Smart home cooperative alarm method and system

Abstract

The application discloses a kind of intelligent home cooperative alarm method and system, belong to intelligent home security and multi-sensor information fusion technical field.The application is modulated to generate dynamic basic probability distribution with space priori and time period priori at edge gateway side, according to double threshold, it is divided into three branches of high, medium and low, the confidence branch is verified by event type perception across device, and secondary fusion in cloud is synthesized final confidence according to continuous cloud weight, then the severity score is obtained by multiplying final confidence, event type weight and family vulnerability coefficient, and multi-channel accessibility screening, utility selected signal, confirmation retransmission, local degradation and concurrent redundancy are graded and distributed according to the score.The application can inhibit context-type interference false alarm, give reliable fusion result when evidence conflicts, consider alarm real-time, accuracy and weak network availability, and realize reliable differentiated delivery with low communication overhead.

Description

Technical Field

[0001] This invention belongs to the field of smart home security and multi-sensor information fusion technology, and particularly relates to a smart home collaborative alarm method and system. Background Technology

[0002] Smart home security systems typically rely on multiple sensors deployed in the home environment to identify security incidents such as fires, gas leaks, falls, or unauthorized intrusions by detecting changes in abnormal physical quantities. However, existing smart home alarm methods still suffer from a number of technical shortcomings that are difficult to address in engineering practice, necessitating systematic improvements.

[0003] First, most existing solutions rely on fixed thresholds from a single sensor or simple AND / OR combinations of multiple sensors to trigger alarms. This approach fails to fully utilize the differences in sensor distribution within a home and prior knowledge of the probability of events occurring at different times. For example, cooking fumes in the kitchen can cause smoke sensors to briefly spike, moisture in the bathroom can be misinterpreted as a fire, and isolated infrared triggers from people at night are often misdiagnosed as intrusions due to pet activity or heat sources. Furthermore, the lack of semantic modulation tailored to sub-areas and time periods means that the same set of sensor readings is assigned the same confidence level in different scenarios, making it impossible to effectively suppress context-dependent interference and resulting in a high false alarm rate.

[0004] Secondly, when multiple sensors provide evidence simultaneously, existing multi-source information fusion methods are prone to producing counterintuitive fusion results under highly conflicting evidence conditions. The classic Dempster combination rule, when dealing with two contradictory pieces of evidence, has a normalization factor approaching zero, causing the fused low-probability events to be unreasonably amplified, resulting in the so-called "counterintuitive amplification" phenomenon. For example, kitchen fumes might cause a smoke sensor to indicate a fire tendency, while temperature and gas sensors strongly support a normal state; simply applying Dempster combination might erroneously output a high confidence level for a fire. While some existing improvements, such as arithmetic mean pre-fusion, can partially alleviate conflict, they lack a confidence weighting mechanism based on evidence distance, failing to quantitatively distinguish reliable from unreliable evidence based on group consensus. Therefore, they struggle to provide robust decisions under complex conditions with heterogeneous sensors and contradictory readings.

[0005] Furthermore, existing smart home alarm systems often exhibit two extremes in their edge computing and cloud-based collaborative models. One approach uploads all sensor data to the cloud for centralized fusion and decision-making. While this leverages cloud computing resources to improve recognition accuracy, it heavily relies on a stable network connection. Weak or outages, or cloud response timeouts, prevent real-time alarms, significantly increasing end-to-end latency and failing to meet the millisecond-level response requirements for emergencies like fires. The other approach completes the decision-making entirely at the home edge gateway. While avoiding network dependence, it is limited by local computing power and single-perspective on-site evidence, resulting in insufficient ability to identify complex events or critical states. This can lead to misclassifying real events as suspicious events and failing to trigger an alarm, or misclassifying critical interference as real events and sending false alarms. Both models lack a tiered collaborative mechanism that maintains real-time performance while utilizing cloud-based secondary verification, resulting in a trade-off between real-time alarm performance and accuracy.

[0006] Furthermore, existing alarm systems lack differentiated verification strategies for different event types during the event confirmation phase. When the edge judgment is within the middle confidence interval, the system often indiscriminately requests supplementary data from all available sensors, or simply repeats sampling from the same type of sensors, failing to customize a supplementary device list based on the specific physical diagnostic logic of events such as fires, falls, and intrusions. For example, the key confirmatory evidence for a fall event should be whether the person under supervision has lost their responsiveness and whether the millimeter-wave radar shows a sudden drop, while the key confirmatory evidence for a fire event is the concurrent increase in smoke, temperature, and gas concentration in adjacent sub-areas. Existing solutions lack this cross-device verification design for event type awareness, resulting in insufficient effectiveness of secondary confirmation and potentially increasing unnecessary privacy risks by requesting irrelevant audio and video data.

[0007] At the alarm distribution level, existing technologies typically rely on a single wireless channel (such as home Wi-Fi) to push alarm messages to a fixed recipient (such as the homeowner's mobile phone). This single-channel, single-receiver mode is highly susceptible to alarm failures due to channel congestion, router power outages, broadband interruptions, or the recipient being offline. Even if some solutions employ primary and backup channel switching, it is merely a simple fault-based replacement, failing to dynamically calculate utility scores based on the real-time latency, transmission success rate, and bandwidth of each channel, and further failing to incorporate the severity of the event and the vulnerability of family members (such as the number of elderly, children, or patients with chronic diseases) into the quantitative criteria for graded triggering. Therefore, low-severity events may receive the same push intensity and channel strategy as high-severity events, causing disruption to normal life, while high-severity events face the risk of delivery failure in weak network environments due to improper channel selection, lacking multi-layered delivery guarantees from momentary congestion to prolonged network outages.

[0008] In summary, existing smart home alarm methods have significant technical shortcomings in areas such as interference suppression, conflict evidence fusion, real-time edge-cloud collaboration, differentiated cross-device verification, and hierarchical multi-channel distribution, making it difficult to simultaneously meet the engineering requirements of low false alarms, low false alarms, high real-time performance, and high reliable delivery. Therefore, it is necessary to propose a smart home collaborative alarm method and system that can systematically solve the above problems. Summary of the Invention

[0009] This invention provides a smart home collaborative alarm method and system, aiming to solve the problems in the prior art such as false alarms caused by contextual interference, unreliable fusion results when evidence conflicts, difficulty in balancing alarm real-time performance and accuracy, and low alarm delivery rate in weak network environments.

[0010] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0011] S1: Configuration Phase. This phase establishes a configuration identification framework, several sub-regional divisions, several time-period divisions, and a configuration baseline for use in subsequent steps on the edge gateway side.

[0012] Specifically, S11: The physical space of the home is pre-divided into sub-areas such as kitchen, living room, bedroom, bathroom, balcony and corridor, and the day is divided into time periods such as late night, early morning, morning, noon, afternoon and evening.

[0013] S12: Establish an identification framework These correspond to the assumptions of fire, gas leak, fall, intrusion, and normal events, respectively.

[0014] S13: Deploy multiple sensors in each sub-area and on the person being monitored, enabling each sensor to communicate with the edge gateway.

[0015] S14: Calibrate the normalized decision threshold and membership function steepness coefficient for each sensor, configure a spatial prior probability table for each sub-region, and configure a time-period prior probability table for each time period.

[0016] S15: The above parameters, as well as edge high threshold, edge low threshold, cloud confirmation threshold, event type perception verification strategy table, virtual sensor mapping table, event type weight table, home vulnerability weight, hierarchical receiver table, receiver-channel reachability matrix, channel utility scoring parameters, and concurrent redundancy threshold, are distributed to the edge gateway in the form of a configuration package and solidified to form a configuration baseline that can be updated online.

[0017] S2: Data Acquisition and Normalization. The edge gateway periodically acquires raw data from each sensor, and obtains the normalized feature values ​​of each sensor in the current frame through feature extraction and amplitude normalization.

[0018] Specifically, S21: The edge gateway uses a sampling period Poll each sensor, read the raw data from each sensor, and extract the sensor feature values.

[0019] S22: For each sensor feature value, perform linear normalization according to its calibrated lower and upper limits to obtain the normalized feature value. And truncate it to the interval [0,1].

[0020] S23: Perform a length of [length] on the normalized eigenvalues. The moving average filter is applied, and the normalized feature values ​​of each sensor in the current frame and the current sampling time are written into the current frame data buffer.

[0021] S3: Dynamic Basic Probability Assignment Generation. Based on the normalized eigenvalues ​​and the decision parameters of each sensor for each event hypothesis in the configuration baseline, the support of each sensor for each event hypothesis is calculated. The spatial prior probability of the sub-region where the sensor is located and the temporal prior probability of the current time period are used as multiplicative modulation factors to normalize and obtain the dynamic basic probability assignment of the sensor in the current frame.

[0022] Specifically, S31: Determine the time period to which it belongs based on the current sampling time, and determine the sub-region to which each sensor belongs based on the current sub-region to which it belongs.

[0023] S32: For each sensor and each event, assume that the support is calculated using the Sigmoid function. .

[0024] S33: Spatial prior probability of query sub-regions in self-configured baselines Prior probability of time period The dynamic basic probability allocation is calculated using the following formula: ,in This is the normalization factor.

[0025] S34: The results proportionally Scaling, and identifying the entire set of elements in the frame. Assigning residual uncertain mass , making Output the complete dynamic basic probability distribution.

[0026] S4: Conflict Adaptive Suppression and Evidence Fusion. Based on the Jaccard matrix on the power set of the recognition framework, the Jousselme evidence distance between each dynamic basic probability assignment is calculated pairwise. The credibility of each piece of evidence is calculated inversely based on the distance matrix and normalized into weights. The weighted sum is used to obtain representative evidence. The representative evidence is then continuously combined according to the Dempster combination rule to obtain fused evidence and marginal confidence. The abnormal event hypothesis with the maximum value on the fused evidence is recorded as the marginal decision category.

[0027] Specifically, S41: Jaccard matrix on the power set focal set of the identification framework is permanently stored on the edge gateway side. Its elements .

[0028] S42: Count the number of valid pieces of evidence related to fusion. ;when When this piece of evidence is used as representative evidence, proceed directly to S46; when Calculate the Jousselme evidence distance between any two pieces of evidence. .

[0029] S43: Calculate the credibility of each piece of evidence using the following formula. .

[0030] S44: Normalize credibility into weights If the denominator is zero, then let each .

[0031] S45: Press The weighted summation yields representative evidence.

[0032] S46: When The evidence will be combined with itself in succession according to Dempster's combination rule. This time, evidence of fusion was obtained. ;when Direct order Record edge confidence. Marginal Judgment Category An abnormal event hypothesis is used to obtain the maximum value.

[0033] S5: Edge dual-threshold classification. The edge confidence level is compared with the high and low edge thresholds in the configuration baseline, and the current decision is classified into one of the following branches: high confidence direct alarm, medium confidence cross-device verification, or low confidence suspicious event recording.

[0034] Specifically, S51: Read edge high threshold With low threshold at the edge ,and .

[0035] S52: When At that time, it will be classified into the high-confidence direct alarm branch. and Write it to the queue of events to be alerted.

[0036] S53: When At that time, it is classified into the cross-device verification branch triggered by the medium confidence.

[0037] S54: When When this happens, the event is classified as a low-confidence suspicious event and the event information is written to the suspicious event log without triggering an alarm.

[0038] S6: Cross-device verification and cloud-based secondary fusion of the centrally trusted branch. For the centrally trusted branch, the event type perception verification strategy table is queried according to the edge decision category. A differentiated supplementary sampling request is initiated to the supplementary device. The supplementary evidence is converted into virtual sensor evidence and then sent to the cloud collaboration platform via an encrypted link for secondary evidence fusion. The cloud collaboration platform synchronously outputs the cloud confidence score and cloud weight coefficient. The edge gateway weights the edge confidence score and the cloud confidence score according to the cloud weight coefficient to synthesize the final confidence score, and determines the confirmation status of this event based on the cloud confirmation threshold in the configuration baseline.

[0039] Specifically, S61: For the high-confidence direct alarm branch, set the final confidence level... And then switch to S7.

[0040] S62: For the central confidence branch, based on Retrieve the verification window from the event type-aware verification strategy table List of Supplementary Equipment The strategy table assigns differentiated supplementary equipment to each of the following events: fire, gas leak, fall, and intrusion.

[0041] S63: The edge gateway initiates a supplementary sampling request to the supplementary device and collects supplementary evidence within the verification window; it converts the non-original reading type supplementary evidence into virtual sensor evidence according to the virtual sensor mapping table; it merges the original evidence and virtual sensor evidence to form an enhanced evidence set, which is then sent to the cloud via an encrypted link.

[0042] S64: The cloud-based collaborative platform executes S3 and S4 again on the enhanced evidence set to obtain cloud-fused evidence. With cloud confidence Simultaneously, a cloud weighting coefficient is obtained based on the completeness of supplementary evidence returned, cloud response latency, and link quality. And simultaneously distribute it to the edge gateway; when the cloud is unreachable, the edge gateway will Forced to set to 0.

[0043] S65: Edge gateways synthesize final confidence levels using the following formula: .

[0044] S66: Will Confirm thresholds with the cloud Comparison: If If the event is confirmed, it is transferred to S7; otherwise, it is marked as an event pending review and stored on the cloud without triggering an alarm.

[0045] S7: Branch Convergence and Event Severity Scoring. For high-confidence branches, the marginal decision category is used as the confirmed event type, and the marginal confidence level is used as the final confidence level. For events confirmed via the cloud in medium-confidence branches, the cloud-confirmed event type is used as the confirmed event type. The event type weight is queried based on the confirmed event type, and the family vulnerability coefficient is calculated based on the vulnerable population within the family. The event severity score is obtained by multiplying the final confidence level, the event type weight, and the family vulnerability coefficient.

[0046] Specifically, S71: Branch decision and confirmation event type selection: high confidence branches are... As a confirmation event type The cloud-confirmed Zhongzhixin branch The abnormal event that takes the maximum value above is assumed to be... Low-confidence branches and events that fail to pass confirmation terminate this step and skip S8.

[0047] S72: Self-configured baseline read event type weight table, by Query event type weight .

[0048] S73: Self-configuring baseline reading of elderly population Number of children Number of patients with chronic diseases and the corresponding vulnerability weights , , Calculate the family vulnerability coefficient .

[0049] S74: Calculate the event severity score using the following formula: .

[0050] S8: Multi-channel hierarchical distribution and ACK retransmission. Based on event severity score, confirmed event type, and hierarchical receiver table, activate the corresponding hierarchical receiver; filter the reachable and available wireless channels for each activated receiver according to the receiver-channel reachability matrix; calculate the channel utility score based on latency, transmission success rate, and bandwidth; select the channel in descending order of score to send the alarm message and start the ACK waiting timer; for transmissions that time out and are not acknowledged, switch to the second-best channel for retransmission; if no second-best channel is available, handle as a fallback; for events with an event severity score not lower than the concurrent redundancy threshold, simultaneously send the same alarm message on the two most reachable and available channels with the highest utility scores.

[0051] Specifically, S81: Self-configuring baseline reads the hierarchical receiver table, for each receiver... Determine the activation flag when Otherwise, the value is 0. Tiered receivers include the homeowner's smart terminal, the homeowner's voice / sound and light alarm, the property management / neighbor terminal, and the emergency response terminal.

[0052] S82: Read the receiver-channel reachability matrix, for each active receiver and each radio channel. (Including at least two of Wi-Fi, 4G / 5G, NB-IoT, and LoRa) Filter for reachable and available channels; for each reachable and available channel, start from the nearest The moving average of the RTT (Round Time) is calculated for each transmission. Success rate index weighted moving average and current bandwidth estimates Calculate the channel utility score .

[0053] S83: Press The system selects a channel in descending order, sends an alarm message to the active receiver, and starts an ACK waiting timer. If no ACK is received within the timeout period, it switches to the second-best channel for retransmission. If no second-best channel is available, the data is added to the retransmission queue, and a new channel is selected according to the backoff period. If all reachable and available channels fail consecutively for more than [time period missing], the system will continue to operate. At this time, the local voice / sound and light alarm will be triggered to perform a local downgrade alarm.

[0054] S84: Update the weighted moving average of the transmission success rate index after each transmission is completed. .

[0055] S85: When (Concurrent redundancy threshold) When each activated receiver is in its The same alarm message is sent simultaneously on the top two reachable and available channels, and delivery is considered complete once either channel receives an ACK.

[0056] This invention also provides a smart home collaborative alarm system implementing the above method, comprising: a perception layer, an edge gateway, a cloud collaborative platform, a multi-channel access module, and hierarchical receivers. The perception layer includes multiple types of sensors deployed in various sub-areas and on the personal devices of the monitored individuals. The edge gateway includes, in sequence, a data acquisition and feature normalization module, a dynamic basic probability allocation generation module, a conflict adaptive suppression and evidence fusion module, an edge dual-threshold decision module, a cross-device verification request module, a final confidence synthesis module, a severity scoring module, a multi-channel selection and retransmission control module, and a configuration storage module. The cloud collaborative platform includes a two-level DS fusion service module, a cloud weight generation module, an evidence chain disk storage module, and a remote configuration distribution module. The multi-channel access module includes at least two of the following independent transceiver units: Wi-Fi, 4G / 5G, NB-IoT, and LoRa. The hierarchical receivers include the homeowner's smart terminal, the homeowner's voice / sound and light alarm, the property management / neighbor terminal, and an emergency linkage terminal. The modules interact and control each other according to the described method steps to achieve smart home collaborative alarm functionality.

[0057] The beneficial effects of this invention are as follows:

[0058] This invention introduces spatial and temporal prior probabilities as multiplicative modulation factors in the generation of dynamic basic probability allocation. It embeds the historical occurrence tendencies of each sub-region and each time period into the evidence construction, allowing the same set of sensor readings to be assigned different evidence distributions in different sub-regions and time periods. Typical context-dependent interferences in home scenarios, such as cooking fumes in the kitchen, water vapor in the bathroom, and isolated human infrared triggering at night, are naturally suppressed, reducing false alarms caused by scene context without sacrificing the ability to detect real events. Verification of the overall decision quality shows that all embodiments using this spatiotemporal prior modulation method achieve higher detection capabilities at the same false alarm level, and their overall decision quality is superior to the comparison method without this modulation.

[0059] This invention calculates the Jousselme distance between pieces of evidence in the evidence fusion stage using the Jaccard matrix on the power set of the identification frame as the geometric basis. Based on the distance matrix, the credibility of each piece of evidence is calculated and normalized into a weight. This weighted average first yields representative evidence that reflects the consensus of the group, and then the Dempster combination rule is used to continuously self-synthesize this representative evidence. This invention first disperses the conflicts between pieces of evidence into the construction of representative evidence, and then performs combination on the already converging representative evidence, avoiding counterintuitive amplification phenomena. This ensures that the fusion result remains consistent with common sense even under conditions of heterogeneous sensors and contradictory readings. This invention also provides a fallback solution for two extreme cases: only one valid piece of evidence and all pieces of evidence being completely contradictory pairwise. This ensures that the fusion process does not degrade abnormally under any evidence scale. Verification results on the overall decision quality show that the implementation using this conflict adaptive suppression fusion method significantly outperforms the comparison methods of directly applying classical combination rules and simple arithmetic mean.

[0060] This invention discretizes each edge decision into three branches: high-confidence direct alarm, medium-confidence cross-device verification triggered by high and low edge thresholds, and low-confidence suspicious event recording. This decouples the real-time performance and accuracy of alarms. High-confidence events receive immediate responses without waiting for cloud access. Medium-confidence events undergo cross-device supplementary sampling based on event type awareness and secondary fusion with the cloud before confirmation. Low-confidence events are only recorded locally for auditing purposes. For medium-confidence events, this invention selects supplementary devices based on event type and incorporates supplementary evidence into the same fusion thread after virtualizing the supplementary evidence into a single sensor. For supplementary devices involving audio and video, only existence results or motion features are collected without uploading the original audio and video, improving the effectiveness of secondary verification while protecting family privacy. This invention employs a convex combination method that weights the edge confidence and cloud confidence using continuous cloud weighting coefficients when synthesizing the final confidence level. Compared to simply taking the larger value, this method raises the confidence level above the confirmation threshold to correctly confirm the real event when there is strong support in the cloud, and lowers the confidence level below the confirmation threshold to reject false high confidence on the edge side when there is strong rejection in the cloud. It possesses true bidirectional adjustment and secondary rejection capabilities. Furthermore, when the cloud is unreachable, the cloud weighting coefficient degenerates to zero, causing the final confidence level synthesis to automatically revert to a pure edge decision, ensuring the system's availability under weak network and network outage conditions. Verification of end-to-end latency and the final confidence level synthesis method shows that the three-level branches retain an immediate high-confidence response channel, and the weighted synthesis can provide a decision consistent with the true value under various typical operating conditions. This corresponds one-to-one with the tiered decision, cross-device verification, and weighted synthesis techniques of this invention, demonstrating that the above methods can indeed bring the beneficial effects of timely response, accurate decision, and availability under weak network conditions.

[0061] This invention multiplies the final confidence level, event type weight, and family vulnerability coefficient to obtain an event severity score. It then constructs a family vulnerability coefficient based on the number of elderly people, children, and patients with chronic diseases within the family, quantifying the family context and incorporating it into alarm distribution decisions. This automatically amplifies event severity when vulnerable individuals are present in the family, reflecting a people-centered, differentiated protection approach. In the distribution phase, this invention filters reachable and usable channels based on the receiver-channel reachability matrix. It calculates channel utility scores based on latency, transmission success rate, and bandwidth, selecting the best channel accordingly. For unacknowledged transmissions that time out, it switches to a suboptimal channel for retransmission. When all channels fail consecutively, it triggers local degradation of the user's voice and audio-visual systems. For high-severity events, it performs concurrent redundant transmissions on the channels with the highest utility, thus constructing a multi-layered delivery guarantee from momentary congestion to long-term network outages. Because multiple wireless channels are approximately independent at the physical and link layers, their concurrent reachability is superior to any single channel. Therefore, this invention can maintain a high alarm delivery success rate with lower communication overhead under medium-to-high link packet loss conditions. The verification results of multi-channel delivery performance show that the embodiment of the multi-channel hierarchical distribution mechanism of the present invention can still maintain a high delivery rate and lower communication overhead than the single-channel scheme under weak network and high packet loss conditions. Attached Figure Description

[0062] Figure 1 The charts show the overall decision quality comparison between Examples 1 to 4 and Comparative Examples 1 and 2. (a) is a comparison chart of ROC curves between Examples 1 and Comparative Examples 1 and 2, and (b) is a bar chart comparing the false alarm rate and false negative rate of Examples 1 to 4 and Comparative Examples 1 and 2 at their respective operating points.

[0063] Figure 2 The diagram shows the end-to-end delay comparison of Examples 1, 4 and Comparative Example 4. (a) is a comparison of the cumulative probability distribution curves of the end-to-end delay of the three methods, and (b) is a comparison of the P50, P95 and P99 delay bars of the three methods.

[0064] Figure 3 The chart shows a comparison of the final confidence level synthesis methods of Example 1 and Comparative Example 3. (a) is a bar chart comparing the final confidence level of Example 1 (weighted synthesis) and Comparative Example 3 (maximum value synthesis) under four typical working conditions. (b) is a bar chart comparing the decision accuracy of Example 1 and Comparative Example 3 under a mixture of four working conditions.

[0065] Figure 4 The following is a comparison chart of the multi-channel delivery performance of Example 1, Example 4 and Comparative Example 4. (a) is a curve comparison chart of the alarm delivery success rate of the three methods as a function of the link packet loss rate, and (b) is a bar chart comparison chart of the 24-hour cumulative uplink bytes of the three methods.

[0066] Figure 5 This is a structural diagram of the smart home collaborative alarm system of the present invention. Detailed Implementation

[0067] The present application will now be described in detail through exemplary embodiments. However, it should be understood that, without further description, elements, structures, and features in one embodiment may be advantageously incorporated into other embodiments.

[0068] This invention first completes the featureization and normalization of the perceived data, the generation of dynamic basic probability allocation, and the Dempster-Shafer evidence fusion under conflict adaptive suppression at the edge gateway to obtain the edge confidence. Secondly, the current decision is classified into three branches according to the dual threshold, and cross-device supplementary sampling and cloud-based secondary fusion are initiated for the middle confidence branch to perceive the event type. The edge gateway weights the edge confidence and the cloud confidence according to the cloud weight coefficient to obtain the final confidence. Then, the event severity score is obtained by multiplying the final confidence, the event type weight, and the family vulnerability coefficient. The event severity score, the receiver association weight, and the trigger threshold jointly determine which graded receivers are activated. Finally, the receiver-channel reachability matrix and the channel utility score are used for signal selection and ACK retransmission, and concurrent redundant transmission is enabled when the severity is high.

[0069] Step S1 is the deployment phase and the solidification configuration stage after initial power-on. Its sub-steps S11 to S15 jointly complete the distribution and disk write of all static parameters. In step S11, the home physical space is divided into functional areas: kitchen, living room, bedroom, bathroom, balcony, and corridor. The day is also divided into nighttime, early morning, morning, noon, afternoon, and evening based on event distribution characteristics. This division reflects both the distribution characteristics of sensors such as smoke, gas, acceleration, and human infrared sensors in the physical space, and the differences in the probability of the same type of event occurring at different times, serving as the semantic prior for the subsequent dynamic basic probability allocation in step S3. Step S12 establishes the recognition framework based on Dempster-Shafer evidence theory. ,in Represents the recognition framework; Indicates the assumption of a fire event; This indicates a hypothesis about a gas leak event; This indicates the hypothesis of a fall event; Indicates the hypothesis of an intrusion event; This represents the normal event hypothesis; the elements of the identification framework are mutually exclusive to form a complete event space, and the quality of evidence is distributed across all its subsets, i.e., the power set.

[0070] In step S13, photoelectric smoke sensors, temperature sensors, carbon monoxide sensors, and combustible gas sensors are deployed in the kitchen according to typical event types in the sub-areas; millimeter-wave radar and emergency buttons are deployed in the bathroom and bedroom; human infrared sensors are deployed in the living room and corridor; and door and window magnetic sensors and combustible gas sensors are deployed at the balcony and door / window locations. These sensors are also connected to the wearable accelerometer worn by the ward. The wearable sensors are not fixed to a specific sub-area; their sub-area in step S31 will be dynamically determined based on the ward's most recent location result or the linkage result with a nearby fixed sensor. In step S14, each sensor... Assume for each event Define a set of decision parameters, including a normalized decision threshold. steepness coefficient of membership function ,in Indicates the sensor number; Indicates the event hypothesis number; Indicates sensor Hypothesis of the event The normalized threshold is determined and its value is taken from . interval; Indicates sensor Hypothesis of the event The membership function steepness coefficient is set and can take positive or negative values. Positive values ​​indicate that the hypothesis of the event is supported when the sensor reading increases, while negative values ​​indicate that the hypothesis of the event is not supported when the sensor reading increases. In the same step, each sub-region is also... Configure its assumptions for each event Spatial prior probability Then, a spatial prior probability table is generated by normalizing the rows, for each time period. Configure its assumptions for each event Prior probability of time period And normalize by row to form a prior probability table for each time period, where Indicates the sub-region number; This indicates the time period number. In step S15, the normalized judgment threshold, membership function steepness coefficient, spatial prior probability table, time period prior probability table, edge high threshold, edge low threshold, cloud confirmation threshold, event type perception verification strategy table, virtual sensor mapping table, event type weight table, home vulnerability weight, hierarchical receiver table, receiver-channel reachability matrix, channel utility scoring parameters, and concurrent redundancy threshold are distributed to the edge gateway in the form of a configuration package via the cloud collaboration platform and written to its readable and writable storage area, forming an online updatable configuration baseline. The channel utility scoring parameters further include a preset lower delay limit, a preset upper allowable delay limit, reference bandwidth, delay term weight, success rate term weight, bandwidth term weight, channel statistical window length, initial RTT, initial transmission success rate, initial bandwidth estimate, ACK waiting time, smoothing coefficient of exponentially weighted moving average, and the maximum number of consecutive failures allowed on the same channel. After step S1 is completed, the edge gateway has the complete operational conditions required to complete edge decision and alarm without cloud access.

[0071] Step S2 is the periodic real-time data acquisition and preliminary shaping stage, completed by its sub-steps S21 to S23. In step S21, the edge gateway uses a sampling period... Poll each sensor and perform a poll on each sensor Read its raw data, where Indicates the sampling period (ms). The preferred value range is 50ms to 200ms; for amplitude-type sensors, i.e., analog sensors such as temperature, smoke, carbon monoxide concentration, and combustible gas concentration, the raw data is directly taken as the scalar raw reading; for multidimensional or sequential sensors, i.e., triaxial acceleration, millimeter-wave radar altitude or attitude sequences, and video frame sequences, sensor feature values ​​are extracted from the raw data according to a preset feature extraction algorithm. For example, the composite magnitude of triaxial acceleration, the magnitude of sudden drops in radar altitude, and the binary representation of target presence in the video are used as features. Indicates sensor The characteristic value; for discrete quantities such as magnetic opening and closing of doors and windows, and opening and closing of gas valves, the characteristic value The state mapping value is directly taken. The two-stage abstract design of first extracting features from the raw data and then normalizing it is to enable subsequent steps to use a unified normalization framework for multi-dimensional sequences, state quantities, and amplitude quantities with completely different structures, avoiding the need to design decision logic separately for each type of sensor. This unified interface is also a prerequisite for virtualizing the non-raw reading evidence of the supplementary device into a virtual sensor in the subsequent S63 step.

[0072] In step S22, for each sensor characteristic value According to its calibrated lower limit of characteristic value With upper limit of eigenvalues Linear normalization is performed to obtain normalized eigenvalues. and normalized eigenvalues Cut off to Interval normalization is performed using the following formula:

[0073] ;

[0074] in Indicates sensor Normalized eigenvalues; Same as before; Indicates sensor The lower bound of eigenvalues ​​and Same unit; Indicates sensor upper bound of eigenvalues ​​and Same unit. The normalized eigenvalues ​​are processed in step S23. Execution length is The moving average filter is used to suppress glitches and single-point abrupt noise, where Indicates the length (frames) of the moving average filter window. The preferred value range is an odd number between 3 and 7; after filtering, the normalized feature values ​​of each sensor in the current frame and the current sampling time are written into the current frame data buffer of the edge gateway for use in step S3.

[0075] Step S3 is the generation stage of the dynamic basic probability assignment, completed by its sub-steps S31 to S34. In step S31, the time period to which the sample belongs is determined by looking up a table based on the current sampling time written in step S23. Based on each sensor The current sub-region is determined by looking up the table. The sub-region to which the wearable sensor belongs is determined according to the positioning and linkage results described in step S13. In step S32, for each sensor... With each event assumption The support for the hypothesis about this event is calculated using a membership function in the form of a sigmoid function, as follows:

[0076] ;

[0077] in Indicates sensor In normalized eigenvalues The following assumptions about events Support and take the value of interval; Represents the natural exponential function; , , Same as before. The Sigmoid function is the standard logistic function, which in... equal The transition value is 0.5. The absolute value controls the steepness of the transition zone. The sign controls the monotonic direction, when When taking a positive value, the support varies. An increase in the sensor reading corresponds to an abnormal reading, such as smoke in a fire. When the value is negative, the support varies. Increasing the value of a sensor increases the corresponding reading, which in turn supports normal sensor-event pairs, such as smoke for normal events. The design, which allows for both positive and negative directions, avoids the logical reversal that a higher abnormal reading equates to a higher support for normal events.

[0078] In step S33, the sub-region is queried from the self-configured baseline. Spatial prior probability With the time period Prior probability of time period And calculate the sensor according to the following formula Assuming events in the current frame Dynamic basic probability allocation:

[0079] ;

[0080] in Indicates sensor Hypothesis of the event The dynamic basic probability allocation; Indicates sensor The normalization factor and its value ; This indicates the summation of hypotheses for each event within the recognition framework; other symbols remain the same as before. Here, spatial prior probability and temporal prior probability are used as multiplicative modulation factors. The historical occurrence tendency of such events in this sub-region and the occurrence tendency of such events in this time period are embedded into the construction of the dynamic basic probability allocation. This modulation causes the same set of sensor readings to produce different evidence distributions in different sub-regions and at different times, naturally suppressing context-dependent interferences common in home scenarios such as cooking fumes in the kitchen, water vapor in the bathroom, and isolated human infrared triggering at night. In step S34, the obtained... proportionally Scaling and recognizing the entire set of elements of the frame Assigning residual uncertain mass To meet ,in Represents the residual uncertain mass used to avoid the zero-mass trap and preferably takes the value of interval; Indicates sensor In identifying all elements of the frame set The basic probability allocation on the [database / database]. The zero-mass trap is a known problem in the classic Dempster combinatorial rule, namely, when two pieces of evidence simultaneously give a mass of 0 at a focal point, that focal point will remain 0 forever after fusion, losing its update space and affecting the entire set. Retaining a small amount of uncertain quality can effectively avoid this problem.

[0081] Step S4 is a fusion stage based on Jousselme evidence distance-based conflict adaptive suppression and Dempster combined fusion, completed by its sub-steps S41 to S46. In step S41, the edge gateway side solidifies the storage identification framework. Jaccard matrix on the power set focal set Its elements take values ​​according to the following formula:

[0082] ;

[0083] in Represents the Jaccard matrix; This represents the focal element corresponding to a row in the matrix; This represents the focal element corresponding to a column of the matrix; Indicates the number of elements in the set; This represents the intersection operation of sets; This represents the union operation of sets. The Jaccard matrix characterizes the similarity between any two focal elements within the recognition framework at the element level and is the geometric basis of the Jousselme evidence distance.

[0084] In step S42, the number of valid pieces of evidence contributing to the fusion in this judgment is first counted. ,when When the value equals 1, this piece of evidence is directly used as the representative evidence. Then jump to step S46, when For any two lines, when the number is not less than 2 and Calculate the Jousselme evidence distance using the following formula:

[0085] ;

[0086] in Indicates sensor With sensors Jousselme distance between pieces of evidence; Indicates the number of valid pieces of evidence contributing to the integration; Indicates and A different sensor number; and These represent the sensors. With sensors The dynamic basic probability assignment is used as a column vector on the focal element set; superscript This represents the matrix transpose operation; Represents the arithmetic square root operation; Same as before. The Jousselme evidence distance is a strict distance metric and takes values ​​of . The interval is set to 0 when two pieces of evidence are completely consistent and 1 when the two pieces of evidence do not intersect. Compared with indicators such as Tessem distance and Hamming distance, which only consider the difference between focal elements with the same name, Jousselme evidence distance incorporates the subset overlap relationship between focal elements into the distance calculation through the Jaccard matrix. It can correctly reflect the fact that two pieces of evidence do not conflict when the focal element names are different but the internal elements highly overlap, which is especially suitable in multi-event hypothetical scenarios.

[0087] In step S43, the credibility of each piece of evidence is calculated using the following formula:

[0088] ;

[0089] in Indicates sensor The credibility of the evidence; Indicates to and Different sensor numbers Summation; all other symbols are the same as before. The geometric meaning of this expression is to sum the values ​​of the sensor. The credibility of an evidence is determined by its average distance-approximation relative to other evidence. Evidence closer to the group average has higher credibility, while evidence further away from the group has lower credibility. This is based on the conflict suppression principle of group consistency. In step S44, credibility is normalized to weights using the following formula:

[0090] ;

[0091] in Indicates sensor Normalized weights of evidence; Indicates to From 1 to Summation; other symbols are the same as before; when When the weight is 0, the weights of each piece of evidence are determined. Pick To avoid a zero denominator, equal-confidence processing is performed. This catch-all rule covers the extreme case where all evidence is completely contradictory. In step S45, the representative evidence is obtained by weighted summation of the dynamic basic probability assignments using the following formula:

[0092] ;

[0093] in This indicates that the representative evidence is in Jiao Yuan. The basic probability distribution on; Indicates sensor The dynamic basic probability allocation in the focal element The value of ; Indicates to From 1 to Summation; other symbols are the same as before.

[0094] In step S46 when A value not less than 2 will represent evidence. Combined with itself according to Dempster's combination rules This time, evidence of fusion was obtained. ,when When equal to 1, directly set equal The core of Dempster's combination rule is based on two pieces of evidence. and Synthesize according to the formula

[0095] ;

[0096] in This indicates that the fusion evidence is in the non-empty focal element. The basic probability distribution on; and Two pieces of evidence indicating participation in the combination; and These represent two focal elements that traverse the focal element set; Represents the conflict factor; empty set Not assigned quality and agreed upon Set to 0; Indicates that all equal The focal element Summation; Indicates that all focal pairs of empty sets Summation. This step replaces Murphy's simple arithmetic mean with a confidence-weighted average. The weighting coefficients are calculated from the Jousselme evidence distance, making the representative evidence obtained through weighted fusion conceptually equivalent to the optimal representative under group consensus. Step S46 concludes with the fused evidence. The marginal confidence level is determined by the maximum value among the four event hypotheses: fire, gas leak, fall, and intrusion.

[0097] ;

[0098] Record Abnormal events that reach the maximum value are assumed to be edge decision categories. ,in Indicates fusion of evidence; This indicates the fusion of evidence in the event hypothesis. The basic probability distribution on; Indicates marginal confidence; This represents the operation to find the maximum value; Indicates the edge decision category; the remaining symbols are the same as before.

[0099] Step S5 is a dual-threshold edge trimming stage. Step S51 reads the edge high threshold from the self-configured baseline. With low threshold at the edge and guarantee Greater than ,in Indicates the edge high threshold and preferably takes the value of interval; Indicates the low threshold of the edge and preferably takes the value of Interval. In step S52 when Not less than This judgment will be categorized under the high-confidence direct alarm branch. and Write the event to be alarmed into the queue for direct consumption in steps S7 and S8. In step S53, when... In The interval will be used to categorize this judgment into the medium confidence trigger cross-device verification branch to introduce more evidence. In step S54, when Less than This judgment was categorized into the low-confidence suspicious event branch, and the event information was only written to the suspicious event log of the edge gateway without issuing an external alarm. The introduction of the three branches decouples alarm latency from alarm accuracy. The high-confidence path is responsible for millisecond-level response, the medium-confidence path is responsible for secondary confirmation, and the low-confidence path is responsible for false alarm suppression and traceable auditing, avoiding frequent false triggers or missed triggers in critical states due to a one-size-fits-all threshold.

[0100] Step S6 is the cross-device verification, cloud-based secondary fusion, and final confidence level synthesis stage under the medium confidence branch, completed by its sub-steps S61 to S66. In step S61, the high confidence direct alarm branch directly sets the final confidence level. equal It then proceeds to the S7 step without waiting for the cloud, among which... This indicates the final confidence level; this bypass ensures the real-time performance of high-confidence events. In step S62, the medium-confidence trigger cross-device verification branch is set according to... Retrieve the corresponding verification window from the event type-aware verification strategy table. List of Supplementary Equipment ; Strategy table for fire events By binding smoke, temperature, and carbon monoxide sensors in adjacent sub-regions with gas valve status data, gas leak events can be detected. Bind adjacent sub-regions to collect data on combustible gas, carbon monoxide, and temperature sensors and gas valve status to monitor fall events. Binding wearable accelerometers, millimeter-wave radar altitude sequences, and TTS (Text-to-Speech) tasks with the monitored person's voice response to intrusion events. The system integrates door and window magnetic and vibration sensors, homeowner's mobile phone geofence, and presence detection with the in-home camera; among which... Indicates event type The corresponding verification window(s); Indicates event type Corresponding supplementary equipment list; The optimal value varies depending on the event type. For fire, gas, and intrusion, the optimal value range is 3 to 8 seconds. For falls, since confirmation needs to be completed before the person under supervision falls into a more serious condition, the optimal value range is 2 to 5 seconds. The physical diagnosis logic differs significantly for different event types. The key to fire is the concurrent support of smoke, temperature, and gas; the key to falls is the concurrent support of sudden descent posture and lack of response; and the key to intrusion is the concurrent support of door opening, person presence, and homeowner absence. Customizing supplementary equipment lists according to event type ensures that the cost of evidence collection is spent in the direction most conducive to disambiguation during each secondary confirmation.

[0101] In step S63, the edge gateway directs... Each supplementary device initiating a supplementary sampling request and in The system collects supplementary evidence. For supplementary devices involving audio and video, only the existence detection results or motion feature vectors are collected, without uploading the original audio and video to meet family privacy requirements. Then, based on the virtual sensor mapping table, each non-original reading type of supplementary evidence, such as TTS response status, camera existence binary, and geofence binary, is converted into corresponding virtual sensor evidence. This data structure is made consistent with the normalized feature values, normalized judgment thresholds, and membership function steepness coefficients output in step S2. This evidence is then merged with the original evidence to form an enhanced evidence set, which is then sent to the cloud collaboration platform via TLS or an equivalent encryption authentication mechanism. TLS, or Transport Layer Security, is a transport layer encryption protocol defined by the Internet Engineering Task Force in RFC 8446. For bandwidth-constrained links such as NB-IoT and LoRa, DTLS or an authorized lightweight encrypted channel can be used as an equivalent mechanism. The virtual sensor arrangement incorporates non-standard supplementary devices into a unified DS fusion framework, avoiding the need to design separate decision rules for each new device. In step S64, the cloud collaboration platform executes steps S3 and S4 again on the enhanced evidence set to obtain cloud-fused evidence. With cloud confidence ,in This indicates evidence fused in the cloud; This represents the confidence level in the cloud; simultaneously, the cloud collaboration platform calculates the cloud weight coefficient based on the completeness of supplementary evidence returned, cloud response latency, and link quality. And , as well as The data is simultaneously distributed to the edge gateway, where Represents the cloud weight coefficient and its value is in The timeout period; when the cloud collaboration platform becomes unreachable within the preset timeout period, the edge gateway will... Forced to set to 0; The continuous design reflects both link reachability and the sufficiency of the secondary confirmation evidence. Compared to simplifying cloud contributions to a binary 0 / 1 switch, the continuous design... In intermediate scenarios such as incomplete supplementary evidence, slow cloud response, or link jitter, a more reasonable discount ratio can be provided. In the S65 step, the edge gateway synthesizes the final confidence level using the following formula:

[0102] ;

[0103] The symbols are the same as before. The weighted fusion method differs from the simple maximum value method; it allows the cloud to adjust the fusion method when supplementary evidence clearly does not support the current event. Pull down to This effectively acts as a secondary veto, directly suppressing marginal false high confidence caused by kitchen fumes or bathroom moisture. In step S66, [the following is implied:] Confirm thresholds with the cloud In comparison, when Not less than If the event is confirmed, proceed to step S7; otherwise, mark the event as pending review and store it on the cloud collaboration platform without issuing an alarm. This indicates that the threshold is confirmed and the optimal value is selected from the cloud. Interval.

[0104] Step S7 is the branch convergence and event severity scoring stage, completed by its sub-steps S71 to S74. Step S71 performs branch determination and confirms the selection of the event type, directly alarming branches with high confidence are... As the type of confirmed event in this instance, the events confirmed via the cloud in the cross-device verification branch triggered by Zhongzhixin are represented by cloud-fused evidence. The abnormal event that reaches the maximum value is assumed to be the confirmation event type for this operation. For low-confidence records of suspicious event branches and events awaiting review that have not passed cloud confirmation, this step is terminated directly and step S8 is skipped; let the confirmation event type for this operation be... ,in This indicates the confirmed event type. The high-confidence path directly uses the edge decision category to avoid introducing cloud trajectories when it is already highly self-consistent. The cloud-based confirmation path selects the event type based on the maximum fusion point of the evidence set itself. Both paths logically close under different branches. In step S72, the self-configured baseline reads the event type weight table and... Query event type weight ,in This indicates the event type weight corresponding to the current event; the event type weight table assigns values ​​to four types of events: fire, gas leak, fall, and intrusion, and decreases monotonically from high to low according to the urgency of the event, reflecting the industry consensus that fire is more urgent than intrusion under the same confidence level.

[0105] In step S73, the self-configured baseline reads the number of elderly people in the household. Number of children Number of patients with chronic diseases and the corresponding vulnerability weights , , Calculate the family vulnerability coefficient using the following formula:

[0106] ;

[0107] in This represents the family vulnerability coefficient; Indicates the number of elderly people in the family (in persons); Indicates the number of children in the family (in persons); This indicates the number of people in a family who have chronic diseases. This represents the incremental vulnerability weight for each elderly person; This represents the incremental vulnerability weight for each child; This represents the incremental vulnerability weight for each chronic disease patient. , , Preferred value is Interval. This formula is constructed using a baseline of 1 plus a cumulative increment, allowing for situations where there are no elderly or children ill in the family. A rollback to 1 aligns with the traditional approach, especially when vulnerable populations are present. The elevation automatically amplifies the subsequent severity score, quantifying the home context—where elderly people and children are at home—which is closely related to alarm distribution, and incorporating it into the distribution decision. In step S74, the event severity score is calculated using the following formula:

[0108] ;

[0109] in This represents the severity score of the event; other symbols are the same as before. The three-factor multiplication structure unifies the confidence level (the probability of the event occurring), the event type weight (the urgency of the event itself), and the family vulnerability level (the vulnerability of the affected family) into a scalar score as the quantitative input for the graded distribution in step S8. This score is also intuitive and interpretable; an increase in any of the three factors will raise the score, and a value of 0 in any of them will cause the score to be 0 and stop the outgoing data, which facilitates the tracing of alarm decisions during review and after-sales debugging.

[0110] Step S8 is the multi-channel hierarchical distribution and ACK retransmission handover phase, completed by its sub-steps S81 to S85. In step S81, the self-configured baseline reads the hierarchical receiver table and performs processing for each receiver. Determine its activation flag using the following formula:

[0111]

[0112] in Indicates the recipient's ID; Indicates the recipient With confirmation event type The weight of the receiver association between them; Indicates the recipient The trigger threshold; Indicates the recipient The activation flag is as follows; other symbols are the same as before. The tiered receivers include the homeowner's smart terminal, the homeowner's voice / sound and light alarm, the property management / neighbor terminal, and the emergency linkage terminal. The tiered receiver table is configured with receiver association weights according to the combination of receiver and event type. This allows the same receiver to have different activation tendencies under different event types. In a fall event, family members' voice communication should be prioritized; in a fire event, emergency response should be prioritized; and in an intrusion event, the homeowner's smart terminal should be prioritized. Differentiated association weights based on event type avoid the awkward situation where a fixed global ranking is unreasonable across different events. In step S82, the self-configured baseline reads the receiver-channel reachability matrix, and for each activated receiver... With each wireless channel Read reachability flag , will satisfy Equal to 1 and a binary indicator of channel availability A channel equal to 1 is considered to be for the receiver. Reachable and usable channels; where The wireless channel number is taken from at least two channels actually integrated in Wi-Fi, 4G / 5G, NB-IoT, and LoRa. Indicates the recipient With wireless channel A binary flag indicating reachability between them, with 1 for reachable and 0 for unreachable; Indicates channel Availability binary indicator. Reachability matrix. The introduction of this technology solves the natural preference of different receivers for different channels, such as the inability of homeowner smart terminals to use LoRa, the reliance of emergency response terminals on cellular networks, and the use of local buses for home audio and visual signals.

[0113] For each reachable and available channel Since recently The moving average of the RTT (Round Time) is calculated for each transmission. Weighted moving average with success rate index And read its current bandwidth estimate In the channel Insufficient historical sending count The initial RTT, initial transmission success rate, and initial bandwidth estimate from the configuration baseline are used in the calculation, and the channel utility score is calculated using the following formula:

[0114] ;

[0115] Combining instructions for unavailable channels or unreachable receivers Equal to 0; where Indicates the channel statistics window length (times) and preferably takes the value of interval; Indicates channel RTT moving average (ms); Indicates channel The index-weighted moving average of the success rate of data transmission; Indicates channel Current bandwidth estimate (kbps); Indicates channel Utility score; Indicates the preset lower limit of delay (ms) and is used to avoid Division by zero when taking 0; Indicates the preset allowable latency limit (ms); Indicates the reference bandwidth (kbps); , , These represent the weights of the latency term, the success rate term, and the bandwidth term, respectively, and their sum is equal to 1. This indicates the operation of taking the larger of the two values; other symbols are the same as before. This scoring transforms the channel selection problem into a weighted sum of three intuitively observable quantities: delay, reliability, and bandwidth. Weighting and normalization ensure... Comparability; Introduction In order to be in To prevent fractional explosion in cases where the value suddenly becomes 0 or abnormally small, an engineering approach is implemented to ensure the robustness of the algorithm.

[0116] In step S83, press Descending order channel selection and to After each receiver with a value of 1 sends an alarm message, it starts an ACK waiting timer. If no ACK is received when the ACK waiting timer expires, the transmission is marked as a failure on this channel. If the activated receiver has a reachable and available channel with a second-best utility score, it switches to that second-best channel for retransmission. If no second-best channel that meets the reachability and availability conditions exists, the alarm message is written to the retransmission queue, and the utility score calculation and channel selection in step S82 are re-executed according to a preset backoff period. If all reachable and available channels of the activated receiver fail consecutively for more than [time period missing], the transmission continues. This will trigger the local voice / visual alarm to perform a local downgrade alarm and record the communication failure status; among which... This represents the maximum number of consecutive failures allowed on the same channel, preferably an integer between 2 and 4. The three-level fallback mechanism of suboptimal switching, backoff retry, and local degradation covers all abnormal operating conditions from instantaneous congestion to long-term global network outages, providing an engineered safety net to ensure that at least someone notices the problem under any circumstances. In step S84, after each transmission, the transmission success rate index-weighted moving average is updated using the following formula:

[0117] ;

[0118] in Indicates the sequence number sent; Indicates channel No. The index-weighted moving average of the success rate of transmission before the first transmission; Indicates channel No. The index-weighted moving average of the success rate of transmission before the first transmission; Indicates channel No. The ACK binary indicator sent in this transmission is set to 1 for success and 0 for failure. This represents the smoothing coefficient of the exponentially weighted moving average and takes a value greater than 0 and less than 1. The preferred value range is The remaining symbols are the same as before. The exponentially weighted moving average assigns greater weight to recent samples and geometrically decaying weights to historical samples, which can quickly track recent changes in channel quality while preserving long-term statistical stability, exhibiting better real-time response characteristics compared to directly taking the arithmetic mean. In step S85, when the event severity score... Not less than the concurrency redundancy threshold Concurrent redundant transmission is enabled at the time, and for each activated receiver in its The same alarm message is sent simultaneously on the top two reachable and available channels. The alarm is considered delivered upon receiving an ACK on either channel. Represents the concurrency redundancy threshold and the preferred value is... The interval. This mechanism offers a reasonable trade-off for high-severity events, allowing for a higher delivery power limit with minimal additional overhead, representing a reasonable compromise for engineering requirements that cannot tolerate failures in high-severity events. At this point, steps S1 to S8 and their sub-steps on the method side have been completed.

[0119] The following describes the system components that implement the above method. (Refer to...) Figure 5 The intelligent home collaborative alarm system of the present invention comprises five primary structures: a perception layer, an edge gateway, a cloud collaborative platform, a multi-channel access module, and a hierarchical receiver. The perception layer includes various types of sensors deployed in sub-areas such as the kitchen, living room, bedroom, bathroom, balcony, and corridor, as well as on the personal devices of the monitored individual. These sensors are connected to the edge gateway via a wired bus or low-power wireless link, serving as the system's input. The edge gateway is the core computing node of this system. Internally, it is arranged in the signal flow direction as follows: a data acquisition and feature normalization module, a dynamic basic probability allocation generation module, a conflict adaptive suppression and evidence fusion module, an edge dual-threshold decision module, a cross-device verification request module, a final confidence synthesis module, a severity scoring module, and a multi-channel selection and retransmission control module—a total of eight functional modules. A configuration storage module is also provided in parallel to serve as the source of parameters and strategies for each functional module. The configuration storage module establishes unidirectional read connections with the dynamic basic probability allocation generation module, the edge dual-threshold decision module, the cross-device verification request module, the final confidence synthesis module, the severity scoring module, and the multi-channel selection and retransmission control module, respectively. This allows each functional module to obtain the necessary decision parameters, prior probability tables, thresholds, policy tables, weight tables, and reachability matrices at any time without interrupting the main process. The output of the edge dual-threshold decision module is further branched to the cross-device verification request module, the final confidence synthesis module, and the suspicious event log storage inside the edge gateway, enabling the three branches to be implemented simultaneously in both downstream processes and local auditing directions.

[0120] The data acquisition and feature normalization module internally includes a sensor access unit, a feature extraction unit, and an amplitude normalization and filtering unit. The input end of the sensor access unit is connected to the sensing layer, and the output end is connected to the feature extraction unit. The output end of the feature extraction unit is connected to the amplitude normalization and filtering unit. The feature extraction unit is used to convert the raw data from multidimensional or sequential sensors into scalar feature values ​​according to a preset algorithm, and its feature extraction parameters are read from the configuration storage module. The function of the data acquisition and feature normalization module is to shape the heterogeneous raw data from the sensing layer into a unified format input required by the subsequent dynamic basic probability allocation generation module, and to incorporate multidimensional sequences, state quantities, and amplitude quantities into the same processing interface, providing a consistent evidence source for subsequent DS evidence fusion.

[0121] The dynamic basic probability allocation generation module internally includes a membership degree calculation unit, a priori lookup table unit, and a BPA normalization unit. The input of the membership degree calculation unit is connected to the output of the data acquisition and feature normalization module, outputting the support degree of each sensor for each event hypothesis in Sigmoid form. The input of the priori lookup table unit is connected to the configuration storage module, retrieving the spatial prior probability and time-period prior probability based on the current sampling time and the current sub-region. The outputs of both the membership degree calculation unit and the priori lookup table unit are connected to the input of the BPA normalization unit. The BPA normalization unit multiplies the output of the membership degree calculation unit and the output of the priori lookup table unit, normalizes the result by a normalization factor, and then normalizes the result within the entire recognition framework. The system retains residual uncertainties and outputs the dynamic basic probability assignment for each sensor in the current frame. The dynamic basic probability assignment generation module embeds semantic priors into the evidence construction, providing DS fusion with evidence sources that are aware of spatial and temporal context, which helps to suppress context-dependent interference commonly found in home scenarios.

[0122] The conflict adaptive suppression and evidence fusion module internally includes a Jaccard matrix storage unit, a Jousselme distance calculation unit, a credibility weighting unit, a Dempster combination unit, and a single evidence / full conflict anomaly processing unit. The Jaccard matrix storage unit permanently stores the Jaccard matrix on the power set focal element set of the recognition framework. The input of the Jousselme distance calculation unit is connected to the output of the dynamic basic probability allocation generation module and the Jaccard matrix storage unit, calculating the Jousselme evidence distance pairwise under conditions of two or more pieces of evidence. The credibility weighting unit... The input of the weighting unit is connected to the output of the Jousselme distance calculation unit, which calculates the credibility of each piece of evidence based on the distance matrix and normalizes it into weights. The input of the Dempster combination unit is connected to the output of the credibility weighting unit, and the output is connected to the edge dual-threshold decision module. The single evidence / full conflict anomaly processing unit is bypassed to the Jousselme distance calculation unit and the credibility weighting unit, respectively. When the number of valid evidence pieces is 1 or the sum of credibility is 0, the representative evidence generation result is directly bypassed to the Dempster combination unit, skipping the Jousselme distance and credibility normalization calculation. The conflict adaptive suppression and evidence fusion module as a whole realizes DS evidence fusion under conflict adaptive suppression, ensuring that the fusion process does not undergo abnormal degradation under all evidence scales, and providing edge confidence with group consistency constraints for subsequent decisions.

[0123] The edge dual-threshold decision module has a dual-threshold comparator and maintains three outputs: an alarm event queue, an event flag to be reviewed, and a suspicious event log. Its function is to discretize the continuous edge confidence into three branches according to two preset thresholds and guide each branch to its respective downstream module or local storage, so that subsequent processes can adopt differentiated processing costs for events of different levels.

[0124] The cross-device verification request module incorporates an event type-aware verification strategy table and a virtual sensor mapping table. The event type-aware verification strategy table binds differentiated supplementary device lists and verification windows to fire events, gas leak events, fall events, and intrusion events, respectively. The virtual sensor mapping table stores the mapping relationship between non-raw reading data returned by supplementary devices and normalized feature values, normalized judgment thresholds, and membership function steepness coefficients, enabling supplementary evidence to follow the same DS fusion mainline as original evidence. The cross-device verification request module also includes a TTS response triggering subunit. This subunit controls and connects to the user's voice / sound and light alarm to initiate a voice greeting to the ward, collect response status, and send the response status back to the cloud collaboration platform as supplementary evidence in the fall event branch of the cross-device verification. This subunit transforms the ward's response signal into virtual sensor evidence that can be incorporated into DS fusion with minimal hardware cost, providing strong support for reducing the cost of misjudgments and omissions.

[0125] The input of the final confidence synthesis module is connected to the edge confidence output of the edge dual-threshold decision module, the cloud confidence output of the cloud collaboration platform, and the cloud weight coefficient output of the cloud collaboration platform. The output is connected to the severity scoring module. The final confidence synthesis module weights the edge confidence and the cloud confidence according to the cloud weight coefficient to synthesize the final confidence score, and forcibly sets the cloud weight coefficient to 0 when the cloud collaboration platform is unreachable. The module is positioned on the edge gateway side rather than the cloud side to ensure that the final confidence synthesis result can still be completed independently by the edge gateway when the cloud is unreachable, maintaining system availability under weak network or network outage conditions.

[0126] The severity scoring module internally includes an event type weight lookup table unit, a family vulnerability calculation unit, and a severity synthesis unit. The input of the family vulnerability calculation unit is connected to the family member information storage area within the edge gateway. The family member information storage area maintains a small table based on the number of elderly people, children, and patients with chronic diseases in the family and their corresponding weights. The input of the severity synthesis unit is connected to the output of the event type weight lookup table unit, the output of the family vulnerability calculation unit, and the output of the final confidence synthesis module, with the output connected to the multi-channel selection and retransmission control module. The severity scoring module compresses multi-source signals into a single scalar score as input for subsequent distribution decisions. Its function is both to quantify the family context in the alarm decision and to completely decouple the distribution logic from the confidence synthesis logic.

[0127] The multi-channel selection and retransmission control module internally includes a channel utility scoring unit, a channel selection unit, a receiver-channel reachability filtering unit, an ACK monitoring unit, a retransmission scheduling unit, a local degradation triggering unit, and a concurrent redundancy triggering unit. The receiver-channel reachability filtering unit reads the receiver-channel reachability matrix from the self-configuration storage module and filters reachable and usable wireless channels for each activated receiver for use by the channel utility scoring unit and the channel selection unit. The channel utility scoring unit calculates a utility score for reachable and usable channels based on three indicators: latency, success rate, and bandwidth. The channel selection unit selects channels in descending order of utility score. The ACK monitoring unit is connected to the feedback terminals of each wireless transceiver unit in the multi-channel access module. The system includes a channel utility scoring unit that drives the online updating of the weighted moving average of the transmission success rate index for each channel; a retransmission scheduling unit that switches to the suboptimal channel or writes an alarm message into the retransmission queue and re-enters the channel selection process according to a preset backoff period when an ACK times out; a local degradation triggering unit whose output is connected to the user's voice / audio-visual alarm to trigger a local degradation alarm when all reachable and available channels of the activated receiver fail consecutively for more than a preset limit; and a concurrent redundancy triggering unit whose input is connected to the output of the severity scoring module, and whose output is connected to the two reachable and available wireless transceivers with the highest utility scores in the multi-channel access module to initiate concurrent redundancy transmission when the severity score is not less than the concurrent redundancy threshold. The multi-channel selection and retransmission control module as a whole implements multi-layered delivery assurance, including on-demand channel selection, retransmission with ACK, local degradation for consecutive failures, and concurrent redundancy for high severity.

[0128] The multi-channel access module includes at least two of the following independent transceiver units: Wi-Fi transceiver unit, 4G / 5G transceiver unit, NB-IoT transceiver unit, and LoRa transceiver unit. The Wi-Fi unit handles high-speed, high-efficiency local and internet access; the 4G / 5G unit handles public network mobile access and cross-domain routing; the NB-IoT unit handles power-sensitive small data uplinks; and the LoRa unit handles long-distance, low-power fallback within buildings. The multi-channel selection and retransmission control module establishes control and ACK feedback connections with each of the aforementioned wireless transceiver units, ensuring that any channel status change is instantly fed back to the channel selection logic. The physical characteristics of the multiple channels are naturally complementary, resulting in overall reachability superior to any single channel after combination.

[0129] The cloud-based collaborative platform communicates with the edge gateway via a multi-channel access module. Internally, it includes a secondary DS fusion service module, a cloud-based weight generation module, an evidence chain storage module, and a remote configuration distribution module. The input of the secondary DS fusion service module is connected to a supplementary device in the perception layer via a cross-device verification request module, and its output is connected to the final confidence synthesis module for distributing cloud-based confidence scores. The input of the cloud-based weight generation module is connected to both the secondary DS fusion service module and the link quality monitoring port of the cloud-based collaborative platform, and its output is connected to the final confidence synthesis module for distributing cloud-based weight coefficients. The evidence chain storage module forms a traceable evidence chain based on timestamps, sub-region numbers, sensor numbers, normalized feature values, dynamic basic probability allocation, confidence weights, Jousselme distance matrix, fusion confidence scores, channel selection, and ACK status. This traceable evidence chain does not contain original audio / video data; for supplementary devices involving audio / video, it only stores existence detection results, motion feature vectors, or response status to meet home privacy requirements. The remote configuration distribution module is connected to the configuration storage module via the multi-channel access module to distribute new configuration packages online to update the configuration baseline. The cloud-based collaborative platform communicates with the edge gateway via TLS or an equivalent encryption authentication mechanism. The cloud-based collaborative platform serves two main purposes: firstly, it undertakes secondary evidence fusion and cross-household verification in computationally intensive and cross-family scenarios; secondly, it assumes comprehensive responsibilities such as system-level evidence archiving, online parameter updates, and third-party system access.

[0130] The tiered receivers include the homeowner's smart terminal, the homeowner's voice / sound and light alarm, the property management / neighbor terminal, and the emergency response terminal. The homeowner's smart terminal serves as the primary receiver for regular alarms, the homeowner's voice / sound and light alarm serves as the final fallback for local visual alarms and local degradation paths, the property management / neighbor terminal serves as the unit for coordinated response within the same building, and the emergency response terminal serves as the access point for social assistance. The aforementioned tiered receivers establish alarm receiving communication connections with the edge gateway and the cloud-based collaborative platform via a multi-channel access module. An ACK feedback transmission link is provided between the edge gateway and each of the tiered receivers, ensuring that each alarm delivery can be accurately detected by the multi-channel selection and retransmission control module, which then updates the channel quality estimate accordingly.

[0131] Example 1: This example uses a three-bedroom, two-living-room house as an application case. The permanent residents are two middle-aged adults and one 75-year-old. One edge gateway, one sensor cluster, and four wireless channels (Wi-Fi, 4G / 5G, NB-IoT, and LoRa) are deployed.

[0132] In step S1, complete the division of the kitchen, living room, master bedroom, secondary bedroom, bathroom, balcony, and corridor into 7 sub-areas and 6 time periods: late night, early morning, morning, noon, afternoon, and evening. Then, establish the recognition framework in step S12. According to step S13, complete the deployment and access of 21 nodes, including photoelectric smoke sensors, temperature sensors, CO sensors, combustible gas sensors, human infrared sensors, millimeter-wave radar sensors, door and window magnetic sensors, and wearable accelerometer sensors, and calibrate the normalized judgment thresholds of each sensor for each event hypothesis. steepness coefficient of membership function A priori probability table is configured, and the configuration baseline is distributed by the cloud collaboration platform in the form of a configuration package according to S15, with a high threshold at the edge. Take 0.85, low threshold at the edge. Take 0.50, and confirm the threshold in the cloud. Set the smoothing coefficient to 0.80. Take 0.20, residual uncertain mass Set the threshold to 0.02 for concurrent redundancy. Take 0.85. In step S2, the edge gateway uses a sampling period of... Take 100ms and execute steps S21, S22, and S23 respectively to complete the raw data acquisition, feature extraction and normalization, and window length adjustment. A moving average filter of 5 is applied. In step S3, steps S31 to S34 respectively complete the sub-regional table lookup, Sigmoid membership calculation, dynamic basic probability assignment generation under spatial-temporal prior modulation, and residual uncertainty mass assignment. In step S4, steps S41 to S46 respectively complete Jaccard matrix storage, Jousselme evidence distance calculation, confidence normalization weighting, representative evidence generation, Dempster combination rule continuous self-synthesis, and marginal confidence and marginal decision category recording. In step S5... and Complete the three-level branch determination. In step S6, the cross-device verification branch triggered by the medium confidence level is set according to... Retrieve the corresponding supplementary equipment list and verification window from the policy table. The verification window is 5 seconds for fire, 5 seconds for gas leak, 3 seconds for fall, and 5 seconds for intrusion. After virtualizing the supplementary evidence into sensors, it is sent to the cloud via a TLS 1.3 encrypted link. The edge gateway then uses a weighted composite formula to obtain the final confidence level and... Determine the confirmation status. In step S7, follow steps S71 to S74 to complete the selection of the confirmation event type, the lookup of the event type weight table, the calculation of the family vulnerability coefficient, and the weighting of the elderly vulnerability. The score is calculated using a value of 0.20 and an event severity score. In step S8, steps S81 to S85 are followed to complete receiver activation, reachability filtering, channel utility scoring, ACK waiting time set to 1 second, and timeout switching to suboptimal. After three consecutive failures, the user's voice / sound and light levels will be downgraded or higher. The system can be transmitted concurrently and redundantly through the two most accessible and available channels, ranking first in utility score.

[0133] Example 2 differs from Example 1 in that the family composition consists of two middle-aged adults, two elderly people over 70 years old, and two children under 8 years old. The vulnerability weight of the elderly is adjusted. The child vulnerability weight is set to 0.25. Set the weight to 0.20 for chronic disease vulnerability. Take 0.15, and the rest of the configuration is the same as in Example 1.

[0134] Example 3 differs from Example 1 in that it uses the edge height threshold. Equal to 0.90, low threshold at the edge Equals 0.55, cloud-confirmed threshold. Equal to 0.85, smoothing coefficient Take 0.30, residual uncertain mass Set the threshold to 0.05 for concurrent redundancy. The value is set to 0.95, which corresponds to the false alarm suppression preference in high security-sensitive scenarios such as hospitals or luxury apartments. The other configurations are the same as in Example 1.

[0135] Example 4 differs from Example 1 in that it simulates a weak network environment, where the Wi-Fi channel is unavailable for an extended period, the primary channel is NB-IoT with LoRa as a backup, and the edge gateway switches to another channel when the cloud collaboration platform is unreachable. Forced to zero and independently completes the final confidence synthesis, taking the edge high threshold. Equal to 0.80, low threshold at the edge Equal to 0.40, smoothing coefficient Take 0.10, residual uncertain mass Take 0.01, and the rest of the configuration is the same as in Example 1.

[0136] Comparative Example 1: This comparative example replaces step S4 with directly applying the original data according to the classic Dempster combinatorial rule. The method of sequential synthesis of basic probability allocation does not calculate the Jousselme evidence distance, does not perform credibility weighting, and does not construct representative evidence; the rest of the configuration is the same as in Example 1.

[0137] Comparative Example 2 replaces step S4 with the simple arithmetic mean method, i.e., for The basic probability distributions are then calculated by taking the equally weighted arithmetic mean and then synthesizing them using Dempster. The confidence level was reduced by 1 time, without using Jousselme evidence distance to calculate the confidence level. The rest of the configuration was the same as in Example 1.

[0138] Comparative Example 3 replaces the weighted composition formula in step S65 with the maximum value formula. equal The remaining configurations are the same as in Example 1.

[0139] Comparative Example 4 replaces steps S82 to S85 with sending alarm messages using only a single Wi-Fi channel, without ACK waiting timer, suboptimal handover, concurrent redundancy, or local degradation, with the remaining configurations the same as in Example 1.

[0140] Experiment 1 uses the methods of Examples 1 to 4 and Comparative Examples 1 and 2 to verify the overall decision quality and calibrate the detection thresholds of each method. The mixed sample includes 1000 positive samples covering five categories: open flame, smoldering, gas leak, fall, and intrusion; and 5000 negative samples covering typical interferences such as oil fumes, water vapor, pet activity, and isolated human infrared triggering. For each method, the false positive rate (FPR) and detection rate (TPR) sequences are scanned on the same batch of samples to obtain the threshold values, and ROC curves are plotted. The area under the curve (AUC) is calculated using the trapezoidal integral method, and the confusion matrix is ​​calculated at each method's operating point to obtain the false positive rate (FAR) and false negative rate (MAR). Since Examples 2 to 4 share the dynamic basic probability allocation generation method in step S3 and the conflict adaptive suppression fusion method in step S4 of this invention with Example 1, differing only in vulnerability weights, threshold values, or network morphology, their ROC curves highly overlap with those of Example 1. This experiment... Figure 1 In section (a), only the ROC curves of Example 1 and Comparative Examples 1 and 2 are plotted, and... Figure 1 In section (b), the FAR and MAR of Examples 1 to 4 and Comparative Examples 1 and 2 at their respective operating points are supplemented in a bar chart format so that the parameter differences of each example can be reflected in the decision index.

[0141] Experimental results are as follows Figure 1 As shown. In Figure 1 Figure (a) shows three ROC curves for Example 1, Comparative Examples 1 and 2, and a random guessed baseline. The AUC for Example 1 is 0.974, while the AUCs for Comparative Examples 1 and 2 are 0.882 and 0.913, respectively. Figure 1 Figure (b) shows a histogram comparison of FAR and MAR at their respective operating points for six methods, namely Examples 1 to 4 and Comparative Examples 1 and 2. The FARs of Examples 1 to 4 are 1.2%, 1.5%, 0.6%, and 1.8%, respectively, and the MARs are 0.8%, 0.7%, 1.5%, and 1.4%, respectively. The FARs of Comparative Examples 1 and 2 are 5.8% and 4.1%, respectively, and the MARs are 3.2% and 2.5%, respectively.

[0142] from Figure 1It can be seen that the ROC curve of the method of the present invention is above the classical Dempster direct synthesis method and the Murphy arithmetic mean method throughout the process. The AUC of Example 1 is 0.092 and 0.061 higher than that of Comparative Examples 1 and 2, respectively. Moreover, the FAR and MAR of Example 1 are 4.6 and 2.4 percentage points lower than those of Comparative Example 1, and 2.9 and 1.7 percentage points lower than those of Comparative Example 2. Figure 1 (b) further shows that the FAR of Examples 1 to 4 is no higher than 1.8% and the MAR is no higher than 1.5% under different vulnerability weights and threshold values, which are comprehensively better than Comparative Examples 1 and 2.

[0143] In step S3 of this invention, spatial prior probability and time-period prior probability are used as multiplicative modulation factors to embed the historical occurrence tendency of such events in the sub-region and time period into the construction of the basic probability allocation, thereby suppressing context-related interferences such as kitchen fumes, bathroom moisture, and pet activities. In step S4 of this invention, the confidence level is calculated by back-calculating the Jousselme evidence distance and then weighted accordingly to obtain representative evidence for Dempster self-synthesis. Conceptually, this is equivalent to the optimal representative under the constraint of group consistency, avoiding the counterintuitive amplification phenomenon commonly seen in the classic Dempster combination rule for high-conflict evidence. It pushes the detection capability under the same false alarm rate to the upper left of the ROC curve, indicating that the dynamic basic probability allocation generation method and the conflict adaptive suppression fusion method described in steps S3 and S4 of this invention can both produce reasonable marginal confidence levels under heterogeneous sensor combinations.

[0144] Experimental Example 2 uses the methods of Examples 1, 4, and Comparative Example 4 to verify the end-to-end delay. Based on the alarm system response delay requirement (i.e., the fire alarm delay should not exceed 10 seconds), the end-to-end link delay of 1000 alarms is used as the evaluation object. This delay is calculated from the start of sampling in S21 until the receiver receives the alarm message or the local degradation trigger ends. The delay distribution is generated according to a log-normal model, and P50, P95, and P99 are calibrated. Since Comparative Examples 1 to 3 only replace the fusion method in step S4 or the synthesis method in step S65 of this invention, respectively, without changing the three-level branch in step S5 or the multi-channel distribution process in step S8, their end-to-end delay curves highly overlap with Example 1. Therefore, this experimental example... Figure 2 Only three methods are compared: Example 1, Example 4, and Comparative Example 4. Example 4 represents an edge autonomous mode detached from the cloud, while Comparative Example 4 represents a single Wi-Fi channel mode without handover or redundancy. Experimental results are as follows: Figure 2 As shown. In Figure 2 Figure (a) shows three end-to-end cumulative delay probability distribution curves (Examples 1, 4, and Comparative Example 4), the horizontal reference line (P95), and the vertical reference line for the 10s delay upper limit required by GB50116-2013. Figure 2 Figure (b) shows a histogram comparison of the P50, P95, and P99 delays for the three methods, with the vertical axis on a logarithmic scale and a 10-second upper limit reference line marked. The P50, P95, and P99 for Example 1 are 180ms, 1.20s, and 2.63s, respectively; for Example 4, they are 85ms, 0.30s, and 0.51s, respectively; and for Comparative Example 4, they are 320ms, 6.00s, and 20.19s, respectively.

[0145] from Figure 2 It can be seen that P50, P95 and P99 of Example 4 are significantly lower than those of other methods, corresponding to its edge autonomous form that is detached from the cloud; P95 and P99 of Example 1 are both within the upper limit of 10s, while P95 of Comparative Example 4 has reached 6s and P99 has further broken through 10s to reach 20.19s.

[0146] The three-branch design in step S5 of this invention decouples the high-confidence path from the medium-confidence path. High-confidence events directly enter steps S7 and S8 without waiting for the cloud, preserving a millisecond-level response channel. The multi-channel concurrent selection, ACK switching suboptimal, and concurrent redundancy in steps S82 to S85 convert the physical characteristics of multiple channels into natural redundancy in latency, ensuring that temporary failures of a single channel under packet loss conditions do not block overall delivery. The transmission layer retry latency of a single-channel scheme exhibits geometric expansion, with a typical retry backoff exponent of 2, and cumulative retries causing P99 latency to reach tens of seconds. In contrast, the overall delivery latency of the multi-channel concurrent scheme is limited by the available channel that returns the fastest ACK, and the lower bound of channel independence guarantees rapid convergence at the tails of P95 and P99. This indicates that the multi-channel hierarchical distribution method in steps S5, S82, S83, and S85 of this invention can still meet the hard constraints on alarm latency under medium to high packet loss conditions.

[0147] Experiment Example 3 verifies the final confidence level synthesis method using the methods of Example 1 and Comparative Example 3. It constructs different combinations of edge confidence and cloud confidence levels under the confidence-triggered cross-device verification branch in four typical operating conditions. Operating condition A is a true positive with strong cloud support, and therefore... Equal to 0.78 Equal to 0.96 Equals 0.85; Condition B is a true positive and has moderate support in the cloud, so take... Equals 0.76 Equal to 0.88 Equals 0.70; Condition C is a false positive for oil fume and has a strong negative result in the cloud, so take... Equal to 0.82 Equal to 0.04 Equals 0.75; Condition D is a weak network true positive, take... Equal to 0.85 The value equals 0. The true value for conditions A, B, and D is "alarm should be triggered," while the true value for condition C is "alarm should not be triggered."

[0148] Calculate according to the weighted synthesis formula of step S65 of the present invention and the maximum value formula of Comparative Example 3 respectively. and The accuracy rate was calculated by comparing a value of 0.80 to the accuracy rate obtained from 4000 tests conducted under the above four operating conditions in a 1:1:1:1 ratio. The experimental results are as follows: Figure 3 As shown. In Figure 3 Figure (a) shows Example 1 and Comparative Example 3 under four operating conditions. The values, true values ​​for each operating condition, and cloud-based confirmation threshold reference lines can be read from Example 1. The values ​​for operating conditions A, B, C, and D are 0.933, 0.844, 0.235, and 0.850 respectively. (Comparative Example 3) The values ​​are 0.816, 0.760, 0.820, and 0.850, respectively. Figure 3 Figure (b) shows the decision accuracy of Example 1 and Comparative Example 3 under a mixed 4-condition system. The accuracy of Example 1 is 98.5%, and the accuracy of Comparative Example 3 is 49.0%.

[0149] from Figure 3 It can be seen that the weighted synthesis formula of the present invention gives a decision consistent with the true value under all four working conditions, while the maximum value formula of Comparative Example 3 fails to produce a result consistent with the true value under working condition B. The false negative occurred because the value was 0.760, which is below the cloud confirmation threshold, under operating condition C. The false alarm occurred because the value of 0.820 was higher than the cloud confirmation threshold. This illustrates that the weighted synthesis formula of this invention, relative to the maximum value formula, can effectively handle situations with strong cloud support. Elevate the confirmation threshold to the cloud to correctly confirm the event, and also allow for strong denial in the cloud. Pull down to the cloud to confirm below the threshold to suppress false positives in typical scenarios such as cooking fumes.

[0150] The weighted synthesis formula in step S65 of this invention is a linear convex combination, which has bidirectional adjustment capability for edge confidence and cloud confidence; while the maximum value formula in Comparative Example 3 is a monotonically non-decreasing operator, which only retains the upward lifting channel and loses the ability to suppress false high confidence at the edge. Edge decisions and cloud secondary decisions can be regarded as two related pieces of evidence for the same event. The linear convex combination is mathematically equivalent to the weighted fusion of two uncertain pieces of evidence in the Bayesian framework, reflecting the semantics of collaborative verification; the maximum value operator is equivalent to the backup enhancement semantics, which cannot provide correction for the edge decision when the edge and cloud orientations are inconsistent. Under operating condition D Taking 0, both methods degenerate into using 0. The direct judgment yielded consistent results, reflecting the rationality of the weighted composition formula under boundary values. The above analysis proves that the weighted composition method described in step S65 of this invention is effective. When the boundary values ​​are 0 and 1, it degenerates into reasonable forms of pure edge decision and pure cloud decision respectively, and the values ​​are stable.

[0151] Example 4 verifies the multi-channel delivery performance using the methods of Examples 1 and 4, and Comparative Example 4. A 0% to 50% link packet loss rate scan is constructed to simulate a typical urban residential environment with mixed operating conditions of Wi-Fi, 4G / 5G, NB-IoT, and LoRa. For each packet loss rate value, 5000 alarm transmission tests are performed to obtain the delivery success rate. The cumulative uplink bytes over 24 hours are used as the communication overhead proxy quantity. Since Examples 2 and 3 share four channels and the distribution process of step S8 of this invention with Example 1, their delivery success rate curves highly overlap with those of Example 1. Therefore, this example... Figure 4 The comparison only includes three methods: Example 1, Example 4, and Comparative Example 4. Example 4 is a weak network configuration with only two channels remaining: NB-IoT and LoRa. Comparative Example 4 is a configuration with a single Wi-Fi channel and no ACK waiting, no suboptimal handover, no concurrent redundancy, and no local degradation.

[0152] Experimental results are as follows Figure 4 As shown. In Figure 4 Figure (a) shows the delivery success rate curves as a function of packet loss rate for three examples (Example 1, Example 4, and Comparative Example 4), along with a reference line for a 30% packet loss rate. It can be seen that at a 30% packet loss rate, the delivery rate for Example 1 is 99.0%, the delivery rate for Example 4 is 95.0%, and the delivery rate for Comparative Example 4 drops to 60.0%. Figure 4 Figure (b) shows a histogram comparison of the 24-hour cumulative uplink bytes for the three methods, with readable Example 1 at 12.5 MB, Example 4 at 18.5 MB, and Comparative Example 4 reaching 25.6 MB. Figure 4 It can be seen that the multi-channel reachability screening plus utility scoring plus ACK switching suboptimization plus concurrent redundancy plus local degradation mechanism of the present invention can still maintain a delivery rate no less than the requirements of GB / T26875.1-2011 for user information transmission reliability under medium and high packet loss conditions, and achieve a higher delivery rate with lower communication overhead.

[0153] The utility scoring formula in step S82 of this invention integrates three intuitive and observable measurements—latency, reliability, and bandwidth—into a single selection criterion. The three-level fallback mechanism in step S83 guides instantaneous congestion, long-term partial network outages, and network-wide failures in a single channel to three independent paths: suboptimal handover, backoff retry, and local degradation, respectively. The concurrent redundancy in step S85 of this invention trades controllable additional channel overhead for an increased delivery rate ceiling for high-severity events. Multiple channels are approximately independent at both the physical and link layers, and their concurrent success rate follows a probability of 1 minus the joint probability of all channels failing simultaneously, exponentially approaching 1 with the number of channels, far exceeding the 1 minus the packet loss rate for a single channel. Simultaneously, ACK handover avoids invalid retries on temporarily faulty channels, thereby reducing byte overhead proxy volume. Example 4 demonstrates that even in a weak network condition with only NB-IoT and LoRa channels available, the delivery rate remains at 95.0% with 30% packet loss, reflecting the reachability matrix of step S82 of this invention. The utility scoring formula is universally applicable under multi-channel heterogeneous characteristics; the cumulative uplink bytes in Example 4 are higher than those in Example 1 because the LoRa link is limited by the single packet payload and needs to fragment the alarm message, and the accumulation of fragment headers brings fixed overhead. This shows that the multi-channel hierarchical distribution method described in step S8 of the present invention can still meet the reliability requirements under medium to high packet loss conditions, and has low communication overhead.

[0154] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A smart home collaborative alarm method, characterized in that, Includes the following steps: S1: Solidify the configuration of the identification framework, sub-region division, time period division, and configuration baseline on the edge gateway side; S2: The edge gateway periodically collects sensor data, and obtains the normalized feature values ​​of each sensor in the current frame through feature extraction and normalization; S3: Calculate the support of each sensor for each event hypothesis based on the normalized eigenvalues ​​and the decision parameters in the configuration baseline. Use the spatial prior probability of the sub-region where the sensor is located and the time prior probability of the current time period as multiplicative modulation factors to normalize and obtain the dynamic basic probability allocation. S4: Calculate the Jousselme evidence distance between each dynamic basic probability assignment based on the Jaccard matrix on the power set of the recognition framework, back-calculate the credibility of each piece of evidence and normalize it into a weight, sum the weighted evidence to obtain representative evidence, and then continuously combine the representative evidence according to the Dempster combination rule to obtain fused evidence and marginal confidence, and record the marginal decision category. S5: Compare the edge confidence with the edge high threshold and edge low threshold in the configuration baseline, and classify this decision into the high confidence direct alarm branch, the medium confidence cross-device verification trigger branch, or the low confidence suspicious event recording branch; S6: For the central confidence branch, query the verification strategy table according to the edge decision category, initiate differentiated supplementary sampling to the supplementary device, virtualize the supplementary evidence into a sensor and send it to the cloud for secondary fusion, and output the cloud confidence and cloud weight coefficient from the cloud; the edge gateway weights the edge confidence and cloud confidence according to the cloud weight coefficient to synthesize the final confidence, and determines the confirmation status according to the cloud confirmation threshold. S7: For high-confidence branches, use the edge decision category as the confirmation event type and the edge confidence level as the final confidence level; For events confirmed via the cloud in the Zhongzhixin branch, the event type confirmed via the cloud shall be used as the confirmation event type; Based on the confirmed event type, query the event type weight; calculate the family vulnerability coefficient based on the vulnerable population within the family; and multiply the final confidence score, event type weight, and family vulnerability coefficient to obtain the event severity score. S8: Activate receivers based on event severity score, confirmed event type, and hierarchical receiver table; filter reachable and available channels according to receiver-channel reachability matrix, calculate channel utility score based on latency, success rate, and bandwidth, select channels in descending order of score to send alarms and initiate ACK waiting, if no acknowledgment is received after timeout, switch to the second-best channel for retransmission, and handle as a fallback if no second-best channel is available; for events with severity score not lower than the concurrent redundancy threshold, send the same alarm message concurrently on the two most reachable channels with the highest scores.

2. The smart home collaborative alarm method according to claim 1, characterized in that: Step S1 includes: S11: Pre-divide the physical space of the home into sub-areas such as kitchen, living room, bedroom, bathroom, balcony, and corridor; pre-divide the day into periods such as late night, early morning, morning, noon, afternoon, and evening; S12: Establish the recognition framework ;in Represents the recognition framework; Indicates the assumption of a fire event; This indicates a hypothesis about a gas leak event; This indicates the hypothesis of a fall event; Indicates the hypothesis of an intrusion event; This indicates a normal event assumption; S13: Deploy at least one of the following in each sub-region and on the person under guardianship: photoelectric smoke sensor, temperature sensor, carbon monoxide sensor, combustible gas sensor, human infrared sensor, millimeter-wave radar, door and window magnetic sensor, and wearable accelerometer, so that each sensor is connected to the edge gateway via a wired bus or low-power wireless link; for wearable sensors deployed on the person under guardianship, determine the sub-region to which they belong based on the most recent location result of the person under guardianship associated with them or the linkage result with fixed sensors; S14: For each sensor Assume for each event Calibration normalization judgment threshold steepness coefficient of membership function For each sub-region Configure its assumptions for each event Spatial prior probability The spatial prior probability table is then formed by normalizing the rows; for each time period... Configure its assumptions for each event Prior probability of time period The prior probability table for the time period is then formed by normalizing the rows; where Indicates the sensor number; Indicates the event hypothesis number; Indicates the sub-region number; Indicates the time period number; Indicates sensor Hypothesis of the event The normalization threshold is set to a value of . interval; Indicates sensor Hypothesis of the event The membership function steepness coefficient takes a positive or negative value, and the sign is used to characterize whether the increase in sensor reading provides positive or negative support for the hypothesis about the event. Subregion Hypothesis of the event Spatial prior probability; Indicates time period Hypothesis of the event Prior probability of the time period; S15: The normalized judgment threshold, the membership function steepness coefficient, the spatial prior probability table, the time-period prior probability table, the edge high threshold, the edge low threshold, the cloud confirmation threshold, the event type perception verification strategy table, the virtual sensor mapping table, the event type weight table, the home vulnerability weight, the hierarchical receiver table, the receiver-channel reachability matrix, the channel utility scoring parameters, and the concurrent redundancy threshold are distributed to the edge gateway in the form of a configuration package via the cloud collaboration platform and stored in a fixed manner to form the configuration baseline that can be updated online; wherein the channel utility scoring parameters include a preset lower delay limit, a preset upper allowable delay limit, a reference bandwidth, a delay item weight, a success rate item weight, a bandwidth item weight, a channel statistical window length, an initial RTT, an initial transmission success rate, an initial bandwidth estimate, an ACK waiting time, a smoothing coefficient of an exponentially weighted moving average, and the maximum number of consecutive failures allowed on the same channel. Step S2 includes: S21: The edge gateway uses a sampling period Poll each sensor, for each sensor Read its raw data; for amplitude-type sensors, the raw data is the scalar raw reading; for multidimensional or sequential sensors, extract the sensor feature values ​​from the raw data according to a preset feature extraction algorithm. For state-type sensors, the sensor characteristic values Take its state mapping value; where Indicates the sampling period; Indicates sensor eigenvalues; S22: For each of the sensor feature values According to its calibrated lower limit of characteristic value With upper limit of eigenvalues Perform linear normalization to obtain normalized eigenvalues. and the normalized eigenvalues Cut off to The interval; where Indicates sensor Normalized eigenvalues; Indicates sensor The lower bound of the eigenvalues; Indicates sensor The upper limit of eigenvalues; S23: For the normalized eigenvalues Execution length is A moving average filter is applied to suppress single-point noise, and the normalized feature values ​​of each sensor in the current frame are written together with the current sampling time into the current frame data buffer of the edge gateway; wherein This indicates the length of the moving average filter window.

3. The smart home collaborative alarm method according to claim 2, characterized in that, Step S3 includes: S31: Determine the time period to which the current sampling time belongs by looking up the table. According to each of the aforementioned sensors The current sub-region is determined by looking up the table. ; S32: For each sensor With each event assumption The support level is calculated using the following formula: ； in Indicates sensor In normalized eigenvalues The following assumptions about events Support, valued as interval; Represents the natural exponential function; other symbols are defined as before; when When the value is positive, the support varies. Increase and increase, when When the value is negative, the support varies. Increase and decrease; S33: Query the sub-region from the configuration baseline. Spatial prior probability With the time period Prior probability of time period Calculate the sensor using the following formula Assuming events in the current frame Dynamic basic probability allocation: ； in Indicates sensor Hypothesis of the event The dynamic basic probability allocation value; Indicates sensor The normalization factor, whose value is ; This indicates a summation of hypotheses about each event within the recognition framework; the definitions of other symbols are the same as before. S34: The results proportionally Scaling, and within the entire set of elements of the recognition framework. Assigning residual uncertain mass , making Output sensor The complete dynamic basic probability allocation in the current frame; where This represents the residual uncertain mass used to avoid the zero-mass trap; Indicates sensor In identifying the complete set of elements of the frame The basic probability allocation.

4. The smart home collaborative alarm method according to claim 3, characterized in that, Step S4 includes: S41: Store the identification frame on the edge gateway side. Jaccard matrix on the power set focal set Its elements take values ​​according to the following formula: ； in Represents the Jaccard matrix; This represents the focal element corresponding to a row in the matrix; This represents the focal element corresponding to a column of the matrix; Indicates the number of elements in the set; This represents the intersection operation of sets; This represents the union operation of sets; S42: Count the number of valid pieces of evidence involved in the fusion in this judgment. ;when When the value equals 1, this piece of evidence shall be directly used as the representative evidence. And jump to step S46; when When the number is not less than 2, for those involved in this judgment Any two of the dynamic basic probability assignments and The Jousselme evidence distance is calculated using the following formula: ； in Indicates sensor With sensors Jousselme distance between pieces of evidence; Indicates the number of valid pieces of evidence contributing to the integration; Indicates and A different sensor number; and These represent the sensors. With sensors The dynamic basic probability assignment is considered as a column vector on the focal element set; superscript This represents the matrix transpose operation; This symbol represents the arithmetic square root operation; the definitions of other symbols are the same as before. S43: Calculate the credibility of each piece of evidence using the following formula: ； in Indicates sensor The credibility of the evidence; Indicates to and Different sensor numbers Summation; other symbols are defined as before; S44: Normalize the confidence level into weights using the following formula: ； when When the weight equals 0, let the weights of each piece of evidence be... All Trusted processing; among which Indicates sensor Normalized weights of evidence; Indicates to From 1 to Summation; other symbols are defined as before; S45: The representative evidence is obtained by weighted summation of the dynamic basic probabilities according to the following formula: ； in This indicates that the representative evidence is in Jiao Yuan. The basic probability distribution on; Indicates sensor The dynamic basic probability allocation in the focal element The value of ; Indicates to From 1 to Summation; other symbols are defined as before; S46: When When the number is not less than 2, the representative evidence will be... It is combined with itself continuously according to the Dempster combination rule. The fusion evidence was obtained this time. ;when When it equals 1, directly set Record the fusion of evidence The marginal confidence level is determined by the maximum value among the four event hypotheses: fire, gas leak, fall, and intrusion. ； Record The abnormal event that achieves the maximum value is assumed to be the edge decision category. ;in Indicates fusion of evidence; This indicates the fusion of evidence in the event hypothesis. The basic probability distribution on; Indicates marginal confidence; This represents the operation to find the maximum value; This indicates the edge decision category; the definitions of the remaining symbols are the same as before.

5. The smart home collaborative alarm method according to claim 4, characterized in that, Step S5 includes: S51: Read the edge height threshold from the configuration baseline. With the edge low threshold The edge height threshold Greater than the lower edge threshold ;in Indicates the edge high threshold; Indicates a low threshold at the edge; S52: When Not less than At that time, this judgment will be categorized into the high-confidence direct alarm branch, and the edge judgment category will be... With the edge confidence Write to the queue of events to be alerted; S53: When In When the interval is reached, this judgment will be classified into the aforementioned confidence-triggered cross-device verification branch; S54: When Less than When this decision is made, it is classified into the low-confidence record suspicious event branch, and the event information is written to the suspicious event log of the edge gateway without issuing an alarm; this step outputs the branch identifier of this decision.

6. The smart home collaborative alarm method according to claim 5, characterized in that, Step S6 includes: S61: For the high-confidence direct alarm branch, set the final confidence level... equal Then proceed directly to step S7; where Indicates the final confidence level; S62: For the medium confidence trigger cross-device verification branch, based on the edge decision category Retrieve the corresponding verification window from the event type-aware verification strategy table. List of Supplementary Equipment The event type perception verification strategy table is as follows: for fire events , This includes data acquisition from smoke sensors, temperature sensors, carbon monoxide sensors, and gas valve status in adjacent sub-regions; and monitoring of gas leak events. , This includes data acquisition from combustible gas sensors, carbon monoxide sensors, temperature sensors, and gas valve status in adjacent sub-regions; and data acquisition for fall events. , This includes wearable accelerometers, millimeter-wave radar altitude sequences, and text-to-speech (TTS) tasks for monitored individuals; for intrusion incidents. , This includes magnetic sensors for doors and windows, vibration sensors, geofencing via the homeowner's mobile phone, and presence detection of in-home cameras; among which... Indicates event type The corresponding verification window; Indicates event type Corresponding supplementary equipment list; S63: The edge gateway sends a list of supplementary devices. Each supplementary device initiates a supplementary sampling request in the verification window. Internally collect supplementary evidence; for supplementary devices involving audio and video, only collect existence detection results or motion feature vectors without uploading the original audio and video; convert each non-original reading type of supplementary evidence into the corresponding virtual sensor evidence according to the virtual sensor mapping table, the virtual sensor evidence having the same data structure as the normalized feature value, normalized judgment threshold, and membership function steepness coefficient output in step S2; merge the original evidence and the virtual sensor evidence to form an enhanced evidence set, and upload it to the cloud collaboration platform through TLS or an equivalent encryption authentication mechanism; S64: The cloud-based collaborative platform performs steps S3 and S4 again on the enhanced evidence set to obtain cloud-fused evidence. With cloud confidence Meanwhile, the cloud-based collaborative platform calculates a cloud weight coefficient based on the completeness of supplementary evidence returned, cloud response latency, and link quality. ,Will , as well as The data is simultaneously sent to the edge gateway; if the cloud collaboration platform becomes unreachable within a preset timeout, the edge gateway will... Forced to set to 0; where This indicates evidence fused in the cloud; Indicates confidence level in the cloud; This represents the cloud-based weighting coefficient, with values ​​ranging from... interval; S65: The edge gateway synthesizes the final confidence level using the following formula: ； S66: The final confidence level Confirm the threshold with the cloud Comparison: If Not less than If the event is confirmed, proceed to step S7; otherwise, mark the event as pending review and store it on the cloud collaboration platform without issuing an alarm. This indicates the cloud confirmation threshold; this step outputs the final confidence level. The status of this incident has been confirmed.

7. The smart home collaborative alarm method according to claim 6, characterized in that, Step S7 includes: S71: Branch Decision and Confirmation Event Type Selection: For the high-confidence direct alarm branch, select the edge decision category. As the type of confirmed event in this instance; for the events confirmed via the cloud in the cross-device verification branch triggered by the central confidence, the cloud-fused evidence is used. The abnormal event that reaches the maximum value is assumed to be the confirmation event type for this time; for the low-confidence record suspicious event branch and the event to be reviewed that has not been confirmed by the cloud, this step is terminated directly and the subsequent step S8 is skipped; the confirmation event type for this time is denoted as ;in Indicates the type of confirmation event; S72: Read the event type weight table from the configuration baseline, and confirm the event type. Query the weight of the event type ;in This indicates the event type weight corresponding to the current event; the event type weight table assigns values ​​to four types of events: fire, gas leak, fall, and intrusion, with values ​​decreasing monotonically from high to low according to the urgency of the event. S73: Read the number of elderly people in the household from the configuration baseline. Number of children Number of patients with chronic diseases and the corresponding vulnerability weights , , The family vulnerability coefficient is calculated using the following formula: ； in This represents the family vulnerability coefficient; Indicates the number of elderly people in the family; Indicates the number of children in a family; This indicates the number of people with chronic diseases in the family; This represents the incremental vulnerability weight for each elderly person; This represents the incremental vulnerability weight for each child; This represents the incremental vulnerability weight for each chronic disease patient. S74: Calculate the severity score of the event using the following formula: ； in This represents the event severity score; the definitions of other symbols are the same as before; the output of this step is the event severity score used as the input for the graded trigger quantization in step S8. .

8. The smart home collaborative alarm method according to claim 7, characterized in that, Step S8 includes: S81: Read the hierarchical receiver table from the configuration baseline, for each receiver Determine its activation flag using the following formula: ； ； in Indicates the recipient's ID; Indicates the recipient With the confirmed event type The weight of the receiver association between them; Indicates the recipient The trigger threshold; Indicates the recipient The activation flag is defined as follows; other symbols are defined as before; the hierarchical receivers include the homeowner's smart terminal, the homeowner's voice / sound and light alarm, the property management / neighbor terminal, and the emergency linkage terminal; the hierarchical receiver table is configured with the associated weights of the receivers according to the combination of receiver and event type. ; S82: Read the receiver-channel reachability matrix from the configuration baseline; for each activated receiver... With each wireless channel Read reachability flag , will satisfy And the channel availability binary indicator The channel is identified as being for the receiver. Reachable and available channels; for each of the aforementioned reachable and available channels Since recently The moving average of the RTT (Round Time) is calculated for each transmission. Weighted moving average with success rate index and read its current bandwidth estimate. In the channel Insufficient historical sending count The initial RTT, initial transmission success rate, and initial bandwidth estimate from the channel utility scoring parameters are used in the calculation; the utility score is calculated using the following formula: ； Combining instructions for unavailable channels or unreachable receivers ;in Indicates the wireless channel number. Taken from no fewer than two channels actually integrated in Wi-Fi, 4G / 5G, NB-IoT and LoRa; Indicates the recipient With wireless channel A binary flag indicating reachability between them, with 1 for reachable and 0 for unreachable; Indicates channel Availability binary indicator; This indicates the length of the channel statistics window; Indicates channel RTT moving average; Indicates channel The index-weighted moving average of the success rate of data transmission; Indicates channel The current bandwidth estimate; Indicates channel Utility score; This indicates the preset lower limit of delay; This indicates the preset allowable delay limit; Indicates the reference bandwidth; , , These represent the weights of the latency term, the success rate term, and the bandwidth term, respectively. and and The sum is 1; This operation represents taking the larger of the two values; S83: Utility rating as stated Channels are selected in descending order, and the activation flag is then used. An alarm message is sent to each receiver with a value of 1, and the ACK waiting timer is started. When the ACK waiting timer expires without receiving an ACK, the current transmission is marked as a failure on this channel. If the activated receiver has a reachable and available channel with a second-best utility score, it switches to that second-best channel for retransmission. If no second-best channel that meets the reachability and availability conditions exists, an alarm message is written to the retransmission queue, and the utility score calculation and channel selection are re-executed according to a preset backoff period. If all reachable and available channels of the activated receiver fail consecutively for more than [time period missing], the transmission continues. At this time, the local voice / sound and light alarm is triggered to perform a local downgrade alarm and record the communication failure status; wherein Indicates the maximum number of consecutive failures allowed on the same channel; S84: After each transmission is completed, update the weighted moving average of the transmission success rate index according to the following formula: ； in Indicates the sequence number sent; Indicates channel No. The index-weighted moving average of the success rate of transmission before the first transmission; Indicates channel No. The index-weighted moving average of the success rate of transmission before the first transmission; Indicates channel No. The ACK binary indicator sent in the next transmission is set to 1 for success and 0 for failure. This represents the smoothing coefficient of the exponentially weighted moving average, with a value greater than 0 and less than 1; the definitions of other symbols are the same as before. S85: When the severity score of the event is given Not less than the concurrency redundancy threshold Concurrent redundant transmission is enabled, and each activated receiver is given a utility score. The same alarm message is sent simultaneously on the top two reachable and available channels. The alarm is considered delivered upon receiving an ACK on either channel. This indicates the concurrency redundancy threshold.

9. A smart home collaborative alarm system implementing the method of any one of claims 1 to 8, characterized in that, include: The sensing layer, edge gateway, cloud collaboration platform, multi-channel access module, and hierarchical receiver; The perception layer includes various types of sensors deployed in each sub-area and on the personal devices of the monitored person. These various sensors are connected to the edge gateway via wired bus or low-power wireless link. The edge gateway includes a data acquisition and feature normalization module, a dynamic basic probability allocation generation module, a conflict adaptive suppression and evidence fusion module, an edge dual-threshold decision module, a cross-device verification request module, a final confidence synthesis module, a severity scoring module, and a multi-channel selection and retransmission control module, which are connected in sequence by signals. The edge gateway further includes a configuration storage module, which establishes readout connections with the dynamic basic probability allocation generation module, the edge dual threshold decision module, the cross-device verification request module, the final confidence synthesis module, the severity scoring module, and the multi-channel selection and retransmission control module, respectively. The output of the edge dual-threshold decision module is connected to the cross-device verification request module, the final confidence synthesis module, and the suspicious event log storage inside the edge gateway, respectively. The multi-channel access module includes at least two of the following: a Wi-Fi transceiver unit, a 4G / 5G transceiver unit, an NB-IoT transceiver unit, and a LoRa transceiver unit. The multi-channel selection and retransmission control module establishes control connections and ACK feedback connections with each of the wireless transceiver units. The cloud-based collaborative platform is communicatively connected to the edge gateway via the multi-channel access module. The cloud-based collaborative platform includes a secondary DS fusion service module, a cloud-based weight generation module, an evidence chain storage module, and a remote configuration distribution module. The input of the secondary DS fusion service module is connected to a supplementary device in the perception layer via the cross-device verification request module, and its output is connected to the final confidence synthesis module to distribute cloud-based confidence scores. The input of the cloud-based weight generation module is connected to both the secondary DS fusion service module and the link quality monitoring port of the cloud-based collaborative platform, and its output is connected to the final confidence synthesis module to distribute cloud-based weight coefficients. The remote configuration distribution module is connected to the configuration storage module via the multi-channel access module. The tiered receivers include a homeowner's smart terminal, a homeowner's voice / sound and light alarm, a property management / neighbor terminal, and an emergency linkage terminal. Each tiered receiver establishes an alarm receiving communication connection with the edge gateway and the cloud collaboration platform via the multi-channel access module. An ACK feedback transmission link is provided between the edge gateway and each tiered receiver.

10. The smart home collaborative alarm system according to claim 9, characterized in that: The data acquisition and feature normalization module includes a sensor access unit, a feature extraction unit, and an amplitude normalization and filtering unit. The input end of the sensor access unit is connected to the sensing layer, and the output end is connected to the feature extraction unit. The output end of the feature extraction unit is connected to the amplitude normalization and filtering unit. The feature extraction unit is used to extract the corresponding sensor feature values ​​from the raw data output by the sensing layer. The feature extraction parameters of the feature extraction unit are stored in the configuration storage module. The dynamic basic probability assignment generation module includes a membership degree calculation unit, a prior lookup table unit, and a BPA normalization unit. The input of the membership degree calculation unit is connected to the output of the data acquisition and feature normalization module, the input of the prior lookup table unit is connected to the configuration storage module, and the outputs of both the membership degree calculation unit and the prior lookup table unit are connected to the input of the BPA normalization unit. The conflict adaptive suppression and evidence fusion module includes a Jaccard matrix storage unit, a Jousselme distance calculation unit, a credibility weighting unit, a Dempster combination unit, and a single evidence / full conflict anomaly processing unit. The Jaccard matrix storage unit stores the Jaccard matrix on the power set focal element set of the recognition framework; the input of the Jousselme distance calculation unit is connected to the output of the dynamic basic probability allocation generation module and the Jaccard matrix storage unit respectively; the input of the confidence weighting unit is connected to the output of the Jousselme distance calculation unit; the input of the Dempster combination unit is connected to the output of the confidence weighting unit, and the output is connected to the edge dual threshold decision module; the single evidence / full conflict anomaly processing unit is bypassed to the Jousselme distance calculation unit and the confidence weighting unit respectively, and when the number of valid evidence is 1 or the sum of confidence is 0, the result of generating the representative evidence is directly bypassed to the Dempster combination unit; The cross-device verification request module has a built-in event type perception verification strategy table and a virtual sensor mapping table. The event type perception verification strategy table binds different supplementary equipment lists and verification windows for fire events, gas leak events, fall events, and intrusion events respectively. The virtual sensor mapping table stores the mapping relationship between the non-raw reading data returned by the supplementary equipment and normalized feature values, normalized judgment thresholds, and membership function steepness coefficients. The cross-device verification request module also includes a TTS response triggering subunit. The TTS response triggering subunit controls the connection to the user's voice / sound and light alarm, which is used to initiate a voice greeting to the person under supervision and collect the response status as supplementary evidence when the fall event is triggered by the cross-device verification branch of Zhongzhixin and send it back to the cloud collaboration platform. The input of the final confidence synthesis module is connected to the edge confidence output of the edge dual-threshold decision module, the cloud confidence output of the cloud collaboration platform, and the cloud weight coefficient output of the cloud collaboration platform, respectively. The output is connected to the severity scoring module. The final confidence synthesis module is used to weight the edge confidence and the cloud confidence according to the cloud weight coefficient to synthesize the final confidence, and to force the cloud weight coefficient to 0 when the cloud collaboration platform is unreachable. The severity scoring module includes an event type weight lookup table unit, a family vulnerability calculation unit, and a severity synthesis unit. The input of the family vulnerability calculation unit is connected to the family member information storage area in the edge gateway. The input of the severity synthesis unit is connected to the output of the event type weight lookup table unit, the output of the family vulnerability calculation unit, and the output of the final confidence synthesis module, respectively. The output is connected to the multi-channel selection and retransmission control module. The multi-channel selection and retransmission control module includes a channel utility scoring unit, a channel selection unit, a receiver-channel reachability filtering unit, an ACK monitoring unit, a retransmission scheduling unit, a local degradation triggering unit, and a concurrent redundancy triggering unit. The receiver-channel reachability filtering unit reads the receiver-channel reachability matrix from the self-configured storage module and filters reachable and usable wireless channels for each activated receiver for use by the channel utility scoring unit and the channel selection unit. The ACK monitoring unit is connected to the feedback ports of each wireless transceiver unit in the multi-channel access module to drive the channel utility scoring unit to update the weighted moving average of the transmission success rate index of each channel online. The output of the local degradation triggering unit controls the connection to the user's voice / sound and light alarm to trigger a local degradation alarm when all reachable and usable channels of the activated receiver fail consecutively for more than a preset limit. The input of the concurrent redundancy triggering unit is connected to the output of the severity scoring module, and the output is connected to the two reachable and usable wireless transceiver units with the highest utility scores in the multi-channel access module. The edge gateway communicates with the cloud collaboration platform via TLS or an equivalent encryption authentication mechanism. The evidence chain storage module forms a traceable evidence chain based on timestamps, sub-region numbers, sensor numbers, normalized feature values, dynamic basic probability allocation, confidence weights, Jousselme distance matrix, fusion confidence, channel selection, and ACK status. The traceable evidence chain does not contain the original audio and video data. For supplementary devices involving audio and video, only the existence detection results, motion feature vectors, or response status are stored.

Images