Rainfall detection and energy efficiency control alarm system based on edge side scatter analysis

Abstract

The application discloses a rain event detection and energy efficiency control alarm system based on edge side deviation analysis, and belongs to the technical field of alarm systems. The system comprises a first environment signal collector, a second environment signal collector, a dynamic background reference model, a statistical anomaly analysis and physical characteristic analysis unit, a historical event library, and a three-stage energy efficiency scheduling mechanism. When the first environment signal meets a wake-up threshold, the second environment signal is collected to generate a cooperative trigger collection signal. The dynamic background reference model is called to process the cooperative trigger collection signal to generate a dynamic background fusion signal. The cooperative trigger collection signal and the dynamic background fusion signal are subjected to statistical anomaly analysis and physical characteristic analysis to generate a deviation index. The historical event library is called to calibrate the deviation index to generate a current detection confidence, which drives the three-stage energy efficiency scheduling mechanism to switch modes. The alarm decision method adopts cooperative collection, fusion statistical deviation and physical recovery characteristics for deviation analysis, and combines the historical event library for confidence calibration, so that the rain event alarm with high precision and low false alarm can be realized, and the energy efficiency of the system is optimized.

Description

Technical Field

[0001] This invention relates to the field of alarm system technology, and in particular to an alarm system for pre-rainfall detection and energy efficiency control based on edge-side separation analysis. Background Technology

[0002] Currently, pre-rainfall detection plays a crucial role in meteorological monitoring, disaster prevention and mitigation, and precision agriculture. Traditional monitoring methods mainly rely on deploying costly large-scale weather radars or spatially limited rain gauges, which often struggle to balance spatial coverage accuracy with real-time early warning. With the development of edge computing and IoT technologies, utilizing distributed micro-sensors for pre-rainfall monitoring has become a new research hotspot.

[0003] In related technologies, Chinese invention patent with announcement number CN119964345B discloses a parking lot emergency early warning method and system based on multi-point edge computing, including: firstly, collecting multimodal environmental data and video streams in different areas of the parking lot and obtaining the collected data; then, performing time synchronization and dynamic threshold filtering on the collected data and obtaining preprocessed data; and finally, performing spatiotemporal feature alignment and noise filtering on the preprocessed data and obtaining a three-dimensional environmental situation map.

[0004] However, the aforementioned solutions still exhibit significant limitations when targeting complex outdoor environments or remote deployment scenarios with limited energy supply. Firstly, these technologies focus on structured scenarios with ample power supply, such as parking lots. Their data acquisition and processing mechanisms often operate continuously or at high frequencies, lacking multi-level energy efficiency scheduling and deep sleep / wake-up mechanisms tailored to outdoor environmental characteristics. This results in extremely low energy efficiency ratios for edge nodes during long-term operation, making it difficult to meet the requirements for long-term maintenance-free deployment. Furthermore, outdoor environments are highly dynamic and subject to random interference; for example, sudden strong winds or changes in intense sunlight can generate significant physical noise. The dynamic threshold filtering schemes used in these technologies primarily analyze statistical anomalies in signals, i.e., deviation, while ignoring the unique restorative characteristics of raindrop particle echoes in terms of physical laws. This makes it impossible to effectively filter out environmental noise interference through signal rebound patterns. More importantly, the baseline models of existing systems are often static, making it difficult to adapt to background drift caused by day-night cycles or seasonal changes. They lack incremental learning capabilities based on historical events and confidence level calibration, causing the system's detection accuracy to gradually decline during long-term operation. Therefore, a rainfall detection system that can couple discrete analysis and refined energy efficiency management is needed to resolve the contradiction between high-precision detection and low-power operation in complex field environments. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an alarm system for impending rainfall detection and energy efficiency control based on edge-side deviance analysis. This system employs a method that uses a low-power sensor to wake up a high-precision sensor for collaborative data acquisition, fuses statistical deviation and physical recovery features for deviance analysis, and combines a historical event database for confidence calibration. This approach enables high-precision, low-false-alarm alarms for impending rainfall events and optimizes the system's energy efficiency.

[0006] The above objectives can be achieved through the following approach: The pre-rainfall detection and energy efficiency control alarm system based on edge-side deduplication analysis includes a sleep / wake-up and collaborative acquisition module. In the monitoring sleep mode, the module acquires a first environmental signal through a thin-film vibration sensor. When the disturbance characteristics of the first environmental signal meet a preset wake-up threshold, the edge analysis node is activated and a miniature radar sensor is triggered to acquire a second environmental signal in a directional manner. The signals are then fused to generate a collaborative trigger acquisition signal. The dynamic background fusion module is used to call a dynamic background baseline model that characterizes the normal state of the environment, and uses this model to process the collaboratively triggered acquisition signal to generate a dynamic background fusion signal; The separation and recovery analysis module is used to perform statistical anomaly analysis and physical feature analysis on the collaboratively triggered acquisition signal and the dynamic background fusion signal, and to generate a separation and recovery index to characterize the signal deviation and recovery degree; The confidence calibration module is used to call a preset historical event library containing historical event records, calibrate the deviance index, and generate the current detection confidence level. The energy efficiency scheduling switching module is used to drive the three-level energy efficiency scheduling mechanism to switch between monitoring sleep mode, edge feature warning mode and full-speed communication alarm mode based on the current detection confidence level.

[0007] Optionally, the sleep / wake-up and collaborative acquisition module includes: The feature extraction unit is used to extract features from the first environmental signal and generate thin film vibration features; The mode determination unit is used to determine whether the vibration characteristics of the thin film conform to the triggering mode used to identify the precursor of rainfall; The radar control unit is used to send a trigger command to the miniature radar sensor when the determination is met, to control it to start and perform directional scanning to acquire the second environmental signal; The Doppler analysis unit is used to perform Doppler frequency analysis on the second environmental signal and extract radar echo features characterizing particle velocity and radial distance. The synchronization cascade unit is used to synchronize and align the first environmental signal and the second environmental signal in the time domain, and to cascade the feature vectors of the thin film vibration characteristics and the radar echo characteristics to generate a cooperative trigger acquisition signal.

[0008] Optionally, the dynamic background blending module includes: The probability distribution unit is used to call the multivariate probability distribution model trained by the collaborative trigger acquisition signal during the historical normal period as the dynamic background benchmark model; A vector input unit is used to input the feature vector of the cooperatively triggered acquisition signal at the current moment into the dynamic background reference model; The projection calculation unit is used to calculate the projection features of the feature vector in the normal signal space defined by the model, and generate a dynamic background fusion signal.

[0009] Optionally, the discrete analysis module includes: The statistical deviation unit is used to calculate the Mahalanobis distance between the collaboratively triggered acquisition signal and the dynamic background fusion signal, and generate a statistical deviation. The physical recovery unit is used to extract the envelope attenuation parameter and short-time autocorrelation coefficient of the collaboratively triggered acquisition signal to generate a physical recovery degree that characterizes the rebound law of the physical signal. The weighted fusion unit is used to generate fusion weights for fusing statistical deviation and physical recovery based on preset stability parameters that characterize the stability of the current environmental background, and to perform weighted fusion of statistical deviation and physical recovery to generate a deviation and recovery index.

[0010] Optionally, the weight fusion unit specifically includes: Obtain the current environmental wind speed and light intensity parameters, and combine them as the stability parameters; When the stability parameter indicates that the environmental physical noise is higher than the noise threshold, the fusion weight contribution of the physical resilience is increased; When the stability parameter indicates that the environmental physical noise is lower than the noise threshold, the fusion weight contribution of the statistical deviation is increased.

[0011] Optionally, the confidence calibration module includes: An event recording unit is used to retrieve historical event records from the historical event database based on the current environmental context; The confidence mapping unit is used to establish a mapping relationship between the deviation index and the confidence calibration based on historical event records, and to generate a preliminary confidence value; The trend correction unit is used to obtain the changing trend of the deviation index, correct the initial confidence value, and generate the current detection confidence.

[0012] Optionally, the energy efficiency scheduling switching module includes: The trend prediction unit is used to generate short-term trend predictions based on the historical sequence of the de-complexity index. The mode-driven unit, specifically, shuts down the micro radar sensor and the high-energy-consuming edge computing module and enters the monitoring sleep mode when the current detection confidence level is continuously lower than a preset first threshold and the short-term trend prediction indication event probability is lower than a preset probability threshold; when the current detection confidence level is between the first threshold and a preset second threshold, or when the short-term trend prediction indication event probability shows an upward trend, it enters the edge feature warning mode; when the current detection confidence level exceeds the second threshold, it enters the full-speed communication alarm mode and sends alarm information outward.

[0013] Optionally, the energy efficiency scheduling switching module further includes: The data encapsulation unit is used to encapsulate the collaborative trigger acquisition signal that triggered this alarm, the degree of separation index, and the final judgment result into a new event data packet while sending alarm information; An incremental learning unit is used to incrementally update the dynamic background baseline model using new event data packets; An adaptive adjustment unit is used to adaptively adjust the fusion weight generation logic used to generate the de-complexity index based on the final judgment result in the new event data packet.

[0014] Optionally, the energy efficiency scheduling switching module further includes: The energy monitoring unit is used to monitor its own energy status in real time and generate energy status parameters. The scheduling and adjustment unit is used to adjust the first threshold and the second threshold in the three-level energy efficiency scheduling mechanism according to the energy state parameters.

[0015] Compared with the prior art, the present invention has the following advantages: 1. By monitoring the sleep mode and co-triggered acquisition mechanism, the system operates in a low-power sleep mode most of the time. Only when the initial sensor detects a disturbance that meets the preset conditions will the high-power micro radar and edge computing module be awakened to perform in-depth analysis. After the event ends, the system will quickly return to sleep mode, thereby reducing the average power consumption of the system, extending the battery life in the absence of external power supply, and improving the deployment flexibility and long-term operation capability of the alarm system.

[0016] 2. A method for generating the restitution index is proposed. This method integrates statistical deviation analysis based on a dynamic background benchmark model and restitution analysis based on the physical characteristics of the signal itself, and dynamically adjusts the fusion weights according to the real-time environmental noise level. This multi-dimensional and adaptive analysis paradigm can effectively distinguish between signal changes caused by real precursors of impending rainfall and noise interference caused by gusts, physical impacts, etc., thereby improving the robustness of alarm decisions and effectively reducing the false alarm rate of the alarm system.

[0017] 3. A closed-loop adaptive learning mechanism based on a historical event database has been established. The system can not only calibrate the current detection confidence level through historical experience, but also incrementally update the internal dynamic background benchmark model and fusion weight generation logic based on the final event confirmation result after an alarm is issued. This enables the alarm system to continuously learn from actual operational experience, gradually adapt to long-term changes in the deployment environment, achieve self-optimization and continuous improvement of performance, and ensure the stability and accuracy of long-term detection.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the framework of the pre-rainfall detection and energy efficiency control alarm system based on edge-side separation analysis according to an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the distance-Doppler distribution characteristics of rainfall particles in a rainfall detection and energy efficiency control alarm system based on edge-side decoupling analysis according to an embodiment of the present invention.

[0022] Figure 3 A schematic diagram of the dynamic background model PCA projection space model of the pre-rainfall detection and energy efficiency control alarm system based on edge-side complex analysis according to an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of the mapping relationship between energy state and alarm threshold adjustment in a pre-rainfall detection and energy efficiency control alarm system based on edge-side deduplication analysis according to an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0025] Reference Figure 1 One embodiment of the present invention proposes an alarm system for impending rainfall detection and energy efficiency control based on edge-side deviance analysis. The system employs a method that uses a low-power sensor to wake up a high-precision sensor for collaborative data acquisition, fuses statistical deviation and physical recovery features for deviance analysis, and combines a historical event database for confidence calibration. This method enables high-precision, low-false-alarm alarms for impending rainfall events and optimizes the system's energy efficiency.

[0026] The system described in this embodiment specifically includes: The sleep / wake-up and collaborative acquisition module is used to acquire a first environmental signal through a thin-film vibration sensor in monitoring sleep mode. When the disturbance characteristics of the first environmental signal meet the preset wake-up threshold, the edge analysis node is activated and the micro radar sensor is triggered to acquire a second environmental signal in a directional manner, and the signals are fused to generate a collaborative trigger acquisition signal. The dynamic background fusion module is used to call a dynamic background baseline model that characterizes the normal state of the environment, and uses this model to process the collaboratively triggered acquisition signal to generate a dynamic background fusion signal; The separation and recovery analysis module is used to perform statistical anomaly analysis and physical feature analysis on the collaboratively triggered acquisition signal and the dynamic background fusion signal, and to generate a separation and recovery index to characterize the signal deviation and recovery degree; The confidence calibration module is used to call a preset historical event library containing historical event records, calibrate the deviance index, and generate the current detection confidence level. The energy efficiency scheduling switching module is used to drive the three-level energy efficiency scheduling mechanism to switch between monitoring sleep mode, edge feature warning mode and full-speed communication alarm mode based on the current detection confidence level.

[0027] Specifically, a low-power continuous monitoring mechanism utilizes initial environmental disturbance signals captured by preliminary sensors as wake-up beacons. Once these beacons meet specific conditions, the system immediately activates high-precision sensors for directional depth information acquisition, fusing the two heterogeneous sensor information into a cohesive, multi-dimensional event feature signal. Subsequently, a dynamic benchmark model characterizing the normal state of the environment is invoked to generate a background reference corresponding to the current event feature signal in real time. By comparing the statistical deviation between the event signal and this dynamic background reference, and combining this with the physical recovery characteristics of the event signal itself, a comprehensive deviation index is generated. This deviation index is calibrated by comparing it with a knowledge base containing historical event experience, thus obtaining a more statistically significant current detection confidence level. This confidence level ultimately drives a three-level energy efficiency scheduling mechanism, enabling intelligent switching of the system's power consumption state under different risk levels. Through monitoring sleep and co-triggered mechanisms, the system remains in a low-power state most of the time, only activating high-power analysis and sensing modules when suspicious precursors are detected, effectively extending the device's battery life. The application of discrete-time analysis (DMA) enables the system to go beyond simple signal amplitude anomaly judgment. By comprehensively analyzing the statistical distance of signal deviation from normal patterns and the signal's own physical dynamic behavior, it can more reliably distinguish between genuine impending rainfall and random environmental noise such as gusts and physical impacts, thereby significantly reducing the false alarm rate and improving detection robustness. The confidence calibration process based on a historical event database endows the system with adaptive and learning capabilities, allowing it to correct current judgments based on historical experience, further improving decision-making accuracy. Finally, the three-level energy efficiency scheduling mechanism directly links detection results with energy consumption management, achieving refined and intelligent management of system energy while ensuring timely early warning.

[0028] Optionally, the sleep / wake-up and collaborative acquisition module includes: The feature extraction unit is used to extract features from the first environmental signal and generate thin film vibration features; The mode determination unit is used to determine whether the vibration characteristics of the thin film conform to the triggering mode used to identify the precursor of rainfall; The radar control unit is used to send a trigger command to the miniature radar sensor when the determination is met, to control it to start and perform directional scanning to acquire the second environmental signal; The Doppler analysis unit is used to perform Doppler frequency analysis on the second environmental signal and extract radar echo features characterizing particle velocity and radial distance. The synchronization cascade unit is used to synchronize and align the first environmental signal and the second environmental signal in the time domain, and to cascade the feature vectors of the thin film vibration characteristics and the radar echo characteristics to generate a cooperative trigger acquisition signal.

[0029] Specifically, the first environmental signal acquired by the thin-film vibration sensor is processed. Its energy distribution within a specific frequency band, such as 5 Hz to 50 Hz, is analyzed using a short-time Fourier transform, and combined with the root mean square amplitude in the time domain, thin-film vibration characteristics are extracted. The trigger mode is preset with a set of logical rules. For example, if the energy in the thin-film vibration characteristics continuously exceeds 150% of a dynamically adjusted baseline noise threshold based on historical data within a duration window, such as 3 to 5 seconds, the characteristic is deemed to meet the trigger mode for identifying precursory rainfall. This mode aims to filter out instantaneous, non-continuous physical impacts, focusing on persistent vibrations caused by wind field changes or early sparse raindrops. Once the determination is met, the edge analysis node immediately sends a trigger command containing scanning parameters to the miniature radar sensor via a serial communication interface, instructing it to wake from sleep mode and perform a directional scan to acquire the second environmental signal, i.e., the radar's raw echo data. The miniature radar sensor, typically a 24 GHz frequency-modulated continuous wave radar, can concentrate energy to detect tiny particles in a specific direction through directional scanning. The system performs Doppler frequency analysis on the acquired second environmental signal, i.e., performs a Fast Fourier Transform on the radar received signal to generate a range-Doppler map. Doppler frequency analysis is a signal processing technique used to analyze the radial velocity of a target. By measuring the frequency shift between the transmitted signal and the echo signal, it accurately calculates the target's velocity. The system extracts radar echo features from this map, mainly including the average Doppler velocity, velocity standard deviation, and particle number density of particles detected within a specific range, such as 5 to 50 meters. To achieve data fusion, the first and second environmental signals are synchronized in the time domain. Specifically, the moment the miniature radar sensor is triggered is used as the reference timestamp. A small time window, such as 500 milliseconds before and after this timestamp, is extracted from the first environmental signal. The extracted thin-film vibration features are then concatenated with the radar echo features at the same timestamp. The concatenation operation is defined as connecting the two feature vectors end-to-end to form a higher-dimensional vector. ; Here, S represents the final generated collaborative trigger acquisition signal, a fused feature vector; Fv represents the thin-film vibration feature vector extracted from the first environmental signal, whose internal elements are normalized values ​​such as energy and amplitude; Fr represents the radar echo feature vector extracted from the second environmental signal, whose internal elements are normalized values ​​such as velocity and particle density. Through this cascaded operation, the system generates a collaborative trigger acquisition signal containing information from both sensors. Figure 2The figure shows the range-Doppler feature distribution map acquired by the miniature radar sensor in this embodiment of the invention. The horizontal axis represents the detection range, the vertical axis represents the radial velocity of the particles, and the shades of the cloud map color represent the energy intensity of the particles at the corresponding spatiotemporal nodes. This visually demonstrates how the system extracts particle density and velocity features from the radar echo to generate a radar echo feature vector.

[0030] For example, the system uses a thin-film vibration sensor with a sampling frequency of 100Hz. During STFT processing, a Hanning window with a length of 256 points is selected. The energy value is obtained by calculating the mean of the squares of the amplitudes at each frequency point within a specific frequency band. To ensure signal stability, the 150% threshold for the trigger mode is based on three standard deviations of the ambient background noise. (In principle) the reference noise threshold is set as follows: The judgment threshold is fixed at 1. The miniature radar uses FMCW modulation, with 512 FFT points and a Doppler frequency of [missing information]. The relationship with radial velocity v follows the formula Among them, the speed of light Center frequency When feature vectors are concatenated, the thin film features... for like radar characteristics for like The cascaded collaborative trigger acquisition signal S is a four-dimensional vector. Normalization is achieved by mapping to the [0,1] interval using the Min-Max algorithm, ensuring dimensional uniformity. The logic of cascaded operations lies in representing the coupling relationship between the kinetic energy of rainfall particles (radar characteristics) and the elastic collision energy (film characteristics) in a high-dimensional space.

[0031] Optionally, the dynamic background blending module includes: The probability distribution unit is used to call the multivariate probability distribution model trained by the collaborative trigger acquisition signal during the historical normal period as the dynamic background benchmark model; A vector input unit is used to input the feature vector of the cooperatively triggered acquisition signal at the current moment into the dynamic background reference model; The projection calculation unit is used to calculate the projection features of the feature vector in the normal signal space defined by the model, and generate a dynamic background fusion signal.

[0032] Specifically, the first step is to call the stored dynamic background benchmark model. This dynamic background benchmark model is a multivariate probability distribution model trained through unsupervised learning based on co-triggered acquisition signals from historical normal periods, such as a dimensionality-reduced subspace model constructed using principal component analysis (PCA). The training data for this model comes from a large number of co-triggered acquisition signal samples collected over long periods, such as weeks or even months, under various weather conditions with no confirmed rainfall, including sunny, cloudy, and different wind speeds. By performing principal component analysis on these sample data, principal component vectors that can explain more than 95% of the data variance are extracted, collectively forming a low-dimensional subspace representing the normal signal variation pattern. The feature vector of the co-triggered acquisition signal generated in the previous steps at the current moment is input into this dynamic background benchmark model. The core operation is to calculate the projection features of this feature vector in the normal signal space defined by the model, and then reconstruct it based on this. ; In this formula, It is the generated dynamic background fusion signal, which is a feature vector with the same dimension as the input signal. The vector represents the feature vector of the cooperatively triggered acquisition signal input at the current moment, which integrates the features of thin-film vibration and radar echo. P represents the dynamic background benchmark model, which is a projection matrix composed of k principal component vectors, trained from historical normal signal data. It is the transpose of matrix P. The engineering significance of this formula lies in the fact that it first projects the current signal onto a normal subspace defined by P, and then reconstructs the projection result back to the original signal space, obtaining... That is The best approximation in normal mode. For example... Figure 3 The diagram shows a schematic representation of the PCA projection space of the dynamic background baseline model in this embodiment of the invention. The plane in the diagram represents the low-dimensional subspace of principal components obtained from training with historical normal signals. The upper vector represents the currently input collaboratively triggered acquisition signal, and the lower vector represents the dynamic background fusion signal after its projection and reconstruction in the normal subspace. The residual between the two vectors reflects the degree of abrupt change in the current signal relative to the normal environment.

[0033] For example, the dynamic background benchmark model employs PCA dimensionality reduction, the core of which lies in mapping highly correlated features to uncorrelated features through orthogonal transformation. The projection matrix P is composed of the first three eigenvectors after the eigenvalue decomposition of the covariance matrix, with dimensions of... ,satisfy Unit array. If the current input signal... for ,go through The coordinates projected onto the lower-dimensional space are Through the formula Perform back-projection reconstruction and calculate... for The reconstructed signal Random components that do not belong to the normal subspace, i.e., those not belonging to sunny weather or simple wind noise patterns, are removed. The basis of this reconstruction algorithm is that the projection loss of normal background signals onto the principal component subspace is minimal, while abnormal signals, such as the initial pulse of rainfall, produce large residuals during reconstruction. Through linear mapping of matrix operations, the system achieves real-time simulation of expected signals for normal environments.

[0034] Optionally, the discrete analysis module includes: The statistical deviation unit is used to calculate the Mahalanobis distance between the collaboratively triggered acquisition signal and the dynamic background fusion signal, and generate a statistical deviation. The physical recovery unit is used to extract the envelope attenuation parameter and short-time autocorrelation coefficient of the collaboratively triggered acquisition signal to generate a physical recovery degree that characterizes the rebound law of the physical signal. The weighted fusion unit is used to generate fusion weights for fusing statistical deviation and physical recovery based on preset stability parameters that characterize the stability of the current environmental background, and to perform weighted fusion of statistical deviation and physical recovery to generate a deviation and recovery index.

[0035] Specifically, firstly, the Mahalanobis distance between the collaboratively triggered acquisition signal and the dynamically fused background signal is calculated as the statistical deviation. Mahalanobis distance is an effective distance metric that considers the correlation between features and can standardize the assessment of the degree to which data points deviate from the distribution center. This calculation requires the covariance matrix of historical normal samples, which can be obtained when training the dynamic background benchmark model. Simultaneously, the system independently analyzes the physical properties of the collaboratively triggered acquisition signal to generate a physical restoring degree. This process includes two parts: first, extracting the signal envelope attenuation parameter, quantifying its attenuation rate by calculating the time constant required for the signal envelope to attenuate to a specific proportion, such as 37%, after a local peak; second, calculating its short-time autocorrelation coefficient at multiple delay steps to assess the signal's periodicity and randomness. The system maps these two physical parameters to a 0-1 range using a preset normalization function, such as the Sigmoid function, and comprehensively generates a physical restoring degree that characterizes the rebound pattern of the physical signal. Next, a preset stability parameter characterizing the stability of the current environmental background is introduced, and dynamic fusion weights are generated based on this parameter. This weight is used to dynamically adjust the contribution ratio of statistical deviation and physical recovery to the final indicator. Based on the fusion weight, statistical deviation and physical recovery are weighted and fused to generate the deviation / recovery index: ; Where L is the final generated decomposition index; The statistical deviation, calculated using Mahalanobis distance, is derived from the current collaboratively triggered acquisition signal and the dynamic background fusion signal. The normalized physical recovery is extracted from the intrinsic physical features of the collaboratively triggered acquisition signal; w is the fusion weight, the value of which is dynamically determined by a preset stability parameter and ranges from 0 to 1.

[0036] For example, statistical deviation Using Mahalanobis distance as an example: Given the inverse of the historical background covariance matrix diagonal array mean vector for Current signal Then the square distance If the deviation value is 0.05, the calculated value is... Taking the square root yields Physical recovery Extracting the envelope attenuation constant Short-time autocorrelation coefficient Through the Sigmoid function Mapping, where For gain factor, With center bias, the normalized result is calculated. The value is 0.92. If the current fusion weight w = 0.6, then the distance from the fusion index is... The formula defines the dimensionless restitution index as 1. Its scientific basis is that statistical deviation reflects the degree of dispersion of the signal relative to the background, while physical restitution reflects the physical reset characteristics of the signal after stress removal. The weighted sum of the two can distinguish between random noise and rainfall processes with physical regularity.

[0037] Optionally, the weight fusion unit specifically includes: Obtain the current environmental wind speed and light intensity parameters, and combine them as the stability parameters; When the stability parameter indicates that the environmental physical noise is higher than the noise threshold, the fusion weight contribution of the physical resilience is increased; When the stability parameter indicates that the environmental physical noise is lower than the noise threshold, the fusion weight contribution of the statistical deviation is increased.

[0038] Specifically, the system first acquires real-time wind speed parameters (typically in meters per second) using an external meteorological sensor, and light intensity parameters (typically in lux) using a photosensor. To capture the stability of the light intensity, the system calculates the standard deviation of the light intensity over a short time window, such as 10 seconds, rather than the instantaneous value. Then, the normalized wind speed parameters are combined with the standard deviation of the light intensity to generate a single-dimensional stability parameter. This stability parameter is a comprehensive indicator designed to quantify the degree of environmental background disturbance caused by physical factors such as wind and light. This stability parameter is compared with a preset noise threshold, which is calibrated based on a large amount of historical data and used to distinguish between a calm and noisy environment. When the stability parameter indicates that the environmental physical noise is higher than the noise threshold, such as in windy weather, signal disturbances caused by physical impacts are more prevalent. In this case, the statistical deviation may be misjudged due to the instability of the background itself, while the rebound pattern of the physical signal, i.e., physical resilience, better reflects the true physical nature of the event. Therefore, the fusion weight contribution of physical resilience is increased. Conversely, when the stability parameter indicates that the environmental physical noise is below the noise threshold, such as on a windless or lightly windy night, the system considers the background to be very clean and stable. In this case, the statistical difference between the collaboratively triggered acquisition signal and the dynamically fused background signal—the statistical deviation—can more sensitively detect weak precursors of anomalies. Therefore, the system increases the fusion weight contribution of the statistical deviation. This dynamic adjustment process is implemented through a mapping function that maps the stability parameter to fusion weights: ; In this formula, W is the final generated fusion weight, with a value range of 0 to 1. E is the stability parameter of the current environment, calculated from wind speed and light intensity parameters. f is a preset nonlinear mapping function, typically an inverse sigmoid function, designed to output a W value close to 1 when E is low, and a W value close to 0 when E is high.

[0039] For example, the environmental stability parameter E is calculated using a linear combination method: Normalized wind speed Standard deviation of illumination After normalization, the result is 0.15, yielding E = 0.22. The weight mapping function f(E) uses the formula... In this example, since E=0.22 is lower than the preset noise threshold of 0.4, the calculated value is... This result significantly increases the statistical deviation of the system in low-noise environments. The weight is 93%, which aligns with the design principle of pursuing high sensitivity in a quiet environment. When E increases to 0.6 (strong wind), W rapidly decreases to 0.04, and the system then relies on physical restoring properties. The parameters of this inverse sigmoid function... This ensures rapid switching of weights near the noise threshold, demonstrating the algorithm's nonlinear adaptive adjustment capability to environmental changes.

[0040] Optionally, the confidence calibration module includes: An event recording unit is used to retrieve historical event records from the historical event database based on the current environmental context; The confidence mapping unit is used to establish a mapping relationship between the deviation index and the confidence calibration based on historical event records, and to generate a preliminary confidence value; The trend correction unit is used to obtain the changing trend of the deviation index, correct the initial confidence value, and generate the current detection confidence.

[0041] Specifically, firstly, based on the current environmental context, such as the current wind speed level, temperature range, and time of day (day or night), a conditional search is performed from a pre-set historical event database to obtain historical event records similar to the current scenario. This historical event database is a long-term maintained database that stores complete data for every past triggering event, including the coordinated trigger acquisition signal at the time, the calculated deviance index, and the final judgment result, such as actual impending rainfall, strong wind interference, and false alarms. Using these retrieved historical event records, a dynamic mapping relationship between the deviance index and confidence calibration is constructed. This mapping relationship is essentially a non-linear function, typically obtained by fitting logistic regression or support vector regression to historical data points. The system uses the deviance index value calculated at the current moment as input and calculates the preliminary confidence value through this mapping relationship. Simultaneously, it acquires the historical sequence within the most recent time window, such as the past 30 seconds, and calculates its slope through linear regression analysis of this sequence data to quantify the changing trend of the deviance index. Finally, based on the obtained changing trend of the deviance index, the system dynamically corrects the preliminary confidence value to generate the final current detection confidence. ; in, It is the generated current detection confidence level; It is the initial confidence value obtained by calibrating the mapping relationship; The value is the normalized trend of the deviation index, which is calculated from the regression slope of the historical series of the deviation index. A positive value indicates an upward trend, and a negative value indicates a downward trend. k is a preset correction coefficient, usually between 0.1 and 0.5, used to control the influence of the trend on the final confidence level.

[0042] For example, the system obtains the distance-to-recovery index sequence with a sampling step size of 1 second over the past 30 seconds and performs linear regression using the least squares method: Calculate the slope m. If the slope m = 0.015, it indicates that the index is trending upwards. Normalize it to [-1, 1] to obtain... If the initial confidence value obtained through regression fitting using a historical event database... The correction coefficient k=0.2 is set based on experience, balancing the influence of the current value and the trend, thus determining the final detection confidence level. The basis of this correction formula is that a continuously rising indicator trend represents the consistency of the evolution of impending rainfall events over time, and positive correction can compensate for the low confidence level caused by errors in a single measurement. If the trend is downward... If it is negative, then The error rate will decrease accordingly, effectively reducing the false alarm rate caused by sudden single interference.

[0043] Optionally, the energy efficiency scheduling switching module includes: The trend prediction unit is used to generate short-term trend predictions based on the historical sequence of the de-complexity index. The mode-driven unit, specifically, shuts down the micro radar sensor and the high-energy-consuming edge computing module and enters the monitoring sleep mode when the current detection confidence level is continuously lower than a preset first threshold and the short-term trend prediction indication event probability is lower than a preset probability threshold; when the current detection confidence level is between the first threshold and a preset second threshold, or when the short-term trend prediction indication event probability shows an upward trend, it enters the edge feature warning mode; when the current detection confidence level exceeds the second threshold, it enters the full-speed communication alarm mode and sends alarm information outward.

[0044] Specifically, firstly, based on the historical sequence of the deviance index within the most recent time window, such as the past 5 minutes, time series analysis methods such as the autoregressive integral moving average model are used to generate a short-term trend prediction for a short period of time, such as the next 1 to 2 minutes. Then, rule-based decision logic is executed to drive mode switching. First, when the current detection confidence level is continuously below a preset first threshold (e.g., 0.3) for a period of time, such as 60 consecutive seconds, and the short-term trend prediction shows that the deviance index is below a preset probability threshold during the prediction period, the system determines that the current environment is stable and there are no signs of an event. At this time, the system will trigger an instruction to enter a monitoring sleep mode. Specific actions include shutting down the RF front-end and baseband circuit of the miniature radar sensor through the power management unit, and putting the edge computing core processor into deep sleep, leaving only the low-power wake-up circuit of the thin-film vibration sensor operational. Second, when the current detection confidence level is between the first threshold and a preset second threshold (e.g., 0.8), or when the short-term trend prediction clearly indicates that the deviance index shows a continuous upward trend (e.g., the prediction slope is greater than a preset slope threshold), the system determines that there is a potential risk in the environment or that an event is developing. At this point, the system will enter edge feature alert mode. In this mode, the miniature radar sensor and edge computing module operate at full power, and the frequency of data acquisition and analysis may be increased for intensive feature monitoring and confidence level updates. Third, when the current detection confidence level momentarily or continuously exceeds the second threshold, the system determines that the probability of an impending rainfall event is extremely high and immediately issues an alarm. At this time, the system will switch to full-speed communication alarm mode, immediately activating a high-power wireless communication module such as 4G or NB-IoT, and sending alarm information containing the current timestamp, device ID, detection confidence level, and key feature vectors to the designated cloud platform or user terminal according to a preset protocol.

[0045] For example, the trend prediction unit calls the ARIMA(1,1,1) model, whose mathematical expression is: ,in , If we predict the mean value of L over the next 2 minutes... And currently .because If the value remains below the first threshold of 0.3 for 60 consecutive seconds, and the predicted value is below the probability threshold of 0.2, the mode driving unit issues a command to control the power management chip, such as TI's TPS series, to cut off the LDO power supply to the miniature radar's RF terminal, and the system enters a monitoring sleep mode. Upon reaching 0.5, which falls between the second threshold of 0.3 and 0.8, the system immediately activates the edge computing core, such as the ARM Cortex-M4F, at full speed, entering edge feature alert mode. The mode switching is based on Markov chain state transition logic. By setting dual thresholds and time lags, the system is prevented from frequently oscillating at critical points, thereby maximizing energy savings while ensuring monitoring sensitivity.

[0046] Optionally, the energy efficiency scheduling switching module further includes: The data encapsulation unit is used to encapsulate the collaborative trigger acquisition signal that triggered this alarm, the degree of separation index, and the final judgment result into a new event data packet while sending alarm information; An incremental learning unit is used to incrementally update the dynamic background baseline model using new event data packets; An adaptive adjustment unit is used to adaptively adjust the fusion weight generation logic used to generate the de-complexity index based on the final judgment result in the new event data packet.

[0047] Specifically, while the system enters full-speed communication alarm mode based on high current detection confidence and sends alarm information outward, the edge analysis node immediately performs data encapsulation. The system encapsulates the original collaborative trigger acquisition signal that triggered this alarm, the calculated separation degree index sequence, and the final judgment result field to be filled into a structured new event data packet, and temporarily stores it in local non-volatile memory. This final judgment result can be issued by the cloud server after cross-validation based on meteorological data from a wider range within a preset time after the alarm is issued, such as 30 minutes, or filled in after confirmation by local ground-based real-time equipment such as simple rain gauges. Once the final judgment result is obtained, the system will execute two parallel adaptive adjustment actions. First, the system incrementally updates the dynamic background benchmark model using the new event data packet. Specifically, if the final determination result is a false alarm, the system will treat the co-triggered acquisition signal corresponding to the event as a new "normal but highly noisy" sample, and use an incremental principal component analysis algorithm to slightly update the mean vector and covariance matrix of the original principal component subspace with this new sample. This allows the new noise pattern to be included in the normal background without completely reconstructing the model. Secondly, based on the final determination result in the new event data packet, the system adaptively adjusts the fusion weight generation logic used to generate the de-recovery index. ; Where β represents a key adjustable parameter in the aforementioned weight mapping function, such as the slope or offset of the function. These are the adjusted new parameters. These are the old parameters. Δβ is an adjustment factor generated based on the final judgment result; for example, if a false alarm caused by strong wind occurs, the system records the stability parameters at that time and determines that it is due to the fusion weight w being too high. In this case, Δβ is a negative value, so that the weight mapping function outputs a lower fusion weight under similar stability parameters, thus relying more on physical restoring degree rather than statistical deviation in similar strong wind scenarios in the future.

[0048] For example, when the cloud feedback indicates a false alarm and the stability parameter E=0.6, indicating strong winds, the system performs adaptive adjustments. The old weight mapping function offset parameter... Adjustment factor Calculated based on the deviation ratio, the formula is as follows: Set the learning rate If we expect w to be 0.1 in a similar scenario, but w is currently 0.3, then... After the update This adjustment shifts the w=f(E) curve to the right, resulting in a smaller output w under the same wind conditions in the future, thus increasing the weight of physical resilience. The incremental learning unit then utilizes the new cooperative acquisition signal vector. Update the mean vector : Among them, the forgetting factor This adjustment mechanism is based on feedback control theory. It corrects model deviations through closed-loop feedback, enabling the system to evolve in response to the environmental characteristics of a specific deployment location.

[0049] Optionally, the energy efficiency scheduling switching module further includes: The energy monitoring unit is used to monitor its own energy status in real time and generate energy status parameters. The scheduling and adjustment unit is used to adjust the first threshold and the second threshold in the three-level energy efficiency scheduling mechanism according to the energy state parameters.

[0050] Specifically, the system monitors the battery's output voltage and charging / discharging current in real time using a built-in power management chip or coulomb counter. By integrating the current over time and combining it with the battery's temperature and aging model, the system accurately calculates the battery's remaining charge capacity and converts it into a standardized state of energy (SGE) parameter. This parameter is a normalized value between 0 and 1, where 0 represents the battery is about to be depleted and 1 represents a fully charged battery. Subsequently, based on this SGE parameter, the system dynamically adjusts the first and second thresholds in the three-level energy efficiency scheduling mechanism. This adjustment strategy aims to appropriately lower the thresholds to improve sensitivity when energy is abundant, and raise the thresholds to reduce power consumption when energy is scarce. ; In this formula, This represents the adjusted new threshold, which can be either the first threshold or the second threshold. This represents the factory reference value for that threshold. This represents the current normalized energy state parameter; M represents a preset maximum adjustment coefficient used to limit the maximum range of threshold increase, such as 30% to 50% of the baseline threshold. The squared term is used to implement a more drastic conservative adjustment strategy when the battery level is extremely low. For example, when the energy state parameter shows a battery level above 90%, the system may lower the first threshold from the baseline of 0.3 to 0.28, while when the battery level is below 20%, the system may raise the first threshold to 0.45 to avoid accelerating energy depletion due to frequent entry into edge warning modes. Figure 4 The diagram shows the mapping relationship between the dynamic alarm threshold and the normalized energy state parameter in an embodiment of the present invention. The diagram illustrates that the system maintains a low trigger threshold to maximize sensitivity when the battery is fully charged, while nonlinearly increasing the threshold when the battery is low (i.e., when the normalized energy state parameter is close to 0). This significantly reduces power consumption and extends standby time by sacrificing some sensitivity.

[0051] For example, if the energy monitoring unit obtains the remaining battery capacity as 200mAh and the full charge capacity as 2000mAh using a coulomb counter, then the normalized energy parameter... Set the baseline value for the first threshold. The maximum adjustment range is M = 0.2. Substitute these values ​​into the formula to calculate the new threshold: This means that in low-power conditions, the confidence requirement for the wake-up alert mode has increased from 0.3 to 0.462, reducing the operating time of high-power modules. The algorithm has a dimension of 1, and its design is based on the nonlinear characteristics of the square term: when the power is sufficient... At low power levels, the threshold is only slightly adjusted to 0.302, having almost no impact on sensitivity; however, when the power level enters the critical region, the threshold is increased dramatically, forcing the system into an extremely low power state. Through this energy-sensing dynamic threshold logic, the system can effectively extend standby time by more than 40% under extreme power conditions.

[0052] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0053] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A rainfall detection and energy efficiency control alarm system based on edge-side decoupling analysis, characterized in that, The system includes: The sleep / wake-up and collaborative acquisition module is used to acquire a first environmental signal through a thin-film vibration sensor in monitoring sleep mode. When the disturbance characteristics of the first environmental signal meet the preset wake-up threshold, the edge analysis node is activated and the micro radar sensor is triggered to acquire a second environmental signal in a directional manner, and the signals are fused to generate a collaborative trigger acquisition signal. The dynamic background fusion module is used to call a dynamic background baseline model that characterizes the normal state of the environment, and uses this model to process the collaboratively triggered acquisition signal to generate a dynamic background fusion signal; The separation and recovery analysis module is used to perform statistical anomaly analysis and physical feature analysis on the collaboratively triggered acquisition signal and the dynamic background fusion signal, and to generate a separation and recovery index to characterize the signal deviation and recovery degree; The confidence calibration module is used to call a preset historical event library containing historical event records, calibrate the deviance index, and generate the current detection confidence level. The energy efficiency scheduling switching module is used to drive the three-level energy efficiency scheduling mechanism to switch between monitoring sleep mode, edge feature warning mode and full-speed communication alarm mode based on the current detection confidence level.

2. The pre-rainfall detection and energy efficiency control alarm system based on edge-side decoupling analysis according to claim 1, characterized in that, The sleep / wake-up and collaborative acquisition module includes: The feature extraction unit is used to extract features from the first environmental signal and generate thin film vibration features; The mode determination unit is used to determine whether the vibration characteristics of the thin film conform to the triggering mode used to identify the precursor of rainfall; The radar control unit is used to send a trigger command to the miniature radar sensor when the determination is met, to control it to start and perform directional scanning to acquire the second environmental signal; The Doppler analysis unit is used to perform Doppler frequency analysis on the second environmental signal and extract radar echo features characterizing particle velocity and radial distance. The synchronization cascade unit is used to synchronize and align the first environmental signal and the second environmental signal in the time domain, and to cascade the feature vectors of the thin film vibration characteristics and the radar echo characteristics to generate a cooperative trigger acquisition signal.

3. The pre-rainfall detection and energy efficiency control alarm system based on edge-side decoupling analysis according to claim 1, characterized in that, The dynamic background blending module includes: The probability distribution unit is used to call the multivariate probability distribution model trained by the collaborative trigger acquisition signal during the historical normal period as the dynamic background benchmark model; A vector input unit is used to input the feature vector of the cooperatively triggered acquisition signal at the current moment into the dynamic background reference model; The projection calculation unit is used to calculate the projection features of the feature vector in the normal signal space defined by the model, and generate a dynamic background fusion signal.

4. The pre-rainfall detection and energy efficiency control alarm system based on edge-side decoupling analysis according to claim 1, characterized in that, The complex analysis module includes: The statistical deviation unit is used to calculate the Mahalanobis distance between the collaboratively triggered acquisition signal and the dynamic background fusion signal, and generate a statistical deviation. The physical recovery unit is used to extract the envelope attenuation parameter and short-time autocorrelation coefficient of the collaboratively triggered acquisition signal to generate a physical recovery degree that characterizes the rebound law of the physical signal. The weighted fusion unit is used to generate fusion weights for fusing statistical deviation and physical recovery based on preset stability parameters that characterize the stability of the current environmental background, and to perform weighted fusion of statistical deviation and physical recovery to generate a deviation and recovery index.

5. The pre-rainfall detection and energy efficiency control alarm system based on edge-side decoupling analysis according to claim 4, characterized in that, The weight fusion unit specifically includes: Obtain the current environmental wind speed and light intensity parameters, and combine them as the stability parameters; When the stability parameter indicates that the environmental physical noise is higher than the noise threshold, the fusion weight contribution of the physical resilience is increased; When the stability parameter indicates that the environmental physical noise is lower than the noise threshold, the fusion weight contribution of the statistical deviation is increased.

6. The pre-rainfall detection and energy efficiency control alarm system based on edge-side decoupling analysis according to claim 1, characterized in that, The confidence calibration module includes: An event recording unit is used to retrieve historical event records from the historical event database based on the current environmental context; The confidence mapping unit is used to establish a mapping relationship between the deviation index and the confidence calibration based on historical event records, and to generate a preliminary confidence value; The trend correction unit is used to obtain the changing trend of the deviation index, correct the initial confidence value, and generate the current detection confidence.

7. The pre-rainfall detection and energy efficiency control alarm system based on edge-side decoupling analysis according to claim 1, characterized in that, The energy efficiency scheduling switching module includes: The trend prediction unit is used to generate short-term trend predictions based on the historical sequence of the de-complexity index. The mode-driven unit, specifically, shuts down the micro radar sensor and the high-energy-consuming edge computing module and enters the monitoring sleep mode when the current detection confidence level is continuously lower than a preset first threshold and the probability of the short-term trend prediction event is lower than a preset probability threshold; when the current detection confidence level is between the first threshold and a preset second threshold, or when the probability of the short-term trend prediction event shows an upward trend, it enters the edge feature warning mode. When the current detection confidence level exceeds the second threshold, the system enters the full-speed communication alarm mode and sends alarm information outward.

8. The pre-rainfall detection and energy efficiency control alarm system based on edge-side decoupling analysis according to claim 7, characterized in that, The energy efficiency scheduling switching module also includes: The data encapsulation unit is used to encapsulate the collaborative trigger acquisition signal that triggered this alarm, the degree of separation index, and the final judgment result into a new event data packet while sending alarm information; An incremental learning unit is used to incrementally update the dynamic background baseline model using new event data packets; An adaptive adjustment unit is used to adaptively adjust the fusion weight generation logic used to generate the de-complexity index based on the final judgment result in the new event data packet.

9. The pre-rainfall detection and energy efficiency control alarm system based on edge-side decoupling analysis according to claim 7, characterized in that, The energy efficiency scheduling switching module also includes: The energy monitoring unit is used to monitor its own energy status in real time and generate energy status parameters. The scheduling and adjustment unit is used to adjust the first threshold and the second threshold in the three-level energy efficiency scheduling mechanism according to the energy state parameters.

Images