Home water and heat safety intelligent protection system and method based on double ai engine cooperation

Abstract

This invention discloses a smart home water and heating safety protection system and method based on dual AI engine collaboration, belonging to the field of smart home water and heating safety protection technology. The system includes: a hardware sensing and execution unit deployed in the water and heating pipelines, integrating multiple types of sensors and a smart shut-off valve with built-in torque feedback; a dual AI engine collaborative control unit, comprising a user-end water supply AI engine, a heating AI engine, and a collaborative arbitration module, respectively realizing functions such as leak classification response, device fingerprint recognition, out / back home logic, multi-evidence fusion leak detection, blockage diagnosis and quantification, valve maintenance, and health prediction; and a cloud-based intelligent platform that achieves model self-evolution through federated learning. The advantages of this invention are: through innovations such as differentiated collaboration of dual AI engines, multi-dimensional feature fusion, device fingerprint recognition, multi-evidence fusion decision-making, predictive maintenance, and three-level power supply protection, it solves the problems of existing products such as single function, high false alarm rate, lack of predictive ability, and inability to coordinate water and heating supply.

Description

Technical Field

[0001] This invention relates to a smart home water and heat safety protection system and method based on dual AI engine collaboration, belonging to the field of smart home water and heat safety protection technology. Background Technology

[0002] With the improvement of residents' living standards, household water supply and heating pipes have become essential infrastructure in modern residences. However, the safety risks of these pipe systems continue to trouble users: leaks in water supply pipes can lead to flooded floors, damaged walls, neighborhood disputes, and even serious property damage; leaks, blockages, or inefficiencies in heating pipes not only waste energy but can also affect the safety of winter heating. Statistics show that household leaks are one of the main causes of property insurance claims, with a single incident potentially resulting in economic losses ranging from thousands to tens of thousands of yuan.

[0003] Currently, protective products for household water and heating safety on the market have the following technical defects: (1) Single and independent functions: Most existing products are water leakage alarms, mechanical shut-off valves or simple thermostats, which can only achieve single-point monitoring or single function, and cannot provide unified management and coordinated protection for water supply and heating systems. Users need to install different devices separately, which increases the complexity of deployment and the cost of use.

[0004] (2) Slow response speed and high false alarm rate: Traditional mechanical water valves rely on manual operation and cannot achieve automatic and fast shut-off; simple electronic alarms use fixed threshold judgment, which are easily affected by environmental interference and false alarms occur frequently, causing users to experience "alarm fatigue" and delaying the handling when a real leak occurs.

[0005] (3) Lack of intelligent identification capability: The existing solution cannot distinguish between normal water use and abnormal leakage, and cannot identify specific water-using equipment. This leads to frequent accidental shutdowns during the deployment of defenses outside the area due to normal equipment operation (such as toilet water replenishment and water purifier flushing), which seriously affects the user experience.

[0006] (4) No predictive maintenance function: Traditional solutions are all passive response type, which can only issue alarms after leakage or failure occurs, and cannot provide early warning of potential risks such as pipeline blockage and valve aging. Users lose the initiative in maintenance and often only carry out maintenance after the failure occurs.

[0007] (5) Lack of monitoring methods for heating systems: Existing heating safety products are mostly focused on the control of the boiler itself, lacking the ability to diagnose system-level faults such as leaks in heating pipelines, filter blockages, and pipe siltation, and thus cannot achieve full life-cycle health management of the heating system.

[0008] To address the aforementioned technical issues, there is an urgent need for a smart home water and heating safety protection solution that can achieve unified protection for water and heating systems, possess intelligent identification and predictive maintenance capabilities, and enable collaborative decision-making. Summary of the Invention

[0009] To overcome the shortcomings of existing technologies, this invention provides a home water and heating safety intelligent protection system and method based on dual AI engine collaboration. The technical solution of this invention is as follows: A smart home water and heating safety protection system based on dual AI engine collaboration includes: The hardware sensing and execution unit, deployed in the household water supply pipeline and heating pipeline, includes: dual-path pressure sensors (P1, P2) on the water supply side, ultrasonic flow meter (Fw), temperature sensor (Tw), intelligent water supply shut-off valve (V) and first ambient temperature sensor (Te1). The heating side includes dual-path pressure sensors (Ph1, Ph2), supply water temperature sensor (Ts), return water temperature sensor (Tr), second ambient temperature sensor (Te2), first hot water intelligent shut-off valve (Vw), second hot water intelligent shut-off valve (Vh), first ultrasonic flow meter (U1), second ultrasonic flow meter (U2), and supply water filter (L). A water supply intelligent shut-off valve (V) is installed on the water supply pipeline. A first hot water intelligent shut-off valve (Vw) and a second hot water intelligent shut-off valve (Vh) are installed on the heating pipeline. The water supply intelligent shut-off valve (V), the first hot water intelligent shut-off valve (Vw), and the second hot water intelligent shut-off valve (Vh) are all equipped with a torque feedback unit for real-time monitoring of motor current and judgment of valve resistance. The dual AI engine collaborative control unit includes a water supply AI engine module, a heating AI engine module, and a collaborative arbitration module; wherein, the water supply AI engine module integrates a leakage classification response module, an equipment fingerprint recognition module, and an out / back home logic module, which are used to process the sampling data on the water supply side in real time; The heating AI engine module integrates a multi-evidence fusion leak detection module, a blockage diagnosis and quantification module, and a valve health prediction module, which are used to fuse and analyze multi-source heterogeneous data on the heating side. The collaborative arbitration module is connected to the water supply AI engine module and the heating AI engine module through an internal data bus, and is used to dynamically schedule the computing resources of the two engines, arbitrate data conflicts, and prioritize control commands according to the preset scenario strategy. The cloud-based intelligent platform is connected to the dual AI engine collaborative control unit via an encrypted communication link. It is used to receive encrypted model gradient data uploaded from several edge devices, perform federated learning to aggregate and update the global model, and securely distribute the global model parameters to the edge devices via OTA.

[0010] The leakage classification response module includes: The feature calculation unit takes the pressure sequence P(t), flow sequence F(t), and device identification result from the device fingerprinting module as inputs within the time window Δt, and calculates the following features: The negative rate of change of pressure characteristic F1 = sigmoid(-k·dP / dt), where dP / dt is the derivative of pressure with respect to time, k is the scaling factor, and the sigmoid function maps the result to the range of 0 to 1; Unsupervised anomaly score F2 based on the isolated forest algorithm; The device matching confidence complement F3 = 1 - C, where C is the device identification confidence level from 0 to 1. When F3 is close to 1, the probability of leakage increases. Based on the Poisson distribution, the temporal context anomaly F4 = -log(P) poisson (k; λ)), where λ is a parameter representing the user's historical water usage habits; specifically, using a Poisson distribution to model the probability P of the current water usage event occurring at the current time point based on the user's historical water usage habits. poisson The higher the F4 value, the higher the temporal anomaly of the event; The leakage probability calculation unit is used to calculate the leakage probability based on the weighted summation formula P. leak =Σ(w i ·F i Calculate the leakage probability, where the weight w i F was obtained by training on a historical leak dataset. i This represents the i-th feature value used to calculate the leakage probability; The hierarchical response decision unit has a built-in hierarchical decision tree, based on the leakage probability P. leak The value triggers different levels of response actions: If P leak If the value is ≥0.85, it is determined to be an L5 level emergency leak. The shutdown command is immediately output to the first intelligent shut-off valve (Vw), and at the same time, an audible and visual alarm is triggered and information is pushed to all emergency contacts. If 0.70 ≤ P leak If the value is less than 0.85, it is judged as a Level 4 serious leak, a local alarm is output, and a shutdown command is automatically output if no cancellation command is received from the user within the preset waiting time. If 0.50 ≤ P leak If the value is less than 0.70, it is determined to be a Level 3 warning, and a strong reminder will be output to the APP and the interactive terminal will be controlled to display a flashing yellow light. If 0.30≤P leak If the value is less than 0.50, it is determined to be an L2 level warning, and an APP notification message is output. If P leakIf the value is less than 0.30, it is judged as a normal or slight fluctuation at level L0 / L1, and only the data is recorded.

[0011] The device fingerprint recognition module includes: The template library construction unit is used to record the water usage cycle for each water-using device and extract a 12-dimensional feature vector from the flow-time curve of each water usage cycle. The 12-dimensional feature vector includes at least: start-up slope, steady-state mean, stop slope, water usage duration, total water consumption, peak flow, and energy proportion of each frequency band after wavelet packet decomposition, forming a feature cluster template library for the device. A real-time feature extraction unit is used to extract the same 12-dimensional feature vector after a water usage event is detected; The matching and recognition unit is used to calculate the distance between the real-time feature vector and all feature vectors in the template library using the dynamic time warping algorithm, and select the K closest samples for voting, outputting the device type and its confidence level C; The feedback correction unit is used to feed back the confidence level C to the leakage classification response module in real time to adjust the F3 characteristic value, thereby correcting the leakage probability P. leak .

[0012] The intelligent logic module for going out / returning home includes: The off-duty determination unit is used to determine the duration of no water usage, T. dry Furthermore, smart home signals are used to comprehensively determine whether a user is away from home, including a threshold T for the duration of no water usage. adaptive Dynamically adjusted based on users' historical travel habits; When the user leaves the premises, the outbound arming unit outputs a command to close the smart shut-off valve (V) and records the initial pressure P0 at the moment of closure. The homecoming identification unit is used to monitor the pressure change curve when the valve is closed. When a pressure drop event is detected, a hidden Markov model is used to classify the pressure drop pattern. The input of the hidden Markov model is the pressure drop rate sequence, and the output is the probability of "turning on the tap when returning home" or "slow leakage in the pipe". If the probability of "turning on the tap when returning home" is greater than 0.8, the automatic valve opening unit will output a command to open the intelligent water shut-off valve (V); otherwise, it will determine that there is a leak, keep the valve closed and trigger an alarm.

[0013] The multi-evidence fusion leakage detection module is constructed based on Dempster-Shafer evidence theory or weighted Bayesian networks, and includes: The first evidence source calculation unit calculates the basic probability allocation m1 based on the evidence of flow imbalance, m1 = f(ΔF, duration), where ΔF is the difference between supply and return water flow and duration is the duration. The second evidence source calculation unit calculates the basic probability allocation m2 based on the pressure differential attenuation evidence, where m2 = g(d(ΔP)). sys ) / dt), where ΔP sys Let d(ΔP) be the system pressure difference. sys ) / dt is the rate of change of pressure difference; The third evidence source calculation unit calculates the basic probability allocation m3 based on the temperature field anomaly evidence, where m3 = h(ΔT). actual , ΔT expected ), where ΔT actual To measure the temperature difference between the supply and return water, ΔT expected This is the theoretical temperature difference based on a thermodynamic model; The fourth evidence source calculation unit calculates the basic probability allocation m4 based on the ultrasonic acoustic fingerprint evidence. m4 is obtained by classifying the audio spectrograms collected by the first ultrasonic flow meter (U1) and the second ultrasonic flow meter (U2) based on the convolutional neural network CNN. The convolutional neural network CNN includes an input layer, two convolutional layers, a pooling layer, a fully connected layer and a Softmax output layer connected in sequence, and outputs the probability of the leakage category. The evidence synthesis unit is used to synthesize m1, m2, m3, and m4 using Dempster's combination rule, and to calculate the confidence intervals of the three hypotheses: leakage, normal, and uncertainty. The leakage detection unit is used to determine that a leakage has occurred when the confidence level of the leakage hypothesis is greater than a preset threshold of 0.75, and outputs a shut-off command to the first hot water intelligent shut-off valve (Vw) and the second hot water intelligent shut-off valve (Vh).

[0014] The blockage diagnosis and quantification module includes: The filter clogging index calculation unit is used to calculate the filter clogging index I according to the following formula. filter : I filter = 0.5×(ΔP filter / ΔP0) + 0.3×(1-F actual / F0) + 0.2×T rend , Wherein, ΔP filter F represents the pressure difference across the filter, where ΔP0 is the baseline pressure difference under clean filter conditions. actual F0 is the actual flow rate, and T is the baseline flow rate under clean conditions. rend This is a pressure difference change trend factor based on historical data; The pipeline siltation index calculation unit is used to establish the system hydraulic model ΔP. sys = R×F 2 F is the volumetric flow rate of circulating water in the pipeline. The normal range of the system resistance coefficient R is determined by fitting operational data. min, R max ], and calculate the current drag coefficient R. current Corresponding siltation index I pipe = (R current - R min ) / (R max -R min ); The valve health prediction module includes: The current curve acquisition unit is used to acquire the motor current curve I(t) at a sampling rate of not less than 1kHz each time the intelligent shut-off valve is activated. The feature extraction unit is used to extract the following features from the motor current curve I(t): Time-domain characteristics: Peak current I peak Average current I mean Current variance σ 2 ; Shape characteristics: Dynamic time warping (DTW) distance between the current and a pre-stored standard health current profile; Frequency domain characteristics: the main frequency components and their amplitudes extracted after performing a fast Fourier transform on I(t); The health score calculation unit has a built-in support vector regression (SVR) model. It takes the extracted features as input and outputs a valve health score S from 0 to 100. valve ; The remaining useful life prediction unit records the historical sequence of health scores and uses exponential smoothing or linear regression extrapolation to predict the time required for the health score to drop to a preset failure threshold, which is the remaining useful life.

[0015] The collaborative arbitration module integrates a scenario strategy library, which includes at least the following: The water supply outage mode strategy is triggered when the water supply AI engine module detects that the continuous period of no water usage exceeds T. adaptive Furthermore, the smart home system prompts the user to leave home; the action executed is as follows: the collaborative arbitration module issues an instruction to the water supply AI engine module to close the smart water supply shut-off valve (V) and enter pressure monitoring; Summer heating mode strategy: When the heating AI engine module detects that the flow rates of the first ultrasonic flow meter (U1) and the second ultrasonic flow meter (U2) are both zero, the second ambient temperature sensor (Te2) has been above 25°C for a week and there is no heating demand, the heating AI engine switches to low-power monitoring mode to release NPU computing resources; the collaborative arbitration module dynamically allocates the released resources to the water supply AI engine module. Winter heating mode strategy: When the heating AI engine module detects that the flow rate of the first ultrasonic flow meter (U1) or the second ultrasonic flow meter (U2) is greater than zero, the inlet water temperature sensor (T1) is more than 2°C higher than the outlet water temperature sensor (T2), and there is a clear heating demand, the heating AI engine module exits the low-power monitoring mode and resumes full-function monitoring; the collaborative arbitration module reallocates computing resources.

[0016] It also includes an integrated health management module, which calculates the system's total health score S according to the following formula. total : S total = 0.4×S water + 0.4×S heat + 0.2×S common , Among them, S water The water supply health score is calculated by the water supply AI engine module based on the number of historical leakage events, valve health scores, and pipeline aging coefficients. S heat The heating health score is calculated by the heating AI engine based on the blockage diagnosis index, thermal efficiency, and valve health score. S common The health score for public infrastructure is calculated based on a comprehensive assessment of backup battery health, storage media lifespan, and network communication quality. The integrated health management module also based on S total The system generates predictive maintenance suggestions or alarm messages, which are then displayed via an interactive terminal. The cloud-based intelligent platform, in conjunction with the dual AI engine collaborative control layer, implements a federated learning self-evolution mechanism, specifically including: The local update unit, deployed on edge devices, is used to incrementally train a lightweight model locally using the event data after detecting a local false alarm or missed alarm event, and generate model gradient update ΔW. An encrypted upload unit is used to encrypt ΔW and upload it to the cloud intelligent platform via the TLS 1.3 protocol; The cloud aggregation unit receives encrypted gradient updates from multiple edge devices, decrypts and aggregates them, and generates global model parameters W. global ; The secure distribution unit is used to control W. global Digital signature encryption is performed and distributed to all edge devices via OTA. After verifying the signature, the edge devices update their local models, thus completing the collective evolution. The hardware sensing and execution unit also includes a multi-level power protection unit, which includes a main power AC220V input, a 12V / 10Ah backup lithium battery and a supercapacitor. The supercapacitor is used to provide the energy required for the final shutdown action when both the main power and the backup battery fail, ensuring that an emergency shutdown can still be completed in extreme power outage conditions. The intelligent water supply shut-off valve (V), the first intelligent hot water shut-off valve (Vw), and the second intelligent hot water shut-off valve (Vh) are all integrated with manual mechanical operating mechanisms. In the event of a power outage, network outage, or abnormality in the electronic control system that prevents automatic shut-off, the valves can be opened and closed manually.

[0017] A protection method for a home water and heating safety intelligent protection system based on dual AI engine collaboration includes the following steps: Step S1: The water supply AI engine module processes the sampling data on the water supply side in real time, executes leakage classification response, equipment fingerprint recognition, and intelligent logic for going out / coming home, and generates water supply side status data and control decisions. Step S2: The heating AI engine module performs fusion analysis on the multi-source heterogeneous data of the heating side, performs multi-evidence fusion leak detection, blockage diagnosis and quantification, and valve health prediction, and generates heating side status data and control decisions. Step S3: The collaborative arbitration module dynamically schedules the computing resources of the water supply AI engine module and the heating AI engine module according to the preset scenario strategy, arbitrates data conflicts and control command priorities, and outputs collaborative control commands. Step S4: The cloud-based intelligent platform receives encrypted model gradient data uploaded from multiple edge devices, performs federated learning to aggregate and update the global model, and sends the optimized global model parameters to the edge devices via OTA. Step S5: The integrated health management module calculates the total health score of the system based on the status data of the water supply side and the heating side, and generates predictive maintenance suggestions or alarm information, which are displayed through the interactive terminal.

[0018] The advantages of this invention are: 1. Addressing the differences in physical characteristics and protection requirements between water supply and heating systems, a pioneering "dual AI engine collaborative" architecture is adopted. The water supply AI engine, based on a real-time operating system, focuses on millisecond-level high-speed sampling and rapid response, ensuring shutdown within 3 seconds in emergencies such as pipe bursts. The heating AI engine, based on a high-performance heterogeneous platform, focuses on complex pattern recognition and deep diagnostics, enabling refined functions such as leak detection, blockage location, and health assessment. The two engines operate independently yet collaboratively, dynamically scheduling resources through a collaborative arbitration module, resolving the technical contradiction that a single AI model cannot simultaneously meet real-time requirements and complexity.

[0019] 2. On the water supply side, a four-dimensional weighted scoring system is used, incorporating negative pressure change rate, isolated forest anomaly score, equipment confidence complement, and temporal context anomaly degree, combined with a random forest classifier, to accurately calculate the leakage probability. Through an L5 to L0 graded response mechanism, it can shut down within seconds in case of an emergency leak, and only record warnings for minor anomalies, avoiding frequent false shutdowns that disrupt users' lives.

[0020] 3. By constructing a 12-dimensional feature template library based on flow-time curves for water-using equipment such as toilets, showers, and washing machines, dynamic time warping and the K-nearest neighbor algorithm are used to identify equipment types, and the confidence level is fed back to the leak detection module in real time. When a device is identified as a high-confidence, normal device, the leakage probability weight is automatically reduced, fundamentally solving the industry problem of accidental shutdown due to normal equipment operation during field deployment.

[0021] 4. Combining an adaptive threshold for inactivity time with smart home system judgment, it accurately identifies the user's absence status, automatically shuts off the water supply valve, and enters pressure monitoring. The homecoming recognition uses a Hidden Markov Model to classify pressure drop patterns, accurately distinguishing between "water usage upon returning home" and "slow leakage." When the probability of automatically opening the valve is greater than 0.8, water supply is restored imperceptibly, ensuring both safety and improving user experience.

[0022] 5. By fusing four heterogeneous evidence sources—flow imbalance, pressure differential attenuation, temperature field anomalies, and ultrasonic acoustic signature CNN classification—a comprehensive decision is made through confidence interval calculation. Even if a single piece of evidence is affected by interference or noise, high reliability can still be maintained through multi-evidence fusion, significantly improving the accuracy and robustness of leak detection in heating systems.

[0023] 6. The degree of filter clogging is quantified using the filter clogging index formula, based on the system hydraulic model ΔP. sys = R·F 2 By fitting the resistance coefficient within the normal range and calculating the pipe siltation index, a quantitative assessment of the system's flow resistance can be achieved. Simultaneously, by comparing the low-frequency noise energy spectrum using dual ultrasonic microphones, the blockage point can be accurately located near the water supply or return end, providing a basis for precise maintenance.

[0024] 7. Acquire valve operating current curves at a sampling rate of no less than 1kHz, extract time-domain, shape, and frequency-domain features, output a health score of 0-100 using a support vector regression model, and predict remaining useful life based on historical score sequences. This non-invasive current analysis allows for the early identification of mechanical failure trends such as valve jamming and wear, transforming reactive maintenance into predictive maintenance and reducing the risk of sudden failures.

[0025] 8. The collaborative arbitration module incorporates scenario strategies such as an "away mode" and a heating season mode: In "away mode," the intelligent water supply shut-off valve is automatically closed and pressure monitoring is initiated to avoid leakage risks during absence; in summer when there is no heating demand, the heating AI engine module switches to a low-power monitoring mode, releasing NPU computing resources for the water supply AI engine module to perform data mining, achieving dynamic optimization of computing resource utilization; in winter when there is heating demand, the heating AI engine module resumes full-function monitoring to ensure heating safety. Through differentiated scenario scheduling, collaborative protection and energy-saving efficiency of the water and heating system are achieved.

[0026] 9. Edge devices perform incremental training based on local false positive / false negative events, uploading only encrypted model gradients to the cloud. The cloud aggregates gradients from multiple devices to generate a global model, which is then distributed via OTA. This enables the collective intelligent evolution of tens of millions of devices while protecting user privacy, making the system increasingly intelligent with use and continuously improving detection accuracy and adaptability.

[0027] 10. It adopts a three-level power supply system with a main power supply of AC220V, a 12V / 10Ah backup lithium battery, and a supercapacitor. The supercapacitor can provide the energy required for the final shutdown when both the main power supply and the backup battery fail, ensuring that an emergency shutdown can still be completed in extreme situations such as power outages and fires, thus maximizing safety protection.

[0028] In summary, this invention solves long-standing technical problems such as limited functionality, high false alarm rate, lack of predictive ability, and inability to collaborate in existing products through a series of innovative technologies, including a dual AI engine collaborative architecture, multi-feature fusion detection, device fingerprint recognition, multi-evidence fusion decision-making, and predictive maintenance. This significantly improves the safety and intelligence level of home water and heating systems. Attached Figure Description

[0029] Figure 1 This is a block diagram of the main structure of the protection system of the present invention.

[0030] Figure 2 yes Figure 1 Block diagram of the dual AI engine collaborative control unit.

[0031] Figure 3 This is a schematic diagram of the main structure of the protection system of the present invention.

[0032] Figure 4 This is a flowchart illustrating the protection method of the present invention. Detailed Implementation

[0033] The present invention will be further described below with reference to specific embodiments, and the advantages and features of the present invention will become clearer as a result. However, these embodiments are merely exemplary and do not constitute any limitation on the scope of the present invention. Those skilled in the art should understand that modifications or substitutions can be made to the details and form of the technical solutions of the present invention without departing from the spirit and scope of the present invention, but all such modifications and substitutions fall within the protection scope of the present invention.

[0034] See Figures 1 to 4 This invention relates to a smart home water and heating safety protection system based on dual AI engine collaboration, comprising: The hardware sensing and execution unit is deployed in the household water supply pipeline and heating pipeline, including: dual pressure sensors on the water supply side (first water supply side pressure sensor P1 and second water supply side pressure sensor P2 respectively), ultrasonic flow meter Fw, temperature sensor Tw, intelligent water supply shut-off valve V and first ambient temperature sensor Te1. The heating side dual-path pressure sensors (namely, the first heating side dual-path pressure sensor Ph1 and the second heating side dual-path pressure sensor Ph2, the supply water temperature sensor Ts, the return water temperature sensor Tr, the second ambient temperature sensor Te2, the first hot water intelligent shut-off valve Vw, the second hot water intelligent shut-off valve Vh, the first ultrasonic flow meter U1 and the second ultrasonic flow meter U2). A first ambient temperature sensor Te1, a second water supply side pressure sensor P2, a water supply intelligent shut-off valve V, an ultrasonic flow meter Fw, a first water supply side pressure sensor P1, and a temperature sensor Tw are sequentially installed on the water supply pipeline. The first heating-side dual-path pressure sensor Ph1, filter L, first hot water intelligent shut-off valve Vw, first ultrasonic flow meter U1 and water supply temperature sensor Ts are sequentially installed on the water inlet pipe of the heating pipeline. A second heating-side dual-path pressure sensor Ph2, a return water temperature sensor Tr, a second ultrasonic flow meter U2, a second hot water intelligent shut-off valve Vh, and a second ambient temperature sensor Te2 are sequentially installed on the outlet pipe of the heating pipeline. The water supply intelligent shut-off valve V, the first hot water intelligent shut-off valve Vw, and the second hot water intelligent shut-off valve Vh are all equipped with a torque feedback unit for real-time monitoring of motor current and judgment of valve resistance. The dual AI engine collaborative control unit 2 includes a water supply AI engine module 21, a heating AI engine module 22, and a collaborative arbitration module 23. The water supply AI engine module 21 integrates a leak classification response module, an equipment fingerprint recognition module, and an out / back home logic module for real-time processing of sampled data from the water supply side. The heating AI engine module 22 integrates a multi-evidence fusion leak detection module, a blockage diagnosis and quantification module, and a valve health prediction module for fusion analysis of multi-source heterogeneous data from the heating side. The collaborative arbitration module 23 communicates with both the water supply AI engine module and the heating AI engine module via an internal data bus, and dynamically schedules the computing resources of the two engines, arbitrates data conflicts, and prioritizes control commands according to preset scenario strategies. The cloud-based intelligent platform 3 is connected to the dual AI engine collaborative control unit via an encrypted communication link. It is used to receive encrypted model gradient data uploaded from several edge devices, perform federated learning to aggregate and update the global model, and securely distribute the global model parameters to the edge devices via OTA.

[0035] This invention addresses the fundamental differences between water supply and heating systems by establishing a water supply AI engine module 21 and a heating AI engine module 22 within the dual AI engine collaborative control unit 2. The water supply AI engine module 21, based on a real-time operating system, focuses on the real-time processing of high-speed sampled data from the water supply side, ensuring shutdown within 3 seconds in emergencies such as pipe bursts. The heating AI engine module 22, based on a high-performance heterogeneous platform, focuses on the complex fusion analysis of multi-source heterogeneous data from the heating side, enabling refined functions such as leak detection, blockage location, and health assessment. The collaborative arbitration module 23 dynamically schedules the computing resources of both engines through an internal data bus, arbitrating control command priorities. This satisfies both the stringent response speed requirements of the water supply side and the diagnostic depth needs of the heating side, resolving the technical contradiction that a single AI model cannot simultaneously handle real-time performance and complexity.

[0036] The device fingerprint recognition module in the water supply AI engine module 21 builds a feature template library based on flow-time curves for water-using equipment such as toilets, showers, and washing machines. It uses dynamic time warping and K-nearest neighbor algorithms to identify equipment types, and the confidence level is fed back to the leak classification response module in real time, fundamentally solving the industry problem of accidental shutdown due to normal equipment operation during field deployment. The multi-evidence fusion leak detection module in the heating AI engine module 22 is based on the Dempster-Shafer evidence theory. It integrates four heterogeneous evidence sources—flow imbalance, pressure differential attenuation, temperature field anomalies, and ultrasonic acoustic signature CNN classification—for comprehensive decision-making, maintaining high reliability even if a single piece of evidence is interfered with.

[0037] The intelligent water supply shut-off valve V, the first intelligent shut-off valve Vw, and the second hot water intelligent shut-off valve Vh all have built-in torque feedback units that monitor motor current and determine valve resistance in real time. The valve health prediction module in the heating AI engine module 22 analyzes the time-domain, shape, and frequency-domain characteristics of the current curve to output a health score and predict the remaining useful life. The blockage diagnosis and quantification module quantitatively assesses the system's flow resistance through filter blockage index and pipe siltation index. The cloud-based intelligent platform 3, through a federated learning mechanism, aggregates the model gradients of tens of millions of edge devices while protecting privacy, generates a global model, and then distributes it via OTA, enabling the system's collective intelligence to continuously evolve, making the devices smarter with use.

[0038] The collaborative arbitration module 23 has a built-in library of scenario strategies such as outing mode and summer mode: automatically shut off the water supply valve when going out; when there is no heating demand in summer, the heating AI engine module 22 releases NPU computing resources to the water supply AI engine module 21 for deep data mining to achieve dynamic optimization of computing resources.

[0039] The integrated health management module calculates the system's total health score based on the health scores of water supply, heating, and public infrastructure, and generates predictive maintenance recommendations. Hardware sensing and execution unit 1 employs a three-tiered power supply system: a main AC220V power supply, a 12V / 10Ah backup lithium battery, and a supercapacitor. The supercapacitor provides the energy needed for a final shutdown when both the main power supply and backup battery fail, ensuring an emergency shutdown can still be completed in extreme power outage situations, maximizing safety protection.

[0040] The leakage classification response module includes: The feature calculation unit takes the pressure sequence P(t), flow sequence F(t), and device identification result from the device fingerprinting module as inputs within the time window Δt, and calculates the following features: The negative rate of change of pressure characteristic F1 = sigmoid(-k·dP / dt), where dP / dt is the derivative of pressure with respect to time, k is the scaling factor, and the sigmoid function maps the result to the range of 0 to 1; Unsupervised anomaly score F2 based on the isolated forest algorithm; The device matching confidence complement F3 = 1 - C, where C is the device identification confidence level from 0 to 1. When F3 is close to 1, the probability of leakage increases. Based on the Poisson distribution, the temporal context anomaly F4 = -log(P) poisson (k; λ)), where λ is a parameter representing the user's historical water usage habits; specifically, using a Poisson distribution to model the probability P of the current water usage event occurring at the current time point based on the user's historical water usage habits. poisson The higher the F4 value, the higher the temporal anomaly of the event; The leakage probability calculation unit is used to calculate the leakage probability based on the weighted summation formula P. leak =Σ(w i ·F i Calculate the leakage probability, where the weight w i F was obtained by training on a historical leak dataset. i This represents the i-th feature value used to calculate the leakage probability; The hierarchical response decision unit has a built-in hierarchical decision tree, based on the leakage probability P. leak The value triggers different levels of response actions: If P leak If the value is ≥0.85, it is determined to be an L5 level emergency leak. The shutdown command is immediately output to the first intelligent shut-off valve (Vw), and at the same time, an audible and visual alarm is triggered and information is pushed to all emergency contacts. If 0.70 ≤ P leak If the value is less than 0.85, it is judged as a Level 4 serious leak, a local alarm is output, and a shutdown command is automatically output if no cancellation command is received from the user within the preset waiting time. If 0.50 ≤ P leak If the value is less than 0.70, it is determined to be a Level 3 warning, and a strong reminder will be output to the APP and the interactive terminal will be controlled to display a flashing yellow light. If 0.30≤P leak If the value is less than 0.50, it is determined to be an L2 level warning, and an APP notification message is output. If P leak If the value is less than 0.30, it is judged as a normal or slight fluctuation at level L0 / L1, and only the data is recorded.

[0041] The advantages of the leakage classification response module are as follows: This module employs four dimensions of features: negative pressure change rate F1, isolated forest anomaly score F2, equipment confidence complement F3, and time context anomaly degree F4. It comprehensively characterizes water use events from four perspectives: physical change, pattern deviation, equipment type, and time habits. By using weighted fusion to calculate the leakage probability, it avoids the shortcomings of single threshold judgment being susceptible to interference and significantly improves the detection accuracy.

[0042] This module constructs F=1-C based on the confidence level C output by the device fingerprint recognition module, and combines it with the Poisson distribution time anomaly level F4 to intelligently distinguish between normal water use and abnormal leaks. When the device is identified with high confidence and the time is consistent with usual patterns, both F3 and F4 are low, automatically reducing the probability of leakage and preventing accidental shutdown due to normal device operation.

[0043] This module sets five response levels from L5 to L0 based on the probability of leakage, ranging from emergency shutdown to data logging only. This ensures both second-level response in the event of a major leak and avoids overreacting to minor anomalies. The action thresholds for each level are optimized using historical data to achieve the best balance between safety and convenience.

[0044] A family was out during the day on a weekday, and no one was home. At 3 PM, the toilet's inlet valve malfunctioned, causing a continuous, low-flow-rate leak (approximately 0.3 L / min); meanwhile, the kitchen dishwasher started running normally according to its scheduled time.

[0045] The feature calculation unit receives data: The dual-channel pressure sensors on the water supply side detected a slow decreasing trend in the pressure sequence P(t), and the ultrasonic flow meter Fw detected a total flow rate of approximately 1.2 L / min from the two water supply channels. The device fingerprint recognition module analyzed the flow-time curve in real time and successfully identified the dishwasher's characteristic waveform (confidence level C1=0.96) by matching the template library using a dynamic time warping algorithm. At the same time, it also identified a continuous and stable unknown waveform (confidence level C2=0.1).

[0046] The feature calculation unit calculates feature values ​​for each of the two water usage paths separately: For normal water usage in dishwashers: F1=0.2 (pressure decreases gradually); F2=0.3 (The traffic pattern matches the historical normal data); F3 = 1 - 0.96 = 0.04 (high confidence level identifies as a known device); F4 calculates the probability P that washing dishes at 3 PM aligns with user habits based on a Poisson distribution. poisson (k; λ)=0.8, F4=-log(0.8)=0.1; Leakage probability P leak =0.2×0.2+0.2×0.3+0.2×0.04+0.3×0.1=0.04+0.06+0.008+0.03=0.138<0.30, judged as L0 level normal, only record data.

[0047] For persistent toilet leaks: F1=0.35 (pressure decreases slowly); F2=0.75 (this flow pattern is abnormal); F3 = 1 - 0.1 = 0.9 (cannot be identified as a known device); F4 calculates that the probability of continuous and stable water use at 3 PM is extremely low, P poisson (k; λ)=0.02, F4=-log(0.02)=1.70; Leakage probability P leak=0.2×0.35+0.2×0.75+0.3×0.9+0.3×1.70=0.07+0.15+0.27+0.51=1.0, after normalization P leak =0.92≥0.85.

[0048] Comprehensive judgment by the hierarchical response decision-making unit: Normal water usage in the dishwasher is classified as Level L0, and only data is recorded without triggering any intervention. Continuous water leakage in the toilet is classified as Level L5 emergency leak, and a shut-off command is immediately sent to the intelligent water supply shut-off valve V. At the same time, an audible and visual alarm is triggered, and an alarm message "Continuous water leakage detected in the bathroom; the main valve has been automatically shut off" is pushed to the user via the APP.

[0049] After receiving the push notification, the user remotely checked the monitoring system to confirm the leak and contacted the property management for on-site repair. Because the system accurately distinguishes between normal dishwasher water usage and abnormal toilet leaks, it avoids disrupting the user's life by accidentally shutting off the system, and automatically shuts off the system in the early stages of a leak, preventing floor soaking and property damage caused by continuous leakage.

[0050] The device fingerprint recognition module includes: The template library construction unit is used to record the water usage cycle for each water-using device and extract a 12-dimensional feature vector from the flow-time curve of each water usage cycle. The 12-dimensional feature vector includes at least: start-up slope, steady-state mean, stop slope, water usage duration, total water consumption, peak flow, and energy proportion of each frequency band after wavelet packet decomposition, forming a feature cluster template library for the device. A real-time feature extraction unit is used to extract the same 12-dimensional feature vector after a water usage event is detected; The matching and recognition unit is used to calculate the distance between the real-time feature vector and all feature vectors in the template library using the dynamic time warping algorithm, and select the K closest samples for voting, outputting the device type and its confidence level C; The feedback correction unit is used to feed back the confidence level C to the leakage classification response module in real time to adjust the F3 characteristic value, thereby correcting the leakage probability P. leak .

[0051] The device's fingerprint recognition module records multiple normal water usage cycles for each water-using device. It extracts 12-dimensional feature vectors from the flow-time curve, including start-up slope, steady-state mean, stop slope, water usage duration, total water consumption, peak flow rate, and the energy proportion of each frequency band after wavelet packet decomposition. These vectors form a feature cluster template library for the device. These multi-dimensional features comprehensively characterize the device's water usage fingerprint from multiple dimensions, including morphology, energy, and statistics, significantly improving the accuracy and robustness of device identification.

[0052] The algorithm uses dynamic time warping to calculate the distance between the real-time feature vector and all feature vectors in the template library. This can effectively address issues such as inconsistent water waveform lengths and shape expansions caused by water pressure fluctuations and differences in usage habits for the same device. Combined with the K-nearest neighbor voting mechanism, the algorithm outputs the device type and confidence level C, achieving an accuracy rate of 96.8%.

[0053] The device identification confidence level C is fed back to the leak classification response module in real time to construct the device matching confidence complement F3 = 1 - C. When a device is identified as a known normal device with high confidence, F3 approaches 0, and the leak probability automatically decreases; when a device cannot be identified or the confidence level is low, F3 approaches 1, and the leak probability increases. This closed-loop feedback mechanism enables the system to intelligently distinguish between normal water use and abnormal leaks, fundamentally solving the false alarm problem.

[0054] A household has water-using appliances such as a toilet, washing machine, dishwasher, and shower installed in their home. After the system is installed for the first time, the device's fingerprint recognition module enters learning mode.

[0055] The template library construction unit records multiple normal water usage cycles for each device: 3 normal toilet refills (approximately 2 minutes each, flow rate 0.8 L / min), 1 standard washing cycle for the washing machine (including multiple stages of water intake, washing, and drainage), 1 dishwasher run (multiple water intake stages), and 2 shower runs (with significant flow rate fluctuations). A 12-dimensional feature vector is extracted from the flow-time curve of each water usage cycle, including the start-up slope (reflecting the water flow impact velocity when the device is turned on), steady-state mean (reflecting the average flow rate during device operation), stop slope (reflecting the water flow decay rate when the device is turned off), water usage duration, total water consumption, peak flow rate, and the energy proportion of each frequency band after wavelet packet decomposition (reflecting the flow fluctuation characteristics of different frequency bands). These feature vectors constitute the feature cluster template library for each device.

[0056] Scenario 1: Normal water usage identification One evening at 8 PM, a user turned on the shower. The real-time feature extraction unit detected the water usage event and extracted the same 12-dimensional feature vector. The matching and recognition unit used the dynamic time warping algorithm to calculate the distance between the real-time feature vector and all feature vectors in the template library. It found that the DTW distance with the samples in the shower template library was the smallest (average distance 0.15). The five closest samples were selected for voting, and the device type was output as "shower" with a confidence level C=0.97.

[0057] The feedback correction unit feeds back C=0.97 to the leakage classification response module in real time, which is used to construct the device matching confidence complement F3=1-0.97=0.03. The leakage probability P is then calculated in conjunction with other characteristics. leak =0.12 < 0.30, which is considered normal water usage, and only data is recorded. The user is unaware of this, and the shower proceeds normally.

[0058] Scenario 2: Accurate Identification and False Alarm Elimination of Abnormal Events At 3 p.m. the following afternoon, the toilet inlet valve became slightly stuck, causing the water replenishment time to be abnormally extended to 8 minutes (normally 2 minutes), while the flow rate remained at 0.8L / min.

[0059] The real-time feature extraction unit detected a water usage event lasting 8 minutes and extracted a 12-dimensional feature vector: the starting slope was close to that of the toilet template (0.92), the steady-state mean was consistent with that of the toilet template (0.8L / min), but the water usage duration feature deviated significantly (8 minutes vs. the template library mean of 2 minutes), and the total water usage feature (6.4L vs. the template library 1.6L) was also obviously abnormal.

[0060] The matching and identification unit used the DTW algorithm to calculate and found that the event matched well with the initiation slope and steady-state mean of some samples in the toilet template library (the distance was small). However, due to the significant deviation in duration and total water consumption, the overall DTW distance increased (average distance 0.65). The K-nearest neighbor voting results showed that among the 5 nearest neighbor samples, 3 were toilets (small distance) and 2 were unknown (large distance). The output device type was "toilet", but the confidence level C decreased to 0.6.

[0061] The feedback correction unit feeds back C=0.6 to the leakage grading response module in real time, and F3=1-0.6=0.4. The leakage grading response module calculates the leakage probability by combining factors such as pressure changes and time context. Because F3 is moderately high and the time anomaly is relatively high, the final P... leak =0.72, triggering an L4 level critical leak alarm. After receiving a notification via the app, the user remotely checked and confirmed the toilet malfunction, then contacted a repair technician to handle the issue, preventing further leakage.

[0062] Scenario 3: Adaptive Recognition by New Devices One day, a user purchased a new robot vacuum cleaner with an automatic mop cleaning function, which generates water for about 1.5 minutes at a flow rate of 0.5L / min.

[0063] The real-time feature extraction unit detected the water usage event and extracted a 12-dimensional feature vector. The matching and recognition unit matched it with the existing template library and found that the DTW distance to all known devices was large (minimum distance 1.2). K-nearest neighbor voting could not classify it, and the output device type was "unknown" with a confidence level C=0.05.

[0064] The feedback correction unit feeds back C=0.05 to the leak classification response module, and F3=0.95. The leak classification response module calculates based on other characteristics; because F3 is extremely high and this period aligns with users' daily habits (F4 is lower), P... leak=0.48, triggering an L2 level alert. The app sends a notification: "Unknown water-using device detected. Please confirm if this is a newly added device." After user confirmation, the system adds the water usage cycle to the template library, creating a new device fingerprint for the robot vacuum. The next time the device runs, the recognition confidence will be significantly improved, and alerts will no longer be triggered.

[0065] The intelligent logic module for going out / returning home includes: The off-duty determination unit is used to determine the duration of no water usage, T. dry Furthermore, smart home signals are used to comprehensively determine whether a user is away from home, including a threshold T for the duration of no water usage. adaptive Dynamically adjusted based on users' historical travel habits; When the user is out of the house, the armed unit outputs a command to close the intelligent water supply shut-off valve V and records the initial pressure P0 at the moment of closure. The homecoming identification unit is used to monitor the pressure change curve when the valve is closed. When a pressure drop event is detected, a hidden Markov model is used to classify the pressure drop pattern. The input of the hidden Markov model is the pressure drop rate sequence, and the output is the probability of "turning on the tap when returning home" or "slow leakage in the pipe". The automatic valve opening unit outputs a command to open the intelligent water supply shut-off valve V if the probability of "turning on the tap upon returning home" is greater than 0.8, according to the Hidden Markov Model. Otherwise, it determines a leak, keeps the valve closed, and triggers an alarm. After the deployment unit is activated, it closes the intelligent water supply shut-off valve V and records the initial pressure P0 at the moment of closure. When the output is less than 0.8, it opens the intelligent water supply shut-off valve V, pressurizes it, and then closes it. It maintains pressure and continues to observe until normal water use of the equipment is eliminated (such as the expansion valve of a water heater opening or a robot vacuum cleaner using water). Otherwise, it determines a leak, keeps the valve closed, and triggers an alarm (the above logic is applicable to weekly nighttime pressure testing for minor leaks).

[0066] This smart logic module for going out / returning home uses the duration of no water usage and the smart home system to determine whether a user has gone out, avoiding misjudgment based on a single condition. Meanwhile, a threshold T for the duration of no water usage is also included. adaptive Based on users' historical travel habits, the system can dynamically adjust to adapt to the different work and rest patterns of different users (such as office workers, freelancers, and retirees), ensuring that the deployment is triggered in a timely manner when going out, while avoiding accidental entry into the out mode due to short-term lack of water (such as during a nap).

[0067] The pressure change curve is monitored when the valve is closed. When a pressure drop event is detected, a Hidden Markov Model (HMM) is used to classify the pressure drop rate sequence. This model can learn the temporal pattern differences between the rapid, step-like pressure drop caused by "turning on the tap when returning home" and the gradual, continuous pressure drop caused by "slow pipe leakage," and output the probability of the two types of events to achieve accurate differentiation. This avoids misjudging a leak as returning home and automatically opening the valve, which could lead to increased losses.

[0068] The system automatically opens the valve when the probability of a user returning home is greater than 0.8, allowing the user to resume normal water use without any intervention, providing a seamless experience. When a leak is detected, the valve remains closed and an alarm is triggered to prevent the leak from escalating. From automatic arming when the user is away and automatic recognition upon returning home to alarms for abnormalities, a complete intelligent protection loop is formed, ensuring both safety and convenience.

[0069] One family user is a working professional who leaves early and returns late on weekdays, with irregular weekend schedules. The system's outing / returning home intelligent logic module continuously learns and adapts to this user's travel habits.

[0070] Scenario 1: Automatic arming when leaving home on weekdays At 8:15 AM on Monday, the user left home for work. The water supply AI engine module detected that the period of no water usage had reached 45 minutes, at which point the no-water usage time threshold T was reached. daptive The interval was dynamically adjusted to 50 minutes using a reinforcement learning algorithm (based on the user's historical average departure time from home on weekdays, which is between 8:10 and 8:20). The result matched the threshold, confirming that the user had left home.

[0071] The outbound arming unit immediately outputs a command to close the intelligent water supply shut-off valve V. The valve closes within 2 seconds, and the initial pressure P0 = 320 kPa is recorded at the moment of closure. The system then enters outbound arming mode. At this time, even if a water pipe bursts at home, there will be no damage because the main valve is closed.

[0072] Scenario 2: Automatic valve opening upon returning home on a weekday At 18:45 that evening, the user returned home from get off work. The homecoming recognition unit continuously monitored the pressure change curve with the valve closed. When the user turned on the tap in the kitchen to wash vegetables, the pressure sensor detected that the pressure dropped rapidly from 320kPa to 280kPa, forming a stepped decline curve.

[0073] The homecoming recognition unit collects a sequence of pressure drop rates (-15 kPa / s for 2 seconds, -8 kPa / s for 3 seconds, and -3 kPa / s for 5 seconds), and inputs this sequence into a Hidden Markov Model (HMM). Through prior training, the model learns the difference between the typical pattern of "turning on the tap upon returning home" (rapid initial drop followed by a gradual slowdown) and the typical pattern of "slow pipe leakage" (continuous, gradual drop). The model calculates the probability as follows: "turning on the tap upon returning home" probability 0.95, "slow pipe leakage" probability 0.05.

[0074] Upon receiving the result, the automatic valve-opening unit, noting that the probability of "turning on the tap upon returning home" is greater than 0.8, immediately outputs a command to open the intelligent water supply shut-off valve V. The valve automatically opens, restoring water supply. The user seamlessly completes the process of turning on the valve upon returning home and resumes normal water use.

[0075] Scenario 3: Avoiding accidental deployment during weekend naps At 2:00 PM on Saturday, the user took a 2-hour nap without using any water. The outing detection unit detected 120 minutes of no water usage, and the current time was a weekend afternoon, which is inconsistent with the user's historical travel habits (the user is usually at home in the afternoon on weekends). The reinforcement learning algorithm dynamically adjusts T... adaptive The time limit is automatically increased to 180 minutes on weekend afternoons (based on the user's weekend routine). Therefore, although the inactivity time has reached 120 minutes and has not exceeded the dynamic weekend threshold, the result is determined as "not out," and the outactivity defense is not triggered. The user can use water normally after waking up without any interference.

[0076] Scenario 4: Real Leakage Identification and Alarm During a user's absence, a slight connection in the home's water supply pipes loosened due to temperature changes, causing a slow leak. The system was in an off-duty armed state, with the valves closed and the pressure sensor continuously monitoring.

[0077] When the leak causes the pressure to slowly decrease from 320 kPa to 305 kPa (the rate of decrease is approximately -0.05 kPa / s over 10 minutes), the homecoming detection unit detects the pressure drop event and inputs the pressure drop rate sequence into the Hidden Markov Model. The model identifies that the sequence exhibits a continuously and gradually decreasing characteristic, which highly matches the training sample of "slow pipe leakage," and outputs: "Homecoming and turning on the tap" probability 0.05, "slow pipe leakage" probability 0.95.

[0078] The automatic valve-opening unit received the results and, because the probability of "turning on the tap upon returning home" was less than 0.8, determined it to be a leak, kept the valve closed, and triggered an alarm. The system immediately pushed an alarm message to the user via the app: "A slow leak was detected in the pipe while you were away. The valve has been kept closed. Please check it promptly." After receiving the notification, the user contacted the property management for an on-site inspection, found a loose connection, and repaired it in time, preventing the leak from expanding due to the automatic valve opening upon returning home or causing greater damage if it went unnoticed for a long time.

[0079] The multi-evidence fusion leakage detection module is constructed based on Dempster-Shafer evidence theory or weighted Bayesian networks, and includes: The first evidence source calculation unit calculates the basic probability allocation m1 based on the evidence of flow imbalance, m1 = f(ΔF, duration), where ΔF is the difference between supply and return water flow and duration is the duration. The second evidence source calculation unit calculates the basic probability allocation m2 based on the pressure differential attenuation evidence, where m2 = g(d(ΔP)). sys ) / dt), where ΔP sys Let d(ΔP) be the system pressure difference. sys ) / dt is the rate of change of pressure difference; The third evidence source calculation unit calculates the basic probability allocation m3 based on the temperature field anomaly evidence, where m3 = h(ΔT). actual , ΔT expected ), where ΔT actual To measure the temperature difference between the supply and return water, ΔT expected This is the theoretical temperature difference based on a thermodynamic model; The fourth evidence source calculation unit calculates the basic probability allocation m4 based on the ultrasonic acoustic fingerprint evidence. m4 is obtained by classifying the audio spectrum maps collected by the first ultrasonic flow meter U1 and the second ultrasonic flow meter U2 based on the convolutional neural network CNN. The convolutional neural network CNN includes an input layer, two convolutional layers, a pooling layer, a fully connected layer and a softmax output layer connected in sequence, and outputs the probability of the leakage category. The evidence synthesis unit is used to synthesize m1, m2, m3, and m4 using Dempster's combination rule, and to calculate the confidence intervals Bel and Pl for the three hypotheses of "leakage", "normal", and "uncertain". The leakage detection unit is used to determine that a leakage has occurred when the confidence level of the "leakage" hypothesis is greater than a preset threshold of 0.75, and outputs a shut-off command to the first hot water intelligent shut-off valve (Vw) and the second hot water intelligent shut-off valve Vh.

[0080] The multi-evidence fusion leak detection module is based on the Dempster-Shafer evidence theory, fusing four physically distinct and independent evidence sources: flow imbalance, pressure differential attenuation, temperature field anomaly, and ultrasonic acoustic signature. Each evidence source perceives leak characteristics from different dimensions, and the confidence intervals of each hypothesis are synthesized through evidence theory. Even if one piece of evidence deviates due to environmental interference or sensor noise, other evidence can still correct the final decision, significantly improving the robustness and reliability of leak detection.

[0081] The system employs Dempster-Shafer evidence theory to calculate the confidence intervals for the three hypotheses: "leakage," "normal," and "uncertain," rather than using simple binary judgment. When there is conflicting evidence or insufficient information, the confidence of the "uncertain" hypothesis increases, and the system can adopt conservative strategies (such as delayed confirmation or raising the warning level) to avoid misjudgment or missed judgment due to insufficient evidence. Shutdown is only executed when the confidence of the "leakage" hypothesis clearly exceeds the threshold of 0.75, minimizing the risk of erroneous actions.

[0082] A family uses an underfloor heating system, with the heating pipes buried beneath the floor. During winter operation, a small leak occurred in one of the heating pipes due to accidental damage from a nail gun used during floor renovations, causing hot water to slowly seep out.

[0083] The first evidence source calculation unit (flow imbalance): The ultrasonic flow meter Fh monitors the supply and return water flow rates and calculates the flow difference ΔF. In the initial stage of minor leakage, the supply and return water flow difference is only 0.1 L / min, and the duration is short. The evidence source calculates that m1 has a low basic probability allocation to the "leakage" hypothesis (m1...). leak =0.2), with a higher allocation to the "normal" assumption (m1) normal =0.7), "uncertain" is assigned 0.1. It is difficult to determine leakage based solely on traffic evidence.

[0084] Second evidence source calculation unit (pressure differential attenuation): Pressure differential ΔP monitored by dual-path pressure sensors Ph1 and Ph2 on the heating side. sys A small leak causes the system pressure differential to decrease slowly, with the rate of change of pressure differential d(ΔP) sys The decay rate is approximately -0.05 kPa / h, and the decay rate is extremely low. The evidence source calculates m2... leak =0.15, m2 normal =0.75, m2 uncertain =0.1. The differential pressure evidence is also insufficient to confirm a leak.

[0085] Third evidence source calculation unit (temperature field anomaly): Supply water temperature sensor Ts and return water temperature sensor Tr monitor the supply and return water temperature difference ΔT. actual The second ambient temperature sensor, Te2, monitors the ambient temperature. The theoretical temperature difference ΔT is calculated based on a thermodynamic model. expected (Considering factors such as heating load and ambient temperature). Minor leaks cause some heat to be lost with the leaking water; the measured supply and return water temperature difference ΔT actual Compared to the theoretical value ΔT expected It is about 1.2℃ lower than normal. Evidence of the temperature field anomaly was calculated to be m3. leak =0.45, m3 normal =0.35, m3 uncertain =0.20. This evidence strengthens the support for a leak, but still does not reach the decision threshold.

[0086] The fourth evidence source calculation unit (ultrasonic acoustic signature): The first ultrasonic flow meter U1 and the second ultrasonic flow meter U2 are installed on the surface of the supply and return water pipes respectively to collect audio signals. The audio signals are converted into a spectrum and then input into a convolutional neural network (CNN) (input layer → two convolutional layers → pooling layer → fully connected layer → softmax output layer). The CNN model is trained on a large number of leak acoustic signature samples and can identify specific frequency turbulent noise generated by minute leaks. In this detection, the CNN output leak category probability was 0.85, normal probability was 0.10, and uncertain probability was 0.05. The ultrasonic acoustic signature evidence was calculated to yield m4. leak =0.85, m4 normal =0.05, m4 uncertain=0.10.

[0087] Evidence synthesis unit: The four evidence sources m1, m2, m3, and m4 are synthesized using Dempster's combination rule. The calculation process is as follows: Initial trust assignment: Four sources of evidence provide the degree of support for the three hypotheses of "leakage", "normal", and "uncertainty".

[0088] Iterative synthesis is performed using Dempster's combination rules, taking into account the degree of conflict between pieces of evidence.

[0089] Synthesis results: The confidence intervals for the "leak" hypothesis were calculated to be Bel(leak) = 0.82 and Pl(leak) = 0.88; for the "normal" hypothesis, Bel(normal) = 0.08 and Pl(normal) = 0.15; and for the "uncertain" hypothesis, Bel(uncertain) = 0.04 and Pl(uncertain) = 0.10.

[0090] Leakage detection unit: Receives the synthesis results. The confidence level of the "leakage" hypothesis Bel(leak) = 0.82 > the preset threshold of 0.75. It determines that a leak has occurred and immediately outputs a shut-off command to the second hot water intelligent shut-off valve Vh and the first hot water intelligent shut-off valve Vw to close the main valve of the heating system. At the same time, it triggers an alarm and pushes information to the user's mobile phone and control panel.

[0091] After receiving an alarm stating "A minor leak has been detected in the underfloor heating system, and it has been automatically shut down," the user contacted professionals for repairs. Because the system accurately identified and shut down the leak in its early stages, it prevented widespread floor flooding, moldy walls, and wasted heat energy caused by continued leakage, saving tens of thousands of yuan in potential repair costs from a single incident. If relying solely on a single source of evidence (such as flow rate or pressure difference), this minor leak might not be detected for weeks or even months, leading to serious consequences.

[0092] The blockage diagnosis and quantification module includes: The filter clogging index calculation unit is used to calculate the filter clogging index I according to the following formula. filter : I filter = 0.5×(ΔP filter / ΔP0) + 0.3×(1-F actual / F0) + 0.2×T rend , Wherein, ΔP filter F represents the pressure difference across the filter, where ΔP0 is the baseline pressure difference under clean filter conditions. actual F0 is the actual flow rate, and T is the baseline flow rate under clean conditions. rend This is a pressure difference change trend factor based on historical data; The pipeline siltation index calculation unit is used to establish the system hydraulic model ΔP. sys = R×F 2 F is the volumetric flow rate of circulating water in the pipeline. The normal range of the system resistance coefficient R is determined by fitting operational data. min , R max ], and calculate the current drag coefficient R. current Corresponding siltation index I pipe = (R current - R min ) / (R max -R min ); The blockage diagnosis and quantification module uses three parameters: the pressure difference across the filter, the ratio of actual flow rate to a baseline value, and historical trends, weighted by formula I. filter =0.5×(ΔP filter / ΔP0)+0.3×(1-F actual / F0)+0.2×T rend Calculate the filter clogging index. This model considers both the deviation between the current pressure difference and flow rate, as well as long-term trends, and can accurately quantify the continuous state of the filter from clean to completely clogged, avoiding misdiagnosis and missed diagnosis caused by simple threshold judgment.

[0093] Establish a system hydraulic model ΔP sys =R×F 2 By fitting long-term operating data, the normal range of the system drag coefficient R is determined. min , R max ], and calculate the current drag coefficient R. current Corresponding siltation index I pipe =(R current -R min ) / (R max -R min This method transforms the complex problem of pipeline siltation into a quantifiable change in resistance coefficient, enabling dynamic monitoring and trend warning of system flow resistance.

[0094] By using a first ultrasonic flow meter U1 and a second ultrasonic flow meter U2 to receive the low-frequency turbulent noise energy spectrum, it is observed that blockage leads to an increase in local flow velocity and enhanced turbulence, resulting in a significant increase in noise energy near the blockage point. This method can roughly locate the blockage point without disassembling the pipeline, providing a basis for precise maintenance and significantly reducing the difficulty and cost of repairs.

[0095] For example, if a household's underfloor heating system has been running for 5 years, the user may recently notice a slower rate of indoor heating and increased energy consumption, but the system is still functioning normally. The blockage diagnosis and quantification module of the heating AI engine will then initiate regular self-checks.

[0096] Filter clogging index calculation unit analysis: Collect the pressure difference ΔP before and after the filter. filter =15kPa, while the reference pressure difference ΔP0 under the clean state of the filter screen is 8kPa, the ratio is 1.875; Current actual traffic F actual =0.8m 3 / h, clean state baseline flow rate F0=1.2m 3 / h, 1-F actual / F0=1-0.67=0.33; Based on historical data, the pressure differential change trend factor T over the past 3 months rend =0.6 (continuously rising); Substitute into the formula to calculate the filter clogging index: I filter =0.5×1.875+0.3×0.33+0.2×0.6=0.9375+0.099+0.12=1.1565.

[0097] Because I filter If the value is >1.0, the system determines that the filter is nearly completely clogged and records the status as "severely clogged filter".

[0098] Analysis of the pipeline siltation index calculation unit: Establish a system hydraulic model ΔP sys =R×F 2 The normal range of the drag coefficient R is fitted using historical operating data. The drag coefficient R in the initial stage (clean state) of the system. min =8 kPa / (m 3 / h) 2 The historical maximum permissible drag coefficient R max =15 kPa / (m 3 / h) 2 .

[0099] Current measured data: Total system voltage drop ΔP sys =45kPa, current flow rate F current =0.8m 3 / h, calculate the current drag coefficient R current =ΔP sys / F current 2 =45 / 0.64=70.3 kPa / (m 3 / h) 2 (Due to severe filter clogging causing a sharp increase in local pressure drop, this filter pressure drop is already included in the total system pressure drop, therefore R) current (far exceeding the normal range).

[0100] To separately assess internal pipe buildup (excluding the influence of filters), the system employs differential analysis to fit the pipe's resistance coefficient R based on branch pipe pressure differential data (not detailed in the formula). pipe =12.5 kPa / (m 3 / h) 2 .

[0101] Calculate the pipeline siltation index: I pipe =(12.5-8) / (15-8)=4.5 / 7=0.64.

[0102] The results showed that there was moderate siltation inside the pipe (I pipe =0.64), but has not yet reached the severe congestion threshold.

[0103] Overall diagnostic conclusion: Filter clogging index I filter If the value is 1.16, the filter is severely clogged and needs to be cleaned or replaced immediately. Pipeline siltation index I pipe =0.64, indicating moderate siltation inside the pipe; pipe cleaning is recommended. The above diagnostic results are summarized in the integrated health management module, and the heating health is divided into S... heat The system's overall health score dropped from 85 to 62. The interactive terminal displayed "Heating system health has declined: the filter is severely clogged and the pipes are moderately silted up. Maintenance is recommended as soon as possible," and generated a predictive maintenance work order, which was then pushed to the user's app.

[0104] After receiving the notification, the user contacted a professional maintenance technician. Following the location prompts, the technician first checked the distributor filter, finding it completely clogged with impurities. After cleaning the filter, the system pressure differential returned to normal, and the flow rate increased to 1.1 m³ / h. 3 / h. It is also recommended that users schedule a full pipe cleaning next quarter.

[0105] The valve health prediction module includes: The current curve acquisition unit is used to acquire the motor current curve I(t) at a sampling rate of not less than 1kHz each time the intelligent shut-off valve is activated. The feature extraction unit is used to extract the following features from the motor current curve I(t): Time-domain characteristics: Peak current I peak Average current I mean Current variance σ 2 ; Shape characteristics: Dynamic time warping (DTW) distance between the current and a pre-stored standard health current profile; Frequency domain characteristics: the main frequency components and their amplitudes extracted after performing a fast Fourier transform on I(t); The health score calculation unit has a built-in support vector regression (SVR) model. It takes the extracted features as input and outputs a valve health score S from 0 to 100. valve ; The remaining useful life prediction unit records the historical sequence of health scores and uses exponential smoothing or linear regression extrapolation to predict the time required for the health score to drop to a preset failure threshold, which is the remaining useful life.

[0106] The valve health prediction module acquires the valve operating current curve I(t) at a high sampling rate of no less than 1kHz, and extracts features from three dimensions: time domain (peak value, mean, variance), shape (DTW distance from the standard curve), and frequency domain (FFT main frequency component). The time domain features reflect the magnitude and fluctuation of the current, the shape features quantify the degree of waveform distortion, and the frequency domain features reveal the vibration frequency changes caused by mechanical wear. These three features complement each other to form a comprehensive characterization of the valve's health status, laying the foundation for accurate assessment.

[0107] A Support Vector Regression (SVR) model is employed, using extracted multidimensional features as input to output a continuous health score from 0 to 100. The SVR model exhibits excellent nonlinear fitting ability and generalization performance, and can learn the complex mapping relationship between features and health status under different failure modes. The output score intuitively reflects the current state of the valve (100 points for ideal health, 0 points for complete failure), facilitating user understanding and system decision-making.

[0108] By recording historical sequences of health scores and extrapolating the trend of score changes using exponential smoothing or linear regression, the remaining useful life (RUL) is predicted when the health score drops to a preset failure threshold (e.g., 60 points). This method transforms discrete health scores into remaining time with clear physical meaning, providing a direct basis for predictive maintenance and enabling users to schedule repairs in advance to avoid unexpected failures.

[0109] In a household water and heating safety intelligent protection system, the intelligent water supply shut-off valve V is installed on the main water supply pipe and has been operating normally for 3 years. The valve health prediction module continuously monitors the health status of this valve.

[0110] Initial action record (initial installation): The current curve acquisition unit acquires the motor current curve I(t) at a sampling rate of 1kHz when the valve first actuates. The feature extraction unit extracts: Time-domain characteristics: Peak current I peak =0.85A, average current Current variance σ 2 =0.012; Shape characteristics: The DTW distance calculated with the pre-stored standard health current curve (factory calibration curve of the same model new valve) is 0.15; Frequency domain characteristics: Perform FFT transformation on I(t) to extract the main frequency component 50Hz and its amplitude 0.3A, and the second harmonic 100Hz with an amplitude of 0.08A.

[0111] The health score calculation unit inputs the above features into the SVR model and outputs a health score S. valve =98 points. The remaining life expectancy prediction unit records this score as a health benchmark.

[0112] Routine testing two years later: The system recorded a total of 850 valve actuations. The current curve characteristics acquired during this actuation are as follows: Peak current rises to I peak =1.12A (an increase of 31.8%), average current (Growth of 30.8%), Current variance σ 2 =0.045 (increased by 2.75 times); The DTW distance increased to 0.58 (significant waveform distortion). FFT analysis showed that the amplitude of the main frequency 50Hz dropped to 0.18A, while obvious high-frequency components at 150Hz and 200Hz (amplitudes of 0.15A and 0.12A) appeared, indicating that mechanical wear was intensifying and generating high-frequency vibration noise. The SVR model output a health score of Svalve=76 after inputting features.

[0113] Trend analysis two and a half years later: The valve has a cumulative total of 1100 actuations. Key characteristics of this operation: Peak current I peak =1.35A, mean variance σ 2 =0.068; DTW distance: 0.82; The high-frequency components are further enhanced, with an amplitude of 0.22A at 150Hz and 0.19A at 200Hz.

[0114] SVR model outputs health score S valve =68 points.

[0115] The remaining useful life prediction unit retrieved the historical score sequence: [98, 92, 87, 83, 76, 72, 68] (time span of 3 years), and used exponential smoothing to extrapolate the trend (with a smoothing coefficient α=0.3) to predict the future score change curve. The system presets a fault threshold of 60 points (below this value indicates that the valve is at risk of jamming and needs to be replaced immediately). The extrapolation results show that the health score is expected to drop to 60 points in 8 months, i.e., the remaining useful life RUL = 8 months.

[0116] The integrated health management module receives valve health prediction results: Current health score: 68 (yellow warning zone); Remaining useful life: 8 months; Based on the valve's importance (main water supply valve) and historical operating frequency, a predictive maintenance recommendation is generated: "The health of the main water supply valve is declining, and it is expected to reach the failure threshold in 8 months. It is recommended to arrange preventive replacement within 6 months." The system displays the suggestion via an interactive terminal and pushes it to the user's app. After receiving the reminder, the user schedules a professional appointment four months later for preventative valve replacement. During the replacement, it was discovered that the original valve's internal seals were severely worn and the gearbox grease had dried out. Continued use would likely result in jamming or incomplete shut-off within six months.

[0117] After the new valve is installed, the health baseline is automatically re-established, and a new monitoring cycle begins. This early warning and intervention avoids the risk of flooding that could result from sudden malfunctions, and saves users the additional costs and time associated with emergency repairs.

[0118] The user's valve health score sequence, characteristic data, and actual replacement records (wear level confirmed during replacement) were anonymized and then encrypted before being uploaded to the cloud-based intelligent platform.

[0119] The collaborative arbitration module integrates a scenario strategy library, which includes at least the following: The "away mode" strategy is triggered when the water supply AI engine module detects that there has been no water usage for more than T consecutive hours. adaptive Furthermore, the smart home system prompts the user to leave home; the action executed is as follows: the collaborative arbitration module issues an instruction to the water supply AI engine module to close the smart water supply shut-off valve V and enter pressure monitoring; Summer heating mode strategy: When the heating AI engine module detects that the flow rates of the first ultrasonic flow meter U1 and the second ultrasonic flow meter U2 are both zero, the second ambient temperature sensor Te2 remains above 25°C for a week and there is no heating demand, the heating AI engine switches to low-power monitoring mode to release NPU computing resources; the collaborative arbitration module dynamically allocates the released resources to the water supply AI engine module. Winter heating mode strategy: When the heating AI engine module detects that the flow rate of the first ultrasonic flow meter U1 or the second ultrasonic flow meter U2 is greater than zero, the inlet water temperature sensor T1 is 2°C higher than the outlet water temperature sensor T2, and there is a clear heating demand, the heating AI engine module exits the low-power monitoring mode and resumes full-function monitoring; the collaborative arbitration module reallocates computing resources.

[0120] The collaborative arbitration module can dynamically adjust the allocation of computing resources between the water supply AI engine and the heating AI engine based on seasonal changes, system load, and user status. During the winter when heating demand is high, it ensures both engines operate at full speed; during the summer when there is no heating demand, it switches the heating AI engine to a low-power monitoring mode, freeing up valuable NPU computing resources for the water supply AI engine to perform deep data mining and model optimization, thus maximizing the utilization of hardware computing power.

[0121] It has a built-in library of scenario strategies, including an "away mode" and winter / summer heating modes, which can automatically execute preset coordinated actions based on multi-dimensional triggering conditions (water usage behavior, ambient temperature, seasonal characteristics). When away from home, it controls the intelligent shut-off valve of the water supply and the operation of the heating system, realizing joint prevention and control of the water and heating systems. This ensures both safety and energy conservation, and is more intelligent and efficient than controlling a single system independently.

[0122] The system monitors the control commands of both engines in real time via an internal data bus and arbitrates command conflicts according to preset priority rules (such as leak shutdown > daily adjustment > model training). When the water supply AI engine issues a pipe burst shutdown command, even if the heating AI engine is performing complex calculation tasks, this module can ensure that the shutdown command receives the highest priority and is executed immediately, guaranteeing a second-level response to emergency events and avoiding system chaos caused by multiple concurrent commands.

[0123] A family has been using a smart home water and heating safety protection system based on dual AI engine collaboration for a year. The collaborative arbitration module continuously schedules the dual engines to work together and records user behavior and environmental changes.

[0124] Scenario 1: Triggering the Out-of-Town Mode Trigger condition monitoring: Water supply AI engine module 21 detected that the continuous period of no water usage reached T. adaptive =50 minutes (dynamically adjusted based on historical habits), while the smart home signal indicates that the user has left home.

[0125] Outing mode trigger: The collaborative arbitration module 23 receives data from both engines, determines that the outing mode trigger conditions are met, and immediately executes the preset strategy: The AI ​​engine module 21 is instructed to close the intelligent shut-off valve V and enter the pressure monitoring state (record the closing pressure P0=320kPa and continuously monitor pressure changes).

[0126] Synergistic effect: Even if a water pipe bursts while the user is away from home, there will be no water damage because the main valve is closed; when the user returns home from get off work, the homecoming recognition unit detects the pressure drop and automatically opens the valve, achieving a seamless experience.

[0127] Scenario 2: Winter Heating Season During the harsh winter, the heating AI engine module 22 operates at full power, continuously monitoring the heating system's status. One day, the heating AI engine module 22 detected a filter blockage index I. filter =0.85, requiring blockage diagnosis calculations (a surge in instantaneous computing power demand). At this time, the water supply AI engine module 21 has a low load (only performing routine basic monitoring tasks, without high-load calculations), and there is idle NPU computing power.

[0128] The collaborative arbitration module 23 detected a real-time load change: the heating AI engine module's computing power demand surged, while the water supply AI engine module had idle resources. It immediately triggered a dynamic resource scheduling strategy, temporarily allocating some computing resources from the water supply AI engine module to assist the heating AI engine module in completing the congestion diagnosis calculation. After the calculation was completed, the resources were automatically reclaimed, restoring their respective default resource allocations.

[0129] Scenario 3: Summer Low Power Mode (July) In the sweltering heat of July, the second ambient temperature sensor Te2 showed temperatures above 30°C for a week, and the heating AI engine module 22 detected no heating demand.

[0130] Summer mode triggered: The collaborative arbitration module 23 receives the "long-term no heating demand" signal from the heating AI engine and verifies the condition that the ambient temperature is continuously higher than 25°C, and determines that the summer mode is triggered.

[0131] The command is given to the heating AI engine module 22: switch to "low-power monitoring mode". The heating AI engine suspends all active monitoring tasks, retaining only basic communication and wake-up functions, thus reducing power consumption.

[0132] The NPU computing resources released by the heating AI engine module are dynamically allocated to the water supply AI engine module 21 by the collaborative arbitration module.

[0133] After the water supply AI engine module gains additional computing power, it launches the "deep data mining" task: to conduct offline model training and optimization on the historical water supply data accumulated over the past six months (including water usage patterns, equipment fingerprint characteristics, pressure fluctuation patterns, etc.), and to update the local equipment fingerprint template library and leak detection model parameters.

[0134] Scenario 4: Command Priority Arbitration (Emergency Events) Late one summer night, while users were asleep, the water supply AI engine module 21 was utilizing its idle computing power at night to perform deep mining of historical data (a high-load computing task), consuming most of the NPU resources.

[0135] At 3 a.m., the kitchen water supply pipe suddenly burst due to pressure fluctuations. The leak classification response module of the water supply AI engine module 21 detected the sudden drop in pressure and surge in flow, and immediately generated an emergency shutdown command.

[0136] At this time, the water supply AI engine module itself is under high load due to deep mining tasks, and the shutdown command needs to be urgently sent to the water supply intelligent shut-off valve V. The collaborative arbitration module 23 detects the L5 level emergency shutdown command through the internal data bus and immediately executes priority arbitration: Suspend the deep mining task currently being performed by the water supply AI engine module to free up communication bus resources; Ensure that the shut-off command receives the highest priority and is immediately transmitted to the intelligent water supply shut-off valve V; At the same time, a coordination command to "pause non-urgent tasks" is sent to the heating AI engine module 22 to prevent it from occupying bus resources.

[0137] The shut-off command was delivered to the valve within 0.5 seconds, and the shut-off was completed within 2.8 seconds, successfully preventing a flooding accident. After the emergency was handled, the collaborative arbitration module resumed the deep excavation task.

[0138] It also includes an integrated health management module, which calculates the system's total health score S according to the following formula. total : S total = 0.4×S water + 0.4×S heat + 0.2×S common , Among them, S water The water supply health score is calculated by the water supply AI engine module based on the number of historical leakage events, valve health scores, and pipeline aging coefficients. S heat The heating health score is calculated by the heating AI engine based on the blockage diagnosis index, thermal efficiency, and valve health score. S common The health score for public infrastructure is calculated based on a comprehensive assessment of backup battery health, storage media lifespan, and network communication quality. The integrated health management module also based on S total The system generates predictive maintenance suggestions or alarm messages, which are then displayed via an interactive terminal. The cloud-based intelligent platform, in conjunction with the dual AI engine collaborative control layer, implements a federated learning self-evolution mechanism, specifically including: The local update unit, deployed on edge devices, is used to incrementally train a lightweight model locally using the event data after detecting a local false alarm or missed alarm event, and generate model gradient update ΔW. An encrypted upload unit is used to encrypt ΔW and upload it to the cloud intelligent platform via the TLS 1.3 protocol; The cloud aggregation unit receives encrypted gradient updates from multiple edge devices, decrypts and aggregates them, and generates global model parameters W. global ; The secure distribution unit is used to control W. global Digital signature encryption is performed and distributed to all edge devices via OTA. After verifying the signature, the edge devices update their local models, thus completing the collective evolution. The hardware sensing and execution unit also includes a multi-level power protection unit, which includes a main power supply AC220V input, a 12V / 10Ah backup lithium battery, and a supercapacitor. The supercapacitor is used to provide the energy required for the final shutdown action when both the main power supply and the backup battery fail, ensuring that an emergency shutdown can still be completed in extreme power outage conditions.

[0139] The combined advantages of the integrated health management module, the cloud-based federated learning self-evolution mechanism, and the multi-level power protection unit are as follows: The integrated health management module constructs a water supply health system. water Heating Health S heat Public infrastructure health segment common The three-dimensional evaluation system calculates a total system health score (Stotal) through weighted fusion. The water supply health score integrates the number of historical leakage events, valve health scores, and pipe aging coefficients; the heating health score integrates the blockage diagnosis index, thermal efficiency, and valve health scores; and the public infrastructure health score integrates backup battery health, storage media lifespan, and network communication quality. This model transforms the dispersed component states into unified health indicators, based on S... total Value generation hierarchical maintenance recommendations (such as S) total A score of ≥90 indicates excellent health, requiring only routine monitoring; 75-89 indicates good health, requiring attention; 60-74 indicates a warning, requiring maintenance; and <60 indicates a serious health condition, requiring immediate repair. The health status of the home's water and heating system is presented visually through an interactive terminal, making it easy to understand at a glance.

[0140] A cloud-based intelligent platform and edge devices collaborate to implement a federated learning self-evolution mechanism. The local update unit is deployed on each home edge device. When a false positive or false negative event is detected locally, it uses the event data to incrementally train a lightweight model (such as a random forest for leak detection or a device fingerprint KNN template library), generating a model gradient update ΔW, which is then encrypted using TLS 1.3 and uploaded to the cloud. The cloud aggregation unit receives encrypted gradients from tens of millions of devices, decrypts and aggregates them to generate the global model parameters W. global The data is then encrypted with a digital signature and distributed via OTA. Edge devices verify the signature and update their local models, completing a collective intelligent evolution. This mechanism, while protecting user privacy, makes the system increasingly intelligent with use, continuously improving detection accuracy.

[0141] The hardware sensing and execution unit incorporates multi-level power protection: When the main AC220V power supply is normal, the system operates at full capacity; after a power outage, a 12V / 10Ah backup lithium battery automatically takes over, supporting core monitoring and communication functions for over 72 hours; in extreme cases where both the main power supply and backup battery fail (such as fire or lightning strikes causing dual power supply damage), the supercapacitor can respond in milliseconds, providing the energy needed for a final emergency shutdown, ensuring that the intelligent water supply shut-off valve V, the first hot water intelligent shut-off valve Vw, or the second hot water intelligent shut-off valve Vh complete their shut-off actions, maximizing safety protection and preventing catastrophic consequences due to power outages preventing valve closure. The intelligent water supply shut-off valve V, the first hot water intelligent shut-off valve Vw, and the second hot water intelligent shut-off valve Vh all integrate manual mechanical operating mechanisms, allowing for manual operation of the valves in case of power outages, network outages, or electronic control system malfunctions that prevent automatic shut-off.

[0142] A high-end residential community has uniformly installed a smart home water and heating safety protection system based on dual AI engine collaboration, covering a total of 500 households. After the system was put into operation, the integrated health management, federated learning self-evolution, and multi-level power supply protection mechanisms worked together to demonstrate an outstanding level of intelligence.

[0143] Scenario 1: System Overall Health Score Early Warning and Predictive Maintenance In a user's home, the integrated health management module periodically calculates the system's total health score: Water supply AI engine module 21 provides the following data: 0 historical leakage events, and a health score of S for the first intelligent shut-off valve Vw. valve =82 points (output by the valve health prediction module), pipe aging coefficient 0.15 (based on cumulative flow and pipe life model), and S is calculated comprehensively. water =85 points.

[0144] Heating AI Engine Module 22 provides data: Blockage Diagnostic Index (Moderately clogged filter), thermal efficiency η=0.88 (slightly lower than the benchmark of 0.92), second hot water intelligent shut-off valve Vh health score S valve =78 points, calculated comprehensively, S heat =76 points.

[0145] Public infrastructure monitoring: Backup battery health (based on charge / discharge cycle count and internal resistance change) is 82%, storage media lifetime (based on write volume and bad block count) is 90%, network communication quality (packet loss rate 0.3%) is 95%, and the comprehensive calculation yields S. common =88 points.

[0146] Substitute into the formula: S total=0.4×85+0.4×76+0.2×88=34+30.4+17.6=82 points (good range, but close to the warning threshold).

[0147] The integrated health management module is based on S total Based on the score of 82 and its sub-items, a targeted maintenance suggestion was generated: "The heating system health score is low (76 points), mainly due to moderate filter blockage and decreased valve health. It is recommended to clean the heating system filters and check the valve status within one month." This suggestion was displayed through the interactive terminal and pushed to the user's APP.

[0148] The user scheduled a professional maintenance appointment. After the maintenance personnel cleaned the filter, I... filter The efficiency dropped to 0.15, but rebounded to 0.91; simultaneously, the valves were inspected, lubricating grease was added, and the valve health score improved to 85 points. The following month, S... heat S rose back to 88 points total The score rose to 88, and the system returned to normal. Thanks to the early warning, the serious consequences of a completely clogged filter leading to poor heating performance and no heating at all during the winter were avoided.

[0149] Scenario 2: Level 3 power supply protection completes emergency shutdown during extreme power outages. One night, a large-scale power outage occurred in the area due to a substation malfunction. The main AC220V power supply to the homes of the residents was interrupted, and the backup lithium batteries automatically and seamlessly took over. The system switched to a low-power mode: the interactive terminal screen was turned off, and non-essential computing tasks were suspended, but the core monitoring and communication modules (Wi-Fi / 4G) continued to operate.

[0150] After a 30-hour power outage, the backup lithium battery was depleted. At this point, if a water pipe burst, the valve would be unable to shut off, with potentially disastrous consequences. However, the system's multi-stage power backup system served as a final line of defense. In the final moments before the lithium battery voltage drops to the cutoff voltage, the power management circuit triggers the activation of the supercapacitor.

[0151] The 50 joules of energy stored in the supercapacitor can be released within 5 seconds, which is enough to drive the intelligent water supply shut-off valve V, the first hot water intelligent shut-off valve Vw, and the second hot water intelligent shut-off valve Vh to complete a full shut-off action.

[0152] At this time, the pressure sensor of the water supply AI engine module 21 detected abnormal pressure fluctuations caused by a power outage in the neighboring house (water pump shutdown), which could potentially trigger water hammer, but no pipe burst actually occurred in the neighboring house. The system did not trigger shutdown, and the supercapacitor maintained its energy.

[0153] Early the following morning, a water pipe in the user's home froze and burst due to a sudden drop in temperature during a power outage. The pressure sensor detected a sudden pressure drop from 280 kPa to 150 kPa, the flow meter detected a surge in flow, and the leak classification response module calculated P... leak =0.92, indicating an L5 level emergency leak, and immediately outputting a shutdown command. At this time, both the main power supply and the lithium battery have failed, but the supercapacitor instantly releases energy, driving the intelligent water supply shut-off valve V, the first hot water intelligent shut-off valve Vw, and the second hot water intelligent shut-off valve Vh to shut down within 2.1 seconds. At the same time, an alarm message "Emergency Shutdown: Pipe burst detected, valves closed, please handle as soon as possible" is sent to the user via the 4G module (powered by the supercapacitor).

[0154] Upon receiving the message, the user promptly contacted property management to shut off the main water supply, preventing severe flooding that could have occurred if the valve had failed to close due to the power outage. Once power was restored, the system automatically recharged, and the supercapacitor re-stored energy, preparing for the next extreme situation.

[0155] Through predictive maintenance via the integrated health management module, the community prevented 23 potential faults throughout the year; through federated learning and self-evolution, the system's false alarm rate decreased from 1.2 times / household / year to 0.3 times / household / year; and through three-level power supply protection, it successfully completed emergency shutdowns during three extreme power outage events, avoiding potential property damage.

[0156] This invention also relates to a protection method for a home water and heat safety intelligent protection system based on dual AI engine collaboration, comprising the following steps: Step S1: The water supply AI engine module processes the sampling data on the water supply side in real time, executes leakage classification response, equipment fingerprint recognition, and intelligent logic for going out / coming home, and generates water supply side status data and control decisions. Step S2: The heating AI engine module performs fusion analysis on the multi-source heterogeneous data of the heating side, performs multi-evidence fusion leak detection, blockage diagnosis and quantification, and valve health prediction, and generates heating side status data and control decisions. Step S3: The collaborative arbitration module dynamically schedules the computing resources of the water supply AI engine module and the heating AI engine module according to the preset scenario strategy, arbitrates data conflicts and control command priorities, and outputs collaborative control commands. Step S4: The cloud-based intelligent platform receives encrypted model gradient data uploaded from multiple edge devices, performs federated learning to aggregate and update the global model, and sends the optimized global model parameters to the edge devices via OTA. Step S5: The integrated health management module calculates the total health score of the system based on the status data of the water supply side and the heating side, and generates predictive maintenance suggestions or alarm information, which are displayed through the interactive terminal.

[0157] The advantages of step S1 are: millisecond-level real-time processing and accurate identification on the water supply side. This method uses a water supply AI engine module to process water supply side sampling data in real time, integrating leak classification response, equipment fingerprint recognition, and intelligent logic for going out / returning home. The leak classification response uses a four-dimensional feature weighting method to calculate the leak probability, which achieves L5 to L0 classification response. The equipment fingerprint recognition accurately identifies water-using equipment through a 12-dimensional feature vector and the DTW-KNN algorithm, and the confidence level is fed back in real time to correct the leak probability. The going out / returning home logic combines a dynamic threshold for no-water usage time and a hidden Markov model to accurately distinguish between water usage upon returning home and slow leaks.

[0158] The advantages of step S2 are as follows: Multi-source heterogeneous data fusion and deep diagnosis on the heating side. The heating AI engine module performs fusion analysis on multi-source heterogeneous data on the heating side, performing multi-evidence fusion leak detection, blockage diagnosis and quantification, and valve health prediction. Multi-evidence fusion integrates four independent evidence sources—flow imbalance, pressure differential attenuation, temperature field anomaly, and ultrasonic acoustic signature—based on Dempster-Shafer theory, overcoming misjudgments from single sensors through confidence interval calculation. Blockage diagnosis quantitatively assesses system resistance using the filter blockage index formula and hydraulic model, and locates blockage points using dual-microphone noise energy spectrum. Valve health prediction outputs health scores and remaining lifespan through time-domain, shape, and frequency-domain feature extraction of current curves and SVR model. This step ensures clear reliability thresholds for heating leak detection, accurate blockage location, and enables full lifecycle health management of the heating system.

[0159] The advantages of step S3 are: collaborative arbitration enables dynamic resource scheduling and instruction priority assurance. The collaborative arbitration module dynamically schedules dual-engine computing resources according to preset scenario strategies and arbitrates data conflicts and control instruction priorities. In winter outing mode, the water supply valve is automatically shut off and the heating system is reduced to anti-freeze operation. In summer mode, the heating engine's NPU resources are released to the water supply engine for deep data mining. Simultaneously, built-in priority rules ensure that L5-level emergency shutdown instructions receive the highest priority, achieving a dynamic balance between energy saving and safety.

[0160] The advantages of step S4 are: federated learning drives model self-evolution and privacy protection. It receives encrypted model gradient data uploaded from multiple edge devices through a cloud-based intelligent platform, performs federated learning to aggregate and update the global model, and then distributes it via OTA. Each edge device generates a gradient update ΔW based on local false positive / false negative events, which is then uploaded using TLS 1.3 encryption; the cloud aggregates gradients from tens of millions of devices to generate the global model W. global The data is signed, encrypted, and then distributed; edge devices verify and update the local model. This step continuously improves the system's detection accuracy without leaking the user's original data.

[0161] The advantages of step S5 are: integrated health management enables visualized predictive maintenance. The integrated health management module calculates the total system health score based on status data from the water supply and heating sides, and generates tiered maintenance suggestions or alarm information, which is displayed through an interactive terminal. The water supply health score integrates leakage frequency, valve health, and pipe aging; the heating health score integrates blockage index, thermal efficiency, and valve health; and the public health score integrates battery health, storage life, and network quality. This step transforms the dispersed component status into intuitive health indicators, allowing users to view the system status in real time and receive alerts.

[0162] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart home water and heating safety protection system based on dual AI engine collaboration, characterized in that, include: The hardware sensing and execution unit, deployed in the household water supply pipeline and heating pipeline, includes: dual-path pressure sensors (P1, P2) on the water supply side, ultrasonic flow meter (Fw), temperature sensor (Tw), intelligent water supply shut-off valve (V), and first ambient temperature sensor (Te1). The heating side includes dual-path pressure sensors (Ph1, Ph2), supply water temperature sensor (Ts), return water temperature sensor (Tr), second ambient temperature sensor (Te2), first hot water intelligent shut-off valve (Vw), second hot water intelligent shut-off valve (Vh), first ultrasonic flow meter (U1), second ultrasonic flow meter (U2), and supply water filter (L). A water supply intelligent shut-off valve (V) is installed on the water supply pipeline. A first hot water intelligent shut-off valve (Vw) and a second hot water intelligent shut-off valve (Vh) are installed on the heating pipeline. The water supply intelligent shut-off valve (V), the first hot water intelligent shut-off valve (Vw), and the second hot water intelligent shut-off valve (Vh) are all equipped with a torque feedback unit for real-time monitoring of motor current and judgment of valve resistance. The dual AI engine collaborative control unit includes a water supply AI engine module, a heating AI engine module, and a collaborative arbitration module; wherein, the water supply AI engine module integrates a leakage classification response module, an equipment fingerprint recognition module, and an out / back home logic module, which are used to process the sampling data on the water supply side in real time; The heating AI engine module integrates a multi-evidence fusion leak detection module, a blockage diagnosis and quantification module, and a valve health prediction module, which are used to fuse and analyze multi-source heterogeneous data on the heating side. The collaborative arbitration module is connected to the water supply AI engine module and the heating AI engine module through an internal data bus, and is used to dynamically schedule the computing resources of the two engines, arbitrate data conflicts, and prioritize control commands according to the preset scenario strategy. The cloud-based intelligent platform is connected to the dual AI engine collaborative control unit via an encrypted communication link. It is used to receive encrypted model gradient data uploaded from several edge devices, perform federated learning to aggregate and update the global model, and securely distribute the global model parameters to the edge devices via OTA.

2. The intelligent home water and heating safety protection system based on dual AI engine collaboration as described in claim 1, characterized in that, The leakage classification response module includes: The feature calculation unit takes the pressure sequence P(t), flow sequence F(t), and device identification result from the device fingerprinting module as inputs within the time window Δt, and calculates the following features: The negative rate of change of pressure characteristic F1 = sigmoid(-k·dP / dt), where dP / dt is the derivative of pressure with respect to time, k is the scaling factor, and the sigmoid function maps the result to the range of 0 to 1; Unsupervised anomaly score F2 based on the isolated forest algorithm; The device matching confidence complement F3 = 1 - C, where C is the device identification confidence level from 0 to 1. When F3 is close to 1, the probability of leakage increases. Based on the Poisson distribution, the temporal context anomaly F4 = -log(P) poisson (k; λ)), where λ is a parameter representing the user's historical water usage habits; specifically, using a Poisson distribution to model the probability P of the current water usage event occurring at the current time point based on the user's historical water usage habits. poisson The higher the F4 value, the higher the temporal anomaly of the event; The leakage probability calculation unit is used to calculate the leakage probability based on the weighted summation formula P. leak =Σ(w i ·F i Calculate the leakage probability, where the weight w i F was obtained by training on a historical leak dataset. i This represents the i-th feature value used to calculate the leakage probability; The hierarchical response decision unit has a built-in hierarchical decision tree, based on the leakage probability P. leak The value triggers different levels of response actions: If P leak If the value is ≥0.85, it is determined to be an L5 level emergency leak. The shutdown command is immediately output to the first intelligent shut-off valve (Vw), and at the same time, an audible and visual alarm is triggered and information is pushed to all emergency contacts. If 0.70 ≤ P leak If the value is less than 0.85, it is judged as a Level 4 serious leak, a local alarm is output, and a shutdown command is automatically output if no cancellation command is received from the user within the preset waiting time. If 0.50 ≤ P leak If the value is less than 0.70, it is determined to be a Level 3 warning, and a strong reminder will be output to the APP and the interactive terminal will be controlled to display a flashing yellow light. If 0.30≤P leak If the value is less than 0.50, it is determined to be an L2 level warning, and an APP notification message is output. If P leak If the value is less than 0.30, it is judged as a normal or slight fluctuation at level L0 / L1, and only the data is recorded.

3. The intelligent home water and heating safety protection system based on dual AI engine collaboration as described in claim 2, characterized in that, The device fingerprint recognition module includes: The template library construction unit is used to record the water usage cycle for each water-using device and extract a 12-dimensional feature vector from the flow-time curve of each water usage cycle. The 12-dimensional feature vector includes at least: start-up slope, steady-state mean, stop slope, water usage duration, total water consumption, peak flow, and energy proportion of each frequency band after wavelet packet decomposition, forming a feature cluster template library for the device. A real-time feature extraction unit is used to extract the same 12-dimensional feature vector after a water usage event is detected; The matching and recognition unit is used to calculate the distance between the real-time feature vector and all feature vectors in the template library using the dynamic time warping algorithm, and select the K closest samples for voting, outputting the device type and its confidence level C; The feedback correction unit is used to feed back the confidence level C to the leakage classification response module in real time to adjust the F3 characteristic value, thereby correcting the leakage probability P. leak .

4. The intelligent home water and heat safety protection system based on dual AI engine collaboration as described in claim 1, characterized in that, The intelligent logic module for going out / returning home includes: The off-duty determination unit is used to determine the duration of no water usage, T. dry Furthermore, smart home signals are used to comprehensively determine whether a user is away from home, including a threshold T for the duration of no water usage. adaptive Dynamically adjusted based on users' historical travel habits; When the user is out of the house, the armed unit outputs a command to close the smart shut-off valve (V) and records the initial pressure P0 at the moment of closure. The homecoming identification unit is used to monitor the pressure change curve when the valve is closed. When a pressure drop event is detected, a hidden Markov model is used to classify the pressure drop pattern. The input of the hidden Markov model is the pressure drop rate sequence, and the output is the probability of "turning on the tap when returning home" or "slow leakage in the pipe". If the probability of turning on the tap upon returning home is greater than 0.8, the automatic valve opening unit will output a command to open the intelligent water shut-off valve (V); otherwise, it will determine that there is a leak, keep the valve closed, and trigger an alarm.

5. The intelligent home water and heating safety protection system based on dual AI engine collaboration as described in claim 1, characterized in that, The multi-evidence fusion leakage detection module is constructed based on Dempster-Shafer evidence theory or weighted Bayesian networks, and includes: The first evidence source calculation unit calculates the basic probability allocation m1 based on the evidence of flow imbalance, m1 = f(ΔF, duration), where ΔF is the difference between supply and return water flow and duration is the duration. The second evidence source calculation unit calculates the basic probability allocation m2 based on the pressure differential attenuation evidence, where m2 = g(d(ΔP)). sys ) / dt), where ΔP sys Let d(ΔP) be the system pressure difference. sys ) / dt is the rate of change of pressure difference; The third evidence source calculation unit calculates the basic probability allocation m3 based on the temperature field anomaly evidence, where m3 = h(ΔT). actual , ΔT expected ), where ΔT actual To measure the temperature difference between the supply and return water, ΔT expected This is the theoretical temperature difference based on a thermodynamic model; The fourth evidence source calculation unit calculates the basic probability allocation m4 based on the ultrasonic acoustic fingerprint evidence. m4 is obtained by classifying the audio spectrograms collected by the first ultrasonic flow meter (U1) and the second ultrasonic flow meter (U2) based on the convolutional neural network CNN. The convolutional neural network CNN includes an input layer, two convolutional layers, a pooling layer, a fully connected layer and a Softmax output layer connected in sequence, and outputs the probability of the leakage category. The evidence synthesis unit is used to synthesize m1, m2, m3, and m4 using Dempster's combination rule, and to calculate the confidence intervals of the three hypotheses: leakage, normal, and uncertainty. The leakage detection unit is used to determine that a leakage has occurred when the confidence level of the "leakage" hypothesis is greater than a preset threshold of 0.75, and outputs a shut-off command to the first hot water intelligent shut-off valve (Vw) and the second hot water intelligent shut-off valve (Vh).

6. The intelligent home water and heating safety protection system based on dual AI engine collaboration as described in claim 1, characterized in that, The blockage diagnosis and quantification module includes: The filter clogging index calculation unit is used to calculate the filter clogging index I according to the following formula. filter : I filter = 0.5×(ΔP filter / ΔP0) + 0.3×(1-F actual / F0) + 0.2×T rend , Wherein, ΔP filter F represents the pressure difference across the filter, where ΔP0 is the baseline pressure difference under clean filter conditions. actual F0 is the actual flow rate, and T is the baseline flow rate under clean conditions. rend This is a pressure difference change trend factor based on historical data; The pipeline siltation index calculation unit is used to establish the system hydraulic model ΔP. sys = R×F 2 F is the volumetric flow rate of circulating water in the pipeline. The normal range of the system resistance coefficient R is determined by fitting operational data. min , R max ], and calculate the current drag coefficient R. current Corresponding siltation index I pipe = (R current - R min ) / (R max -R min ).

7. The intelligent home water and heating safety protection system based on dual AI engine collaboration as described in claim 1, characterized in that, The valve health prediction module includes: The current curve acquisition unit is used to acquire the motor current curve I(t) at a sampling rate of not less than 1kHz each time the intelligent shut-off valve is activated. The feature extraction unit is used to extract the following features from the motor current curve I(t): Time-domain characteristics: Peak current I peak Average current I mean Current variance σ 2 ; Shape characteristics: Dynamic time warping (DTW) distance between the current profile and a pre-stored standard health current profile; Frequency domain characteristics: the main frequency components and their amplitudes extracted after performing a fast Fourier transform on I(t); The health score calculation unit has a built-in support vector regression (SVR) model. It takes the extracted features as input and outputs a valve health score S from 0 to 100. valve ; The remaining useful life prediction unit records the historical sequence of health scores and uses exponential smoothing or linear regression extrapolation to predict the time required for the health score to drop to a preset failure threshold, which is the remaining useful life.

8. The intelligent home water and heating safety protection system based on dual AI engine collaboration as described in claim 1, characterized in that, The collaborative arbitration module integrates a scenario strategy library, which includes at least the following: The water supply outage mode strategy is triggered when the water supply AI engine module detects that the continuous period of no water usage exceeds T. adaptive Furthermore, the smart home system prompts the user to leave home; the action executed is: the collaborative arbitration module issues an instruction to the water supply AI engine module to close the water supply smart shut-off valve (V) and enter pressure monitoring; Summer heating mode strategy: When the heating AI engine module detects that the flow rates of the first ultrasonic flow meter (U1) and the second ultrasonic flow meter (U2) are both zero, the second ambient temperature sensor (Te2) has been above 25°C for a week and there is no heating demand, the heating AI engine switches to low-power monitoring mode to release NPU computing resources; the collaborative arbitration module dynamically allocates the released resources to the water supply AI engine module. Winter heating mode strategy: When the heating AI engine module detects that the flow rate of the first ultrasonic flow meter (U1) or the second ultrasonic flow meter (U2) is greater than zero, the inlet water temperature sensor (T1) is more than 2°C higher than the outlet water temperature sensor (T2), and there is a clear heating demand, the heating AI engine module exits the low-power monitoring mode and resumes full-function monitoring; the collaborative arbitration module reallocates computing resources.

9. The intelligent home water and heating safety protection system based on dual AI engine collaboration as described in claim 1, characterized in that, It also includes an integrated health management module, which calculates the system's total health score S_total based on the following formula: S total = 0.4×S water + 0.4×S heat + 0.2×S common , Among them, S water The water supply health score is calculated by the water supply AI engine module based on the number of historical leakage events, valve health scores, and pipeline aging coefficients. S heat The heating health score is calculated by the heating AI engine based on the blockage diagnosis index, thermal efficiency, and valve health score. S common The health score for public infrastructure is calculated based on a comprehensive assessment of backup battery health, storage media lifespan, and network communication quality. The integrated health management module also based on S total The system generates predictive maintenance suggestions or alarm messages, which are then displayed via an interactive terminal. The cloud-based intelligent platform, in conjunction with the dual AI engine collaborative control layer, implements a federated learning self-evolution mechanism, specifically including: The local update unit, deployed on edge devices, is used to incrementally train a lightweight model locally using the event data after detecting a local false alarm or missed alarm event, and generate model gradient update ΔW. An encrypted upload unit is used to encrypt ΔW and upload it to the cloud intelligent platform via the TLS 1.3 protocol; The cloud aggregation unit receives encrypted gradient updates from multiple edge devices, decrypts and aggregates them, and generates global model parameters W. global ; The secure distribution unit is used to control W. global Digital signature encryption is performed and distributed to all edge devices via OTA. After verifying the signature, the edge devices update their local models, thus completing the collective evolution. The hardware sensing and execution unit also includes a multi-level power protection unit, which includes a main power AC220V input, a 12V / 10Ah backup lithium battery and a supercapacitor. The supercapacitor is used to provide the energy required for the final shutdown action when both the main power and the backup battery fail, ensuring that an emergency shutdown can still be completed in extreme power outage conditions. The intelligent water supply shut-off valve (V), the first intelligent hot water shut-off valve (Vw), and the second intelligent hot water shut-off valve (Vh) are all integrated with manual mechanical operating mechanisms. In the event of a power outage, network outage, or abnormality in the electronic control system that prevents automatic shut-off, the valves can be opened and closed manually.

10. A protection method based on the home water and heat safety intelligent protection system based on dual AI engine collaboration as described in any one of claims 1 to 9, characterized in that, Includes the following steps: Step S1: The water supply AI engine module processes the sampling data on the water supply side in real time, executes leakage classification response, equipment fingerprint recognition, and intelligent logic for going out / coming home, and generates water supply side status data and control decisions. Step S2: The heating AI engine module performs fusion analysis on the multi-source heterogeneous data of the heating side, performs multi-evidence fusion leak detection, blockage diagnosis and quantification, and valve health prediction, and generates heating side status data and control decisions. Step S3: The collaborative arbitration module dynamically schedules the computing resources of the water supply AI engine module and the heating AI engine module according to the preset scenario strategy, arbitrates data conflicts and control command priorities, and outputs collaborative control commands. Step S4: The cloud-based intelligent platform receives encrypted model gradient data uploaded from multiple edge devices, performs federated learning to aggregate and update the global model, and sends the optimized global model parameters to the edge devices via OTA. Step S5: The integrated health management module calculates the total health score of the system based on the status data of the water supply side and the heating side, and generates predictive maintenance suggestions or alarm information, which are displayed through the interactive terminal.

Images