Adaptive fault diagnosis and early warning method and system for smart elderly care internet of things device

Abstract

This invention provides an adaptive fault diagnosis and early warning method and system for smart elderly care IoT devices, belonging to the field of intelligent diagnostic technology. The method includes: acquiring observation time-series data, scene context data, and business attribute data of IoT devices in a target scenario; decoupling the observation time-series data based on the scene context data to obtain scene disturbance components, device aging drift components, and actual fault components; constructing three-dimensional coupling features based on the scene disturbance components, device aging drift components, actual fault components, and business attribute data; inputting the three-dimensional coupling features of each device as initial features of graph nodes into a graph neural network and reinforcement learning fusion model to obtain fault diagnosis results and fault propagation tracing results; dynamically generating early warning thresholds based on device aging drift degree, business risk score, and time period adjustment factor, and outputting early warning information according to the early warning level. This invention improves emergency response efficiency and operation and maintenance resource allocation efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology, and specifically to an adaptive fault diagnosis and early warning method and system for smart elderly care IoT devices. Background Technology

[0002] As my country's population continues to age, smart elderly care has become a core direction for the transformation and upgrading of the elderly care industry. Internet of Things (IoT) technology, as the core support for smart elderly care, has been widely applied in all aspects of elderly care institutions, including monitoring the elderly's vital signs, managing the living environment, providing safety and emergency protection, and offering daily living services. Existing large-scale chain elderly care institutions have comprehensively deployed various IoT devices such as smart bracelets, human posture monitoring equipment, indoor temperature and humidity sensors, smoke detectors, and emergency call terminals, building a fully covered smart elderly care service system.

[0003] However, in actual operation, fault diagnosis and early warning of IoT devices for elderly care face significant challenges. Existing technologies typically combine fixed threshold judgment with manual inspection and verification. This involves pre-setting fixed alarm thresholds for each type of device, triggering an alarm when the observed value exceeds the threshold, and then having maintenance personnel conduct on-site verification. Some improved solutions introduce a single machine learning model to detect anomalies in the device's raw time-series data. However, these solutions have revealed significant shortcomings in elderly care scenarios.

[0004] First, in elderly care settings, disturbances such as frequent human activity, furniture obstruction, and rapid changes in temperature and humidity are highly intertwined with actual equipment malfunctions in the observed signals. Existing methods only collect the raw time-series parameters of the equipment, lacking contextual semantics. They cannot mathematically separate scene disturbances, equipment aging drift, and actual fault abrupt changes in interpretability. This often leads to misjudging normal fluctuations caused by elderly people's activities as equipment malfunctions, or underreporting slow equipment aging drift as normal fluctuations, resulting in persistently high false alarm and false negative rates.

[0005] Secondly, latent fault samples are scarce and severely imbalanced in elderly care scenarios. Latent fault samples such as equipment data drift and intermittent disconnections account for only a very small proportion and are difficult to label. Existing methods generally lack targeted solutions for the scarcity of latent fault samples and class imbalance. Direct training will cause the model to be severely biased towards normal samples, resulting in a large number of high-risk latent faults being missed.

[0006] Third, most existing diagnostic models are static and fixed, unable to adapt to fluctuations in operational status caused by equipment aging and environmental changes, and lack cross-device correlation analysis capabilities. In elderly care institutions, multiple devices are densely deployed and operate collaboratively. Failure of a single device can easily trigger a chain of anomalies. Traditional methods cannot uncover the fault propagation path between devices, making it difficult to locate the core fault node, and diagnostic strategies cannot be dynamically optimized with changes in the scenario.

[0007] Fourth, existing early warning mechanisms mostly use fixed thresholds and do not dynamically adjust them based on equipment aging, the risk level of elderly care services, and time-based safety priorities. A unified alarm for all faults prevents maintenance personnel from prioritizing high-risk faults, resulting in low emergency response efficiency, a large number of invalid alarms, and unreasonable allocation of maintenance resources. Summary of the Invention

[0008] The purpose of this invention is to provide an adaptive fault diagnosis and early warning method and system for smart elderly care IoT devices, so as to at least solve the problems of existing technologies, such as the high degree of overlap between scene disturbances and equipment faults in the observed signals, the static and fixed diagnostic models and the lack of cross-device correlation analysis capabilities, and the fixed early warning thresholds.

[0009] To achieve the above objectives, a first aspect of the present invention provides an adaptive fault diagnosis and early warning method for smart elderly care IoT devices, the method comprising: Acquire observation time-series data, scene context data, and business attribute data of IoT devices in the target scenario; Based on the scene context data, the observation time series data is decoupled to obtain scene disturbance components, equipment aging drift components and actual fault components. Among them, the scene disturbance components are extracted by sliding window dynamic baseline fitting, and the equipment aging drift components are extracted by adaptive variational mode decomposition after removing the scene disturbance components. The actual fault components are obtained by subtracting the scene disturbance components and equipment aging drift components from the observation time series data. Based on scenario disturbance components, equipment aging drift components, actual fault components, and business attribute data, a three-dimensional coupled feature integrating faults, scenarios, and business risks is constructed. The three-dimensional coupling features of each device are used as the initial features of the graph nodes and input into the graph neural network and reinforcement learning fusion model. Through cross-device associated feature aggregation and dynamic optimization of diagnostic strategies, fault diagnosis results and fault propagation tracing results are obtained. Based on the equipment aging drift degree determined by the equipment aging drift component, the business risk score, and the time period adjustment factor determined by the scenario context data, the early warning threshold is dynamically generated, and early warning information is output according to the early warning level.

[0010] Preferably, the observation time series data includes at least one of sensor sample values, power supply voltage, gateway packet loss rate, signal strength, and sampling timestamp; the scene context data includes at least one of time period, room type, human presence event, door and window opening / closing event, and nursing staff entry / exit event; and the business attribute data includes at least one of care level, equipment business security level, and business risk weight.

[0011] Preferably, before decoupling, the method further includes: using time-weighted Gaussian kernel interpolation to map observation time-series data with different sampling frequencies to a unified time axis, and adaptively setting the time-series smoothing bandwidth according to the equipment type, wherein the time-series smoothing bandwidth of high-risk monitoring equipment is less than that of ordinary environmental equipment.

[0012] Preferably, the sliding window dynamic baseline fitting includes: within a sliding window centered on the current time, performing scene baseline solving based on observed time-series data, and applying smoothing regularization constraints to the scene baseline to determine scene disturbance components, and the range of the sliding window is adaptively determined according to the scene event type in the scene context data; The adaptive variational mode decomposition includes: iteratively updating the center frequency and bandwidth parameters, decomposing the residual signal into modal components of different frequency bands, and retaining the low-frequency components that characterize the slow aging of the equipment; wherein, the residual signal is the signal after removing scene disturbance components from the observation time series data.

[0013] Preferably, the three-dimensional coupling features include: scene disturbance intensity factor, equipment drift aging degree, and business risk score; wherein, the scene disturbance intensity factor is the ratio of the variance of the scene disturbance component within the time window to the reference variance of the normal scene under the same scene event in the same historical period; the equipment drift aging degree is the average absolute deviation of the drift component relative to the equipment's factory standard baseline; the business risk score is calculated based on the correlation tensor between equipment type, failure mode and elderly care business, business security weight, and failure mode coupling coefficient.

[0014] Preferably, the business risk score in the three-dimensional coupling feature is determined in the following way: based on the correspondence between equipment type and failure mode, the preset association tensor between equipment type, failure mode and elderly care business is queried to obtain the coupling strength of the current failure mode to each elderly care business; combined with the security weight of each elderly care business and the historical probability of the current equipment failure causing the failure of the business, the business risk score is calculated.

[0015] Preferably, the graph neural network and reinforcement learning fusion model includes: constructing a heterogeneous network topology graph with IoT devices as nodes and data transmission associations and scene associations between devices as edges; aggregating features of neighboring devices layer by layer through message passing in a multi-layer graph neural network and updating the hidden representation of nodes; constructing a reward function with changes in device operating environment, device aging degree and historical fault diagnosis results as states, and adjusting judgment weights and adjusting detection priorities as actions, and using the reward function to guide the iterative update of the multi-layer graph neural network.

[0016] Secondly, the present invention provides an adaptive fault diagnosis and early warning system for smart elderly care IoT devices, used to implement the aforementioned adaptive fault diagnosis and early warning method for smart elderly care IoT devices, the system comprising: The data sensing module is used to acquire observation time-series data, scene context data, and business attribute data of IoT devices in the target scene; The data decoupling module is used to decouple the observed time series data based on the scene context data to obtain scene disturbance components, equipment aging drift components and actual fault components. Specifically, the scene disturbance components are extracted by sliding window dynamic baseline fitting, and the equipment aging drift components are extracted by adaptive variational mode decomposition after removing the scene disturbance components. The actual fault components are obtained by subtracting the scene disturbance components and equipment aging drift components from the observed time series data. The feature construction module is used to construct a three-dimensional coupled feature that integrates faults, scenarios, and business risks based on scenario disturbance components, equipment aging and drift components, real fault components, and business attribute data. The intelligent diagnostic module is used to input the three-dimensional coupling features of each device as the initial features of the graph nodes into the graph neural network and reinforcement learning fusion model. Through cross-device associated feature aggregation and dynamic optimization of diagnostic strategies, fault diagnosis results and fault propagation tracing results are obtained. The early warning decision module is used to dynamically generate early warning thresholds based on the equipment aging drift degree determined by the equipment aging drift component, the business risk score, and the time period adjustment factor determined by the scenario context data, and output early warning information according to the early warning level.

[0017] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned adaptive fault diagnosis and early warning method for smart elderly care IoT devices.

[0018] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described adaptive fault diagnosis and early warning method for smart elderly care IoT devices.

[0019] The beneficial effects of this invention are: 1. This invention decouples observation time-series data based on scene context data. Scene disturbance components are extracted through sliding window dynamic baseline fitting. After removing scene disturbance components, equipment aging drift components are extracted through adaptive variational mode decomposition. The observation time-series data is then subtracted from both scene disturbance and equipment aging drift components to obtain the true fault components. Thus, interpretable mathematical separation of scene disturbances, equipment aging drift, and true faults is achieved from the data source, avoiding misjudging normal fluctuations caused by elderly activity as faults and preventing the underreporting of slow aging drift as normal fluctuations, significantly reducing both false alarm and false negative rates. 2. This invention uses the three-dimensional coupling features of each device as initial features of graph nodes, inputting them into a graph neural network and reinforcement learning fusion model. Through cross-device associated feature aggregation and dynamic optimization of diagnostic strategies, fault diagnosis results and fault propagation tracing results are obtained. The graph neural network can discover fault propagation paths between multiple devices, realizing associated fault tracing; reinforcement learning uses the device operating environment, aging degree, and historical diagnostic results as states, and uses diagnostic weight adjustment and detection priority adjustment as actions. Moreover, the reward function includes a business risk score item, and sets a larger penalty weight for high-risk device underreporting, so that the model prioritizes minimizing high-risk device underreporting during iteration, dynamically adapting to scene changes, and solving the defects of static and fixed traditional models. 3. This invention dynamically generates early warning thresholds based on equipment aging drift, business risk scores, and time-period adjustment factors determined by scenario context data, and outputs early warning information according to the warning level. The greater the equipment aging, the more relaxed the threshold is to tolerate slow drift; the higher the business risk, the tighter the threshold; the threshold is tightened during high-risk monitoring periods at night and relaxed during daytime periods of personnel activity. This achieves dynamically adjustable and differentiated classification of early warning standards, enabling maintenance personnel to prioritize high-risk faults and improving emergency response efficiency and maintenance resource allocation efficiency. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the adaptive fault diagnosis and early warning method for smart elderly care IoT devices provided by the present invention; Figure 2 This is a block diagram of the adaptive fault diagnosis and early warning system for smart elderly care IoT devices provided by the present invention. Detailed Implementation

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0022] Example 1

[0023] like Figure 1As shown, this embodiment provides an adaptive fault diagnosis and early warning method for smart elderly care IoT devices, including the following steps: Step S1: Obtain observation time series data, scene context data, and business attribute data of IoT devices in the target scene.

[0024] In elderly care facilities, IoT devices deployed at the edge sensing layer include human monitoring devices, environmental monitoring devices, emergency interaction devices, and operation and maintenance support devices. Human monitoring devices include smart wearable bracelets, fall detection terminals, and vital sign sensors, used to collect information on the elderly's heart rate, blood oxygen, activity status, and posture; environmental monitoring devices include temperature and humidity sensors, air quality sensors, smoke sensors, and gas sensors; emergency interaction devices include emergency call buttons and voice help terminals; and operation and maintenance support devices include network gateways, data transmission nodes, and equipment status monitoring terminals.

[0025] Data acquisition is divided into three categories. Device runtime sequence data includes sensor sample values, power supply voltage, gateway packet loss rate, signal strength, and sampling timestamps. Wearable devices sample at a 1Hz cycle, environmental sensors sample sparsely at the minute level, and call terminals report based on event triggers. Scene context event data is recorded according to event triggering methods, including time period (day / night), room type (disabled room, semi-disabled room, self-care room, common activity area), human presence detection events, door / window opening / closing events, and caregiver entry / exit events. Business care attribute data is recorded in a static ledger format, including the room's elderly care level and the equipment's business safety level (Level 1 High Risk, Level 2 Environmental, Level 3 Ordinary).

[0026] In this embodiment, since different devices have large differences in the sampling frequency of the observation time series data, before executing step S2, time-weighted Gaussian kernel interpolation is used to map the observation time series data with different sampling frequencies to a unified time axis, and the time series smoothing bandwidth is adaptively set according to the device type. The time series smoothing bandwidth of high-risk monitoring devices is smaller than that of ordinary environmental devices.

[0027] Specifically, within a sampling window centered at time t, each original sampling point within the window is assigned a Gaussian weight based on its time distance, and the weighted sum is used to obtain the aligned standard observation value. The temporal smoothing bandwidth is adaptively set according to the device type: a smaller bandwidth is used for high-risk monitoring devices such as fall detection terminals and vital sign sensors to ensure abrupt change resolution; a larger bandwidth is used for ordinary environmental devices such as temperature and humidity sensors to suppress noise.

[0028] The alignment mapping formula is as follows:

[0029] In the formula, For a moment Aligned standard observations; For the first Observations from the original non-uniform sampling points; For the first Timestamp of each sampling point; For time The set of valid sampling points within the central sampling window; For timing smoothing bandwidth, the value is adaptively set according to the device type. A smaller value is used for high-risk monitoring devices to ensure resolution of sudden changes, while a larger value is used for devices in ordinary environments to suppress noise.

[0030] This embodiment's time-weighted Gaussian kernel interpolation uses the original sampling points as the weight kernel, maintaining the contribution of the original observations for both event-triggered devices and periodically sampling devices, without falsifying fault characteristics. It achieves high-precision spatiotemporal fusion of heterogeneous devices, preserving latent fault characteristic information while unifying the time axis, providing time-aligned observation sequences for subsequent decoupling modeling. .

[0031] Step S2: Based on the scene context data, decouple the observation time series data to obtain scene disturbance components, equipment aging drift components, and actual fault components; among them, scene disturbance components are extracted by sliding window dynamic baseline fitting, and after removing scene disturbance components, equipment aging drift components are extracted by adaptive variational mode decomposition. The actual fault components are obtained by subtracting the scene disturbance components and equipment aging drift components from the observation time series data.

[0032] In this embodiment of the invention, because existing technologies cannot mathematically distinguish between the three causes of "personnel disturbance, equipment aging, and actual failure," scene disturbances are often misjudged as failures or drift failures are treated as normal fluctuations. Therefore, this embodiment mathematically decomposes the observation signal of the elderly care equipment into the sum of three independent components and a residual term, namely:

[0033] In the formula, For equipment observation sequences; The dynamic disturbances in the elderly care scenario (caused by people moving around, obstruction, and sudden changes in temperature and humidity, etc.) are the scenario disturbance components. This refers to the slow drift item (latent fault) due to equipment aging, i.e., the equipment aging drift component. These are real-world sudden failures (disconnection, damage, sudden changes), i.e., real-world failure components. The residual term is Gaussian noise and satisfies... .

[0034] This embodiment utilizes sliding window dynamic baseline fitting and adaptive variational mode decomposition to separate the three components from the device observation sequence. This achieves interpretable mathematical separation of scene disturbances, device aging drift, and actual faults from the data source, avoiding misjudging normal fluctuations caused by elderly activity as faults and preventing underreporting of slow aging drift as normal fluctuations, significantly reducing both false alarm and false negative rates. Specifically, the sliding window dynamic baseline fitting includes: within a sliding window centered on the current time, performing scene baseline solving based on observed time-series data, and applying smoothing regularization constraints to the scene baseline to determine scene disturbance components, and the range of the sliding window is adaptively determined according to the scene event types in the scene context data.

[0035] The scene baseline is solved as follows:

[0036] In the formula, For time A sliding window centered on the user; For the observation sequence of devices within the window; This is the baseline of the scenario to be solved.

[0037] After solving to the scene baseline and then applying smoothing and regularization constraints, the scene disturbance components are calculated:

[0038] In the formula, For scene disturbance components, To smooth the regularization coefficients, the degree of baseline smoothing is controlled; The baseline gradient term constrains the baseline from excessive smoothing in order to preserve true fault mutation information.

[0039] In this embodiment of the invention, the range of the sliding window is adaptively determined based on the type of scene event in the scene context data. For example, when the scene context data indicates that the current event is "frequent elderly activity", the range of the sliding window is expanded to capture more complete activity disturbances; when it indicates a "quiet night" event, the range of the sliding window is narrowed to improve sensitivity to minor faults.

[0040] Specifically, the adaptive variational mode decomposition includes: iteratively updating the center frequency and bandwidth parameters, decomposing the residual signal into modal components of different frequency bands, and retaining the low-frequency components that characterize the slow aging of the equipment; wherein, the residual signal is the signal after removing scene disturbance components from the observation time series data.

[0041] Among them, the low-frequency component of slow equipment aging is:

[0042] In the formula, This refers to the low-frequency components that cause the equipment to age slowly, i.e., the equipment aging drift components. The low-frequency mode reconstruction operator for adaptive variational mode decomposition iteratively updates the center frequency and bandwidth parameters to reconstruct the residual signal ( The device is decomposed into modal components of different frequency bands, and only the low-frequency components that characterize the slow aging of the device are retained. In this embodiment, the low-frequency mode retention band is adjusted according to the time period information in the scene context data. For example, when the device load is stable at night, a lower frequency band is retained to capture slower aging drift.

[0043] After decomposing the scene disturbance component and the equipment aging drift component, the actual fault mutation component is obtained by subtracting these components from the observation sequence.

[0044] In the formula, This represents the actual fault mutation component.

[0045] Step S3: Based on the scenario disturbance component, equipment aging drift component, actual fault component, and business attribute data, construct a three-dimensional coupled feature that integrates faults, scenarios, and business risks.

[0046] Traditional datasets only use binary labels for equipment normal / faulty, which cannot quantify the severity of faults, the intensity of scene disturbances, and the degree of equipment aging. This embodiment constructs a continuous high-dimensional semantic feature space based on decoupled components, specifically including three quantification factors: scene disturbance intensity factor, equipment drift aging degree, and business risk score.

[0047] Wherein, the scene disturbance intensity factor is the ratio of the variance of the scene disturbance component within the time window to the reference variance of the normal scene under the same scene event in the same historical period:

[0048] In the formula, For a moment Scene disturbance intensity factor; For scene perturbation components in the window within variance; This serves as the reference variance for normal device operation under the same historical timeframe and scenario. This indicates that the current fluctuation is within the normal fluctuation range. This indicates that the fluctuation intensity exceeds the normal baseline of the scenario, and further determination is needed to determine whether it is a real fault.

[0049] Among them, the equipment drift aging degree is the average absolute deviation of the drift component relative to the equipment's factory standard baseline:

[0050] In the formula, For a moment Equipment aging drift; Number of dimensions for key equipment parameters; For the first Dimensional parameters at time The drift component value; For the first The baseline values ​​of the equipment's factory standards for the dimension parameters. The greater the drift, the deeper the equipment's aging and the higher the risk of latent failures.

[0051] The business risk score is calculated based on the correlation tensor between equipment type, failure mode and elderly care business, business security weight and failure mode coupling coefficient. Specifically, according to the correspondence between equipment type and failure mode, the preset correlation tensor between equipment type, failure mode and elderly care business is queried to obtain the coupling strength of the current failure mode to each elderly care business; combined with the security weight of each elderly care business and the historical probability of the current equipment failure causing the failure of the business, the business risk score is calculated.

[0052]

[0053] In the formula, This represents the business risk score corresponding to the current sample. The total number of related elderly care services (including fall detection, emergency call, gas alarm, temperature and humidity monitoring, etc.); For the first Safety weighting of elderly care services (taking Chunxitang as an example, fall monitoring) Emergency call Gas alarm Temperature and humidity monitoring ); The current equipment failure caused the first The historical probability of a monitoring service failure is obtained from historical operation and maintenance records. The fault mode coupling coefficient represents the relationship between the current fault type and the fault mode coupling coefficient. The coupling strength between business items.

[0054] This embodiment quantifies fault hazards, scene disturbance intensity, and equipment aging degree into a continuous high-dimensional label space, endowing the dataset with semantics specific to the elderly care industry. Its advantage lies in upgrading the dataset from a "binary classification label" to a "high-dimensional semantic risk dataset," providing label basis for subsequent business risk-weighted reinforcement learning reward functions and multi-level early warning mechanisms.

[0055] Step S4: Input the three-dimensional coupling features of each device as the initial features of the graph nodes into the graph neural network and reinforcement learning fusion model. Through cross-device associated feature aggregation and dynamic optimization of the diagnostic strategy, obtain the fault diagnosis results and fault propagation tracing results.

[0056] Due to the limitations of existing technologies in adapting to heterogeneous elderly care equipment, difficulties in fault tracing, and model rigidity, this embodiment constructs an adaptive fault diagnosis model that integrates graph neural networks (GNN) and reinforcement learning (RL). This model fully utilizes the association mining capabilities of graph neural networks and the dynamic optimization capabilities of reinforcement learning to adapt to the collaborative operation scenario of multiple elderly care devices.

[0057] Specifically, a heterogeneous network topology diagram for elderly care equipment is constructed, with various IoT devices from Chunxitang serving as network nodes. Data transmission and scene association between devices are set as edges. The initial features of a node consist of the three-dimensional coupled feature vector from step S3. The features of neighboring devices are aggregated layer by layer through a multi-layer graph neural network message passing mechanism. Layer nodes The hidden representation is updated to:

[0058] In the formula, For the first Layer nodes Embedded representation; For nodes The set of neighboring nodes; For the first Layer-learnable weight matrix; For the first Layer bias vector; It is a non-linear activation function. Initial node features. The feature vectors formed by combining the three-dimensional coupling features in step S3 are as follows: .

[0059] In this embodiment, the GNN is used to aggregate neighborhood device information based on the initial features of nodes, learn multi-device topology associations and fault propagation features, complete cross-device fusion and refined classification of fault features, and obtain fault diagnosis results. Then, through node embedding technology, it mines the fault propagation paths between different devices and different protocol networks, realizes associated fault tracing, and obtains fault propagation tracing results.

[0060] The reinforcement learning algorithm in this embodiment achieves adaptive optimization of the diagnostic strategy. It uses changes in the equipment operating environment, equipment aging level, and historical fault diagnosis results as reinforcement learning states. The action space is defined by diagnostic actions (adjusting judgment weights and adjusting detection priorities). Construct the reward function:

[0061] In the formula, For the first Comprehensive rewards for step-by-step diagnostic decision-making; This represents the current diagnostic accuracy rate. False alarm rate; This refers to the false negative rate. The average business risk score for the current batch of samples is determined by step S3. Calculated; , , , Here are the weighting coefficients for each indicator, where Larger values ​​are set for Level 1 high-risk monitoring equipment, so that the model prioritizes minimizing the false negatives of high-risk equipment during iteration, rather than balancing global accuracy. The model continuously iterates and optimizes diagnostic weights and judgment rules. For high-frequency fault scenarios, such as intermittent sensor disconnection, data drift, and abnormal channel interference, the detection priority of corresponding faults is automatically increased, dynamically adapting to changes in scenarios and overcoming the shortcomings of static and fixed traditional models.

[0062] Step S5: Based on the equipment aging drift degree determined by the equipment aging drift component, the business risk score, and the time period adjustment factor determined by the scenario context data, dynamically generate the early warning threshold and output the early warning information according to the early warning level.

[0063] Because existing early warning mechanisms mostly use fixed thresholds, all faults are alarmed in the same way, which makes it impossible for maintenance personnel to prioritize high-risk faults, resulting in low emergency response efficiency, a large number of invalid alarms, and unreasonable allocation of maintenance resources.

[0064] To address this, this embodiment employs a three-tiered adaptive threshold early warning mechanism. First, equipment is categorized based on its functional importance: Level 1 devices are high-risk monitoring equipment, including fall detection terminals, emergency call devices, and vital sign sensors, directly related to the safety of the elderly; Level 2 devices are environmental safety equipment, including smoke and gas sensors, related to public area safety; and Level 3 devices are general monitoring equipment, including conventional temperature, humidity, and air quality sensors. Then, early warning thresholds are dynamically generated based on equipment aging drift, business risk scores, and time-period adjustment factors determined by scenario context data.

[0065] In the formula, For a moment Dynamic early warning threshold; This is the basic threshold for the device; This refers to the equipment aging drift degree; For a moment Business risk score; As the aging release factor increases, the threshold is appropriately relaxed to tolerate slow drift when the equipment aging degree deepens; This is a risk tightening coefficient; the higher the business risk, the tighter the threshold. As a time-period adjustment factor, the high-risk monitoring period at night (22:00-6:00) is set to less than [a certain value]. The value is tightened to a threshold, taking values ​​close to those used during peak daytime activity periods. The value is used to reduce false alarms caused by scene disturbances.

[0066] Therefore, for Level 1 high-risk equipment, the warning threshold is tightened to identify minor anomalies in advance, enabling early warning and early response. For Level 3 ordinary equipment, the threshold is appropriately relaxed to reduce the number of invalid alarms. Furthermore, the greater the equipment's aging, the more relaxed the threshold becomes to tolerate slow drift; the higher the business risk, the tighter the threshold; the threshold is tightened during high-risk monitoring periods at night and relaxed during daytime periods of personnel activity. This achieves dynamic adjustability and differentiated classification of warning standards, enabling maintenance personnel to prioritize high-risk faults and improving emergency response efficiency and maintenance resource allocation efficiency.

[0067] As a further optimization of this embodiment, during the training phase, this embodiment also employs a hierarchical-aware multi-weight imbalanced sampling strategy: In real-world operational data, normal samples account for approximately 95%, overt fault samples for about 4%, and high-risk latent faults (data drift, intermittent disconnections) samples for only about 0.5%, with labels scarce and difficult to annotate. Direct training would cause the model to be severely biased towards normal samples, leading to missed detections of high-risk latent faults. Therefore, this embodiment proposes a hierarchical awareness multi-weight imbalanced sampling strategy, considering the imbalance in sample quantity, business risk, and scarcity of latent features. The sample sampling weight is defined as follows:

[0068] In the formula, The overall sampling weights are used for the samples. , , The fusion hyperparameters for the three types of weights satisfy... The three weights are defined as follows: Sample size balancing weights: defined as the inverse ratio of the number of samples per category.

[0069] In the formula, For category Quantity balance weight; The total number of all samples; For category The number of original samples.

[0070] High-risk business weighting: The business risk score is exponentially amplified to give high-risk samples a higher sampling probability.

[0071] In the formula, High-risk weighting for business operations; Assign a business risk score.

[0072] Latent drift feature scarcity weights: assigning extra weights to samples with drift components.

[0073] In the formula, The weights are for the scarcity of latent drift features; This is the drift bonus coefficient; This is an indicator function that takes the value when the condition is true. Otherwise ; It is a zero vector.

[0074] In terms of sampling strategy, time-preserving oversampling (preserving time-dependent sequential replication enhancement) is used for high-risk latent drift samples; spatiotemporal stratified undersampling (stratified random sampling by facility area and room type) is used for a large number of normal samples; and stationary interpolation augmentation is used for explicit fault samples to avoid disrupting the temporal structure. The reason for this strategy is that ordinary resampling only relies on the number of samples, ignoring business risks and the scarcity of latent features, and cannot solve the core contradiction in the elderly care scenario of "few high-risk samples but extremely high costs".

[0075] This embodiment simultaneously achieves quantity balance, risk balance, and latent feature balance, significantly improving the model's ability to learn from high-frequency, difficult-to-identify faults such as data drift and intermittent disconnections, while preserving the sensor timing dependencies throughout the process.

[0076] When dividing the training set, test set, and validation set during the training phase, a spatiotemporal hierarchical division mechanism of "training in the main hospital area, cross-seasonal validation in the main hospital area, and testing across chain hospitals area" is adopted: the training set takes data from the main hospital area throughout the entire time period, the validation set takes cross-seasonal data from the main hospital area for parameter tuning, and the test set takes unknown scenario data from other chain hospitals area to simulate the real-world deployment scenario of the model in the new hospital area.

[0077] Therefore, this embodiment avoids low-risk samples dominating the test set by stratifying and verifying coverage across different hospital areas, seasons, and care levels, ensuring that the experimental results are consistent with the actual operation and maintenance scenarios of Chunxitang's multiple hospital areas.

[0078] Example 2

[0079] like Figure 2 As shown, this embodiment provides an adaptive fault diagnosis and early warning system for smart elderly care IoT devices, used to implement the adaptive fault diagnosis and early warning method for smart elderly care IoT devices in Embodiment 1. The system includes: The data sensing module is used to acquire observation time-series data, scene context data, and business attribute data of IoT devices in the target scene; The data decoupling module is used to decouple the observed time series data based on the scene context data to obtain scene disturbance components, equipment aging drift components and actual fault components. Specifically, the scene disturbance components are extracted by sliding window dynamic baseline fitting, and the equipment aging drift components are extracted by adaptive variational mode decomposition after removing the scene disturbance components. The actual fault components are obtained by subtracting the scene disturbance components and equipment aging drift components from the observed time series data. The feature construction module is used to construct a three-dimensional coupled feature that integrates faults, scenarios, and business risks based on scenario disturbance components, equipment aging and drift components, real fault components, and business attribute data. The intelligent diagnostic module is used to input the three-dimensional coupling features of each device as the initial features of the graph nodes into the graph neural network and reinforcement learning fusion model. Through cross-device associated feature aggregation and dynamic optimization of diagnostic strategies, fault diagnosis results and fault propagation tracing results are obtained. The early warning decision module is used to dynamically generate early warning thresholds based on the equipment aging drift degree determined by the equipment aging drift component, the business risk score, and the time period adjustment factor determined by the scenario context data, and output early warning information according to the early warning level.

[0080] This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the adaptive fault diagnosis and early warning method for the smart elderly care IoT device in Embodiment 1.

[0081] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the adaptive fault diagnosis and early warning method for smart elderly care IoT devices in Embodiment 1.

[0082] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0083] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0084] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An adaptive fault diagnosis and early warning method for smart elderly care IoT devices, characterized in that, The method includes: Acquire observation time-series data, scene context data, and business attribute data of IoT devices in the target scenario; Based on the scene context data, the observation time series data is decoupled to obtain scene disturbance components, equipment aging drift components and actual fault components. Among them, the scene disturbance components are extracted by sliding window dynamic baseline fitting, and the equipment aging drift components are extracted by adaptive variational mode decomposition after removing the scene disturbance components. The actual fault components are obtained by subtracting the scene disturbance components and equipment aging drift components from the observation time series data. Based on scenario disturbance components, equipment aging drift components, actual fault components, and business attribute data, a three-dimensional coupled feature integrating faults, scenarios, and business risks is constructed. The three-dimensional coupling features of each device are used as the initial features of the graph nodes and input into the graph neural network and reinforcement learning fusion model. Through cross-device associated feature aggregation and dynamic optimization of diagnostic strategies, fault diagnosis results and fault propagation tracing results are obtained. Based on the equipment aging drift degree determined by the equipment aging drift component, the business risk score, and the time period adjustment factor determined by the scenario context data, the early warning threshold is dynamically generated, and early warning information is output according to the early warning level.

2. The adaptive fault diagnosis and early warning method for smart elderly care IoT devices according to claim 1, characterized in that, The observation time series data includes at least one of sensor sample values, power supply voltage, gateway packet loss rate, signal strength, and sampling timestamp; the scene context data includes at least one of time period, room type, human presence event, door and window opening / closing event, and nursing staff entry / exit event; the business attribute data includes at least one of care level, equipment business security level, and business risk weight.

3. The adaptive fault diagnosis and early warning method for smart elderly care IoT devices according to claim 2, characterized in that, Before decoupling, the method further includes: using time-weighted Gaussian kernel interpolation to map observation time-series data with different sampling frequencies to a unified time axis, and adaptively setting the time-series smoothing bandwidth according to the equipment type, with the time-series smoothing bandwidth of high-risk monitoring equipment being less than that of ordinary environmental equipment.

4. The adaptive fault diagnosis and early warning method for smart elderly care IoT devices according to claim 2, characterized in that, The sliding window dynamic baseline fitting includes: within a sliding window centered on the current time, performing scene baseline solving based on observed time series data, and applying smoothing regularization constraints to the scene baseline to determine scene disturbance components; and the range of the sliding window is adaptively determined according to the scene event type in the scene context data. The adaptive variational mode decomposition includes: iteratively updating the center frequency and bandwidth parameters, decomposing the residual signal into modal components of different frequency bands, and retaining the low-frequency components that characterize the slow aging of the equipment; wherein, the residual signal is the signal after removing scene disturbance components from the observation time series data.

5. The adaptive fault diagnosis and early warning method for smart elderly care IoT devices according to any one of claims 1-4, characterized in that, The three-dimensional coupling features include: scene disturbance intensity factor, equipment drift aging degree, and business risk score; wherein, the scene disturbance intensity factor is the ratio of the variance of the scene disturbance component within the time window to the reference variance of the normal scene under the same scene event in the same historical period; the equipment drift aging degree is the average absolute deviation of the drift component relative to the equipment's factory standard baseline; the business risk score is calculated based on the correlation tensor between equipment type, failure mode and elderly care business, business security weight, and failure mode coupling coefficient.

6. The adaptive fault diagnosis and early warning method for smart elderly care IoT devices according to claim 5, characterized in that, The business risk score in the three-dimensional coupling feature is determined in the following way: based on the correspondence between equipment type and failure mode, the preset association tensor between equipment type, failure mode and elderly care business is queried to obtain the coupling strength of the current failure mode to each elderly care business. By combining the safety weights of various elderly care services with the historical probability of current equipment failures causing the failure of these services, a business risk score is calculated.

7. The adaptive fault diagnosis and early warning method for smart elderly care IoT devices according to claim 1, characterized in that, The graph neural network and reinforcement learning fusion model includes: constructing a heterogeneous network topology graph with IoT devices as nodes and data transmission associations and scene associations between devices as edges; aggregating features of neighboring devices layer by layer through message passing in a multi-layer graph neural network and updating the hidden representation of nodes; constructing a reward function with changes in device operating environment, device aging degree and historical fault diagnosis results as states, and adjusting judgment weights and detection priorities as actions, and using the reward function to guide the iterative update of the multi-layer graph neural network.

8. An adaptive fault diagnosis and early warning system for smart elderly care IoT devices, used to implement the adaptive fault diagnosis and early warning method for smart elderly care IoT devices as described in any one of claims 1-7, characterized in that, The system includes: The data sensing module is used to acquire observation time-series data, scene context data, and business attribute data of IoT devices in the target scene; The data decoupling module is used to decouple the observed time series data based on the scene context data to obtain scene disturbance components, equipment aging drift components and actual fault components. Specifically, the scene disturbance components are extracted by sliding window dynamic baseline fitting, and the equipment aging drift components are extracted by adaptive variational mode decomposition after removing the scene disturbance components. The actual fault components are obtained by subtracting the scene disturbance components and equipment aging drift components from the observed time series data. The feature construction module is used to construct a three-dimensional coupled feature that integrates faults, scenarios, and business risks based on scenario disturbance components, equipment aging and drift components, real fault components, and business attribute data. The intelligent diagnostic module is used to input the three-dimensional coupling features of each device as the initial features of the graph nodes into the graph neural network and reinforcement learning fusion model. Through cross-device associated feature aggregation and dynamic optimization of diagnostic strategies, fault diagnosis results and fault propagation tracing results are obtained. The early warning decision module is used to dynamically generate early warning thresholds based on the equipment aging drift degree determined by the equipment aging drift component, the business risk score, and the time period adjustment factor determined by the scenario context data, and output early warning information according to the early warning level.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the adaptive fault diagnosis and early warning method for the smart elderly care IoT device as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the adaptive fault diagnosis and early warning method for smart elderly care IoT devices as described in any one of claims 1-7.

Images