Multi-parameter fusion type fire early warning method and system based on environmental perception

Abstract

The invention relates to the field of fire early warning, and particularly discloses a multi-parameter fusion type fire early warning method and system based on environmental perception, and the method comprises the steps: carrying out the internal correlation and significance analysis of the time series data of a plurality of sensors through a dynamic fusion module; sensor multi-parameter significant fusion coding information capable of highlighting a core fire symptom is generated, so that the problem of importance dynamic change of each sensor in different scenes is solved; and secondly, combining the highly concentrated sensor fusion feature with the independently coded environment context feature to form comprehensive environment-multi-parameter state time sequence combined coding information. Risk assessment is carried out on the joint coding information, and the model can interpret a sensor signal in combination with a specific environment (such as weather and position), so that a real fire risk and normal background interference are accurately distinguished, and finally high-accuracy and high-adaptability fire early warning is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fire warning, and more particularly, to a multi-parameter fusion type fire warning method and system based on environmental perception. BACKGROUND

[0002] Fire is one of the main disasters that threaten public safety and cause property loss. Traditional fire detectors, such as independent smoke or temperature detectors, although play a role in early warning to some extent, their working principle is single, and only rely on whether a single physical quantity (such as smoke concentration or ambient temperature) reaches a preset threshold to trigger an alarm. This way has obvious limitations in complex environments, for example, in the kitchen, smoking area and other scenes, cooking oil fume or cigarette smoke is easy to trigger false alarms of smoke detectors; while in the early stage of fire, some smoldering fires may produce a large amount of toxic gases (such as carbon monoxide) but smoke and temperature changes are not obvious, resulting in a lag in response of temperature or smoke detectors, missing the best evacuation and rescue opportunity.

[0003] In the development process of multi-parameter fire warning technology, it is attempted to fuse multiple sensor data, such as temperature, humidity, carbon monoxide concentration, particulate matter concentration, etc., in order to obtain a more comprehensive characterization of the fire state. However, the existing multi-parameter fusion method often uses simple feature splicing or static weighting strategy when processing multi-source heterogeneous data. For example, preset fixed weight coefficients for different types of sensors are weighted and summed, or the time series data of each sensor is directly spliced into a long vector and input into the model. Such methods have significant technical defects: first, static weighting cannot adapt to the dynamically changing fire scene. The importance of each sensor is dynamically changing in different fire stages and different environments. For example, in the early stage of fire, a small increase in carbon monoxide concentration may be more indicative than temperature change; while in the flammable goods warehouse, the anomaly of volatile organic compound index may be more worthy of attention than other parameters. Static weight cannot capture this dynamic importance difference, resulting in poor model generalization ability. Second, simple feature splicing ignores the key influence of environmental context on the interpretation of sensor data. For example, in the hot summer afternoon, higher ambient temperature is a normal phenomenon, but if in the cold winter night, the same temperature reading may indicate an anomaly. The existing model fails to effectively take into account environmental factors such as light, weather, building type, etc., resulting in a deviation in the interpretation of sensor data, thereby reducing the robustness and accuracy of the warning.

[0004] Therefore, an optimized multi-parameter fusion type fire warning scheme is expected. SUMMARY

[0005] In order to solve the above technical problems, the present application is proposed. The embodiments of the present application provide a multi-parameter fusion type fire warning method and system based on environmental perception.

[0006] According to an aspect of the present application, a multi-parameter fusion fire warning method based on environment perception is provided, which comprises: acquiring a time sequence of an original sensor stream and a time sequence of a context data stream; arranging the time sequence of the original sensor stream into a sensor data time sequence feature matrix according to a parameter sample dimension and a time dimension, and inputting the sensor data time sequence feature matrix into a sequence encoder based on an LSTM model to obtain a set of sensor data time sequence pattern feature encoding vectors; performing multi-parameter fusion on the set of sensor data time sequence pattern feature encoding vectors to obtain a sensor multi-parameter significant fusion encoding vector; performing sequence encoding on the time sequence of the context data stream based on an LSTM model to obtain a context time sequence pattern feature encoding vector; concatenating the sensor multi-parameter significant fusion encoding vector and the context time sequence pattern feature encoding vector to obtain an environment-multi-parameter state time sequence joint encoding vector; calculating a fire risk score based on the environment-multi-parameter state time sequence joint encoding vector; determining whether to generate an alarm signal based on a comparison between the fire risk score and a preset threshold.

[0007] According to another aspect of the present application, a multi-parameter fusion fire warning system based on environment perception is provided, which comprises: an environment data acquisition module, configured to acquire a time sequence of an original sensor stream and a time sequence of a context data stream; a sensor data time sequence encoding module, configured to arrange the time sequence of the original sensor stream into a sensor data time sequence feature matrix according to a parameter sample dimension and a time dimension, and input the sensor data time sequence feature matrix into a sequence encoder based on an LSTM model to obtain a set of sensor data time sequence pattern feature encoding vectors; a multi-parameter sensor data time sequence feature fusion module, configured to perform multi-parameter fusion on the set of sensor data time sequence pattern feature encoding vectors to obtain a sensor multi-parameter significant fusion encoding vector; a context data stream time sequence encoding module, configured to perform sequence encoding on the time sequence of the context data stream based on an LSTM model to obtain a context time sequence pattern feature encoding vector; an environment-multi-parameter state time sequence joint encoding module, configured to concatenate the sensor multi-parameter significant fusion encoding vector and the context time sequence pattern feature encoding vector to obtain an environment-multi-parameter state time sequence joint encoding vector; a fire risk assessment module, configured to calculate a fire risk score based on the environment-multi-parameter state time sequence joint encoding vector; An alarm module is configured to determine whether to generate an alarm signal based on a comparison between the fire risk score and a preset threshold.

[0008] Compared with the prior art, the multi-parameter fusion fire warning method and system based on environment perception provided by the application abandon the simple feature weighting or splicing method of the traditional method, and the key lies in two-step deep fusion. First, a dynamic fusion module is used to analyze the internal correlation and significance of the time series data of multiple sensors, and generate a sensor multi-parameter significant fusion encoding information which can highlight the core fire signs, which solves the problem of dynamic change of importance of each sensor in different scenes. Second, the highly concentrated sensor fusion feature is combined with the independently encoded environment context feature to form a comprehensive environment-multiple parameter state time series joint encoding information. Through risk assessment on the joint encoding information, the model can combine the specific environment (such as weather, location) to interpret the sensor signal, so as to accurately distinguish the real fire risk from the normal background interference, and finally realize the fire warning with high accuracy and strong adaptability. BRIEF DESCRIPTION OF DRAWINGS

[0009] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application, when taken in conjunction with the accompanying drawings. The drawings provided in the disclosure serve to provide a further understanding of the embodiments of the present application, and constitute a part of the specification, which together with the application, serve to explain the present application, and do not constitute a limitation on the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 A flowchart of the multi-parameter fusion fire warning method based on environment perception according to the embodiments of the present application; Figure 2 A data flow diagram of the multi-parameter fusion fire warning method based on environment perception according to the embodiments of the present application; Figure 3 A flowchart of the multi-parameter fusion fire warning method based on environment perception according to the embodiments of the present application, which performs multi-parameter fusion on the set of sensor data time series pattern feature encoding vectors to obtain a sensor multi-parameter significant fusion encoding vector; Figure 4 A flowchart of the multi-parameter fusion fire warning method based on environment perception according to the embodiments of the present application, which performs context correlation importance analysis on the set of sensor data time series pattern feature encoding vectors to obtain a set of sensor data time series pattern feature context correlation attention weights; Figure 5 A block diagram of the multi-parameter fusion fire warning system based on environment perception according to the embodiments of the present application. DETAILED DESCRIPTION

[0011] In the following, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of embodiments of the present application, and thus do not limit the present application, and concretely, the present application can be implemented in various different ways. This application is not also limited to the described examples, but rather, a variety of changes in form and details can be made thereto.

[0012] As used in this application and in the claims, the terms "comprises", "comprising", "includes", "including" or the like can have the meaning ascribed to them in US patent law and can mean the inclusion of strict ingredients or steps or components "not" exclusive of other not specified ingredients, steps or components.

[0013] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or server. The modules are merely illustrative, and different aspects of the system and method can use different modules.

[0014] Flow diagrams are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the operations previously or hereinafter are not necessarily performed in the exact order as shown. Rather, various steps can be handled in reverse order, or at the same time, as desired. Other operations can also be added to, or removed from, these processes, or one or more steps can be removed from these processes.

[0015] In the following, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of embodiments of the present application, and thus do not limit the present application, and concretely, the present application can be implemented in various different ways. This application is not also limited to the described examples, but rather, a variety of changes in form and details can be made thereto.

[0016] In view of the technical problems of low warning accuracy and poor robustness caused by the static and simple fusion strategy and the neglect of environmental context in the existing fire warning technology, in the technical solution of the present application, a multi-parameter fusion type fire warning method based on environmental perception is proposed. The method first obtains two types of key data in parallel: one type is the original sensor stream directly reflecting the fire signs, such as temperature, carbon monoxide concentration, etc.; the other type is the context data stream describing the surrounding environment, such as light intensity, personnel activity state, etc. Subsequently, the long short-term memory network is used to deeply encode the time series data of each sensor to capture its dynamic change pattern and form a set of initial feature vectors. In particular, the present scheme does not directly splice these features, but designs a multi-parameter fusion module. The module evaluates the prominence of individual sensor signals through internal self-importance analysis, and simultaneously uses context correlation analysis based on the converter architecture to mine the mutual influence and synergy between different sensors. By dynamically aggregating these two importance weights, a highly condensed sensor multi-parameter significant fusion encoding vector is generated, which can represent the comprehensive fire risk. At the same time, the environmental context data stream is also independently time-encoded. Finally, the sensor fusion vector containing the core fire signs and the context feature vector representing the environmental state are spliced to form a comprehensive environment-multi-parameter state time series joint encoding, and based on this, the final fire risk score is calculated. After comparing with the preset threshold, it is decided whether to trigger the alarm. In this way, the warning decision considers not only the inherent intensity and relevance of the fire signal, but also the environmental background in which it is located, thereby achieving accurate, dynamic and high-robustness evaluation of fire risk.

[0017] Figure 1 Flow chart of the multi-parameter fusion type fire warning method based on environmental perception according to the embodiment of the present application. Figure 2 Data flow diagram of the multi-parameter fusion type fire warning method based on environmental perception according to the embodiment of the present application. As Figure 1 and Figure 2As shown, the environment-aware multi-parameter fusion fire warning method according to the embodiments of the present application comprises the following steps: S100, acquiring a time sequence of an original sensor stream and a time sequence of a context data stream; S200, arranging the time sequence of the original sensor stream into a sensor data time sequence feature matrix according to a parameter sample dimension and a time dimension, and then inputting the sensor data time sequence feature matrix into a sequence encoder based on an LSTM model to obtain a set of sensor data time sequence mode feature encoding vectors; S300, performing multi-parameter fusion on the set of sensor data time sequence mode feature encoding vectors to obtain a sensor multi-parameter significant fusion encoding vector; S400, performing sequence encoding on the time sequence of the context data stream based on an LSTM model to obtain a context time sequence mode feature encoding vector; S500, performing feature splicing on the sensor multi-parameter significant fusion encoding vector and the context time sequence mode feature encoding vector to obtain an environment-multi-parameter state time sequence joint encoding vector; S600, calculating a fire risk score based on the environment-multi-parameter state time sequence joint encoding vector; and S700, determining whether to generate an alarm signal based on a comparison between the fire risk score and a preset threshold.

[0018] Specifically, in step S100, a time sequence of an original sensor stream and a time sequence of a context data stream are acquired. In particular, in one specific example of the present application, the original sensor stream comprises temperature, humidity, carbon monoxide concentration, fine particulate matter concentration, original infrared signal, wind speed, wind direction, and volatile organic compound index; and the context data stream comprises illumination intensity, whether it is raining, personnel activity state, building material, location type, altitude, and main hazard source type. It should be understood that, due to the inherent ambiguity of single sensor data when it is decoupled from the specific environmental background, for example, a high temperature reading may be caused by a fire, or may be caused by only sunshine or normal equipment operation, which leads to the fact that it is extremely easy to produce false positives or false negatives when only relying on sensor data to determine a fire. Therefore, in the technical solution of the present application, the time sequence of the original sensor stream directly representing the physical and chemical characteristics of a fire, and the time sequence of the context data stream describing the static and dynamic environmental characteristics of a monitoring area are further acquired synchronously, so as to construct a comprehensive and multi-dimensional state information space, and provide a complete data basis for subsequent deep fusion analysis. In this way, it is ensured from the source that the basis for model decision-making is sufficient and has situational awareness, and a solid foundation is laid for subsequent accurate differentiation between real fire risk and environmental interference, and improvement of the overall accuracy and robustness of the warning system.

[0019] More specifically, in a concrete example of this application, the implementation process of this acquisition step includes the deployment of the data acquisition terminal and data aggregation and transmission. First, in the target area to be monitored, a composite detector or distributed sensor network integrating multiple sensors is deployed. This network includes thermocouples or thermistors for measuring ambient temperature, humidity sensors for detecting humidity, electrochemical sensors for detecting carbon monoxide concentration, laser scattering sensors for monitoring fine particulate matter concentration, infrared detectors for capturing thermal radiation anomalies, ultrasonic or mechanical anemometers for acquiring wind speed and direction, and volatile organic compound (VOC) index sensors for monitoring combustible gases or vapors. These sensors continuously acquire data at a preset sampling frequency and convert the measurements into timestamped digital signals, forming a time series of the raw sensor stream. Second, contextual data is acquired in parallel. Meteorological information such as light intensity and rainfall status is obtained through on-site deployed photosensitive sensors, rain sensors, or by calling third-party meteorological service application interfaces. Personnel activity status is determined through passive infrared sensors, video analytics, or access control system logs. Relatively static information such as building materials, location type, altitude, and main hazard types are input and stored as configuration parameters during system initialization, or imported from databases such as Building Information Modeling (BIM). All this dynamic and static contextual information is also formatted as a time-series data stream with timestamps and transmitted to subsequent processing modules along with the raw sensor stream.

[0020] Specifically, in step S200, the time series of the original sensor stream is arranged into a sensor data time series feature matrix according to the parameter sample dimension and the time dimension, and then input into a sequence encoder based on an LSTM model to obtain a set of sensor data time series pattern feature encoding vectors. It should be understood that since the original sensor data exists as instantaneous values ​​at discrete time points, it does not directly reflect key information such as dynamic trends, rates of change, and potential periodicity over time. These time series dependencies are precisely the core basis for distinguishing abnormal patterns from normal environmental fluctuations in the early stages of a fire. Therefore, in the technical solution of this application, the time series data of each sensor is further constructed into a structured feature matrix, and a Long Short-Term Memory (LSTM) network model is used to sequence encode it, thereby deeply mining and learning the long-term and short-term temporal correlations inherent in each sensor data stream. In this way, the original, high-dimensional time series data stream can be transformed into a low-dimensional, concise feature encoding vector set rich in dynamic pattern information, providing high-quality input for subsequent intelligent fusion analysis and effectively improving the model's ability to capture complex time series patterns.

[0021] More specifically, in one specific example of the present application, the encoding process first proceeds with data preprocessing and matrix construction. Specifically, for each kind of raw sensor stream obtained, such as temperature, carbon monoxide concentration, etc., a fixed length T time window is selected, and T consecutive sampling points within the window are taken as an independent sequence sample. The values within the window are normalized, for example using min-max normalization, to eliminate the differences in dimension and value range between different sensor parameters. Subsequently, the N (N is the number of sensor types) normalized sequences of length T are arranged into a sensor data time series feature matrix. Then, each processed sensor sequence is input into a pre-constructed and trained LSTM model-based sequence encoder. The encoder contains one or more LSTM layers, and the gating mechanism (input gate, forget gate, output gate) inside can selectively remember and forget the information in the sequence. When a sensor sequence of length T is input into the LSTM encoder step by step, the encoder will constantly update its internal hidden state. Finally, the hidden state vector at the last time step (T time) is taken as the overall representation of the sensor sequence, i.e. the sensor data time series pattern feature encoding vector. Repeat this process for all N sensor streams, and a set containing N encoding vectors is obtained, each of which condenses the core dynamic time series pattern of the corresponding sensor in the past T time steps.

[0022] Specifically, in step S300, the set of sensor data time series pattern feature encoding vectors is subjected to multi-parameter fusion to obtain a sensor multi-parameter significant fusion encoding vector. It should be understood that, in the set of sensor data time series pattern feature encoding vectors after time series encoding, the indication of each vector to the fire risk is not constant. The importance of one sensor parameter is dual and dynamic: it depends on the mutation degree of its own signal, i.e., the intrinsic significance, such as a sharp rise in temperature; and it depends on its cooperative relationship with other sensor parameter change patterns, i.e., context dependence, such as a slight increase in temperature accompanied by a simultaneous increase in carbon monoxide and fine particulate matter concentration. The traditional fusion method cannot decouple and effectively weigh the two importance, resulting in incomplete characterization of the overall fire situation and easy misjudgment. Therefore, in the technical solution of the present application, the set of sensor data time series pattern feature encoding vectors is further subjected to multi-parameter fusion to construct a mechanism that can dynamically evaluate and aggregate the value of multi-source information. This fusion process evaluates the importance of each sensor data time series pattern feature encoding vector from two dimensions in parallel: first, the intrinsic significance is quickly captured by analyzing the peak value and other features of the vector itself; second, the context-dependent value of the vector in the current multi-parameter combination is evaluated from a global perspective by using a converter architecture-based encoder to deeply mine the correlation between all vectors. Subsequently, the importance weights obtained from the two different perspectives are intelligently fused, and the set of original sensor data time series pattern feature encoding vectors is weighted and summed accordingly. In this way, a sensor multi-parameter significant fusion encoding vector with high information density and strong representation ability can be generated, which can highlight key single abnormal signals and not ignore weak but key fire signs under the synergistic action of multiple parameters, thereby providing a more accurate and robust feature basis for subsequent risk assessment.

[0023] Figure 3 The flowchart of the multi-parameter fusion of the set of sensor data time series pattern feature encoding vectors to obtain a sensor multi-parameter significant fusion encoding vector according to the environment perception-based multi-parameter fusion fire warning method of the embodiments of the present application. As shown in FIG. 3, the process includes the following steps: Figure 3As shown, step S300 includes: S310, performing intrinsic self-importance analysis on the set of sensor data temporal pattern feature encoding vectors to obtain a set of sensor data temporal pattern feature attention weights; S320, performing context-related importance analysis on the set of sensor data temporal pattern feature encoding vectors to obtain a set of sensor data temporal pattern feature context-related attention weights; S330, dynamically aggregating the set of sensor data temporal pattern feature encoding vectors based on the set of sensor data temporal pattern feature attention weights and the set of sensor data temporal pattern feature context-related attention weights to obtain the sensor multi-parameter significant fusion encoding vector.

[0024] Specifically, in this embodiment, step S310, performing an intrinsic self-importance analysis on the set of sensor data time-series pattern feature encoding vectors to obtain a set of sensor data time-series pattern feature attention weights, includes: extracting the peak values ​​of each sensor data time-series pattern feature encoding vector in the set of sensor data time-series pattern feature encoding vectors to obtain a set of sensor data time-series pattern significant feature values; and performing a normalization operation on the set of sensor data time-series pattern significant feature values ​​to obtain the set of sensor data time-series pattern feature attention weights.

[0025] Accordingly, the peak values ​​of each sensor data time-series pattern feature encoding vector in the set of sensor data time-series pattern feature encoding vectors are extracted to obtain the set of significant feature values ​​of the sensor data time-series pattern, expressed by the following formula: ; ; in, It is the set of time-series pattern feature encoding vectors of the sensor data. These are respectively the 1st, 2nd, and 3rd elements in the set of sensor data temporal pattern feature encoding vectors. The and the first Time-series pattern feature encoding vector of sensor data To extract the peak element from the vector, It is the set of significant feature values ​​of the time-series pattern of sensor data.

[0026] It can be understood that, since the sensor data time sequence mode feature encoding vector obtained after time sequence encoding is a high-dimensional complex representation, direct fusion processing has large calculation amount and it is difficult to intuitively measure the emergency degree or signal strength of each sensor mode. In the fire warning scene, a parameter with a sharp change, such as a steep rise in temperature, itself contains very high risk information, and this inherent and independent saliency needs to be quickly and effectively captured. Therefore, in the technical solution of the present application, the peak element, i.e. the maximum value, of each vector in the set of sensor data time sequence mode feature encoding vectors is further extracted as a direct proxy index for measuring the inherent saliency of each sensor time sequence mode. In this way, a quantified salient feature value representing the signal strength of each sensor (such as temperature, carbon monoxide, etc.) dynamic mode can be assigned in a highly efficient manner, thereby providing a preliminary judgment basis for the subsequent attention weight allocation based on the importance of the local perspective and self-contained information.

[0027] Correspondingly, a normalization operation is performed on the set of sensor data time sequence mode salient feature values to obtain the set of sensor data time sequence mode feature attention weights, which is expressed by the following formula: ; wherein, is an activation function, is the i-th element in the set of sensor data time sequence mode salient feature values, is the total number of elements in the set of sensor data time sequence mode salient feature values, is the i-th element in the set of sensor data time sequence mode salient feature values, is the total number of elements in the set of sensor data time sequence mode salient feature values, is the total number of elements in the set of sensor data time sequence mode salient feature values, is the total number of elements in the set of sensor data time sequence mode salient feature values, is the set of sensor data time sequence mode feature attention weights.

[0028] It should be appreciated that, due to the fact that the extracted sensor data time series pattern salient feature value set from the sensor data time series pattern feature encoding vector has different numerical values and physical dimensions, it cannot be directly compared horizontally to judge the relative importance of different sensor signals. For example, the peak value of temperature and the peak value of carbon monoxide concentration are not comparable in numerical value, and directly using these original peak values for weighting will introduce bias. Therefore, in the technical solution of the present application, further normalization operation is performed on the salient feature value set, for example, the Softmax function is applied, so as to convert these discrete and heterogeneous salient values into a probability distribution with a total sum of 1. In this way, a set of sensor data time series pattern feature attention weights can be generated, and each weight value in the set clearly and quantitatively represents the proportion of attention that the corresponding sensor time series pattern should obtain due to its inherent signal strength, providing a local saliency-based, standardized importance measure for subsequent feature aggregation, so that the model can perform differential weighting processing according to the prominence of each parameter signal.

[0029] Figure 4 The flowchart of the context-related importance analysis of the set of sensor data time series pattern feature encoding vectors to obtain the set of sensor data time series pattern feature context-related attention weights according to the environment perception-based multi-parameter fusion fire warning method of the embodiments of the present application is shown in FIG. 8. As shown in FIG. 8, step S320 includes: S321, inputting the set of sensor data time series pattern feature encoding vectors into a context encoder based on a transformer architecture to obtain a set of sensor data time series pattern feature context-related encoding vectors; S322, extracting the peak value of each sensor data time series pattern feature context-related encoding vector in the set of sensor data time series pattern feature context-related encoding vectors to obtain a set of sensor data time series pattern feature context-related representation salient values; and S323, performing normalization operation on the set of sensor data time series pattern feature context-related representation salient values to obtain the set of sensor data time series pattern feature context-related attention weights. Figure 4

[0030] Accordingly, step S321 inputs the set of sensor data time series pattern feature encoding vectors into a context encoder based on a transformer architecture to obtain a set of sensor data time series pattern feature context-related encoding vectors, which can be expressed by the following formula: ; wherein, the context encoder based on a transformer architecture is the set of sensor data time series pattern feature context-related encoding vectors is ​the first, the second, the third and the fourth sensor data time-series pattern feature context-related encoding vectors in the set of sensor data time-series pattern feature context-related encoding vectors, respectively. sensor data time-series pattern feature context-related encoding vectors.

[0031] It can be understood that, due to the judgment of the intrinsic significance of a single sensor time-series pattern, the synergistic effect and complex correlation of different sensor parameters in the development process of a fire are ignored. For example, a weak temperature rising signal itself may be insignificant, but if it is accompanied by a simultaneous slight increase in carbon monoxide and fine particulate matter concentration, the risk weight indicating a fire will increase dramatically. Such a group context-based dependency relationship cannot be captured by analyzing each feature vector independently. Therefore, in the technical solution of the present application, the complete set of sensor data time-series pattern feature encoding vectors is further input into a context encoder based on a transformer architecture, so as to utilize the core self-attention mechanism to penetrate the surface information of the features and deeply model the complex and nonlinear long-distance dependency relationship between the various sensor data time-series patterns. In this way, each isolated sensor feature vector can be converted into a sensor data time-series pattern feature context-related encoding vector that has undergone deep context information interaction and contains global relationships. This vector not only represents the dynamic pattern of each sensor data time-series pattern itself, but more importantly, encodes the role and value it plays in the overall situation of the current set of sensors, providing a global perspective for evaluating its true importance in a specific fire scenario.

[0032] Correspondingly, in steps S322 and S323, the peak values of each sensor data time-series pattern feature context-related encoding vector in the set of sensor data time-series pattern feature context-related encoding vectors are extracted to obtain a set of sensor data time-series pattern feature context-related representation saliency values, and a normalization operation is performed on the set of sensor data time-series pattern feature context-related representation saliency values to obtain a set of sensor data time-series pattern feature context-related attention weights, which can be represented by the following formula: ; ; wherein, is the set of sensor data time-series pattern feature context-related representation saliency values, is the set of sensor data time-series pattern feature context-related attention weights.

[0033] ​It should be appreciated that, since the sensor data time series pattern feature context-related encoding vector obtained after the converter encoder processing contains rich global correlation information, it is still a high-dimensional feature representation and cannot be directly used to quantify the relative importance of each sensor in the current complex situation. A clear mechanism is needed to extract scalar weights representing its strategic value from these context-rich vectors. Therefore, in the technical solution of the present application, firstly, the peak value of each sensor data time series pattern feature context-related encoding vector is extracted to obtain a set of sensor data time series pattern feature context-related significant values, and then a normalization operation is performed on the set of significant values to convert the complex feature vector with global information interaction into a standardized attention weight distribution that can be directly compared. In this way, a set of sensor data time series pattern feature context-related attention weights can be generated, and each weight value in the set no longer reflects the isolated intensity of the sensor signal, but accurately quantifies its importance in the current multi-parameter coordinated fire scene, providing an indispensable global dependency perspective for the final feature fusion.

[0034] Specifically, in the embodiment of the present application, step S330, based on the set of sensor data time series pattern feature attention weights and the set of sensor data time series pattern feature context-related attention weights, dynamically aggregating the set of sensor data time series pattern feature encoding vectors to obtain the sensor multi-parameter significant fusion encoding vector, includes: fusing the set of sensor data time series pattern feature context-related attention weights and the set of sensor data time series pattern feature attention weights to obtain a set of sensor data time series pattern hybrid attention weight values; based on the set of sensor data time series pattern hybrid attention weight values, calculating the weighted sum of the set of sensor data time series pattern feature encoding vectors to obtain the sensor multi-parameter significant fusion encoding vector.

[0035] Correspondingly, the set of sensor data time series pattern hybrid attention weight values obtained by fusing the set of sensor data time series pattern feature context-related attention weights and the set of sensor data time series pattern feature attention weights is expressed as follows: ; wherein, and are learnable scaling factors, controlling the linear mixing ratio and the gating strength, and are radiometric transformation parameters, wherein, is a radiometric transformation weight matrix, is a radiometric transformation bias vector, for element-wise multiplication, for function, for vector concatenation, a set of sensor data time-series pattern hybrid attention weight values.

[0036] It should be understood that, since there are limitations in relying on any single dimension of attention weight alone, only focusing on the intrinsic saliency of the sensor signal itself may ignore the early fire pattern formed by the synergistic effect of multiple parameters, although the signal is weak but indicative; and completely relying on the context association between parameters may not be fast enough or the weight distribution is insufficient when facing an overwhelming and explicit dangerous signal of a certain parameter. Therefore, in the technical solution of the present application, the sensor data time-series pattern feature attention weight derived from the local perspective and the sensor data time-series pattern feature context association attention weight derived from the global perspective are effectively fused to construct a more comprehensive and robust hybrid attention weight system that takes into account both important dimensions. In this way, a set of sensor data time-series pattern hybrid attention weight values can be generated, which can ensure that the model gives sufficient attention to those star features with extremely high signal strength (such as temperature sudden rise), and does not ignore those parameters that play a key role in a specific fire scene combination (such as the synergistic change of humidity and volatile organic compound index), thus achieving the final and most accurate dynamic evaluation of the value of each sensor information.

[0037] Correspondingly, based on the set of sensor data time-series pattern hybrid attention weight values, a weighted sum of the set of sensor data time-series pattern feature encoding vectors is calculated to obtain the sensor multi-parameter salient fusion encoding vector, which is represented by the following formula: ; wherein, is the i-th sensor data time-series pattern feature encoding vector in the set of sensor data time-series pattern feature encoding vectors, is the i-th sensor data time-series pattern feature encoding vector in the set of sensor data time-series pattern feature encoding vectors, is the i-th sensor data time-series pattern hybrid attention weight value in the set of sensor data time-series pattern hybrid attention weight values, is the i-th sensor data time-series pattern hybrid attention weight value in the set of sensor data time-series pattern hybrid attention weight values, is the sensor multi-parameter salient fusion encoding vector.

[0038] It should be understood that, since the previous steps have respectively generated a set of feature encoding vectors representing the dynamic patterns of each sensor and a set of mixed attention weights quantifying the overall importance of these patterns in a specific scenario, but both are still discrete information units, and have not yet formed a unified overall situation representation that can be directly used for final risk judgment. Therefore, in the technical solution of the present application, further based on the set of mixed attention weight values of the sensor data time series pattern, the set of sensor data time series pattern feature encoding vectors is executed weighted sum operation, so as to effectively act the abstract importance score on the specific feature information, thereby aggregating and compressing the dispersed multi-dimensional feature information according to its dynamically evaluated importance. In this way, the final sensor multi-parameter significant fusion encoding vector can be generated, which is a highly condensed single feature representation, the internal structure and numerical distribution of which are dominated by the sensor features with higher mixed attention weights, thereby realizing the focusing and amplification of key fire signs at the information level, providing a high information density and strong discriminability input for the downstream risk assessment, greatly simplifying the final decision-making process and improving its accuracy.

[0039] Specifically, in step S400, the time series of the context data stream is sequentially encoded based on the LSTM model to obtain a context time series pattern feature encoding vector. It should be understood that, since the original context data stream, whether it is a dynamically changing light intensity, a personnel activity state, or a static building material, a location type and other information, is in the form of discrete time series or isolated parameters, and has not formed a feature representation that can uniformly represent the overall state of the environment and its dynamic evolution trend. If not handled, these key background information will not be effectively combined with the deeply fused sensor features, thereby weakening the situational awareness ability of the final decision. Therefore, in the technical solution of the present application, the time series of the context data stream is further sequentially encoded based on the LSTM model, so as to capture and learn the internal patterns of environmental factors changing over time and the potential associations between different static background factors. In this way, the multi-source and heterogeneous context information stream can be converted into a unified and condensed context time series pattern feature encoding vector. This vector not only provides a format-compatible input for the subsequent splicing of the sensor multi-parameter significant fusion encoding vector, but more importantly, it injects key scene understanding ability into the entire early warning system, enabling the final fire risk assessment to be performed in an environment that is fully understood, thereby greatly improving the accuracy and reliability of the early warning decision.

[0040] Specifically, in step S500, the sensor multi-parameter significant fusion encoding vector and the context time sequence mode feature encoding vector are spliced to obtain an environment-multi-parameter state time sequence joint encoding vector. It should be understood that, since the preceding steps have respectively generated the sensor multi-parameter significant fusion encoding vector representing the physical and chemical signs of the fire and the context time sequence mode feature encoding vector describing the monitoring environment background, but the two are still independent of each other in the information level, and have not yet formed a unified view that allows the decision model to simultaneously observe the event itself and its occurrence environment, which limits the comprehensive judgment ability of the model on the fire risk in a complex scene. Therefore, in the technical solution of the present application, the sensor multi-parameter significant fusion encoding vector and the context time sequence mode feature encoding vector are further spliced to integrate the direct fire physical evidence and the indirect environmental context information in the feature space, and to construct a comprehensive and multi-dimensional state descriptor. In this way, an environment-multi-parameter state time sequence joint encoding vector is generated, which provides a complete and unambiguous input for the final fire risk assessment model, so that it can learn and understand the real meaning of a specific sensor reading in a specific environmental background, thereby fundamentally improving the accuracy and context adaptability of the early warning decision.

[0041] It should be understood that, since the sensor multi-parameter significant fusion encoding vector and the context time sequence mode feature encoding vector obtained after independent encoding and fusion respectively represent the physical and chemical characterization of the fire event and the environmental background in which it occurs, the two parts of information are logically separated. If the two are not effectively integrated, the subsequent risk assessment model will not be able to learn the complex interaction between a specific sensor signal pattern and a specific environmental background, for example, it will not be able to distinguish between the similar signals of high temperature and particulate matter generated by normal cooking in the kitchen and real fire, thereby leading to one-sidedness of the decision and high false alarm rate. Therefore, in the technical solution of the present application, the sensor multi-parameter significant fusion encoding vector and the context time sequence mode feature encoding vector are further spliced to construct a unified and comprehensive high-dimensional feature space that can represent both the event itself and its occurrence environment. In this way, an environment-multi-parameter state time sequence joint encoding vector is generated, which places the sensor information directly representing the fire risk and the environmental information providing the basis for discrimination in the same feature dimension, providing a complete and context-aware input for the final risk assessment model, so that it can learn and understand the real meaning of a specific sensor reading in a specific environmental background, thereby enabling the model to learn the deep rules of what kind of environmental and what kind of sensor signal combination really means fire, thereby fundamentally improving the accuracy and context adaptability of the early warning decision.

[0042] More specifically, in one specific example of the present application, the feature concatenation step is performed in a designated feature fusion module. First, the module receives two input vectors: one is a sensor multi-parameter salient fusion encoding vector of dimension D_s, which is the result of weighted aggregation by hybrid attention mechanism, condensing the core dynamic information of all sensor data within the current time window; the other is a context temporal pattern feature encoding vector of dimension D_c, which is the result of LSTM encoding, representing the overall state and change trend of the monitored environment. Then, a vector concatenation operation is performed, which in calculation is represented as concatenating the two vectors head to tail to form a longer vector. For example, all elements of the context temporal pattern feature encoding vector are sequentially appended to the end of the sensor multi-parameter salient fusion encoding vector. Finally, an environment-multi-parameter state temporal joint encoding vector of dimension (D_s + D_c) is generated. The first D_s elements of the joint encoding vector completely retain the physical and chemical characteristics of the fire event, while the last D_c elements completely retain the environmental background characteristics in which it is located, and the two are seamlessly integrated in the new vector space, laying a solid data foundation for subsequent accurate risk calculation.

[0043] Specifically, in step S600, a fire risk score is calculated based on the environment-multi-parameter state temporal joint encoding vector. It should be understood that since the environment-multi-parameter state temporal joint encoding vector is a high-dimensional, complex feature representation that contains both event physical characteristics and environmental background information, it is not a direct risk measure that can be used for decision-making. The system needs a final decision mechanism to convert this complex feature vector into a single, explicit quantitative indicator to achieve automated and standardized alarm triggering. Therefore, in the technical solution of the present application, a fire risk score is further calculated based on the joint encoding vector to perform a nonlinear mapping from the high-dimensional feature space to the one-dimensional risk scale, and all fused information is finally condensed into a numerical value representing the current fire possibility or danger level. In this way, a clear and unambiguous final decision basis can be provided for the entire warning system, so that the subsequent alarm decision can be based on a continuous and comparable scalar value, thereby greatly simplifying the decision logic and ensuring its objectivity and consistency.

[0044] More specifically, in one specific example of the present application, the calculation of the fire risk score is accomplished in a pre-trained risk assessment module, which is structurally a fully connected neural network. First, the risk assessment module receives the environment-multi-parameter state time-series joint encoding vector of dimension (D_s + D_c) as its only input. Then, the input vector is fed into one or more hidden layers of the network. Each hidden layer is composed of a number of neurons, which perform a linear transformation on the features output by the previous layer, followed by a non-linear activation function, such as the rectified linear unit (ReLU), to learn the complex, non-linear dependencies between the elements of the joint encoding vector. Finally, the output layer of the network is designed to contain only one neuron, with a Sigmoid activation function. The Sigmoid function compresses the value passed from the previous layer into the interval of 0 to 1, and its output is the final fire risk score. The score intuitively represents the probability of a fire occurring in the current state.

[0045] Specifically, in step S700, it is determined whether to generate an alarm signal based on a comparison of the fire risk score and a preset threshold. It should be understood that, since the fire risk score generated by the foregoing steps is a continuous scalar between 0 and 1, which only represents the degree of risk possibility and cannot directly constitute an executable alarm action, the system needs an explicit triggering rule to convert this probabilistic output into a determined operation instruction. Therefore, in the technical solution of the present application, the fire risk score is further compared with a preset alarm threshold to establish an objective and standardized decision boundary for the system. In this way, the complex and continuous risk assessment result can be finally converted into a discrete and binary decision instruction, so that the entire early warning system can automatically and explicitly trigger the subsequent alarm process according to the quantified risk level, ensuring the timeliness and reliability of the fire response.

[0046] More specifically, in the embodiments of the present application, based on the comparison of the fire risk score and the preset threshold, it is determined whether to generate an alarm signal, including: in response to the fire risk score being greater than the preset threshold, it is determined to generate an alarm signal. That is, specifically, first, the calculated fire risk score is received in real time from the upstream risk assessment module, and the score is a floating point number. At the same time, the unit reads a preset alarm threshold from its internal storage or the configuration system. The threshold is a key configurable parameter, and its specific value is set according to the characteristics of the deployment environment and the balance requirements of the early warning sensitivity and false alarm rate, for example, a lower threshold is set in high-risk areas such as flammable product warehouses, and a higher threshold is set in areas prone to interference signals such as kitchens. Then, the core comparison operation is performed, that is, it is determined whether the fire risk score is greater than the preset threshold. Finally, based on the comparison of the fire risk score and the preset threshold, it is determined whether to generate an alarm signal, specifically, in response to the fire risk score being greater than the preset threshold, a high level or a specific alarm instruction code is output to determine to generate an alarm signal, which is transmitted to downstream execution devices such as audible and visual alarms, automatic dialing systems or central monitoring platforms, thereby triggering actual alarm actions; otherwise, if the score is not greater than the threshold, a low level or a silent state is output, and the system continues to monitor and evaluate at the next time step.

[0047] In summary, according to the multi-parameter fusion fire warning method based on environment perception according to the embodiments of the present application, the traditional simple feature weighting or splicing method is abandoned, and the key lies in two-step deep fusion: first, a dynamic fusion module is used to analyze the internal correlation and significance of the time series data of multiple sensors, and generate a sensor multi-parameter significant fusion coding information that can highlight the core fire signs, which solves the problem of dynamic change of importance of each sensor in different scenes; second, the highly concentrated sensor fusion feature is combined with the independently coded environment context feature to form a comprehensive environment-multi-parameter state time series joint coding information. By performing risk assessment on this joint coding information, the model can combine the specific environment (such as weather, location) to interpret the sensor signal, thereby accurately distinguishing between real fire risk and normal background interference, and finally realizing high accuracy and strong adaptability of fire warning.

[0048] Further, a multi-parameter fusion fire warning system based on environment perception is also provided.

[0049] Figure 5 A block diagram of the multi-parameter fusion fire warning system based on environment perception according to the embodiments of the present application is shown in FIG. 1. As shown in FIG. 1, the system includes a risk assessment module 101, a dynamic fusion module 102, an environment context coding module 103, a risk evaluation module 104, and an alarm signal generation module 105. Figure 5As shown, the environment-aware multi-parameter fusion fire warning system 100 according to the embodiment of the present application comprises: an environment data acquisition module 110, configured to acquire a time sequence of an original sensor stream and a time sequence of a context data stream; a sensor data time sequence encoding module 120, configured to arrange the time sequence of the original sensor stream according to a parameter sample dimension and a time dimension into a sensor data time sequence feature matrix, and input the sensor data time sequence feature matrix into an LSTM model-based sequence encoder to obtain a set of sensor data time sequence mode feature encoding vectors; a multi-parameter sensor data time sequence feature fusion module 130, configured to perform multi-parameter fusion on the set of sensor data time sequence mode feature encoding vectors to obtain a sensor multi-parameter significant fusion encoding vector; a context data stream time sequence encoding module 140, configured to perform LSTM model-based sequence encoding on the time sequence of the context data stream to obtain a context time sequence mode feature encoding vector; an environment-multi-parameter state time sequence joint encoding module 150, configured to splice the sensor multi-parameter significant fusion encoding vector and the context time sequence mode feature encoding vector to obtain an environment-multi-parameter state time sequence joint encoding vector; a fire risk assessment module 160, configured to calculate a fire risk score based on the environment-multi-parameter state time sequence joint encoding vector; and an alarm module 170, configured to determine whether to generate an alarm signal based on a comparison between the fire risk score and a preset threshold.

[0050] As described above, the environment-aware multi-parameter fusion fire warning system 100 according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with an environment-aware multi-parameter fusion fire warning algorithm, and the like. In one possible implementation, the environment-aware multi-parameter fusion fire warning system 100 according to the embodiment of the present application can be integrated into a wireless terminal as a software module and / or a hardware module. For example, the environment-aware multi-parameter fusion fire warning system 100 can be a software module in an operating system of the wireless terminal, or can be an application program developed for the wireless terminal; of course, the environment-aware multi-parameter fusion fire warning system 100 can also be one of many hardware modules of the wireless terminal.

[0051] The above has described the embodiments of the present disclosure, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes are obvious to those skilled in the art without departing from the scope and spirit of the described embodiments. The selection of the terms used herein is intended to best explain the principles, practical applications, or improvements to the technology in the market of the embodiments, or to enable other ordinary skilled persons in the art to understand the embodiments disclosed herein.

Claims

1. A multi-parameter fusion-based fire early warning method based on environmental perception, characterized in that, include: Acquire the time series of the raw sensor stream and the time series of the context data stream; After arranging the time series of the original sensor stream into a sensor data time series feature matrix according to the parameter sample dimension and the time dimension, it is input into a sequence encoder based on the LSTM model to obtain a set of sensor data time series pattern feature encoding vectors. The set of temporal pattern feature encoding vectors of the sensor data is fused using multiple parameters to obtain a sensor multi-parameter significant fused encoding vector; The time series of the context data stream is sequence encoded using an LSTM model to obtain a context temporal pattern feature encoding vector; The sensor multi-parameter significant fusion coding vector and the context temporal pattern feature coding vector are concatenated to obtain the environment-multi-parameter state temporal joint coding vector. Fire risk scores are calculated based on the joint encoding vector of environment and multi-parameter state time series. Based on the comparison between the fire risk score and the preset threshold, it is determined whether to generate an alarm signal.

2. The multi-parameter fusion fire early warning method based on environmental perception according to claim 1, characterized in that, The raw sensor stream includes temperature, humidity, carbon monoxide concentration, fine particulate matter concentration, raw infrared signal, wind speed, wind direction, and volatile organic compound index.

3. The multi-parameter fusion fire early warning method based on environmental perception according to claim 1, characterized in that, The contextual data stream includes light intensity, whether it is raining, personnel activity status, building materials, location type, altitude, and main hazard type.

4. The multi-parameter fusion fire early warning method based on environmental perception according to claim 1, characterized in that, Multi-parameter fusion is performed on the set of temporal pattern feature encoding vectors of the sensor data to obtain a sensor multi-parameter significant fused encoding vector, including: An intrinsic self-importance analysis is performed on the set of temporal pattern feature encoding vectors of the sensor data to obtain a set of attention weights for the temporal pattern features of the sensor data; A context-related importance analysis is performed on the set of temporal pattern feature encoding vectors of the sensor data to obtain a set of context-related attention weights for the temporal pattern features of the sensor data. Based on the set of attention weights for the temporal pattern features of the sensor data and the set of context-related attention weights for the temporal pattern features of the sensor data, the set of encoding vectors for the temporal pattern features of the sensor data is dynamically aggregated to obtain the sensor multi-parameter saliency fusion encoding vector.

5. The multi-parameter fusion fire early warning method based on environmental perception according to claim 4, characterized in that, An intrinsic self-importance analysis is performed on the set of temporal pattern feature encoding vectors of the sensor data to obtain a set of attention weights for the temporal pattern features of the sensor data, including: The peak values ​​of each sensor data time-series pattern feature encoding vector in the set of sensor data time-series pattern feature encoding vectors are extracted to obtain a set of significant feature values ​​of sensor data time-series patterns; A normalization operation is performed on the set of salient feature values ​​of the time-series pattern of the sensor data to obtain the set of attention weights of the time-series pattern features of the sensor data.

6. The multi-parameter fusion fire early warning method based on environmental perception according to claim 5, characterized in that, A context-related importance analysis is performed on the set of temporal pattern feature encoding vectors of the sensor data to obtain a set of context-related attention weights for the temporal pattern features of the sensor data, including: The set of sensor data temporal pattern feature encoding vectors is input into a context encoder based on a converter architecture to obtain a set of sensor data temporal pattern feature context-associated encoding vectors. The peak value of each sensor data temporal pattern feature context association encoding vector in the set of sensor data temporal pattern feature context association encoding vectors is extracted to obtain a set of significant values ​​of sensor data temporal pattern feature context association representation; A normalization operation is performed on the set of significant values ​​of the temporal pattern feature context association representation of the sensor data to obtain the set of attention weights of the temporal pattern feature context association of the sensor data.

7. The multi-parameter fusion fire early warning method based on environmental perception according to claim 6, characterized in that, Based on the set of attention weights for the temporal pattern features of the sensor data and the set of context-related attention weights for the temporal pattern features of the sensor data, the set of encoding vectors for the temporal pattern features of the sensor data is dynamically aggregated to obtain the sensor multi-parameter salient fusion encoding vector, including: The set of context-related attention weights of the temporal pattern features of the sensor data and the set of attention weights of the temporal pattern features of the sensor data are fused to obtain a set of hybrid attention weight values ​​of the temporal pattern of the sensor data. Based on the set of attention weight values ​​of the temporal pattern of the sensor data, a weighted sum of the set of feature encoding vectors of the temporal pattern of the sensor data is calculated to obtain the sensor multi-parameter significant fusion encoding vector.

8. The multi-parameter fusion fire early warning method based on environmental perception according to claim 1, characterized in that, Based on the comparison between the fire risk score and a preset threshold, it is determined whether to generate an alarm signal, including: in response to the fire risk score being greater than the preset threshold, determining to generate an alarm signal.

9. A multi-parameter fusion fire early warning system based on environmental perception, characterized in that, include: The environmental data acquisition module is used to acquire the time series of raw sensor streams and the time series of context data streams; The sensor data temporal coding module is used to arrange the time series of the original sensor stream into a sensor data temporal feature matrix according to the parameter sample dimension and the time dimension, and then input it into a sequence encoder based on the LSTM model to obtain a set of sensor data temporal pattern feature coding vectors. A multi-parameter sensor data temporal feature fusion module is used to perform multi-parameter fusion on the set of sensor data temporal pattern feature encoding vectors to obtain sensor multi-parameter significant fusion encoding vectors; The context data stream temporal coding module is used to perform LSTM-based sequence coding on the time series of the context data stream to obtain a context temporal pattern feature coding vector. The environment-multi-parameter state-temporal joint coding module is used to perform feature concatenation on the sensor multi-parameter significant fusion coding vector and the context temporal pattern feature coding vector to obtain the environment-multi-parameter state-temporal joint coding vector. The fire risk assessment module is used to calculate fire risk scores based on the environmental-multi-parameter state-time joint coding vector. The alarm module is used to determine whether to generate an alarm signal based on a comparison between the fire risk score and a preset threshold.

Images