Multi-source alarm fusion emergency system linkage disposal method

Abstract

The application relates to a multi-source alarm fusion emergency system linkage disposal method, which comprises the following steps: acquiring multi-dimensional heterogeneous sensing data streams in a monitoring area and generating a basic evidence set object; calculating a global conflict quantitative index and performing adaptive confidence correction on high-conflict evidence; performing weighted fusion on the corrected evidence by using a dynamically updated sensor reliability weight vector to generate a target safety state probability; determining an alarm response level according to the probability and driving equipment cooperative response; and finally iteratively updating the sensor weight based on the execution feedback data. The application has the advantages that precise fusion determination of multi-source heterogeneous evidence under a complex interference environment is realized, and the system has the adaptive adjustment capability for sensor performance changes.

Description

Technical Field

[0001] This application relates to the field of building emergency safety, and in particular to a method for coordinated response of a multi-source alarm fusion emergency system. Background Technology

[0002] With the development of building intelligence technology, emergency systems, as crucial facilities for ensuring building safety, have received widespread attention regarding their reliability and accuracy. Current building safety monitoring technologies often rely on a single type of sensor for signal acquisition, such as smoke detectors or temperature sensors, and typically set fixed numerical thresholds as the basis for alarm triggering. This single-dimensional judgment method is less adaptable to complex environmental changes and is easily affected by environmental factors such as moisture, dust accumulation, or changes in lighting, leading to false alarms or missed alarms.

[0003] To improve detection accuracy, some technical solutions have begun to incorporate multiple types of sensors for monitoring. However, when processing multi-source heterogeneous evidence, data conflicts often occur between different sensors. For example, a smoke sensor may display abnormal values, while video surveillance equipment may not detect corresponding fire characteristics. Traditional evidence fusion algorithms often struggle to effectively eliminate the influence of interfering evidence when dealing with such highly conflicting evidence, easily producing fusion results that deviate significantly from the actual situation, resulting in insufficient accuracy in the final alarm determination.

[0004] Furthermore, existing multi-source fusion technologies typically employ static weight allocation strategies, assuming that all sensors maintain a constant reliability level throughout long-term operation. However, in real-world applications, the physical performance of sensors dynamically degrades with usage time, equipment aging, or environmental contamination. If the system cannot dynamically adjust the weight of sensors in the decision model based on their historical prediction accuracy, degraded sensors will still participate in the calculation with their original high weights, significantly reducing the overall accuracy of the system's judgment. Simultaneously, most existing handling processes lack closed-loop feedback mechanisms, failing to utilize subsequent equipment execution results to back-calibrate the front-end detection model, making it difficult for the system to adapt to constantly changing monitoring environments through self-iteration. Summary of the Invention

[0005] To achieve accurate fusion and judgment of multi-source heterogeneous evidence under complex interference environments, ensure that the system has the ability to adaptively adjust to changes in sensor performance, and build a closed-loop linkage system for mutual verification of perception, decision-making, and execution, this application provides a multi-source alarm fusion emergency system linkage response method.

[0006] This application provides a multi-source alarm fusion emergency system linkage response method, which adopts the following technical solution:

[0007] A method for coordinated response of a multi-source alarm fusion emergency system includes the following steps:

[0008] S1. Acquire multidimensional heterogeneous sensing data streams within the monitoring area, perform anomaly cleaning and feature mapping processing on the multidimensional heterogeneous sensing data streams, and generate a basic evidence set object;

[0009] S2. Input the basic evidence set object into the conflict detection and correction model, calculate the global conflict quantification index, and when the global conflict quantification index exceeds the preset conflict threshold, perform adaptive confidence correction on the basic evidence set object to generate a corrected evidence set object.

[0010] S3. Call the dynamic weight configuration model to obtain the current sensor reliability weight vector, and perform weighted fusion calculation on the corrected evidence set object based on the sensor reliability weight vector to generate the target safety status probability result;

[0011] S4. Map the target safety status probability result to the hierarchical linkage strategy model, match and determine the current alarm response level, generate a collaborative handling control command sequence, and drive the execution equipment group to complete the emergency response;

[0012] S5. Trace the execution feedback status data of the execution device group, construct a historical prediction accuracy evaluation index based on the execution feedback status data, and use an online optimization algorithm to update the sensor reliability weight vector in the dynamic weight configuration model based on the historical prediction accuracy evaluation index.

[0013] Optionally, step S1 includes the following sub-steps:

[0014] S11. Based on the heterogeneous sensor network deployed in the monitoring area, smoke concentration signals, ambient temperature signals, loop current status signals and video monitoring stream data are collected in parallel, and the smoke concentration signals, ambient temperature signals, loop current status signals and video monitoring stream data are aggregated into a multi-dimensional heterogeneous sensing data stream.

[0015] S12. Based on the statistical anomaly detection algorithm, outlier noise points in the multidimensional heterogeneous sensing data stream are removed, and the data after removing outlier noise points is normalized and standardized to generate standardized feature data.

[0016] S13. Based on the preset recognition framework, the standardized feature data are mapped to basic probability allocation functions corresponding to each heterogeneous sensor. The basic probability allocation functions include normal state confidence, fire alarm state confidence, fault state confidence and uncertain state confidence. The basic probability allocation functions are then aggregated to construct a basic evidence set object.

[0017] Optionally, step S2 includes the following sub-steps:

[0018] S21. Receive the basic evidence set object to extract the multiple individual pieces of evidence contained therein, calculate the Jaccard similarity coefficient between any two individual pieces of evidence, and determine the pairwise conflict coefficient based on the complement of the Jaccard similarity coefficient.

[0019] S22. Aggregate all paired conflict coefficients to calculate the global average conflict degree and use it as a global conflict quantification index. When the global average conflict degree exceeds the preset conflict threshold, a nonlinear correction factor is generated based on the global average conflict degree.

[0020] S23. Use a nonlinear correction factor to perform a downgrade operation on the confidence of the basic evidence set object, and redistribute the confidence loss caused by the downgrade operation to the uncertain state confidence to generate a corrected evidence set object.

[0021] Optionally, step S3 includes the following sub-steps:

[0022] S31. Read the current sensor reliability weight vector corresponding to each type of sensor from the dynamic weight configuration model;

[0023] S32. Use the current sensor reliability weight vector to perform a weighted average process on the basic probability allocation function of each individual piece of evidence in the modified evidence set object to generate a weighted evidence set;

[0024] S33. The DS evidence combination rule is used to perform orthogonal sum operation on the weighted evidence set to calculate the fusion confidence of normal state, fire alarm state, fault state and uncertain state within the identification framework, and the fusion confidence corresponding to the fire alarm state is extracted from the fusion confidence as the target safety state probability result.

[0025] Optionally, step S4 includes the following sub-steps:

[0026] S41. Analyze the target safety status probability results and extract the fire alarm status probability value;

[0027] S42. Compare the fire alarm status probability value with the preset four-level response interval threshold to determine the alarm response level, wherein the alarm response level includes a first-level emergency response level, a second-level rapid response level, a third-level routine response level, and a fourth-level monitoring response level.

[0028] S43. Based on the determined alarm response level, retrieve the corresponding device action combination and timing parameters from the preset strategy library, and generate a coordinated handling control instruction sequence that includes action type, execution parameters and priority.

[0029] Optionally, step S43 includes the following sub-steps:

[0030] When the alarm response level is Level 1 Emergency Response, the generated coordinated handling control instruction sequence includes a first instruction to drive the emergency lights to flash at a first preset frequency, and a second instruction to push alarm information to all personnel on the fire platform.

[0031] When the alarm response level is Level II rapid response, the generated coordinated handling control instruction sequence includes a third instruction to drive the emergency lights to flash at a second preset frequency, and a fourth instruction to push alarm information to the fire platform area.

[0032] When the alarm response level is Level 3 (normal response level), the generated coordinated handling control command sequence includes a fifth command that drives the emergency lights to flash at a third preset frequency; wherein, the first preset frequency is greater than the second preset frequency, and the second preset frequency is greater than the third preset frequency.

[0033] Optionally, step S5 includes the following sub-steps:

[0034] S51. Receive execution feedback status data from the execution device group, compare the execution feedback status data with the target safety status probability result, and generate a single prediction accuracy identifier;

[0035] S52. Store the single prediction accuracy identifier in a preset sliding window, calculate the historical average accuracy within the preset sliding window, and use the historical average accuracy as the historical prediction accuracy evaluation index.

[0036] S53. Calculate the error gradient based on the deviation between the historical prediction accuracy evaluation index and the current sensor reliability weight vector, introduce a regularization penalty term to constrain the error gradient, iteratively calculate the updated sensor reliability weight vector, and use the updated sensor reliability weight vector to cover the original parameters in the dynamic weight configuration model.

[0037] In summary, this application includes at least one of the following beneficial technical effects:

[0038] 1. This application constructs a conflict detection and adaptive correction mechanism based on similarity complementarity, which effectively quantifies the degree of conflict between multi-source heterogeneous evidence and performs confidence downgrading processing on highly conflicting evidence. This mechanism effectively avoids the synthesis bias problem that traditional fusion algorithms are prone to when processing highly conflicting evidence, significantly improves the alarm judgment accuracy of the system under complex environmental interference, and reduces the probability of false alarms and missed alarms.

[0039] 2. This application introduces a dynamic sensor weight configuration strategy based on online gradient optimization, which can continuously iteratively update the reliability weights of each sensor using historical prediction accuracy within a sliding window. This technique enables the system to adapt to changes in the monitoring environment and the degradation of sensor performance, avoiding the long-term accuracy decline problem caused by static weight allocation, thereby ensuring the stability and robustness of the multi-source fusion model throughout its entire lifecycle.

[0040] 3. This application establishes a closed-loop control system covering the entire process from perception fusion and decision execution to feedback optimization. By tracing the actual feedback status of the execution equipment to calibrate the front-end weight configuration model, the system achieves self-learning and iterative optimization. At the same time, the probability-based collaborative linkage strategy ensures accurate matching between emergency response measures and actual risk levels, effectively improving the collaborative efficiency and safety of building emergency response. Attached Figure Description

[0041] Figure 1 A flowchart illustrating a multi-source alarm fusion emergency system linkage response method in one embodiment of the present invention is shown. Detailed Implementation

[0042] The present application will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the application and are not intended to limit the scope of the application.

[0043] In the following description, numerous specific details are set forth for purposes of explanation in order to provide a thorough understanding of the inventive concept. As part of this specification, some of the accompanying drawings of this disclosure are block diagrams illustrating structures and devices to avoid complicating the disclosed principles. For clarity, not all features of the actual embodiment need to be described. Furthermore, the language used in this disclosure has been primarily chosen for readability and instructional purposes and may not have been chosen to define or limit the subject matter of the invention, thus requiring the necessary claims to determine such inventive subject matter. References to “an embodiment” or “an embodiment” in this disclosure mean that a particular feature, structure, or characteristic described in connection with that embodiment is included in at least one embodiment, and multiple references to “an embodiment” or “an embodiment” should not be construed as necessarily referring to the same embodiment.

[0044] This application discloses a method for coordinated response of a multi-source alarm fusion emergency system, referring to... Figure 1 This includes the following steps S1-S5.

[0045] S1. Acquire multidimensional heterogeneous sensing data streams within the monitoring area, perform anomaly cleaning and feature mapping processing on the multidimensional heterogeneous sensing data streams, and generate a basic evidence set object.

[0046] In step S1, the multidimensional heterogeneous sensing data stream refers to a collection of sensor signals originating from different physical properties, sampling frequencies, and communication protocols within the monitoring area. Specifically, the physical sources of this data stream include, but are not limited to, photoelectric smoke detectors for detecting particulate matter concentration in space, thermistor temperature sensors for detecting changes in ambient heat, current transformers for monitoring the electrical status of emergency lighting circuits, and video surveillance cameras for capturing real-time images of the scene. Because the data types output by these sensors cover continuous scalar data (such as temperature values ​​and concentration values), discrete state data (such as switching quantities), and unstructured streaming media data (such as video frames), and because the time bases and transmission rates of each sensor differ, this data stream exhibits significant heterogeneous characteristics.

[0047] The basic evidence set object is a standardized data structure generated based on the aforementioned heterogeneous data and adapted for subsequent evidence reasoning model operations. Logically, this object is defined as a data container containing multi-dimensional information, including a unified timestamp, unique sensor identifiers (IDs), normalized feature values, and a mapped initial confidence assignment vector. Generating this basic evidence set object unifies the original heterogeneous signals into mathematical evidence, facilitating subsequent conflict detection and fusion reasoning, and helping to reduce the impact of underlying hardware differences on upper-level algorithms.

[0048] Taking a fire monitoring scenario in a large commercial complex as an example, the system connects in real time to photoelectric smoke detectors, NTC thermistors, Hall current sensors in emergency lighting circuits, and high-definition network cameras deployed in key fire compartments. In step S1, the system first collects in parallel the simulated smoke concentration, ambient temperature, circuit leakage current data, and real-time monitoring video streams uploaded by the aforementioned devices. Subsequently, the system cleans these raw data to remove noise caused by transient electromagnetic interference and maps the cleaned data to a normalized standard feature space. Finally, the system outputs a basic evidence set object containing the current state confidence of each sensor (e.g., the confidence of a smoke sensor displaying a fire alarm is 0.85, the confidence of normal is 0.1, and the confidence of uncertainty is 0.05), which serves as the input data for subsequent fusion calculations.

[0049] Optionally, step S1 may include the following sub-steps S11-S13.

[0050] S11. Based on a heterogeneous sensor network deployed in the monitoring area, smoke concentration signals, ambient temperature signals, loop current status signals and video monitoring stream data are collected in parallel, and the smoke concentration signals, ambient temperature signals, loop current status signals and video monitoring stream data are aggregated into a multi-dimensional heterogeneous sensing data stream.

[0051] In step S11, the selection of each sensor in the heterogeneous sensor network and their communication methods directly determine the real-time performance and reliability of the system data. Specifically, the smoke concentration signal is usually acquired by a photoelectric smoke detector, which works by utilizing the principle of infrared light scattering by smoke particles and transmits the digital concentration value through the Modbus-RTU protocol on the RS-485 physical layer. The ambient temperature signal is acquired by an NTC thermistor sensor and is usually transmitted via Ethernet using the Modbus-TCP protocol to meet the needs of long-distance cabling. The loop current status signal comes from a Hall current sensor connected in series in the emergency lighting loop, used to monitor in real time whether there is an open circuit or short circuit fault in the loop, and its output is usually an analog voltage signal or a digital signal after A / D conversion. The video surveillance stream data is acquired by a high-definition network camera and outputs an H.264 / H.265 encoded video stream through RTSP (Real-Time Streaming Protocol) or ONVIF standard interface.

[0052] The engineering implementation of parallel acquisition and data aggregation relies on multi-threaded I / O multiplexing technology. Edge computing gateways or central servers establish multiple parallel data acquisition channels, each connecting to the interfaces of the different protocols mentioned above. Since the frame rate of video data (e.g., 25fps) is much higher than the polling frequency of environmental sensors (e.g., 1Hz), the system employs a timestamp-based alignment mechanism when aggregating the data into a multi-dimensional heterogeneous sensing data stream. That is, using high-frequency data as a benchmark or setting a uniform time slice (e.g., 500ms), various types of data arriving within the same time window are packaged to form a time synchronization vector containing "smoke-temperature-current-video features," ensuring that subsequent fusion calculations are based on the state at the same physical moment.

[0053] S12. Based on the statistical anomaly detection algorithm, outliers in the multidimensional heterogeneous sensing data stream are removed, and the data after removing outliers is normalized to generate standardized feature data.

[0054] In step S12, a statistical anomaly detection algorithm is used to address transient false alarms caused by sudden environmental changes or electromagnetic interference from the sensor. Normalization processing aims to eliminate the interference of different physical units on the fusion weights. Smoke concentration is measured in ppm or % / m, temperature in degrees Celsius, and video features typically have confidence levels between 0 and 1. For example, mapping temperatures from 0-100°C to decimals from 0 to 1 allows subsequent evidence reasoning to be compared and synthesized on a unified mathematical plane.

[0055] S13. Based on the preset recognition framework, the standardized feature data are mapped to basic probability allocation functions corresponding to each heterogeneous sensor. The basic probability allocation functions include normal state confidence, fire alarm state confidence, fault state confidence and uncertain state confidence. The basic probability allocation functions are then aggregated to construct a basic evidence set object.

[0056] In step S13, the identification framework is defined as {normal, fire alarm, fault, uncertain}. Here, "normal" represents a safe monitoring environment; "fire alarm" represents a clear indication of a fire; "fault" represents a sensor disconnection, readings exceeding the range, or an abnormal emergency lighting circuit; and "uncertain" represents a sensor reading in an ambiguous range or a state where interference makes a clear determination impossible.

[0057] S2. Input the basic evidence set object into the conflict detection and correction model, calculate the global conflict quantification index, and when the global conflict quantification index exceeds the preset conflict threshold, perform adaptive confidence correction on the basic evidence set object to generate a corrected evidence set object.

[0058] In step S2, due to the different mechanisms by which different sensors detect physical quantities, highly inconsistent judgment results may occur between sensors under specific interference environments. The core function of this model is to act as a filter before fusion, identifying and mitigating such source-end conflicts, thereby avoiding distortion of the synthesis results caused by directly fusing highly conflicting evidence.

[0059] The principle behind setting the preset conflict threshold is based on a balance between the system's tolerance for false alarms and its sensitivity. Taking a value of 0.7 as an example, this threshold constitutes the decision boundary between triggering the correction mechanism and direct fusion. When the calculated conflict index is below this threshold, the system considers the opinions of each sensor to be basically consistent and adopts direct fusion to preserve the integrity of the information; when the conflict index exceeds this threshold, the system determines that there is strong interference or equipment failure in the current environment, and the correction mechanism must be intervened.

[0060] Following the aforementioned fire monitoring scenario, suppose that at a certain moment, the evidence output by the photoelectric smoke detector shows a very high fire alarm confidence level, while at the same time, the video surveillance equipment, through image analysis, fails to extract flame or smoke features, and its output evidence shows a very high normal confidence level. In this case, the model will determine that there is a serious conflict between the two pieces of evidence and prevent them from being directly mathematically synthesized. Instead, it will enter a correction process to prevent the system from reaching an incorrect judgment.

[0061] Optionally, step S2 may include the following sub-steps S21-S23.

[0062] S21. Receive the basic evidence set object to extract the multiple individual pieces of evidence contained therein, calculate the Jaccard similarity coefficient between any two individual pieces of evidence, and determine the pairwise conflict coefficient based on the complement of the Jaccard similarity coefficient.

[0063] In step S21, the Jaccard similarity coefficient J is defined as the ratio of the intersection to the union of the probability assignment sets of two pieces of evidence. Compared to Euclidean distance, this coefficient is more sensitive to the degree of overlap of uncertain information between the evidence and can more accurately measure the consistency of the evidence. Specifically, the pairwise conflict coefficient = 1 - J. The larger the value of this coefficient, the more significant the discrepancy in the judgments of the two sensors regarding the state of the monitored object.

[0064] S22. Aggregate all paired conflict coefficients to calculate the global average conflict degree and use it as a global conflict quantification index. When the global average conflict degree exceeds the preset conflict threshold, a nonlinear correction factor is generated based on the global average conflict degree.

[0065] In step S22, the global average conflict degree is the arithmetic mean of the pairwise conflict coefficients between all individual pieces of evidence. This indicator reflects the dispersion of the current multi-source data. The logic for generating a nonlinear correction factor based on the global average conflict degree follows a negative correlation principle, that is, the higher the average conflict degree, the smaller the generated correction factor, and the greater the weakening effect on the evidence. The algorithm introduces a lower bound constraint based on a maximum value function to ensure that the correction factor is not lower than a preset minimum retention threshold, preventing the complete loss of effective information under extreme conflict conditions.

[0066] S23. Use a nonlinear correction factor to perform a downgrade operation on the confidence of the basic evidence set object, and redistribute the confidence loss caused by the downgrade operation to the uncertain state confidence to generate a corrected evidence set object.

[0067] In step S23, the confidence downgrading operation multiplies the original confidence values ​​of each deterministic state in the basic evidence set by a nonlinear correction factor, proportionally compressing the support strength of each piece of evidence for a specific state. To maintain the normalization of the basic probability allocation function, the confidence loss caused by downgrading is fully redistributed to the confidence of uncertain states. When the system detects high conflict, this approach proactively reduces reliance on a single deterministic conclusion and increases the evaluation weight of unknown states, helping to preserve decision-making leeway and avoid logical paradoxes.

[0068] Following the scenario of severe conflict between smoke detectors and video surveillance equipment, the specific calculation process is demonstrated as follows: First, the system calculates the Jaccard similarity coefficient between smoke detector evidence (high fire alarm confidence) and video surveillance evidence (high normal confidence). Since the states they support are mutually exclusive and have minimal overlap, the calculated similarity coefficient approaches zero. Second, the system calculates a pairwise conflict coefficient approaching one, causing the global average conflict level (e.g., 0.9) to exceed the preset threshold of 0.7. Third, the system generates a correction factor (e.g., 0.55) based on the high conflict level. Finally, the system performs a correction operation. The original confidence level of the smoke detector for the fire alarm state (0.8) is multiplied by 0.55, downgrading it to 0.44; similarly, the confidence level of the video equipment for the normal state is also reduced proportionally. The confidence level reduction during the operation is accumulated into the confidence level for uncertain states. In the final generated set of corrected evidence, the confidence levels for both fire alarms and normal states are significantly reduced, while the uncertainty is significantly increased, reducing the interference of high-conflict evidence on subsequent fusion calculations.

[0069] S3. Call the dynamic weight configuration model to obtain the current sensor reliability weight vector, and perform weighted fusion calculation on the corrected evidence set object based on the sensor reliability weight vector to generate the target safety status probability result.

[0070] In step S3, the dynamic weight configuration model serves as a persistent parameter storage and state maintenance unit within the system architecture, maintaining the current reliability score of each sensor in key-value pairs. This model acts as a bridge connecting the backend feedback mechanism and the frontend inference engine, ensuring that each fusion calculation utilizes parameters updated based on the latest historical performance. The sensor reliability weight vector is a vector composed of a set of normalized values, typically ranging from 0 to 1. Each element in this vector physically represents the reliability of the corresponding sensor in the current environment; a higher value indicates higher accuracy and stronger anti-interference capability in past predictions. By weighting the evidence using this weight vector, the performance differences between heterogeneous sensors due to variations in aging, installation location, or detection principles are essentially differentiated at the mathematical level.

[0071] The target safety status probability result is the final decision-making basis output after weighted and orthogonal fusion of multi-source information. Specifically, it is represented by the fusion confidence value of the specific state "fire alarm" in the identification framework. This result compresses multi-dimensional sensor data into a single, quantitative probability index, which serves as the sole input variable for the subsequent hierarchical linkage strategy model to make logical judgments.

[0072] Continuing with the previous scenario, assume that the photoelectric smoke detector is assigned a higher reliability weight (e.g., 0.8) due to its long-term stable operation, while the video surveillance equipment is assigned a lower weight (e.g., 0.4) due to recent frequent light interference. During weighted fusion calculations, although the two may provide conflicting evidence, the smoke detector's dominant weight will lead to a "fire alarm" conclusion, significantly suppressing the influence of the video equipment's "normal" conclusion on the final result. This mechanism ensures that the system's decisions favor sources with better historical performance.

[0073] Optionally, step S3 includes the following sub-steps S31-S33.

[0074] S31. Read the current sensor reliability weight vector corresponding to each type of sensor from the dynamic weight configuration model.

[0075] In step S31, the system uses the unique sensor identifier carried in the basic evidence set object to perform an index query in the dynamic weight configuration model, accurately extracts the real-time weight values ​​corresponding to each sensor participating in the current fusion calculation, and combines them to form a sensor reliability weight vector.

[0076] S32. Use the current sensor reliability weight vector to perform a weighted average process on the basic probability allocation function of each individual piece of evidence in the modified evidence set object to generate a weighted evidence set.

[0077] In step S32, the logic of the weighted average processing is to use the reliability weight of each sensor as a correction coefficient and apply it to its corresponding basic probability allocation function. In specific calculations, the system uses the weight values ​​to reconstruct the reliability of each state in the evidence body, so that the reliability distribution of high-weight evidence is strengthened before synthesis, while the reliability distribution of low-weight evidence is diluted.

[0078] S33. The DS evidence combination rule is used to perform orthogonal sum operation on the weighted evidence set to calculate the fusion confidence of normal state, fire alarm state, fault state and uncertain state within the identification framework, and the fusion confidence corresponding to the fire alarm state is extracted from the fusion confidence as the target safety state probability result.

[0079] In step S33, the core algorithm principle of the DS evidence combination rule is to synthesize the weighted multi-source evidence using an orthogonal summation formula. This algorithm calculates the sum of the products of all possible focal elements within the identification framework for each piece of evidence, and handles the remaining conflict components by introducing a normalization constant (i.e., 1 minus the reciprocal of the conflict coefficient K), ensuring that the sum of the fusion confidence scores in the final output is 1. The fusion confidence score physically represents the system's final degree of certainty regarding whether the current environment is in a "normal," "fire alarm," "fault," or "uncertain" state after comprehensively considering all sensor data and reliability weights. The fusion confidence score for the "fire alarm state" is extracted separately as the probability result of the target safety state because, in emergency response systems, the fire alarm probability is a core key indicator for triggering graded responses and driving physical equipment actions.

[0080] Based on a numerical example, assuming that after correction in S2, the corrected evidence from the smoke sensor is m1={fire alarm: 0.4, normal: 0.1, fault: 0.3, uncertain: 0.2}, and the corrected evidence from the video device is m2={fire alarm: 0.1, normal: 0.4, fault: 0.4, uncertain: 0.1}, the weights read at this point are W1=0.8 (high confidence) and W2=0.5 (medium confidence), respectively. First, in S32, the system uses the weights to weight the two, giving the smoke sensor evidence a larger proportion in the synthesized sequence. Then, in S33, the DS orthogonal sum formula is applied for calculation. Due to the high weight of the smoke sensor, the confidence of the "fire alarm" state in the fused result is significantly improved; for example, the calculated fusion confidence is m3={fire alarm: 0.65, normal: 0.2, fault: 0.1, uncertain: 0.05}. Finally, the system extracts the value 0.65 as the target safety state probability result and sends it to the subsequent classification module.

[0081] S4. Map the target safety status probability result to the hierarchical linkage strategy model, match and determine the current alarm response level, generate a collaborative handling control command sequence, and drive the execution equipment group to complete the emergency response.

[0082] In step S4, the hierarchical linkage strategy model constructs a mapping logic between continuous probability values ​​and discrete physical actions, serving as the core unit for system decision output. Based on the range of the input probability value, the model determines the degree of risk faced by the current environment through table lookup or logical judgment.

[0083] Classifying alarm response levels can address the technical shortcomings of the traditional one-size-fits-all response model in emergency systems. In the traditional model, regardless of the probability of a fire, an alarm is triggered across the entire system, easily causing unnecessary panic or wasting operational resources on low-risk false alarms. By establishing a tiered response mechanism, the system can match the appropriate level of response based on the certainty of the risk, achieving precise resource allocation that prioritizes monitoring in low-probability situations and focuses on full-scale response in high-probability situations.

[0084] The coordinated control command sequence is a structured data packet generated by the system to guide the actions of underlying devices. The data structure of this sequence includes a unique identifier for the target device, the specific action command type (e.g., activate, flash, push), execution parameters (e.g., flashing frequency value, push range list), and timing priorities (e.g., prioritize activating the audible and visual alarm, followed by activating the smoke exhaust fan). This structured command format ensures logical rigor and timing accuracy when multiple devices work collaboratively.

[0085] The physical implementation of the drive execution device group is the process of converting the above logical instructions into physical control signals. For field hardware devices (such as emergency lighting controllers), the system sends Modbus write register instructions through the industrial fieldbus to directly modify the value of the device control word to change its operating status; for software platforms (such as fire protection operation and maintenance centers), the system pushes the encapsulated alarm information packets to the remote server by calling RESTful API interfaces or MQTT message queues to achieve cross-system linkage response.

[0086] Optionally, step S4 includes the following sub-steps S41-S43.

[0087] S41. Analyze the target safety status probability results and extract the fire alarm status probability value.

[0088] In step S41, the data parsing logic reads key fields from the target safety status probability results, removes the confidence components for non-fire alarms (such as faults or normal conditions), and accurately extracts a single floating-point value representing the fire risk. This process ensures that subsequent comparison operations are based on normalized and clearly defined risk probabilities.

[0089] S42. Compare the fire alarm status probability value with the preset four-level response interval threshold to determine the alarm response level, wherein the alarm response level includes a first-level emergency response level, a second-level rapid response level, a third-level routine response level, and a fourth-level monitoring response level.

[0090] In step S42, the numerical division of the four response levels constitutes the quantitative standard for the system's hierarchical decision-making. For example, the system sets a probability value between 0.8 and 1.0 as the Level 1 emergency response level, representing a very high certainty of fire occurrence, requiring immediate evacuation; a probability value between 0.6 and 0.8 as the Level 2 rapid response level, representing significant fire characteristics but still with uncertainty, requiring rapid manual verification; a probability value between 0.4 and 0.6 as the Level 3 routine response level, representing an abnormal environment but not reaching the alarm threshold, requiring continuous monitoring; and below 0.4 as the Level 4 monitoring response level, maintaining routine inspections.

[0091] S43. Based on the determined alarm response level, retrieve the corresponding device action combination and timing parameters from the preset strategy library, and generate a coordinated handling control instruction sequence that includes action type, execution parameters and priority.

[0092] In step S43, a pre-set strategy library stores standard response plans corresponding to different levels. The system uses a lookup table to retrieve the equipment action combinations and timing parameters corresponding to the current level, transforming abstract level instructions into specific control logic. For example, "equipment action combinations" define which equipment needs to intervene, and "timing parameters" define the order of intervention (e.g., first cutting off non-fire-fighting power, then turning on emergency lighting). Simultaneously, the priority definition ensures that in the event of multiple concurrent conflicting instructions, control instructions involving life safety (e.g., initiating smoke extraction) are executed and scheduled with priority over auxiliary notification instructions (e.g., sending SMS messages).

[0093] The calculations are explained using numerical examples. Assuming the fire alarm probability value output in the preceding steps is 0.85, falling within the preset range of [0.8, 1.0], the system determines the current emergency response level to be Level 1. If the output probability value is 0.5, falling within the range of [0.4, 0.6], the system determines it to be a Level 3 conventional response. This hard-determination logic based on numerical ranges ensures the determinism and consistency of the system's response.

[0094] Optionally, step S43 includes the following sub-steps:

[0095] When the alarm response level is Level 1 Emergency Response, the generated coordinated handling control instruction sequence includes a first instruction to drive the emergency lights to flash at a first preset frequency, and a second instruction to push alarm information to all personnel on the fire platform.

[0096] When the alarm response level is Level II rapid response, the generated coordinated handling control instruction sequence includes a third instruction to drive the emergency lights to flash at a second preset frequency, and a fourth instruction to push alarm information to the fire platform area.

[0097] When the alarm response level is Level 3 (normal response level), the generated coordinated handling control command sequence includes a fifth command that drives the emergency lights to flash at a third preset frequency; wherein, the first preset frequency is greater than the second preset frequency, and the second preset frequency is greater than the third preset frequency.

[0098] For Level 1 emergency response, which corresponds to an extremely high fire risk, the first preset frequency (e.g., 4Hz) generated by the system is a high-frequency strobe. This strong visual stimulus can quickly penetrate smoke and raise people's alertness. Combined with the alarm information pushed to all personnel on the fire platform, it achieves full-coverage emergency mobilization and ensures that all relevant personnel immediately enter evacuation or rescue mode.

[0099] For Level 2 rapid response, which corresponds to medium to high risk, the system uses a second preset frequency (e.g., 1Hz) to flash at a medium speed. This is intended to alert on-site personnel to abnormal environmental conditions and simultaneously push alarm information to the fire platform area. Only the security personnel responsible for that area are notified to go to the scene to "confirm the fire," thus avoiding panic among all personnel.

[0100] For Level 3 routine response, this state corresponds to low risk or equipment warning. The system uses a slow flashing at a third preset frequency (e.g., 0.5Hz) to serve as a prompt warning, informing inspection personnel that there are data fluctuations at this location and a routine check is required.

[0101] The relative order of the three preset frequencies (first > second > third) visually establishes an intuitive gradient of crisis. High frequencies correspond to high urgency, while low frequencies correspond to low attention, allowing on-site personnel to intuitively perceive the current level of danger simply by observing the speed of the flashing lights, without needing to check specific data.

[0102] Based on the actual scenario of building fire evacuation, when the probability value is 0.85 and a Level 1 response is triggered, the emergency lights flash at a high frequency of 4Hz, and the people on site are visually impacted and immediately take escape actions; while when the probability value is 0.5 and a Level 3 response is triggered, the emergency lights only flash slowly at 0.5Hz, and the maintenance personnel passing by will realize that the equipment at this point needs to be checked, while ordinary office staff will not run around in a disorderly manner, thus achieving precise guidance for emergency behavior.

[0103] S5. Trace the execution feedback status data of the execution device group, construct a historical prediction accuracy evaluation index based on the execution feedback status data, and use an online optimization algorithm to update the sensor reliability weight vector in the dynamic weight configuration model based on the historical prediction accuracy evaluation index.

[0104] In step S5, the execution feedback status data serves as the truth value basis for the verification system's judgment of correctness. Its acquisition path primarily originates from manual confirmation signals from the fire control room or reset signals from the on-site handling terminal. For example, security personnel press the confirmation button after confirming a fire on-site, or press the reset button after confirming a false alarm. The physical signals generated by these operations are transmitted back by the system and marked as the true label of that alarm event. Updating the sensor reliability weight vector enables the system to adaptively calibrate against sensor aging drift, lens contamination, or changes in environmental background noise, ensuring that the evaluation criteria of the fusion model always remain consistent with the current actual performance of the equipment.

[0105] Taking a complete closed-loop cycle as an example: Suppose a photoelectric smoke detector triggers an alarm due to water vapor interference (the system determines the fire alarm probability to be 0.85), and the linkage system immediately activates the audible and visual alarms. Security personnel arrive at the scene and confirm that no fire has occurred, then input a "false alarm report" command into the maintenance terminal. The system captures this negative feedback data, identifies that the sensor has generated false high-confidence evidence in the current environment, and then triggers an online optimization algorithm to automatically reduce the sensor's reliability weight in the dynamic weight configuration model, thus reducing its decision-making influence in subsequent similar scenarios.

[0106] Optionally, step S5 includes the following sub-steps S51-S53.

[0107] S51. Receive execution feedback status data from the execution device group, compare the execution feedback status data with the target safety status probability result, and generate a single prediction accuracy identifier.

[0108] In step S51, the comparison logic is implemented using a threshold determination method. The system defines that when the target safety status probability result shows a high risk (e.g., greater than 0.8) and the execution feedback status data confirms "real fire alarm," or when the probability result shows a low risk (e.g., less than 0.4) and the execution feedback data confirms "no fire alarm," the prediction is considered successful, and a single prediction accuracy flag with a value of 1 is generated; otherwise, if a missed or false alarm occurs, a flag with a value of 0 is generated.

[0109] S52. Store the single prediction accuracy identifier in a preset sliding window, calculate the historical average accuracy within the preset sliding window, and use the historical average accuracy as the historical prediction accuracy evaluation index.

[0110] In step S52, a preset sliding window (e.g., with a capacity of the most recent 20 events) serves as a cache container for historical data, balancing the impact of "long-term memory" and "recent trends" on the evaluation metrics. The historical average accuracy is obtained by calculating the arithmetic mean of the markers within the window. This metric filters out the interference of single, occasional random errors and sensitively reflects recent trend changes in sensor performance.

[0111] S53. Calculate the error gradient based on the deviation between the historical prediction accuracy evaluation index and the current sensor reliability weight vector, introduce a regularization penalty term to constrain the error gradient, iteratively calculate the updated sensor reliability weight vector, and use the updated sensor reliability weight vector to cover the original parameters in the dynamic weight configuration model.

[0112] In step S53, the formula for calculating the error gradient is typically set as the historical average accuracy minus the current sensor reliability weight. The sign and magnitude of this gradient value indicate the direction and step size of the weight adjustment: if the accuracy is higher than the current weight, the gradient is positive, driving the weight to increase; conversely, it drives the weight to decrease. Introducing a regularization penalty term (such as L2 regularization) to constrain the error gradient can prevent drastic oscillations in the weight values ​​due to overfitting historical data, ensuring the smoothness of model updates.

[0113] The specific update formula for iterative calculation typically employs the gradient ascent method, i.e.: Updated weight = Current weight + Learning rate × (Error gradient - Regularization term). The system uses the calculated new weights to overwrite the original parameters in the dynamic weight configuration model, thus completing the real-time evolution of the model parameters.

[0114] A specific numerical calculation illustrates the weight update process: Assuming a sensor's current reliability weight is 0.5, its historical average accuracy after sliding window statistics is 0.8, the learning rate is set to 0.1, and regularization is ignored. The system first calculates the error gradient as 0.3 (0.8 - 0.5); then calculates the update step size as 0.03 (0.1 × 0.3); finally, the weight is updated to 0.53. Through multiple iterations of this type, the sensor's weight will gradually approach its true accuracy level of 0.8, thus achieving dynamic convergence and calibration of the weights.

[0115] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0116] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0117] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for coordinated response of a multi-source alarm fusion emergency system, characterized in that, Includes the following steps: S1. Acquire multidimensional heterogeneous sensing data streams within the monitoring area, perform anomaly cleaning and feature mapping processing on the multidimensional heterogeneous sensing data streams, and generate a basic evidence set object; S2. Input the basic evidence set object into the conflict detection and correction model, calculate the global conflict quantification index, and when the global conflict quantification index exceeds the preset conflict threshold, perform adaptive confidence correction on the basic evidence set object to generate a corrected evidence set object. S3. Call the dynamic weight configuration model to obtain the current sensor reliability weight vector, and perform weighted fusion calculation on the corrected evidence set object based on the sensor reliability weight vector to generate the target safety status probability result; S4. Map the target safety status probability result to the hierarchical linkage strategy model, match and determine the current alarm response level, generate a collaborative handling control command sequence, and drive the execution equipment group to complete the emergency response; S5. Trace the execution feedback status data of the execution equipment group, construct a historical prediction accuracy evaluation index based on the execution feedback status data, and use an online optimization algorithm to update the sensor reliability weight vector in the dynamic weight configuration model based on the historical prediction accuracy evaluation index. Step S2 includes the following sub-steps: S21. Receive the basic evidence set object to extract the multiple individual pieces of evidence contained therein, calculate the Jaccard similarity coefficient between any two individual pieces of evidence, and determine the pairwise conflict coefficient based on the complement of the Jaccard similarity coefficient. S22. Aggregate all paired conflict coefficients to calculate the global average conflict degree and use it as a global conflict quantification index. When the global average conflict degree exceeds the preset conflict threshold, a nonlinear correction factor is generated based on the global average conflict degree. S23. Use a nonlinear correction factor to perform a downgrade operation on the confidence of the basic evidence set object, and redistribute the confidence loss caused by the downgrade operation to the uncertain state confidence to generate a corrected evidence set object; Step S5 includes the following sub-steps: S51. Receive execution feedback status data from the execution device group, compare the execution feedback status data with the target safety status probability result, and generate a single prediction accuracy identifier; S52. Store the single prediction accuracy identifier in a preset sliding window, calculate the historical average accuracy within the preset sliding window, and use the historical average accuracy as the historical prediction accuracy evaluation index. S53. Calculate the error gradient based on the deviation between the historical prediction accuracy evaluation index and the current sensor reliability weight vector, introduce a regularization penalty term to constrain the error gradient, iteratively calculate the updated sensor reliability weight vector, and use the updated sensor reliability weight vector to cover the original parameters in the dynamic weight configuration model.

2. The multi-source alarm fusion emergency system linkage response method according to claim 1, characterized in that, Step S1 includes the following sub-steps: S11. Based on the heterogeneous sensor network deployed in the monitoring area, smoke concentration signals, ambient temperature signals, loop current status signals and video monitoring stream data are collected in parallel, and the smoke concentration signals, ambient temperature signals, loop current status signals and video monitoring stream data are aggregated into a multi-dimensional heterogeneous sensing data stream. S12. Based on the statistical anomaly detection algorithm, outlier noise points in the multidimensional heterogeneous sensing data stream are removed, and the data after removing outlier noise points is normalized and standardized to generate standardized feature data. S13. Based on the preset recognition framework, the standardized feature data are mapped to basic probability allocation functions corresponding to each heterogeneous sensor. The basic probability allocation functions include normal state confidence, fire alarm state confidence, fault state confidence and uncertain state confidence. The basic probability allocation functions are then aggregated to construct a basic evidence set object.

3. The multi-source alarm fusion emergency system linkage response method according to claim 1, characterized in that, Step S3 includes the following sub-steps: S31. Read the current sensor reliability weight vector corresponding to each type of sensor from the dynamic weight configuration model; S32. Use the current sensor reliability weight vector to perform a weighted average process on the basic probability allocation function of each individual piece of evidence in the modified evidence set object to generate a weighted evidence set; S33. The DS evidence combination rule is used to perform orthogonal sum operation on the weighted evidence set to calculate the fusion confidence of normal state, fire alarm state, fault state and uncertain state within the identification framework, and the fusion confidence corresponding to the fire alarm state is extracted from the fusion confidence as the target safety state probability result.

4. The multi-source alarm fusion emergency system linkage response method according to claim 3, characterized in that, Step S4 includes the following sub-steps: S41. Analyze the target safety status probability results and extract the fire alarm status probability value; S42. Compare the fire alarm status probability value with the preset four-level response interval threshold to determine the alarm response level, wherein the alarm response level includes a first-level emergency response level, a second-level rapid response level, a third-level routine response level, and a fourth-level monitoring response level. S43. Based on the determined alarm response level, retrieve the corresponding device action combination and timing parameters from the preset strategy library, and generate a coordinated handling control instruction sequence that includes action type, execution parameters and priority.

5. The multi-source alarm fusion emergency system linkage response method according to claim 4, characterized in that, Step S43 includes the following sub-steps: When the alarm response level is Level 1 Emergency Response, the generated coordinated handling control instruction sequence includes a first instruction to drive the emergency lights to flash at a first preset frequency, and a second instruction to push alarm information to all personnel on the fire platform. When the alarm response level is Level II rapid response, the generated coordinated handling control instruction sequence includes a third instruction to drive the emergency lights to flash at a second preset frequency, and a fourth instruction to push alarm information to the fire platform area. When the alarm response level is Level 3 (normal response level), the generated coordinated handling control command sequence includes a fifth command that drives the emergency lights to flash at a third preset frequency; wherein, the first preset frequency is greater than the second preset frequency, and the second preset frequency is greater than the third preset frequency.

6. The multi-source alarm fusion emergency system linkage response method according to claim 1, characterized in that, In the step of generating a nonlinear correction factor based on the global average conflict degree, a lower limit constraint based on the maximum value function is introduced to ensure that the generated nonlinear correction factor is not lower than the preset minimum retention threshold.

7. The multi-source alarm fusion emergency system linkage response method according to claim 1, characterized in that, The step of calculating the error gradient based on the deviation between the historical prediction accuracy evaluation index and the current sensor reliability weight vector includes: The error gradient is obtained by subtracting the current sensor reliability weight from the historical average accuracy. The iterative calculation yields an updated sensor reliability weight vector, which is executed using the gradient ascent method. The updated sensor reliability weight vector is equal to the current sensor reliability weight vector plus the product of the learning rate and the difference between the error gradient and the regularization term.

8. The multi-source alarm fusion emergency system linkage response method according to claim 1, characterized in that, S51 includes: When the target safety status probability result shows high risk and the execution feedback status data confirms a real fire alarm, or when the target safety status probability result shows low risk and the execution feedback status data confirms no fire alarm, a single prediction accuracy indicator with a value of 1 is generated. When a missed or false alarm occurs, a single prediction accuracy indicator with a value of 0 is generated.

Images