A nursing behavior supervision method and system based on smart care for the elderly

Abstract

The application relates to the technical field of smart endowment supervision, and discloses a nursing behavior supervision method and system based on smart endowment. The method obtains original signals from a nursing behavior monitoring device, and the original signals are processed by a median filtering algorithm to eliminate random interference; based on the processed signals, time sequence characteristics are analyzed, a change period and a trend direction are extracted, a state classification model is used to determine whether the behavior is in a normal or abnormal stage, and when the behavior is abnormal, an intervention target value is calculated according to preset rules and is dynamically corrected in combination with the trend; a reinforcement learning control algorithm is applied, a supervision action intensity is generated according to the difference between a supervision target value and a real-time monitoring value, and a control instruction is output to an execution unit; a compliance feedback device is used to obtain a compliance index, the compliance index is compared with a preset standard, if the compliance index exceeds the range, a rule adjustment amount is derived by a decision tree algorithm and an adjustment instruction is output; the instruction time sequence is optimized in combination with the response characteristics of the execution unit, the signals and the rules are continuously tracked, and fine nursing behavior supervision is realized.

Description

Technical Field

[0001] This invention relates to the field of smart elderly care supervision technology, specifically to a method and system for supervising nursing behavior based on smart elderly care. Background Technology

[0002] With the accelerating aging of the population, the demand for elderly care services is growing, and the standardization and safety of nursing practices have become important concerns in the field of smart elderly care. Currently, traditional nursing supervision relies heavily on manual recording and regular inspections, which suffers from insufficient real-time performance and limited coverage. Some institutions have attempted to introduce monitoring equipment to collect nursing behavior data, but the raw signals are often affected by environmental interference and equipment vibration, resulting in significant data noise that is difficult to use directly for behavioral analysis.

[0003] In assessing behavioral states, existing methods are mostly based on fixed thresholds or simple rules for classification, lacking in-depth analysis of the time-series characteristics of nursing behaviors and making it difficult to identify gradual abnormalities or complex behavioral patterns. Furthermore, during regulatory interventions, goal setting and instruction generation often rely on empirical values, failing to dynamically adjust based on real-time behavioral trends, which can easily lead to over- or under-intervention.

[0004] Furthermore, the nursing compliance feedback mechanism is inadequate, and rule updates rely heavily on manual assessments, making it difficult to rapidly iterate and optimize based on actual nursing outcomes. These issues collectively result in insufficient precision in nursing supervision, failing to adequately adapt to the personalized nursing needs of different elderly individuals and struggling to cope with the ever-changing nursing scenarios in elderly care. Therefore, constructing a comprehensive supervision methodology encompassing signal processing, behavioral analysis, dynamic intervention, and rule optimization has become a crucial direction for improving the quality of smart elderly care services. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for monitoring nursing behavior based on smart elderly care, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for monitoring nursing behavior based on smart elderly care, the method comprising: The raw nursing behavior signals are acquired from the nursing behavior monitoring equipment, and the median filtering algorithm is used to suppress noise in the signals to eliminate random interference, and the processed nursing behavior signals are output. Based on the processed nursing behavior signals, the time series characteristics of the signals are analyzed, the change cycle and trend direction of nursing behavior are extracted, and the nursing behavior is determined to be in a normal or abnormal stage through a state classification model. When it is identified as an abnormal stage, the required intervention target value is calculated according to the preset nursing supervision rules, and the supervision target is dynamically adjusted in combination with the behavior trend. By applying reinforcement learning control algorithms, the intensity of regulatory actions is generated based on the difference between the regulatory target value and the real-time behavior monitoring value, and control commands are output to the regulatory execution unit. The current nursing compliance indicators are obtained by using a compliance feedback device and compared with the preset standard range to determine whether the regulatory rules need to be updated. If the compliance indicators exceed the standard range, the adjustment amount of the regulatory rules is derived by using a decision tree algorithm based on the difference value, and the adjustment instruction is output to the execution unit. By combining the response characteristics of the regulatory execution unit, the timing output of control instructions is optimized to ensure the continuous and stable regulatory process. Furthermore, by continuously tracking nursing behavior signals and the dynamics of regulatory rules, preset rules are updated in real time to adapt to changes in nursing needs and achieve refined regulation of nursing behavior.

[0007] Preferably, the step of using a median filtering algorithm to suppress noise in the signal, eliminate random interference, and output the processed nursing behavior signal specifically includes: The nursing behavior signal is input into the median filtering algorithm module, and then the median filtering algorithm is applied to perform window sliding processing on the signal to divide the signal into multiple time windows; Within each time window, sort the signal values ​​and select the median, remove outliers, and output the smoothed sub-signal; The smoothed sub-signals are reconstructed into a time series to obtain a noise-suppressed nursing behavior signal. The moving average algorithm is then applied to eliminate residual fluctuations and output a stable nursing behavior signal. Stable nursing behavior signals are transmitted to the feature analysis module as input for nursing behavior feature extraction.

[0008] Preferably, the step of analyzing the time-series characteristics of the processed nursing behavior signals, extracting the change cycle and trend direction of the nursing behavior, and determining whether the nursing behavior is in a normal or abnormal stage through a state classification model specifically includes: The processed nursing behavior signals are decomposed into time series to obtain the trend and periodic components of nursing behavior, and then the key features of nursing behavior, including behavior frequency and behavior intensity, are extracted. Based on the frequency characteristics of the behavior, the nursing behavior cycle is calculated, the start and end points of each cycle are determined, and then based on the intensity characteristics of the behavior, the slope of the behavior trend is calculated, and the behavior change curve is output. By combining the nursing behavior cycle and trend slope, the current behavior status is determined to be in a normal or abnormal stage. When the behavior trend is rising and in the early stage of the cycle, it is identified as an abnormal stage, and when the behavior trend is declining and in the late stage of the cycle, it is identified as a normal stage. Output the behavior state result and associate the behavior state with the timestamp to form a complete state sequence data.

[0009] Preferably, after determining whether the nursing behavior is in a normal or abnormal stage through a state classification model, the method further includes: Based on the preset nursing supervision rules and real-time behavior frequency data, the data is input into the neural network prediction model. The trained neural network prediction model then calculates the intervention target value required for the abnormal stage. Based on the calculated intervention target value, combined with behavioral frequency trend data, and using fuzzy logic algorithms, the output intensity of the regulatory execution unit is dynamically adjusted to gradually approach the target value. During the adjustment process, the frequency of behavior and the actual behavior value are monitored in real time. When the frequency of behavior changes abruptly or the actual behavior value deviates from the target value by more than a threshold, the neural network prediction model is triggered to recalculate the intervention target value. The recalculated intervention target value is input into the fuzzy logic algorithm to iteratively adjust the regulatory execution unit until the actual behavior value stabilizes near the target value.

[0010] Preferably, the application of reinforcement learning control algorithm generates the intensity of regulatory action based on the difference between the regulatory target value and the real-time behavior monitoring value, and outputs control commands to the regulatory execution unit, specifically including: The regulatory target value and real-time behavior monitoring value are obtained, the difference value is calculated, and the difference value is input into the reinforcement learning control algorithm to generate the intensity of the regulatory action based on the reward function and the state space. Based on the intensity of the regulatory action, control commands are generated and output to the regulatory execution unit to control the operation of the execution unit. The regulatory enforcement unit adjusts the intensity of actions and changes actual behavior according to control instructions. By continuously collecting real-time behavioral data and comparing it with regulatory target values, the difference value is updated to form a closed-loop regulatory cycle.

[0011] Preferably, the step of using a compliance feedback device to obtain current nursing compliance indicators, comparing them with a preset standard range, and determining whether regulatory rules need to be updated specifically includes: The compliance indicator data is transmitted to the regulatory system and compared with the preset standard range to determine whether the compliance indicator is within a reasonable range. If the compliance indicator is lower than the lower limit of the standard range, the regulatory system generates an instruction to enhance the intensity of the regulatory rule through the executor. If the compliance indicators exceed the upper limit of the standard range, the regulatory system generates instructions to weaken the intensity of the regulatory rules and reduce regulatory output through the executor. During the rule adjustment process, clustering algorithms are applied to analyze historical compliance data, establish a correlation model between compliance indicators and rule strength, and use the correlation model to predict the optimal rule strength for feedforward regulatory control.

[0012] Preferably, the step of using a decision tree algorithm to derive the adjustment amount of the regulatory rule and outputting the adjustment instruction to the execution unit specifically includes: The compliance feedback device obtains the current compliance indicators in real time and compares them with the preset standard range to determine whether the compliance indicators exceed the standard range. When the compliance indicator exceeds the standard range, the deviation value between the compliance indicator and the standard median is calculated, the deviation value is input into the decision tree algorithm, and the rule adjustment amount is derived based on the preset decision rule base. The rule adjustment amount is converted into an adjustment command, which is then output to the execution unit via a signal converter to drive the execution unit to adjust the rule strength. The execution unit changes the rule application according to the adjustment instructions, thereby optimizing compliance indicators and bringing them back to the standard range.

[0013] Preferably, the optimization of the timing output of control commands based on the response characteristics of the monitoring execution unit to ensure the continuous and stable monitoring process specifically includes: Obtain the response time parameters of the regulatory execution unit, establish a dynamic model of the execution unit, obtain the relationship curve between action intensity and regulatory effect, and calculate the optimal output timing through the dynamic model based on the regulatory objectives and actual behavior values. Using the deviation in action intensity and the deviation in regulatory effect as inputs, a time series optimization algorithm is applied to obtain the timing adjustment amount of the control command; The adjusted control commands are output to the monitoring and execution unit according to the calculation sequence for dynamic adjustment; During the dynamic adjustment process, nursing behavior data and monitoring effects are collected in real time and compared with the target value to generate deviation data. When the deviation exceeds the threshold, the deviation is input into the time series optimization algorithm to iteratively optimize the timing of control commands.

[0014] Preferably, the step of continuously tracking nursing behavior signals and monitoring rules, updating preset rules in real time, and adapting to changes in nursing needs to achieve refined monitoring of nursing behavior specifically includes: The system acquires real-time care behavior signals and current monitoring rules from users, compares them with preset normal ranges, determines whether they deviate from the normal range, triggers an early warning mechanism if they deviate, and determines the extent of rule updates based on the degree of deviation. By applying regression analysis algorithms, based on users' historical nursing behavior data and regulatory rule trends, we can predict future changes in nursing behavior. Based on real-time nursing behavior signals, current regulatory rules, and predicted changes, the regulatory rules are dynamically updated, and new rule settings are generated. The new rules are set and distributed to the monitoring equipment, and the control equipment adjusts the monitoring output according to the new rules. After the rules are updated, nursing behavior signals are continuously monitored to determine whether they have returned to normal. If they have not returned to normal, the rules are updated again. Real-time nursing behavior signals, monitoring rules, and update results are stored in a database to optimize regression analysis algorithms.

[0015] Preferably, the present invention also includes a nursing behavior monitoring system based on smart elderly care, used to implement the nursing behavior monitoring method based on smart elderly care as described above, the system comprising: The behavior signal processing module acquires nursing behavior signals from the nursing behavior monitoring device, applies a medium-value filtering algorithm to suppress noise, and outputs stable nursing behavior signals. The behavior status analysis module extracts time series features based on stable nursing behavior signals, judges the behavior status, and calculates the intervention target value through a neural network prediction model when the behavior status is abnormal. It also applies a fuzzy logic algorithm to dynamically adjust the monitoring output. The monitoring and control module applies a reinforcement learning control algorithm to generate the intensity of monitoring actions based on the difference between the target value and the actual value, and outputs control commands to the monitoring execution unit. The rule optimization module obtains compliance indicators through the compliance feedback device, combines them with preset standard ranges, uses a decision tree algorithm to deduce the rule adjustment amount, and outputs adjustment instructions. The dynamic monitoring module optimizes the timing output by combining the response characteristics of the monitoring execution unit to ensure continuous and stable monitoring. It also dynamically updates preset rules by tracking behavioral signals and rules, and stores the data in the database.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Starting with the processing of raw nursing behavior signals, a median filtering algorithm is used to suppress random interference and reduce the impact of noise on subsequent analysis, making the processed signal more closely resemble the true characteristics of nursing behavior. Based on the processed signal, the periodicity and trend direction of change are extracted by analyzing time series features. A state classification model is then used to identify normal and abnormal stages. Compared with fixed threshold judgment, this approach is better able to capture the dynamic changes in nursing behavior and is adaptable to diverse nursing scenarios.

[0017] When an abnormal phase is identified, the intervention target value is calculated based on preset monitoring rules and dynamically adjusted in conjunction with behavioral trends. This ensures that the intervention target is more closely aligned with the real-time nursing status, avoiding intervention bias caused by static targets. A reinforcement learning control algorithm is applied to generate the intensity of monitoring actions. Control commands are adjusted based on the difference between real-time behavioral monitoring values ​​and target values, making monitoring execution more adaptive and flexibly responding to the reactions and needs of different patients.

[0018] The system utilizes a compliance feedback device to obtain current nursing compliance indicators. After comparing these indicators with preset standard ranges, if they exceed the range, a decision tree algorithm is used to deduce rule adjustment amounts. This drives the regulatory rules to be dynamically updated based on actual nursing outcomes, enhancing the adaptability and flexibility of the rules. The timing of control command output is optimized by combining the response characteristics of the regulatory execution unit, ensuring a continuous and stable regulatory process and reducing interference with the nursing rhythm caused by command fluctuations.

[0019] By continuously tracking nursing behavior signals and regulatory rules, and updating preset rules in real time, the entire regulatory system can gradually adapt to changes in nursing needs, be continuously optimized in long-term application, better integrate into the actual scenarios of smart elderly care, and meet the personalized nursing supervision needs of different elderly people. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the working principle of the nursing behavior monitoring method based on smart elderly care as described in this invention. Figure 2 A flowchart illustrating the median filtering algorithm for signal processing; Figure 3 A flowchart for classifying behavioral states; Figure 4 A flowchart for generating control instructions to reinforce learning; Figure 5 A flowchart for deriving the adjustment amount for decision tree rules. Detailed Implementation

[0021] 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.

[0022] Please see Figure 1 This invention provides a method for monitoring nursing behavior based on smart elderly care, the method comprising: This system achieves refined monitoring of nursing behaviors in smart elderly care scenarios through multi-module collaboration. The system architecture comprises a signal acquisition layer, an analysis and decision-making layer, and an execution feedback layer, forming a closed-loop control system. The signal acquisition layer collects raw signals of nursing behaviors in real time through wearable devices, environmental sensors, and other monitoring devices, including but not limited to parameters such as amplitude, frequency, and duration of movements. The analysis and decision-making layer preprocesses the signals and extracts features to establish a dynamic monitoring model. The execution feedback layer outputs control instructions based on the decision results and achieves rule self-optimization through compliance feedback.

[0023] The system workflow is as follows: Raw nursing behavior signals are processed by median filtering and moving average, then input into the time series analysis module to extract behavioral cycle and trend features. The state classification model determines the current behavioral stage based on feature vectors; abnormal states trigger the intervention target calculation module. The reinforcement learning controller generates control instructions based on the target difference, driving the execution unit to adjust the intensity of nursing behavior. The compliance feedback device monitors indicator deviations in real time and dynamically corrects regulatory rules using a decision tree algorithm. The execution unit response characteristic model optimizes instruction timing to ensure a smooth control process. The historical data storage module continuously accumulates behavioral features and rule adjustment records, supporting online model updates.

[0024] Example 1: See Figure 2 The specific technical details of noise suppression processing for nursing behavior signals are as follows: This process is implemented through a multi-level signal processing architecture. The original nursing behavior signals originate from a multimodal sensor network installed in the nursing environment, including a triaxial accelerometer, a pressure sensor array, and an infrared motion detector. These sensors collect physical quantity data according to a synchronized clock signal and a unified timestamp, forming an initial signal sequence. The signal acquisition terminal uses an industrial-grade analog-to-digital converter chip to convert analog signals into digital signals, achieving a sampling accuracy of 16 bits and a uniform sampling rate of 50 data points per second. The converted digital signals are transmitted to the central processing unit through an isolated communication interface, with a CRC check mechanism used during transmission to ensure data integrity.

[0025] The signal preprocessing stage begins with time alignment to address time discrepancies in data from multiple sensors. The system maintains a global clock server, and each acquisition node synchronizes its local clock via the NTP protocol, keeping the time deviation within 10 milliseconds. For out-of-order data packets caused by network latency, the reordering module rearranges the data sequence based on the timestamp information embedded in the data packets. The aligned signal then enters the amplitude normalization stage, mapping the output values ​​of different sensors to a unified numerical range and eliminating dimensional inconsistencies caused by individual sensor differences. The normalization coefficients are determined through a calibration process; each sensor must complete standard testing before deployment, recording its input-output characteristic curves.

[0026] Median filtering employs a sliding window mechanism for real-time noise suppression. The window width is set to include a time period of 50 consecutive sampling points, equivalent to 1 second of data. Within each sliding window, the algorithm sorts all sampled values ​​based on the signal amplitude. After sorting, the sampled value at the middle position is selected as the output value of that window, i.e., the value of the 25th sampling point. This processing method effectively filters out pulse-type interference in the signal, especially spike noise caused by momentary poor sensor contact or electromagnetic interference. For interference signals with a duration less than half the window width, median filtering can completely eliminate their influence.

[0027] The outlier detection algorithm is based on statistical principles and is used to identify and remove sampled values ​​that significantly deviate from the normal range. Within each filtering window, the system calculates the interquartile range (IQR) of the sampled values, which is the difference between the third quartile and the first quartile. Any sampled point exceeding 1.5 times the IQR of the median is marked as an outlier. Marked outliers do not participate in subsequent calculations; instead, an interpolation algorithm generates replacement values. The outlier detection results are also recorded in metadata for use by the quality assessment module. When consecutive outliers exceed 20% of the window width, the system determines that the signal quality of that segment is unreliable and triggers a signal re-acquisition process.

[0028] The signal reconstruction stage is responsible for restoring the processed discrete data into a continuous time series. A linear interpolation algorithm is used to fill data gaps caused by outlier removal. The algorithm establishes a linear relationship between two adjacent valid sampling points and calculates the theoretical value of the gap location based on the time interval ratio. For data points at the window edges, a mirror extension technique is used to handle boundary effects and avoid introducing artificial jumps. The reconstructed signal enters the moving average processing stage to further smooth random fluctuations. The weighted moving average algorithm uses a Gaussian kernel function to generate weight coefficients, with a time constant set to 0.5 seconds. The current sampling point and the 10 sampling points before and after it participate in the weighted calculation, with sampling points closer to the current point having a larger weight.

[0029] The signal quality assessment module performs a comprehensive inspection of the processed data. The signal-to-noise ratio (SNR) is calculated using a segmented estimation method, dividing the signal into several frames and calculating the signal energy to noise energy ratio for each frame. Continuity testing statistically analyzes the percentage of valid sampling points and calculates the proportion of missing data in the total data. The system sets the SNR threshold to 20 dB, requiring a continuity requirement of a missing rate of less than 5%. Substandard signals trigger a quality alarm and simultaneously initiate a signal re-acquisition process. Quality-qualified data is stored in a circular buffer, designed as a first-in, first-out queue capable of storing 10 minutes of data, providing a stable data flow for subsequent feature analysis modules.

[0030] All parameters in the signal processing process can be adjusted through the configuration interface to adapt to the needs of different application scenarios. The median filter window width can be adjusted from 0.5 seconds to 2 seconds to accommodate different frequencies of nursing care. The sensitivity of anomaly detection is achieved by adjusting the interquartile range coefficient, which can be set from 1.0 to 2.0. The time constant of the moving average algorithm supports dynamic configuration from 0.1 seconds to 1 second. These parameters form a preset scheme library, automatically matching the optimal parameter combination based on sensor type and behavioral characteristics.

[0031] The system maintenance module is responsible for health monitoring and performance optimization of the signal processing link. Processing latency monitoring measures the time required from signal input to output in real time, issuing a performance alarm when it exceeds 100 milliseconds. Memory management employs a garbage collection mechanism, periodically cleaning up temporary data to release resources. Anomaly detection monitors the input-output relationships of each stage, initiating a self-check program when unexpected data changes are detected. An online learning function records the changing trends of signal characteristics, automatically fine-tuning processing parameters to maintain optimal performance.

[0032] The signal output interface supports multiple data formats, including raw sampled value sequences, processed smoothed curves, and feature parameter summaries. Timestamp information is accurate to the millisecond level, maintaining a strict correspondence with the sampled values. Data compression algorithms can be selected in lossless or lossy modes to adapt to different bandwidth transmission environments. Metadata records a complete processing history, including filtered parameters used, outlier statistics, and quality assessment results, providing a basis for subsequent analysis.

[0033] The entire signal processing flow runs under the scheduling of a real-time operating system, ensuring strict time determinism. Computational tasks on the critical path are set to the highest priority to ensure completion within the specified time. Task allocation on multi-core processors employs a static binding strategy to reduce thread switching overhead. Memory access optimization is achieved through data locality design, arranging frequently accessed data structures in adjacent memory regions. This implementation can meet real-time requirements on typical hardware platforms, with processing latency consistently controlled within 50 milliseconds.

[0034] Example 2: See Figure 3 This invention presents a complete technical solution for time-series feature analysis and state recognition of nursing behaviors. Based on preprocessed nursing behavior signals, the solution achieves accurate classification of behavioral states through multi-dimensional feature extraction and pattern recognition. The system employs a hierarchical processing architecture, progressively extracting discriminative feature indicators from the raw signals to construct a state model that reflects the essential characteristics of nursing behaviors.

[0035] The time series decomposition module receives the smoothed signal from the preprocessing stage and decomposes it into three components: a trend component reflecting the long-term direction of behavior, a periodic component reflecting the repetitive patterns of behavior, and a residual component containing random fluctuations and sudden events. Trend extraction uses a locally weighted regression algorithm with an adaptively adjusted analysis window. The window width maintains a fixed proportional relationship with the estimated behavior cycle length. The algorithm calculates a weighted least squares fit within the sliding window, with data points closer to the current point receiving greater weight. The resulting regression curve is output as the trend component. Cycle detection is achieved through frequency domain analysis, calculating the power spectral density of the signal and identifying the frequency components corresponding to significant peaks. For typical turning over nursing behaviors, the system presets a detection frequency range corresponding to cycle lengths between 40 and 180 minutes. The residual component is obtained by subtracting the trend and periodic components from the original signal, reflecting signal variations not explained by the former two.

[0036] The feature engineering phase extracts discriminative index parameters from the decomposition results. Trend features include average slope, rate of change of curvature, and distribution of extreme points. Slope calculation uses a piecewise linear approximation, dividing the trend curve into segments, each lasting at least 5 minutes, and calculating the slope value for each segment. The rate of change of curvature is estimated using second-order difference to reflect the speed of change in trend direction. The distribution of extreme points statistically analyzes the frequency and amplitude differences of peaks and troughs. Periodic features include period length stability, phase consistency, and amplitude uniformity. The stability index measures the coefficient of variation of continuous period lengths, phase consistency assesses the similarity of periodic waveforms, and amplitude uniformity analyzes the distribution of peak heights within a period. Residual features mainly examine standard deviation and autocorrelation characteristics. Standard deviation reflects the intensity of random fluctuations, and autocorrelation reveals potential short-term correlations in the residuals.

[0037] The behavioral state classification model employs a Support Vector Machine (SVM) algorithm, which learns the optimal classification boundary in the feature space during the training phase. The training dataset contains thousands of labeled samples, each corresponding to a nursing behavior record within a specific time period. Sample labeling was completed by a team of nursing experts, determining the behavioral state category for each time period based on video recordings and nursing logs. Feature vector construction comprehensively considers three dimensions of feature indicators: trend, periodicity, and residual, forming a high-dimensional feature space. During model training, kernel function mapping technology is used to transform the original feature space into a higher-dimensional space, making it easier for samples of different categories to be separated by a linear hyperplane. The classifier outputs a binary label, distinguishing between normal and abnormal behavioral states. The determination of abnormal states comprehensively considers the deviation of multiple feature indicators; when the trend slope continuously increases and the periodic phase advances, the system tends to classify it as an abnormal state.

[0038] The intervention target calculation module employs a neural network architecture to establish a non-linear mapping relationship from the current state to the ideal control target. The network input layer receives multi-dimensional feature vectors, including real-time behavioral features, environmental parameters, and user physiological indicators. Behavioral features are derived from the output of the feature engineering stage; environmental parameters encompass sensor data such as room temperature, humidity, and light intensity; and user physiological indicators include wearable device monitoring values ​​such as heart rate and blood oxygen saturation. The network hidden layer uses a recurrent unit structure with a gating mechanism, capable of memorizing historical state information and influencing current decisions. The output layer generates three key control parameters: target behavior intensity, allowable fluctuation range, and target achievement time requirement. Target behavior intensity is a normalized value representing the expected level of nursing behavior activity. The allowable fluctuation range defines a reasonable interval around the target value, and the target achievement time requirement specifies the timeframe within which the system expects to achieve the target.

[0039] A dynamic adjustment mechanism enables the system to optimize its control strategy based on real-time feedback. The fuzzy logic controller receives the target difference signal, which is the deviation between the current behavior intensity and the target value, as well as the trend of this deviation. The controller internally establishes a language rule base, converting precise numerical inputs into fuzzy concepts, such as "the deviation is large and rapidly decreasing." The inference engine activates the corresponding rules based on the current input, generating fuzzy output. The defuzzification process converts the linguistically described control suggestions into precise control intensity values. Control commands are updated every 30 seconds, taking into account the response characteristics of the actuator and the user's adaptability. When the actual behavior intensity deviates from the target value for more than 15 minutes, the system triggers a target recalculation process, where the neural network re-evaluates the ideal target value based on the latest data.

[0040] The status tracking module continuously records the trajectory of behavioral status changes, forming a complete profile of nursing behavior. Each status transition event is accompanied by a detailed timestamp and feature snapshot, stored in a recurring database. The system maintains a status duration counter to track the duration of the current status. When the duration of an abnormal status exceeds a preset threshold, the system escalates the response level and takes more aggressive intervention measures. Historical status sequences can be used to analyze the long-term evolution trend of behavioral patterns, supporting gradual adjustments to nursing plans. The user interface intuitively displays the current status and recent change curves, helping nursing staff understand the basis of the system's decisions.

[0041] The model update mechanism ensures the system can adapt to individual differences and environmental changes. The online learning function continuously monitors classification accuracy and target achievement rate, initiating a model retraining process when performance metrics drop below a threshold. The incremental learning algorithm gradually integrates newly collected labeled samples into the existing model without forgetting existing knowledge. The feature selection module periodically evaluates the discriminative power of each feature, discarding features with low contribution and adding newly discovered significant features. Model version management records the content and effects of each update, supporting rollback to previous stable versions when necessary. User personalization settings are achieved by adjusting model parameters, making system decisions more aligned with the habits and needs of specific users.

[0042] The entire analysis process is implemented on a distributed computing framework, with feature extraction and model inference tasks distributed across different computing nodes for parallel execution. A time-series database optimizes the storage and retrieval efficiency of massive amounts of monitoring data, supporting rapid queries of signal characteristics for any time period. A stream processing engine ensures real-time data analysis, keeping the latency from signal input to status output within acceptable limits. A system resource monitoring module balances the computing load, dynamically allocating more resources to critical tasks during peak periods. A fault-tolerance mechanism handles abnormal situations such as sensor failures or communication interruptions, activating predictive mode to maintain basic system functionality when data is missing.

[0043] Example 3: See Figure 4 This solution focuses on the collaborative working mechanism of reinforcement learning control algorithms and compliance feedback systems to form a dynamic closed-loop nursing behavior monitoring system. It optimizes control strategies through real-time interactive learning and achieves adaptive adjustment of rules through multi-dimensional compliance assessments. The system architecture includes four functional components: state awareness, decision generation, execution control, and effect evaluation. Structured data is exchanged between these components through standardized interfaces.

[0044] The core of the reinforcement learning controller is the deep Q-network algorithm, which is based on the Markov decision process framework. The state space S is defined as a five-dimensional vector: Where Δ represents the percentage difference between the current behavior intensity and the target value, The instantaneous rate of change representing behavioral intensity, φ reflects the phase of the current nursing cycle, ρ records the historical compliance rate, and σ represents the intensity level of the currently applied rule. The action space A is discretized into 10 levels, each corresponding to a different regulatory intensity output, linearly distributed from 0 (no intervention) to 9 (maximum intervention). The state transition function T models the environment's response to the regulated actions, approximated by a neural network. The design of the reward function R comprehensively considers both short-term and long-term goals. , Where α, β, γ, and η are weighting coefficients. The average compliance rate within the sliding window. They represent the values before and after the adjustment of the rule strength respectively. The experience replay buffer stores quadruples (s, a, r, s'), namely state, action, reward, and new state, and the capacity is set to 1000 records. The target network synchronizes parameters with the online network regularly to reduce the bias caused by bootstrapping.

[0045] The rule adjustment mechanism adopts a hierarchical response strategy. The primary adjustment targets short-term fluctuations and is achieved by fine-tuning the existing rule parameters, with the response time controlled within 5 minutes. The intermediate adjustment deals with continuous deviations, recalculates the rule strength coefficient, and the response time is about 15 minutes. The advanced adjustment copes with systematic changes, reconstructs the rule logic structure, and usually takes 1 hour to complete. The adjustment amount is calculated based on the decision tree algorithm, and the feature vector includes the current deviation degree, deviation duration, historical adjustment effect, and user sensitivity score. The decision tree nodes store adjustment strategies, including the amplitude of strength increase or decrease and the expected recovery time. The rule version management system records the complete context of each adjustment, supporting traceability and rollback.

[0046] The actuator interface module implements the conversion of control instructions into physical signals. The digital output channels correspond to discrete regulation levels, and the analog output supports continuous adjustment. The signal conditioning circuit eliminates noise interference, and the isolation protection circuit prevents electrical shocks. The execution status feedback loop monitors the consistency between the actual output and the instruction, and abnormal situations trigger retry or alarm. The timing controller coordinates the linkage operations of multiple actuators to avoid resource conflicts. The priority scheduler processes concurrent instructions to ensure the timely execution of critical operations. The execution log records the detailed information of each regulation, including instruction content, execution time, and actual effect.

[0047] The online learning system continuously optimizes the control strategy. The performance monitoring module tracks key metrics: target achievement rate, compliance stability, and intervention frequency. In the policy evaluation stage, the control effects under different parameter configurations are compared, and a multi-objective optimization method is used to balance various metrics. The trigger conditions for model update include: the performance decreases for 3 consecutive evaluation cycles, or the environmental characteristics change significantly. The incremental learning algorithm integrates new data while retaining existing knowledge, and the network parameters adopt an elastic weight update strategy. The knowledge distillation technology compresses complex models into lightweight versions to adapt to the computing power of edge devices.

[0048] The state visualization interface presents the panoramic view of the system operation. The real-time monitoring area displays the current behavior curve, regulation instructions, and compliance metrics. The historical analysis area shows trend comparisons and event logs. The parameter configuration area allows manual adjustment of algorithm coefficients and thresholds. The warning notification area highlights abnormal states and recommended measures. All visualization elements support interactive exploration, and the data drilling function can view the detailed records at any time point. The interface layout adapts to different screen sizes, and key information is preferentially displayed on mobile devices.

[0049] An anomaly handling mechanism ensures system robustness. Sensor failure detection is achieved through signal integrity checks and heartbeat packet monitoring. Communication interruption handling employs local caching and breakpoint resumption techniques. Algorithm anomaly capture includes numerical overflow checks and convergence monitoring. The failover scheme automatically transfers service load between primary and backup nodes. The recovery strategy distinguishes between temporary errors and systemic failures, implementing different levels of recovery measures. All anomalies are logged for subsequent root cause analysis and preventative improvements.

[0050] The system deployment architecture takes into account the allocation of computing resources. Real-time control components are deployed on edge gateways to meet low-latency requirements. Data analysis components run on cloud servers, utilizing powerful computing capabilities to process complex models. A data synchronization mechanism maintains consistency across nodes, and a timestamp-first strategy is used to resolve conflicts. The resource monitoring module dynamically adjusts task allocation, prioritizing core functions during peak periods. Energy management optimization algorithms extend the battery life of mobile devices and intelligently adjust the sampling frequency based on usage patterns.

[0051] The version upgrade process ensures a seamless transition. A canary release mechanism first pushes the new version to a select group of nodes, and only after verification is it rolled out nationwide. Data format conversion maintains backward compatibility, allowing older versions to read newly generated data. The rollback plan clearly defines the downgrade steps and conditions, ensuring the system remains available at all times. Upgrade logs meticulously record changes for easy issue tracking. The system notifies users in advance of upgrade arrangements, minimizing disruption to nursing workflows.

[0052] Example 4: Reference Figure 5 The collaborative working mechanism of decision tree algorithm and dynamic adjustment of execution unit is demonstrated through specific application scenarios. Taking the monitoring of turning over in nursing care as an example, the system collects user posture data through multi-source sensors, including a pressure distribution sensor array, a triaxial accelerometer, and mattress-embedded sensors. The pressure sensors are arranged in a 20×20 matrix, acquiring a full-body pressure distribution image every 2 minutes. After data preprocessing, posture feature parameters are generated. The accelerometer records the body movement trajectory at a frequency of 10Hz, and the turning angle is calculated through a posture calculation algorithm. The mattress sensor monitors the micro-motion frequency and identifies potential discomfort.

[0053] When handling compliance indicator deviations, the decision tree algorithm first constructs a feature vector table. The table below shows the input features and corresponding adjustment strategies involved in a typical decision-making process:

[0054] The decision tree construction process uses the C4.5 algorithm. Each non-leaf node stores five decision conditions, corresponding to the feature dimensions in the table. The node splitting criterion is based on the information gain ratio, and the minimum number of samples for a leaf node is set to 15. When the turning angle deviation exceeds 15% for 30 minutes and the user's sensitivity score is 0.8, the system may choose the "feedback compensation" strategy with an amplitude coefficient of 1.1. The decision path is recorded in the adjustment log, containing complete information such as triggering conditions, decision process, and execution results.

[0055] The dynamic adjustment of the execution unit is implemented based on a second-order system model. The turning-over assist device, as a typical actuator, has its response characteristics obtained through step response testing. Test data shows that when the input control signal suddenly increases from 0 to 5V, the average response time for the device to complete a 90-degree turning motion is 8 seconds, with overshoot controlled within 15%. Based on these parameters, the system establishes a transfer function model to predict the execution effect under different control commands. The timing optimization algorithm considers the device's motion inertia and mechanical delay, decomposing the turning motion into three stages: acceleration, constant speed, and deceleration, with different time budgets allocated to each stage.

[0056] Control command scheduling employs a three-tiered priority management system. Emergency turning requests (such as those triggered by detected pressure ulcer risk) enter the real-time queue, with a delay of no more than 2 seconds; routine positioning commands enter the general queue and are processed within 10 seconds; parameter calibration commands are stored in the background queue, awaiting execution when the system is idle. The queue monitoring module tracks the depth of each queue in real time, automatically pausing low-priority tasks when the real-time queue backlog exceeds three requests. Command timestamps are synchronized using the IEEE 1588 precise time protocol, with clock deviations between devices less than 1 millisecond.

[0057] Rule version management enables iterative updates. Each adjustment generates a new rule version, recording metadata including adjustment time, decision basis, and execution results. The version rollback function allows reverting to any of the three most recent stable versions. A rule difference comparison tool visually displays changes between versions, helping caregivers understand the adjustment logic. Version compatibility checks ensure new rules match existing equipment configurations, preventing execution failures due to hardware limitations.

[0058] Anomaly handling is designed for typical failure scenarios. When obstruction of the turning device is detected, the system immediately stops the current command, returns to a safe position, and initiates the resistance detection program. In the event of a communication interruption, the edge node continues to operate based on the last valid command, while caching local data. Anomalies in sensor data trigger a triple verification mechanism, comparing the results of pressure distribution, acceleration, and video analysis. All anomalies generate standardized reports, including fault symptoms, handling measures, and repair recommendations.

[0059] The user interface displays the entire adjustment process. The real-time view shows the current body position angle, target value, and progress. The history view presents adjustment records in a timeline format, allowing filtering to view specific types of decision events. The parameter configuration panel allows caregivers to manually fine-tune sensitivity thresholds and response speeds. The alert panel uses color coding to distinguish between different levels of urgency in abnormal conditions, with red indicating situations requiring immediate intervention.

[0060] The data collection system records the entire decision-making chain. Each adjustment cycle saves raw sensor data, feature vectors, decision paths, execution instructions, and final results. Data is anonymized and stored in a knowledge base for subsequent algorithm optimization. The privacy protection module automatically blurs facial features in video data and encrypts and stores personal health information.

[0061] The system integrates and supports various nursing devices. It connects to turning beds, lifting supports, and other devices via a standardized CAN bus interface. The protocol converter is compatible with industrial protocols such as Modbus and Profibus, enabling interoperability between devices from different manufacturers. A security authentication module verifies the legitimacy of each connected device, preventing unauthorized access.

[0062] Performance monitoring continuously evaluates system effectiveness. Key performance indicators include decision accuracy, response latency, and execution success rate. Resource usage monitoring covers dimensions such as CPU load, memory usage, and network traffic. When indicators exceed normal ranges, the system automatically triggers diagnostic procedures to identify potential root causes of problems. Capacity planning tools predict future loads based on historical data and recommend necessary hardware upgrades.

[0063] The maintenance mode supports system debugging and calibration. The engineer interface provides detailed equipment status diagnostics and signal quality analysis. The automatic calibration procedure guides caregivers through sensor zero-point and range adjustments. The log export function supports multiple formats for easy offline analysis and troubleshooting. The remote assistance module allows authorized engineers to securely access the system and make necessary configuration modifications.

[0064] Example 5: A complete adaptive adjustment system was established around the dynamic updating mechanism of nursing behavior supervision rules. This system achieves intelligent optimization of rule parameters by continuously monitoring the interaction between user behavior characteristics and supervision effectiveness. The core architecture includes five functional modules: behavior monitoring, deviation analysis, rule prediction, version management, and effect verification, forming a closed-loop rule evolution process.

[0065] The behavior monitoring module employs multimodal data fusion technology to process real-time care signals. A fiber optic sensor array installed under the mattress collects pressure distribution data 10 times per second, which is then converted into postural status indicators using a pattern recognition algorithm. A ceiling-mounted millimeter-wave radar monitors a wide range of movement trajectories, supplementing the blind spots of the mattress sensors. A wearable inertial measurement unit records detailed limb movements, providing fine parameters such as joint angles and movement speed. An environmental sensor network captures changes in room temperature, humidity, and light intensity, as these external factors can influence user behavior. All sensor data is time-aligned and spatially calibrated to form a unified framework for describing care behavior.

[0066] The deviation detection mechanism employs a three-tiered sensitivity monitoring strategy. Level 1 detection targets instantaneous anomalies, comparing the current sampled value with a preset threshold for immediate deviation. A primary alarm is triggered when the interval between turning over exceeds 3 hours or the angle of a single turn is less than 45 degrees. Level 2 detection analyzes short-term trends, calculating the average behavioral characteristics within a 30-minute sliding window. A medium-level warning is generated if the average activity level is more than 20% below the individual's baseline. Level 3 detection assesses long-term pattern changes, identifying gradual changes in behavioral habits using a 7-day moving average. The results from each level of detection are weighted and integrated to form a comprehensive deviation score, reflecting the severity of the current behavioral abnormality.

[0067] The rule prediction engine employs ensemble learning to analyze historical data. A time-series database stores complete behavioral records and rule application logs for the past 30 days, with a sampling interval of 1 minute. The feature engineering phase extracts key indicators such as periodic patterns, trend slopes, and the distribution of abrupt change points. A random forest algorithm trains 100 decision trees, each using a different subset of data and feature subspace. The prediction objectives include possible behavioral patterns, potential risk points, and optimal intervention timing within the next 6 hours. The model is retrained every 24 hours, incorporating the latest observation data to maintain predictive accuracy.

[0068] The rule generation process employs a differential evolution strategy. Currently valid rules serve as parent rules, generating five sets of candidate new rules through random mutation. Mutation operations include three types: parameter fine-tuning, logical expansion, and conditional recombination. Fitness evaluation considers three dimensions: historical data fit to examine the new rule's explanatory power for past situations, stress testing to assess the rule's robustness in extreme scenarios, and feasibility analysis to calculate the matching degree between required resources and equipment capabilities. The optimal rule selection uses Pareto optimal frontier analysis to balance the scores of different evaluation dimensions. The new rule generation cycle is typically 4 hours, but can be shortened to 30 minutes in emergency situations.

[0069] The version control system enables full lifecycle management of rules. Semantic version numbers identify the nature of each update: major version number changes indicate significant logical adjustments, minor version numbers add corresponding parameter optimizations, and revision numbers record bug fixes. A version dependency graph records inheritance and substitution relationships between rules, supporting parallel testing of multiple versions. A rollback mechanism retains the five most recent stable versions, allowing for rapid recovery in case of system performance degradation. Version difference reports automatically generate change summaries to help nursing staff understand the adjustments.

[0070] The deployment and execution module ensures a smooth transition for rule updates. A canary release strategy applies new rules to 10% of user nodes initially, and the scope is gradually expanded after observing the actual effects. Shadow mode allows new and old rules to run in parallel, comparing the decision differences between the two without affecting actual control. A circuit breaker mechanism suspends the application of new rules in abnormal situations, reverting to the previous stable version. Deployment status monitoring tracks the rule synchronization progress of each node in real time, resolving version inconsistencies. The user notification system informs users in advance of significant rule changes, explaining the expected impact and precautions.

[0071] The effectiveness verification system continuously evaluates the performance of the new rules. Key performance indicators include behavioral achievement rate, intervention frequency, and user comfort score. An A / B testing framework compares the control effects of the new and old rules under the same conditions. Longitudinal analysis tracks long-term behavioral changes of the same user under different rules. Feedback collection channels integrate automated monitoring data and subjective evaluations from caregivers. The verification period is typically set at 24 hours, extended to 72 hours for high-risk rules. Verification results are categorized into three types: clear improvement, no significant difference, and performance decline, guiding subsequent adjustments.

[0072] The knowledge base construction process accumulates experience in rule evolution. Each rule version and its application effect form a case record, annotating success conditions and limiting factors. Similarity retrieval algorithms help discover relevant historical cases in the current scenario. Pattern mining technology identifies high-frequency adjustment paths and effective rule combinations. The knowledge graph establishes a network of associations between nursing behavior characteristics, environmental factors, and optimal rules. Search results are sorted by time weight, with recent cases receiving higher priority.

[0073] An exception handling mechanism ensures the safety of rule updates. A syntax checker verifies the logical integrity of new rules, preventing contradictory conditions. A conflict detector identifies mutually exclusive clauses with the existing rule set. Resource auditing assesses the impact of new rules on system load and rejects requests exceeding capacity. Rollback triggers automatically restore old rules when critical metrics deteriorate. Fault tree analysis locates the root cause of problems during the update process. A security log records all rule change operations, supporting post-event auditing.

[0074] The user interface presents a dynamic visualization of rules. A timeline view shows the evolution of rule versions, highlighting important update nodes. A heatmap displays the application effects of different rules in various scenarios. A comparison panel lists the key parameters and logical differences between old and new rules side-by-side. A simulator predicts the potential impact of specific rule changes. An access control system manages the scope of rule operations for different roles: nursing staff can view rules but cannot modify them; engineers can fine-tune parameters; and only administrators can adjust the logical structure.

[0075] The system maintenance features support long-term stable operation. An automatic cleanup mechanism regularly archives expired rule versions, freeing up storage space. Index optimization maintains efficient knowledge base retrieval. Memory management prevents rule caching from consuming excessive resources. Performance analysis tools identify bottlenecks in the rule engine. A test sandbox provides a secure environment for validating high-risk updates. The maintenance window is set weekly, with non-urgent updates concentrated during this period.

[0076] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring nursing behavior based on smart elderly care, characterized in that, The method includes the following steps: The original nursing behavior signal is obtained from the nursing behavior monitoring device. The nursing behavior signal is the amplitude, frequency and duration parameters of the action collected by wearable device and environmental sensor. The signal is processed by median filtering algorithm to suppress noise and eliminate random interference, and the processed nursing behavior signal is output. Based on the processed nursing behavior signals, the time series characteristics of the signals are analyzed, the change cycle and trend direction of nursing behavior are extracted, and the nursing behavior is determined to be in a normal or abnormal stage through a state classification model. When it is identified as an abnormal stage, the extracted behavioral features are input into a trained neural network prediction model. The required intervention target value is calculated through the neural network prediction model. The intervention target value includes the target behavior intensity, allowable fluctuation range and target achievement time requirement. The monitoring target is dynamically adjusted in combination with the behavior trend. The reinforcement learning control algorithm is applied to generate the intensity of the regulatory action based on the difference between the regulatory target value and the real-time behavior monitoring value. The intensity of the regulatory action is 10 levels after the action space is discretized in the reinforcement learning control algorithm. Each level corresponds to a different control intensity output, from level 0 representing no intervention to level 9 representing a linear distribution of maximum intervention. The control command is output to the regulatory execution unit. The current nursing compliance indicators are obtained using a compliance feedback device. These indicators are historical compliance rates. The data are compared with a preset standard range to determine whether the regulatory rules need to be updated. If the compliance indicators exceed the standard range, the regulatory rule adjustment amount is derived based on the difference value using a decision tree algorithm, and the adjustment instruction is output to the execution unit. By combining the response characteristics of the regulatory execution unit, the timing output of control instructions is optimized to ensure the continuous and stable regulatory process. Furthermore, by continuously tracking nursing behavior signals and the dynamics of regulatory rules, preset rules are updated in real time to adapt to changes in nursing needs and achieve refined regulation of nursing behavior.

2. The method for monitoring nursing behavior based on smart elderly care according to claim 1, characterized in that, The process of using a median filtering algorithm to suppress noise in the signal, eliminate random interference, and output the processed nursing behavior signal specifically includes: The nursing behavior signal is input into the median filtering algorithm module, and then the median filtering algorithm is applied to perform window sliding processing on the signal to divide the signal into multiple time windows; Within each time window, sort the signal values ​​and select the median, remove outliers, and output the smoothed sub-signal; The smoothed sub-signals are reconstructed into a time series to obtain a noise-suppressed nursing behavior signal. The moving average algorithm is then applied to eliminate residual fluctuations and output a stable nursing behavior signal. Stable nursing behavior signals are transmitted to the feature analysis module as input for nursing behavior feature extraction.

3. The method for monitoring nursing behavior based on smart elderly care according to claim 1, characterized in that, The process involves analyzing the time-series characteristics of the processed nursing behavior signals, extracting the change cycle and trend direction of the nursing behavior, and determining whether the nursing behavior is in a normal or abnormal stage through a state classification model. Specifically, this includes: The processed nursing behavior signal is decomposed into a time series to obtain the trend and periodic components of the nursing behavior. Then, the key features of the nursing behavior are extracted, including behavior frequency and behavior intensity. The behavior intensity is the average slope, rate of change of curvature and distribution of extreme points in the trend features extracted from the decomposition results, the period length stability, phase consistency and amplitude uniformity in the periodic features, and the standard deviation and autocorrelation characteristics in the residual features. Based on the frequency characteristics of the behavior, the nursing behavior cycle is calculated, the start and end points of each cycle are determined, and then based on the intensity characteristics of the behavior, the slope of the behavior trend is calculated, and the behavior change curve is output. By combining the nursing behavior cycle and trend slope, the current behavior status is determined to be in a normal or abnormal stage. When the behavior trend is rising and in the early stage of the cycle, it is identified as an abnormal stage, and when the behavior trend is declining and in the late stage of the cycle, it is identified as a normal stage. Output the behavior state result and associate the behavior state with the timestamp to form a complete state sequence data.

4. The method for monitoring nursing behavior based on smart elderly care according to claim 3, characterized in that, After determining whether the nursing behavior is in a normal or abnormal stage using a state classification model, the following steps are also included: Based on the preset nursing supervision rules and real-time behavior frequency data, the data is input into the neural network prediction model. The trained neural network prediction model then calculates the intervention target value required for the abnormal stage. Based on the calculated intervention target value, combined with behavioral frequency trend data, a fuzzy logic algorithm is applied to dynamically adjust the output intensity of the regulatory execution unit to gradually approach the target value. The output intensity is 10 levels after the action space is discretized in the reinforcement learning control algorithm. Each level corresponds to a different control intensity output, from level 0 representing no intervention to level 9 representing the maximum intervention linear distribution. During the adjustment process, the frequency of behavior and the actual behavior value are monitored in real time. The actual behavior value is the current behavior intensity. When the frequency of behavior changes abruptly or the deviation of the actual behavior value from the target value exceeds a threshold, the neural network prediction model is triggered to recalculate the intervention target value. The recalculated intervention target value is input into the fuzzy logic algorithm to iteratively adjust the regulatory execution unit until the actual behavior value stabilizes near the target value.

5. The method for monitoring nursing behavior based on smart elderly care according to claim 1, characterized in that, The application of reinforcement learning control algorithm generates the intensity of regulatory actions based on the difference between the regulatory target value and the real-time behavior monitoring value, and outputs control commands to the regulatory execution unit, specifically including: The regulatory target value and real-time behavior monitoring value are obtained, the difference value is calculated, and the difference value is input into the reinforcement learning control algorithm to generate the intensity of the regulatory action based on the reward function and the state space. Based on the intensity of the regulatory action, control commands are generated and output to the regulatory execution unit to control the operation of the execution unit. The regulatory enforcement unit adjusts the intensity of actions and changes actual behavior according to control instructions. By continuously collecting real-time behavioral data and comparing it with regulatory target values, the difference value is updated to form a closed-loop regulatory cycle.

6. The method for monitoring nursing behavior based on smart elderly care according to claim 1, characterized in that, The process of using a compliance feedback device to obtain current nursing compliance indicators, comparing them with preset standard ranges, and determining whether regulatory rules need to be updated specifically includes: The compliance indicator data is transmitted to the regulatory system and compared with the preset standard range to determine whether the compliance indicator is within a reasonable range. If the compliance indicator is lower than the lower limit of the standard range, the regulatory system generates an instruction to enhance the intensity of the regulatory rule through the executor. If the compliance indicators exceed the upper limit of the standard range, the regulatory system generates instructions to weaken the intensity of the regulatory rules and reduce regulatory output through the executor. During the rule adjustment process, clustering algorithms are applied to analyze historical compliance data, establish a correlation model between compliance indicators and rule strength, and use the correlation model to predict the optimal rule strength for feedforward regulatory control.

7. The method for monitoring nursing behavior based on smart elderly care according to claim 1, characterized in that, The process of using a decision tree algorithm to derive the adjustment amount of the regulatory rule and outputting the adjustment instruction to the execution unit specifically includes: The compliance feedback device obtains the current compliance indicators in real time and compares them with the preset standard range to determine whether the compliance indicators exceed the standard range. When a compliance indicator exceeds the standard range, the deviation between the compliance indicator and the standard median is calculated. The deviation value is input into the decision tree algorithm, and then the rule adjustment amount is derived based on the preset decision rule library. The preset decision rule library is the adjustment strategy stored in the decision tree node, including the intensity increase or decrease and the expected recovery time. The judgment conditions stored in the decision tree node include the deviation duration, deviation degree, historical adjustment effect, time period weight and user sensitivity. The rule adjustment amount is the adjustment magnitude coefficient corresponding to the adjustment strategy selected by the decision tree according to the feature vector, including linear compensation, proportional adjustment, feedback compensation, weighted adjustment or personalized scaling. The rule adjustment amount is converted into an adjustment command, which is then output to the execution unit via a signal converter to drive the execution unit to adjust the rule strength. The execution unit changes the rule application according to the adjustment instructions, thereby optimizing compliance indicators and bringing them back to the standard range.

8. The method for monitoring nursing behavior based on smart elderly care according to claim 1, characterized in that, The optimization of the timing output of control commands by combining the response characteristics of the regulatory execution unit to ensure the continuous and stable regulatory process specifically includes: The response time parameters of the regulatory execution unit are obtained, a dynamic model of the execution unit is established, and the relationship curve between action intensity and regulatory effect is obtained. Based on the regulatory target and actual behavior value, the optimal output timing is calculated through the dynamic model. The dynamic model of the execution unit is a transfer function model obtained based on step response test, which is used to predict the execution effect under different control commands. The relationship curve between action intensity and regulatory effect is the correspondence between the input control signal obtained from the step response test data and the action result of the actuator. The dynamic model is a second-order system model. The optimal output timing is the control command timing obtained by decomposing the execution action into three stages: acceleration, constant speed and deceleration, and allocating time budget to each stage. Using the deviation in action intensity and the deviation in regulatory effect as inputs, a time series optimization algorithm is applied to obtain the timing adjustment amount of the control command; The adjusted control commands are output to the monitoring execution unit according to the calculation sequence for dynamic adjustment. The monitoring effect deviation is generated by real-time collection of nursing behavior data and the deviation data generated by comparing the monitoring effect with the target value. The time series optimization algorithm is an algorithm that predicts the execution effect under different control commands through the dynamic model of the execution unit, and adjusts the timing of the control commands by combining the motion inertia and mechanical delay of the device. The adjusted control commands are output to the monitoring execution unit according to the calculation sequence for dynamic adjustment. During the dynamic adjustment process, nursing behavior data and monitoring effects are collected in real time and compared with the target value to generate deviation data. When the deviation exceeds the threshold, the deviation is input into the time series optimization algorithm to iteratively optimize the timing of control commands.

9. The method for monitoring nursing behavior based on smart elderly care according to claim 1, characterized in that, The method involves continuously tracking nursing behavior signals and monitoring rules, updating preset rules in real time, and adapting to changes in nursing needs to achieve refined monitoring of nursing behavior. Specifically, this includes: The system acquires real-time care behavior signals and current monitoring rules from users, compares them with preset normal ranges, determines whether they deviate from the normal range, triggers an early warning mechanism if they deviate, and determines the extent of rule updates based on the degree of deviation. By applying regression analysis algorithms, based on users' historical nursing behavior data and regulatory rule trends, we can predict future changes in nursing behavior. Based on real-time nursing behavior signals, current regulatory rules, and predicted changes, the regulatory rules are dynamically updated, and new rule settings are generated. The new rules are set and distributed to the monitoring equipment, and the control equipment adjusts the monitoring output according to the new rules. After the rules are updated, nursing behavior signals are continuously monitored to determine whether they have returned to normal. If they have not returned to normal, the rules are updated again. Real-time nursing behavior signals, monitoring rules, and update results are stored in a database to optimize regression analysis algorithms.

10. A nursing behavior monitoring system based on smart elderly care, used to implement the nursing behavior monitoring method based on smart elderly care as described in any one of claims 1 to 9, characterized in that, The system includes: The behavior signal processing module acquires nursing behavior signals from the nursing behavior monitoring device, applies a medium-value filtering algorithm to suppress noise, and outputs stable nursing behavior signals. The behavior status analysis module extracts time series features based on stable nursing behavior signals, judges the behavior status, and calculates the intervention target value through a neural network prediction model when the behavior status is abnormal. It also applies a fuzzy logic algorithm to dynamically adjust the monitoring output. The monitoring and control module applies a reinforcement learning control algorithm to generate the intensity of monitoring actions based on the difference between the target value and the actual value, and outputs control commands to the monitoring execution unit. The rule optimization module obtains compliance indicators through the compliance feedback device, combines them with preset standard ranges, uses a decision tree algorithm to deduce the rule adjustment amount, and outputs adjustment instructions. The dynamic monitoring module optimizes the timing output by combining the response characteristics of the monitoring execution unit to ensure continuous and stable monitoring. It also dynamically updates preset rules by tracking behavioral signals and rules, and stores the data in the database.

Images