation model to determine the risk level of the current event. Based on the risk level determined, corresponding emergency response plans are triggered. For events of high risk or above, the non-dominated sorting genetic algorithm NSGA-II with elitist strategy is used to optimize multi-objective resource scheduling and generate the optimal rescue resource scheduling plan. The application execution layer executes corresponding emergency operations according to the optimal rescue resource scheduling plan, and feeds back the execution results to the cloud decision layer to complete the closed loop of intelligent early warning and handling decision-making; Based on the final fusion confidence level and combined with the elderly person's electronic health record (EHR) data, a dynamic risk assessment system is constructed using a fuzzy comprehensive evaluation model to determine the risk level of the current event, specifically including: Constructing an evaluation index set ,in To integrate anomaly confidence levels in real time, For historical health records risk coefficient, For environmental risk index, In response to resource scarcity; Constructing a set of evaluation statements ,in These represent four assessment levels: low risk, medium risk, high risk, and extremely high risk. The weight vector of each evaluation index is objectively determined using the entropy weight method. ,in These are the weights corresponding to the four evaluation indicators, and the sum of all weights is 1; The membership degree of the i-th indicator to the j-th risk level is calculated using the trapezoidal distribution membership function. Construct the membership matrix Where i is the evaluation indicator number and j is the risk level number. The value range is [0, 1]; The comprehensive evaluation result is obtained by performing fuzzy linear transformation. : ； in, For fuzzy linear composition operators, These represent the overall membership degree of the current event to the four risk levels; Determine the current risk level based on the principle of maximum membership. for: ； in, To obtain the risk level operator corresponding to the maximum value, the evaluation set element corresponding to the maximum comprehensive membership degree is selected as the final risk level.

2. The intelligent early warning and processing decision-making method for elderly care equipment based on the Internet of Things as described in claim 1, characterized in that, The process involves preprocessing the original monitoring data through an edge computing layer, then calling the built-in Long Short-Term Memory (LSTM) autoencoder network to perform local anomaly detection on the preprocessed time-series data, generating anomaly feature vectors and anomaly confidence scores. Specifically, this includes: Kalman filtering at the edge computing layer removes sensor noise from the original monitoring data and normalizes data of different dimensions to the [0, 1] interval. The LSTM autoencoder is invoked to extract and reconstruct features from the normalized time-series data. The state update formula for the LSTM unit is as follows: Forgotten Gate: ; Input Gate: ; ; Cell status update: ; Output gate: ; ; in, is the Sigmoid activation function; tanh is the hyperbolic tangent function; This represents the Hadamard product, which is an element-wise multiplication operation. , , , These are the weight matrices corresponding to the forget gate, input gate, cell state, and output gate, respectively. , , , These are the bias terms corresponding to the forget gate, input gate, cell state, and output gate, respectively. Let be the input vector at time t, with dimension d, where d is the number of physiological features input at a single time. Let be the hidden state at time t. This represents the hidden state at time t-1; Let t represent the cell state at time t. The cell state at time t-1; The mean square error of time series data reconstruction within the sliding window is used as the outlier score. : ； Where L is the time length of the sliding window; This represents the original input data at time t; The reconstructed output data of the autoencoder at time t; Square operations for Euclidean norms; The decision threshold is dynamically generated based on Chebyshev's inequality or Gaussian distribution assumption. Defined as: ； in, This represents the average reconstruction error of a historical normal window. The standard deviation of the reconstruction error of the historical normal window; This is the sensitivity coefficient, with a default value of 3. when When an anomaly is identified, the anomaly confidence level is calculated. ; in, The function is for finding the minimum value; This is a preset scaling factor used to adjust the output range of anomaly confidence.

3. The intelligent early warning and processing decision-making method for elderly care equipment based on the Internet of Things as described in claim 1, characterized in that, The cloud-based decision-making layer employs an improved Dempster-Shafer evidence theory to perform spatiotemporal evidence fusion on multi-source anomalous evidence, obtaining the final fusion confidence score of the current elderly person's state, specifically including: Building an identification framework ,in These correspond to three mutually exclusive elderly states: normal, suspected risk, and emergency help. The value is 3 to identify the number of elements within the frame; For N sensor evidence sources, acquire the data from each sensor. The basic probability assignment (BPA) of the output is denoted as... ,satisfy: ； in, A is an empty set; A is the identification frame. Any subset of is also called a focal element; Let i be the basic probability assignment value of focus element A to evidence source i; The difference between the two pieces of evidence was measured using the Jousselme distance metric. and Jousselme distance between for: ； Where the superscript T denotes the matrix transpose operator; D denotes the matrix transpose operator. The correlation matrix, the elements within the matrix A and B are any two focal elements within the identification frame. Let be the number of elements in the intersection of focal elements A and B. Let f be the number of elements in the union of focal elements A and B; - Assign vector differences to the basic probabilities of the two pieces of evidence; Calculate the support of evidence i : ； Where N is the total number of sensor evidence sources; j is the sequence number of other evidence sources besides evidence source i; Let Jousselme distance be the distance between evidence i and evidence j; Normalization yields the credibility weights : ； Where k is the sequence number of all evidence sources; The sum of support from all sources of evidence; Let be the normalized confidence weight corresponding to evidence source i, and the sum of all weights is 1; The original BPA was corrected by weighted average to obtain the corrected evidence. : ； in, This is the corrected basic probability assignment value corresponding to focal element A; Using Dempster's orthogonal combination rule to analyze the modified evidence Perform N-1 self-fusions to obtain the final fusion result. The formula is: ； in, For Dempster orthogonal combination operators; A1 to A N To identify any focal element within the frame; The intersection of all focal elements is the target focal element A; K is the evidence conflict coefficient, denoted as: ； in, This indicates that the intersection of all focal elements is an empty set, meaning there is a complete conflict between the evidence.

4. The intelligent early warning and processing decision-making method for elderly care equipment based on the Internet of Things as described in claim 1, characterized in that, For events of high risk or above, the non-dominated sorting genetic algorithm NSGA-II with an elitist strategy is used for multi-objective resource scheduling optimization to generate the optimal rescue resource scheduling scheme, specifically including: Construct three conflicting optimization objective functions: Objective 1: Minimize total response time ; Objective 2: Minimize service costs ; Objective 3: Maximize service matching ; Where K is the set of available rescue resources, k is the sequence number of the rescue resource; V is the set of task nodes, i and j are the sequence numbers of the task nodes; The path selection variable is set to 0-1. A value of 1 indicates that the rescue resource k travels from node i to node j, while a value of 0 indicates that the path is not selected. Let be the travel time from node i to node j. For the on-site service time of node i; For the fixed cost of a single service, The cost of service varies per unit distance. The total travel distance for rescue resource k; Assign variables for tasks ranging from 0 to 1. A value of 1 indicates that task i is assigned to rescue resource k, while a value of 0 indicates that it is not assigned. The skill matching degree between rescue resource k and task i, with a value range of [0, 1]; Set constraints, including: uniqueness constraint, each emergency task must be assigned to one and only one rescue resource; golden rescue window constraint, the arrival time of the rescue resource must be within the preset golden rescue time; and capacity constraint, the number of tasks that a single rescue resource can simultaneously handle cannot exceed its maximum capacity. The resource allocation scheme is encoded using an integer permutation encoding method. The population is iteratively optimized through fast non-dominated sorting, crowding distance calculation, elite retention strategy, crossover and mutation operations until the preset iteration termination condition is reached. The knee point solution is selected from the Pareto optimal solution set after iterative convergence as the final optimal rescue resource scheduling scheme.

5. The intelligent early warning and processing decision-making method for elderly care equipment based on the Internet of Things according to claim 1, characterized in that, The edge computing layer and the cloud decision-making layer employ a reflective encrypted communication mechanism, specifically including: The edge gateway always acts as the initiator of communication, while the cloud service acts as the passive responder, thus avoiding security vulnerabilities caused by proactive cloud scanning. The edge gateway and cloud services negotiate session keys using the ECDH elliptic curve Diffie-Hellman key negotiation algorithm. All upstream physiological data are encrypted using the AES-GCM-256 algorithm. The encryption formula is as follows: ； in, For session key based The AES-GCM-256 encryption function, where Data is the original upstream data, Timestamp is the data timestamp, and Signature is the data digital signature. Cipher is the data concatenation operator, and Cipher is the encrypted ciphertext data; Control commands issued from the cloud must carry a digital signature based on a private key. The edge device can only execute the corresponding command after verifying the signature's validity.

6. The intelligent early warning and processing decision-making method for elderly care equipment based on the Internet of Things according to claim 1, characterized in that, Following the step of constructing a dynamic risk assessment system based on the final fusion confidence level and the elderly person's electronic health record (EHR) data, using a fuzzy comprehensive evaluation model to determine the risk level of the current event, the following steps are also included: The comprehensive evaluation result obtained from the fuzzy linear transformation As the state inputs to a Markov decision process, an MDP model is constructed, comprising: State space: ; Action space: ; Transition probability: ; Reward function: ; in, Let t represent the risk state. The risk state at time t+1; The decision action to be performed at time t; To perform the action After that, the risk status changed from Transferred to The probability of; Cost is the health benefit value after the action is executed, and Cost is the system cost of executing the action. , These are the weighting coefficients corresponding to health benefits and system overhead, respectively. Finding the optimal strategy using the Q-Learning algorithm The iterative update formula is: ； in, Let be the action value function for performing action a in state s. The learning rate; Based on the optimal strategy after training, the hierarchical processing architecture automatically selects the decision action with the best long-term return according to the current risk status.

7. The intelligent early warning and processing decision-making method for elderly care equipment based on the Internet of Things according to claim 2, characterized in that, Before the step of calling the LSTM autoencoder to extract and reconstruct features from the normalized time-series data, the following steps are also included: The 32-bit floating-point weight matrix of the LSTM autoencoder is quantized using INT8 weight quantization. The quantization formula is as follows: ； Where q is the quantized 8-bit integer value; r is the original 32-bit floating-point true value; S is the scaling factor; and Z is the zero-point offset. This is the rounding function; The weight matrix is ​​sparsified using the amplitude pruning algorithm, as shown in the formula: ; in, W is the pruned weight matrix; W is the original weight matrix. The preset pruning threshold; For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. The pruned weight matrix is ​​stored using a sparse matrix storage format, and the quantized LSTM autoencoder is accelerated for inference using the neural network processing unit (NPU) built into the edge gateway.

8. The intelligent early warning and processing decision-making method for elderly care equipment based on the Internet of Things according to claim 1, characterized in that, The emergency response plan triggered based on the determined risk level specifically includes: When the risk level is determined to be low, the hierarchical processing architecture records the operation log and silently monitors the elderly person's condition. When the risk level is determined to be medium, a reminder message will be pushed to the guardian's APP, and the device's voice inquiry function will be activated at the same time; When the risk level is determined to be high or extremely high, an automatic dispatch order is triggered to coordinate with 120 emergency services, and the highest level alarm information is sent to the guardian at the same time. For all events that generate service work orders, the subsidy settlement module automatically matches the elderly care level and service duration according to preset policy rules, calculates the service fee and government subsidy deduction amount, and stores the service records and subsidy settlement information together.

9. An intelligent early warning and decision-making system for elderly care equipment based on the Internet of Things, characterized in that, The system is used to execute the IoT-based intelligent early warning and processing decision-making method for elderly care equipment according to any one of claims 1 to 8, the system comprising a sensing access layer, an edge computing layer, a cloud decision-making layer and an application execution layer that are sequentially connected in communication; The perception access layer is used to collect multi-dimensional raw monitoring data and connect the raw monitoring data to the system through a preset communication protocol. The perception access layer includes at least one of smart wearable devices, environmental sensing devices, medical terminals, and security devices. The edge computing layer includes a multi-protocol adaptation module, a data cleaning and normalization module, a local anomaly detection engine, and an encrypted communication module, which are used to preprocess the raw monitoring data and perform local anomaly detection, and to complete encrypted data interaction with the cloud. The cloud-based decision-making layer includes a multi-source evidence fusion module, a comprehensive risk assessment module, a resource scheduling and dispatching module, and a subsidy settlement module, which are used to complete multi-source data fusion, risk level determination, emergency resource scheduling, and service subsidy settlement. The application execution layer includes a guardian APP, a caregiver PDA, a large-screen monitoring center, and a 120 emergency response module, which are used to perform emergency response operations and feed back the execution results to the cloud decision layer.

Images