An early fire detection method and system based on a recurrent neural network

Abstract

The application relates to the technical field of fire detection, and discloses an early fire detection method and system based on a recurrent neural network, which introduces a behavior mode baseline, multi-scale time sequence feature extraction, bidirectional long short-term memory encoding, time sequence attention fusion and a parameter self-adaptive adjustment mechanism based on feedback information, aims to effectively cope with the challenges brought by environmental changes and sensor drift, and significantly improves the accuracy and robustness of early fire detection.

Description

Technical Field

[0001] This application relates to the field of fire detection technology, and more specifically, to an early fire detection method and system based on recurrent neural networks. Background Technology

[0002] With the increasing demands for safe operation in data centers, smart buildings, and industrial facilities, early fire detection technologies based on multi-sensor fusion and recurrent neural networks have been widely researched and applied.

[0003] Existing technical solutions mainly fall into the following categories: The first category is the traditional detection method based on a fixed threshold, which determines whether a fire has occurred based on whether the instantaneous reading of a single type of sensor exceeds a preset threshold, such as using a temperature sensor or a smoke sensor alone to set high and low threshold alarms; the second category is the detection method based on a joint threshold of multiple types of sensors, which compares the weighted sum of readings from multiple types of sensors such as temperature, smoke, and carbon monoxide with a comprehensive threshold; the third category is the end-to-end detection method based on recurrent neural networks, which directly inputs the raw readings or simple statistics of multiple types of sensors into the recurrent neural network, and the network outputs the fire probability; the fourth category is the detection method based on time-series features, which introduces time-series processing techniques such as sliding windows and attention mechanisms on the basis of recurrent neural networks to enhance the capture of long-term dependencies.

[0004] However, the above-mentioned existing technical solutions have the following technical problems in practical applications: (i) Fixed threshold methods have poor adaptability to background drift. Environments such as data centers commonly experience scenarios like power density upgrades, air conditioning mode switching, and the introduction of new equipment, leading to systematic and gradual deviations in background readings from various sensors. When background drift occurs, fixed threshold methods are prone to misinterpreting background drift as abnormal, or misinterpreting weak real fire signals superimposed on the drifting background as normal, thus resulting in false alarms or missed alarms.

[0005] (ii) The joint thresholding method using multiple sensor types is difficult to capture complex temporal correlations. This type of method usually performs weighted aggregation of instantaneous readings from various types of sensors, ignoring the temporal coupling patterns of fire signals among multiple sensors. For example, in the early stages of smoldering fires, volatile organic compounds and ultrafine particles rise synchronously while temperature and carbon monoxide changes lag behind, and in open flame fires, smoke responds before temperature. Therefore, it is difficult to accurately identify the true trend pointing to the fire from a complex background.

[0006] (iii) End-to-end recurrent neural network methods overfit to the training data distribution. This type of method uses raw sensor readings as input, and the model overfits to the background distribution during training. Once the background distribution changes in the deployment environment, the detection performance of the model will drop significantly, and the model lacks an interpretable bias quantification mechanism, making it difficult to pinpoint the cause of the performance degradation.

[0007] (iv) Temporal feature detection methods are difficult to capture multi-scale changes simultaneously. Fire signals have different temporal scale characteristics in different development stages and different types of fires. Fixed sliding window or single-scale temporal processing methods are difficult to capture short-term abrupt changes and long-term drifts at the same time, which limits the ability to identify weak early fire signals.

[0008] (v) Existing methods generally lack feedback and adaptive mechanisms. When false alarms or missed alarms occur, existing methods usually rely solely on manually resetting the threshold or retraining the model. They cannot continuously and adaptively adjust the detection parameters during system operation and are unable to cope with the cumulative deviations caused by long-term environmental changes and long-term sensor aging. Summary of the Invention

[0009] This application discloses an early fire detection method and system based on recurrent neural networks, aiming to solve the problem that early fire detection systems in smart buildings and industrial facilities suffer from decreased detection accuracy and reliability when facing continuous environmental changes and sensor aging and drift, resulting in insufficient ability to identify real fire signals.

[0010] The technical solution of this application is as follows: In a first aspect, this application discloses an early fire detection method based on a recurrent neural network, executed by a computer system, comprising: Acquire sensor readings for each type of sensor, and perform noise reduction and normalization processing on the sensor readings for each type of sensor; each type of sensor includes one or more of the following: temperature sensor, smoke sensor, carbon monoxide sensor, volatile organic compound sensor, and ultrafine particulate matter sensor; Based on the sensor readings of each type of sensor, establish and update the behavioral pattern baseline of each type of sensor in the current environment; the behavioral pattern baseline includes the statistical characteristics and temporal characteristics of the sensor readings. Based on the real-time readings and behavioral pattern baseline of each type of sensor, the deviation information of each type of sensor is calculated; The deviation information is input into a trained recurrent neural network (RNN) model. The multi-scale temporal feature extraction layer of the RNN model adaptively adjusts the sampling weights of multiple sliding windows of different lengths according to the rate of change of the deviation information, and samples the deviation information to obtain multi-scale temporal features. The bidirectional long short-term memory (LSTM) encoding layer of the RNN model performs forward and backward encoding on the multi-scale temporal features to obtain cross-scale interactive forward and backward encodings. The temporal attention fusion layer of the RNN model performs weighted fusion on the forward and backward encodings according to fire dynamics priors, and outputs the anomaly probability of the area where each type of sensor is located. When the probability of an anomaly exceeds a preset threshold, a consistency judgment is performed based on the deviation information to confirm whether there is an abnormal trend pointing to a fire. The consistency judgment includes analyzing the changing trend of the deviation information and verifying whether there is a physical correlation between the deviation information. A fire alarm is triggered based on the consistency assessment result; Based on feedback from fire alarms, adjust the update parameters of the behavioral pattern baseline or the parameters for consistency judgment.

[0011] Secondly, this application also discloses an early fire detection system based on a recurrent neural network, comprising: The data acquisition and preprocessing module is used to acquire sensor readings for each type of sensor and perform noise reduction and normalization processing on the sensor readings for each type of sensor; each type of sensor includes one or more of the following: temperature sensor, smoke sensor, carbon monoxide sensor, volatile organic compound sensor, and ultrafine particulate matter sensor. The behavior pattern baseline establishment and update module is used to establish and update the behavior pattern baseline of each type of sensor in the current environment based on the sensor readings of each type of sensor; the behavior pattern baseline includes the statistical characteristics and temporal characteristics of the sensor readings. The deviation information calculation module is used to calculate the deviation information of each type of sensor based on the real-time readings and behavior pattern baseline of each type of sensor. The anomaly probability assessment module is used to input deviation information into a trained recurrent neural network model. The multi-scale temporal feature extraction layer of the recurrent neural network model adaptively adjusts the sampling weights of multiple sliding windows of different lengths according to the rate of change of the deviation information to sample the deviation information and obtain multi-scale temporal features. The bidirectional long short-term memory encoding layer of the recurrent neural network model performs forward and backward encoding on the multi-scale temporal features to obtain cross-scale interactive forward and backward encoding. The temporal attention fusion layer of the recurrent neural network model performs weighted fusion on the forward and backward encodings according to fire dynamics priors and outputs the anomaly probability of the area where each type of sensor is located. The consistency judgment module is used to perform consistency judgment based on deviation information when the anomaly probability exceeds a preset threshold, to confirm whether there is an abnormal trend pointing to a fire; the consistency judgment includes analyzing the changing trend of deviation information and verifying whether there is a physical correlation between deviation information. The alarm triggering module is used to trigger a fire alarm based on the consistency judgment result; The parameter adjustment module is used to adjust the update parameters of the behavior pattern baseline or the parameters for consistency judgment based on the feedback information from the fire alarm.

[0012] This application discloses an early fire detection method and system based on recurrent neural networks. By introducing behavioral pattern baselines, multi-scale temporal feature extraction, bidirectional long short-term memory encoding, temporal attention fusion, and a parameter adaptive adjustment mechanism based on feedback information, it aims to effectively cope with the challenges brought about by environmental changes and sensor drift, and significantly improve the accuracy and robustness of early fire detection. Attached Figure Description

[0013] Figure 1 This is a schematic flowchart of an early fire detection method based on a recurrent neural network provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of an early fire detection system based on a recurrent neural network provided in an embodiment of the present invention. Detailed Implementation

[0014] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0015] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0016] Traditional early fire detection systems suffer a significant decline in detection accuracy and reliability when faced with continuous environmental changes and sensor aging and drift. This weakens their ability to identify real fire signals, especially in complex environments such as data centers. The combined effect of micro-environmental changes and long-term sensor aging causes systemic background feature drift, making it difficult for models to accurately distinguish weak fire signals from background noise, thus missing the best opportunity for early warning.

[0017] In this regard, refer to Figure 1 , Figure 1 This is a flowchart illustrating an early fire detection method based on a recurrent neural network provided in an embodiment of the present invention, executed by a computer system, including: S11, acquire sensor readings for each type of sensor, and perform noise reduction and normalization processing on the sensor readings for each type of sensor; each type of sensor includes one or more of the following: temperature sensor, smoke sensor, carbon monoxide sensor, volatile organic compound sensor, and ultrafine particulate matter sensor. S12, Based on the sensor readings of each type of sensor, establish and update the behavior pattern baseline of each type of sensor in the current environment; the behavior pattern baseline includes the statistical characteristics and temporal characteristics of the sensor readings; S13, Calculate the deviation information of each type of sensor based on the real-time readings of each type of sensor and the baseline of the behavior pattern; S14, the deviation information is input into the trained recurrent neural network model. The multi-scale temporal feature extraction layer of the recurrent neural network model adaptively adjusts the sampling weights of multiple sliding windows of different lengths according to the rate of change of the deviation information, and samples the deviation information to obtain multi-scale temporal features. The bidirectional long short-term memory encoding layer of the recurrent neural network model performs forward and backward encoding on the multi-scale temporal features to obtain cross-scale interactive forward and backward encoding. The temporal attention fusion layer of the recurrent neural network model performs weighted fusion on the forward and backward encodings according to fire dynamics priors, and outputs the anomaly probability of the area where each type of sensor is located. S15, when the anomaly probability exceeds a preset threshold, a consistency judgment is performed based on the deviation information to confirm whether there is an abnormal trend pointing to a fire; the consistency judgment includes analyzing the changing trend of the deviation information and verifying whether there is a physical correlation between the deviation information. S16, trigger a fire alarm based on the consistency judgment result; adjust the update parameters of the behavior pattern baseline or the parameters of the consistency judgment based on the feedback information of the fire alarm.

[0018] This application aims to effectively address the challenges posed by environmental changes and sensor drift by introducing behavioral pattern baselines, multi-scale temporal feature extraction, bidirectional long short-term memory encoding, temporal attention fusion, and a parameter adaptive adjustment mechanism based on feedback information, thereby significantly improving the accuracy and robustness of early fire detection.

[0019] The core of the method proposed in this application lies in utilizing recurrent neural networks to perform deep analysis of multi-source sensor data to achieve accurate identification of early-stage fires. Here, "each type of sensor" refers to various physical sensors used to monitor environmental parameters, such as temperature sensors, smoke sensors, carbon monoxide sensors, volatile organic compound sensors, and ultrafine particulate matter sensors. These sensors work together to provide comprehensive environmental status information. "Sensor readings" are the raw data collected by these sensors in real time. "Behavioral pattern baseline" refers to the normal fluctuation range and temporal characteristics exhibited by sensor readings under specific environmental conditions, serving as a reference standard for judging anomalies. " Deviation information" is the difference between real-time sensor readings and the behavioral pattern baseline, reflecting the degree of deviation between the current environmental state and the normal state. The "recurrent neural network model" is a deep learning model particularly suitable for processing time-series data, capable of capturing long-term dependencies and complex patterns in the data. The "multi-scale temporal feature extraction layer," "bidirectional long short-term memory encoding layer," and "temporal attention fusion layer" are key components of the recurrent neural network model, responsible for extracting and fusing features from different time scales, directions, and incorporating prior knowledge of fire dynamics, respectively. "Anomaly Probability" is a quantitative indicator output by the model, representing the likelihood of an abnormal event (such as a fire) occurring in the current area. "Consistency Judgment" further verifies the anomaly probability by analyzing the physical correlation of deviation information to eliminate false alarms. "Fire Alarm" is a warning signal issued by the system after confirming a fire risk. "Feedback Information" refers to the response results to the fire alarm, used to guide the adaptive adjustment of system parameters.

[0020] The early fire detection method of this application includes the following key steps in its implementation process: First, sensor readings for each type of sensor need to be acquired and then denoised and normalized. Each type of sensor can include one or more of the following: temperature sensors, smoke sensors, carbon monoxide sensors, volatile organic compound (VOC) sensors, and ultrafine particulate matter (UPM) sensors. For example, a sensor network can be deployed, including multiple temperature sensors to monitor changes in ambient temperature, multiple smoke sensors to detect the concentration of smoke particles in the air, carbon monoxide sensors to detect the concentration of carbon monoxide gas, VOC sensors to detect specific organic gases, and UPM sensors to detect fine particulate matter in the air. These sensors can periodically transmit data to a central processing unit. During the data preprocessing stage, methods such as moving average filtering and Kalman filtering can be used to denoise the sensor readings to eliminate random noise and transient interference. Subsequently, methods such as min-max normalization or Z-score normalization are used to unify sensor readings of different dimensions and ranges into a similar numerical range to facilitate subsequent processing by the neural network model.

[0021] Secondly, based on the sensor readings of each type of sensor after the aforementioned denoising and normalization processes, a baseline for the behavioral patterns of these sensors in the current environment is established and updated. The behavioral pattern baseline includes the statistical and temporal characteristics of the sensor readings. For example, historical sensor data can be used to characterize the statistical characteristics of the sensor readings through statistical analysis methods (such as calculating the mean, standard deviation, quantiles, etc.). Simultaneously, time-series analysis methods such as the autoregressive moving average (ARMA) model and exponential smoothing can be employed to capture the temporal characteristics of the sensor readings, such as periodicity and trends. The behavioral pattern baseline is not static but needs to be dynamically updated according to long-term environmental changes. For example, a sliding window averaging method can be used to periodically calculate the statistical and temporal characteristics of the sensor readings over the most recent period and then perform a weighted average with the old baseline, thereby achieving smooth baseline updates.

[0022] Next, based on the real-time readings of each type of sensor and the established behavioral pattern baseline, deviation information for each sensor type is calculated. This deviation information quantifies the degree of deviation between the real-time readings and the baseline. For example, the real-time sensor readings can be compared to the mean in the behavioral pattern baseline, and the difference or percentage deviation can be calculated. Alternatively, the real-time readings can be input into a baseline-trained prediction model, and the residual between the predicted and actual values ​​can be used as deviation information. Furthermore, the position of the real-time readings within the statistical distribution defined by the baseline can be considered, for example, by calculating their Z-score to reflect their deviation from the baseline mean by a factor of standard deviation.

[0023] Subsequently, the calculated bias information is input into a trained recurrent neural network model. This recurrent neural network model includes a multi-scale temporal feature extraction layer, a bidirectional long short-term memory encoding layer, and a temporal attention fusion layer. Specifically, the multi-scale temporal feature extraction layer adaptively adjusts the sampling weights of multiple sliding windows of different lengths according to the rate of change of the bias information, and samples the bias information to obtain multi-scale temporal features. For example, for rapidly changing bias information, shorter sliding windows can be assigned higher sampling weights; for slowly changing bias information, longer sliding windows can be assigned higher sampling weights. The bidirectional long short-term memory encoding layer performs forward and backward encoding on these multi-scale temporal features to capture the bidirectional dependencies of the data in the time dimension, thus obtaining cross-scale interactive forward and backward encodings. For example, forward encoding can capture dependencies from the past to the present, while backward encoding can capture dependencies from the future to the present. The temporal attention fusion layer then performs weighted fusion of the forward and backward encodings based on prior knowledge of fire dynamics, ultimately outputting the anomaly probability of the area where each type of sensor is located. For example, a weight allocation strategy for the attention mechanism can be designed based on the typical time-series patterns of changes in sensor readings such as temperature, smoke, and carbon monoxide during a fire.

[0024] When the output anomaly probability exceeds a preset threshold, the system performs a consistency judgment based on the deviation information to confirm whether there is an abnormal trend pointing to a fire. The consistency judgment includes analyzing the changing trends of the deviation information and verifying whether there is a physical correlation between the deviation information. For example, if the temperature sensor reading continues to rise, and the smoke sensor reading also rises synchronously, and the carbon monoxide sensor reading also increases, these deviation information are physically correlated and are likely to collectively point to a fire. Conversely, if only a single sensor shows an anomaly, or if multiple sensors show anomalies but their changing trends are physically inconsistent, it may be a false alarm.

[0025] Based on the results of the consensus assessment, the system will trigger a fire alarm. For example, if the consensus assessment confirms an abnormal trend pointing to a fire, an audible and visual alarm will be issued immediately, and a notification will be sent to all personnel.

[0026] Finally, based on feedback from fire alarms, the system adaptively adjusts the update parameters of the behavior pattern baseline or the parameters for consistency judgment. For example, if the alarm is confirmed as a real fire, the system enhances its ability to identify similar patterns, potentially adjusting the sensitivity of baseline updates or the threshold for consistency judgment to make it more sensitive to real fire signals. If the alarm is confirmed as a false alarm, the system learns the characteristics that led to the false alarm and adjusts the parameters to reduce the likelihood of similar false alarms in the future. This feedback mechanism enables the system to continuously learn and adapt to environmental changes, improving the accuracy and robustness of detection.

[0027] The early fire detection method of this application effectively solves the problem of decreased detection accuracy of traditional systems under environmental changes and sensor drift by introducing a behavioral pattern baseline and a parameter adaptive adjustment mechanism based on feedback information.

[0028] This application further proposes that the above-mentioned recurrent neural network model outputs the anomaly probability, uncertainty estimate, anomaly type classification, and anomaly time window prediction; the uncertainty estimate is obtained by the recurrent neural network model through Monte Carlo dropout sampling, including cognitive uncertainty and accidental uncertainty; the consistency judgment is based on the anomaly type classification, the uncertainty estimate, and the anomaly time window prediction, and uses Bayesian inference to perform multidimensional evidence fusion, and judges whether there is an abnormal trend pointing to fire based on the fused posterior probability.

[0029] Specifically, the recurrent neural network model, while outputting the anomaly probability of the area where each type of sensor is located, is also configured to output uncertainty estimation, anomaly type classification, and anomaly time window prediction. Uncertainty estimation aims to quantify the confidence level of the model's predictions and is obtained through Monte Carlo dropout sampling. Monte Carlo dropout sampling refers to randomly shutting down some neurons multiple times during the model's inference phase, resulting in multiple different predictions. The variance or distribution of these results can be used to estimate the model's uncertainty. This uncertainty estimation is further subdivided into cognitive uncertainty and random uncertainty. Cognitive uncertainty reflects the model's insufficient ability to identify patterns not seen in the training data and can usually be reduced by increasing training data or improving the model structure; random uncertainty stems from noise or inherent randomness in the data itself and cannot be completely eliminated even with a perfect model. Furthermore, anomaly type classification indicates that the model can identify whether the current anomaly belongs to a fire, equipment failure, environmental change, or other types of anomalies, providing a more refined basis for subsequent judgments. Anomaly time window prediction indicates the time range within which an anomaly may occur or persist, helping to more accurately locate and respond to it.

[0030] When making consistency judgments, the above scheme no longer relies solely on the changing trends and physical correlations of deviation information, but instead incorporates the aforementioned anomaly type classification, uncertainty estimation, and anomaly time window prediction as multidimensional evidence. This evidence is then fused within a Bayesian inference framework. Bayesian inference is a probabilistic reasoning method that updates beliefs about an event's occurrence by combining prior knowledge and observational evidence. Here, prior knowledge can include historical fire data, sensor characteristics, etc., while observational evidence comes from the outputs of recurrent neural network models. Through Bayesian inference, the fused posterior probability can be calculated, which more comprehensively and robustly reflects the true likelihood of a fire in the current situation.

[0031] The proposed solution significantly enhances the accuracy and robustness of consistency judgment in early fire detection by introducing uncertainty estimation, anomaly type classification, and anomaly time window prediction from the output of a recurrent neural network model, and combining it with Bayesian inference for multidimensional evidence fusion.

[0032] In some preferred embodiments, it is assumed that in a data center environment, the temperature sensor, smoke sensor, and carbon monoxide sensor all show a slight upward trend over a short period of time, but the increase is not significant, while the readings of the volatile organic compound sensor and the ultrafine particulate matter sensor remain stable. In the basic scheme, judging solely based on the trend of deviation information and physical correlation may be insufficient to determine whether it is a fire due to unclear trends or weak correlations.

[0033] However, in this embodiment, the recurrent neural network model outputs not only anomaly probability but also uncertainty estimation, anomaly type classification, and anomaly time window prediction. For example, the model might output an anomaly probability of 0.6 (slightly higher than a preset threshold) but simultaneously output high cognitive uncertainty (e.g., 0.4), indicating insufficient confidence in the model's judgment of the current situation. The anomaly type classification might indicate "environmental fluctuations" rather than "fire," and the anomaly time window prediction might show a short duration of the anomaly. In this case, the consistency judgment module uses this information as multidimensional evidence and fuses it using Bayesian inference. Bayesian inference comprehensively considers the anomaly probability, high uncertainty, non-fire type classification, and short time window prediction. For example, if the prior probability indicates frequent data center environmental fluctuations, then after fusing this evidence, the posterior probability will significantly reduce the likelihood of pointing to a fire. Thus, the system can more accurately determine that the current anomaly is not a real fire, thereby avoiding false alarms.

[0034] Conversely, if the model outputs low uncertainty, high fire type classification probability, and a longer anomaly time window prediction, even if the anomaly probability is slightly low, Bayesian inference may still yield a higher posterior fire probability, thus triggering an alarm in a timely manner and avoiding missed reports. This mechanism of multi-dimensional evidence fusion and uncertainty quantification enables the system to make more informed and reliable decisions when facing ambiguous or complex situations.

[0035] This application further proposes a method for embedding physical information about fire propagation into a recurrent neural network model. By constructing a physical information embedding layer and introducing a physical constraint loss term, the model can better understand and predict the dynamic evolution of a fire. The aforementioned early fire detection method based on recurrent neural networks also includes: A fire propagation physical information embedding layer is constructed, in which the fire heat release rate, smoke spread speed, and toxic gas diffusion coefficient are embedded into the feature space of the recurrent neural network model; The fire propagation physical information embedding layer is used to dynamically adjust the embedding parameters based on the real-time readings of each type of sensor; the recurrent neural network model is used to adjust the network parameters in reverse through physical constraint loss terms so that the network output satisfies the fire dynamics conservation relationship.

[0036] Specifically, the fire propagation physical information embedding layer refers to a specially designed neural network module. Its function is to transform key physical quantities related to fire propagation, such as the fire heat release rate, smoke spread rate, and toxic gas diffusion coefficient, into feature representations that can be understood and utilized by the recurrent neural network model. These physical quantities are core indicators describing fire development and spread. Embedding them into the model's feature space aims to provide the model with deeper physical background knowledge. Among them, the fire heat release rate can be understood as the total heat released by the fire per unit time, a key parameter for measuring fire intensity and development speed; the smoke spread rate refers to the speed at which smoke generated by the fire spreads in space, reflecting the expanding trend of the fire's impact range; and the toxic gas diffusion coefficient describes the diffusion efficiency of toxic gases such as carbon monoxide and volatile organic compounds generated by the fire in the environment. This physical information can provide the recurrent neural network model with physical constraints and guidance that go beyond pure sensor data models.

[0037] In practical applications, the fire propagation physics information embedding layer dynamically adjusts the embedding parameters based on real-time readings from each type of sensor. For example, when the temperature sensor reading rises sharply, the embedding layer can adjust the embedding parameter for the fire heat release rate accordingly to reflect the increase in fire intensity. This dynamic adjustment ensures that the embedded physical information can reflect the fire development status in the current environment in real time. Furthermore, a recurrent neural network model is used to inversely adjust the network parameters through a physical constraint loss term, ensuring that the network output satisfies the fire dynamics conservation relationships. The physical constraint loss term is a special loss function that incorporates fire dynamics principles, such as energy conservation and mass conservation, into the model training process. Its purpose is to force the model's predictions to be physically reasonable, avoiding outputs that contradict actual physical laws. In this way, the model's learning process not only relies on data patterns but is also guided by physical laws, thereby improving the model's generalization ability and reliability.

[0038] The solution proposed in this application constructs a fire propagation physical information embedding layer, which integrates key physical quantities such as fire heat release rate, smoke spread speed, and toxic gas diffusion coefficient into the feature space of a recurrent neural network model. This enables the model to not only capture the temporal patterns at the data level when processing sensor bias information, but also to make judgments based on the physical evolution laws of the fire.

[0039] This application further proposes a method for training the recurrent neural network model, aiming to significantly improve the detection accuracy and robustness of the model by constructing a comprehensive training dataset, employing advanced data processing and augmentation techniques, and incorporating training strategies based on physical laws. The method also includes training the recurrent neural network model, specifically: Construct a training dataset for each type of sensor, including normal operating conditions, early fire operating conditions, and interference operating conditions. Establish the behavior pattern baseline and calculate the deviation information for the sensor readings in each type of sensor training dataset to obtain training samples of deviation information from the same source. The bias information training samples are segmented by a multi-scale sliding window to generate fixed-length time-series samples that cover short-term mutations and long-term drifts; Data augmentation is performed on the fixed-length time series samples based on the physical fire propagation model to generate synthetic samples that conform to the laws of fire dynamics. The main task is trained using focus loss, and the class weights of focus loss are dynamically adjusted according to the output of the physical consistency regularization term so that the network output satisfies the laws of fire dynamics. The recurrent neural network model is trained together to obtain a trained recurrent neural network model.

[0040] Specifically, before training the recurrent neural network model, a comprehensive training dataset needs to be constructed. This dataset is designed to include sensor readings under various operating conditions, such as normal operating conditions, simulated or real early fire conditions, and various interference conditions that may lead to false alarms. Normal operating condition data is used to learn the regular fluctuation patterns of the environment; early fire condition data is used to capture the unique signals during a fire; and interference condition data is used to train the model to distinguish non-fire anomalies. Sensor readings in each type of sensor training dataset are first used to establish the baseline of the behavioral pattern, and the deviation information is calculated based on this baseline. In this way, deviation information training samples with the same origin as the actual detection process can be obtained, ensuring the consistency between the training data and the actual input to the model.

[0041] The deviation information training samples are then segmented using a multi-scale sliding window. The purpose of this step is to extract features at different time scales from the original deviation information sequence. By using multiple sliding windows of varying lengths, a series of fixed-length time-series samples can be generated, simultaneously covering both short-term sudden changes (e.g., instantaneous increases in sensor readings) and long-term trend drifts (e.g., slow changes in environmental parameters). This multi-scale segmentation helps the model capture the performance of fire signals across different time dimensions.

[0042] In practical applications, to address the scarcity of real fire data and enhance the model's generalization ability, data augmentation is performed on the fixed-length time-series samples based on a physical fire propagation model. This physical fire propagation model can simulate the development of a fire under different environments, such as heat release rate, smoke spread rate, and toxic gas diffusion. By applying these physical laws to existing fixed-length time-series samples, synthetic samples that conform to fire dynamics can be generated. These synthetic samples not only increase the diversity of training data but also enable the model to learn deeper physical correlations.

[0043] Furthermore, during training, FocalLoss is employed to train the main task. FocalLoss is an improved cross-entropy loss function designed to address class imbalance, especially in scenarios like early fire detection where negative samples (normal operating conditions) far outnumber positive samples (early fire conditions). By reducing the weight of easily classified samples and increasing the weight of difficult-to-classify samples, FocalLoss allows the model to focus more on learning the hard-to-distinguish features of early fires. In addition, the class weights of FocalLoss are dynamically adjusted based on the output of a physical consistency regularization term. This term evaluates whether the network output satisfies predefined fire dynamics principles, such as the expected physical correlation between different sensor readings during a fire. This dynamic adjustment ensures that the network output is not only statistically accurate but also physically plausible.

[0044] Therefore, by jointly training the recurrent neural network model, it can fully utilize the constructed training dataset, multi-scale temporal features, and physically augmented samples, ultimately resulting in a trained recurrent neural network model. This model can effectively identify early fire signals and possesses strong anti-interference capabilities. Through the above technical solution, this application can significantly improve the performance of recurrent neural network models in early fire detection.

[0045] In some preferred embodiments, training the recurrent neural network model can be specifically implemented as follows: First, collect a training dataset for each type of sensor, including data from several months or even years. This dataset should cover normal operation data of the data center under different seasons and loads, as well as early fire event data obtained through simulation experiments or historical records (e.g., minor smoke or gas leaks caused by cable overheating or server short circuits), and also include data on some common disturbance events (e.g., local temperature fluctuations caused by cleaning agent spraying, air conditioning failures, and temporary increases in particulate matter caused by human activity).

[0046] Next, these raw sensor readings are denoised and normalized, and a behavioral pattern baseline is established based on methods such as moving average or exponentially weighted average. For example, for a temperature sensor, the baseline might be a dynamic temperature range rather than a fixed value. Then, the deviation information of each sensor reading relative to its behavioral pattern baseline is calculated. This deviation information (e.g., the difference between the current temperature and the baseline temperature) constitutes a homogeneous deviation information training sample.

[0047] Subsequently, these bias information training samples are segmented using a multi-scale sliding window. For example, three sliding windows of different lengths can be set: a short window (e.g., 5 minutes) to capture rapidly changing signals; a medium window (e.g., 30 minutes) to capture trends over a medium duration; and a long window (e.g., 2 hours) to capture slowly evolving or accumulating anomalies. Through these windows, multiple fixed-length time-series samples are extracted from the original bias information sequence.

[0048] To enhance the dataset, physical fire propagation models, such as computational fluid dynamics (CFD) models, can be used to simulate the smoke, heat, and gas diffusion processes of different fire sources (e.g., cable fires, liquid fuel fires) in specific data center layouts. These simulation results can then be combined with existing fixed-length time-series samples to generate new synthetic samples. For example, a small fire source can be simulated to gradually develop inside a server rack, and the trajectory of sensor readings at different locations can be predicted, thereby generating early fire samples that conform to physical laws.

[0049] When training the recurrent neural network model, focus loss is used as the primary loss function. For example, the focus parameter γ of the focus loss can be set to 2 to more strongly penalize misclassified hard samples. Simultaneously, a physical consistency regularization term is introduced. This regularization term can be a penalty function that increases the loss when the model's predicted anomalous probability does not conform to the fire dynamics prior (e.g., a sharp increase in smoke concentration without a corresponding change in carbon monoxide concentration). The class weights of the focus loss (e.g., earlier fire samples have higher weights than normal samples) are dynamically adjusted based on the output of this physical consistency regularization term. For example, if the model predicts an anomalous probability that is physically unreasonable at a certain time step, the weight of that sample will be increased, prompting the model to pay more attention to such samples in subsequent iterations.

[0050] Ultimately, through joint training, the recurrent neural network model is able to learn to identify early fire signals from multi-scale temporal features and ensure that its prediction results are physically plausible, thus obtaining a high-precision and robust early fire detection model.

[0051] The steps outlined above for performing a consistency judgment based on deviation information to confirm the existence of an abnormal trend pointing to a fire when the anomaly probability exceeds a preset threshold include: Analyze the deviation information to identify the main source of the current abnormal signal, and determine whether the current situation belongs to the first fire mode or the second fire mode based on the main source. If the current situation is determined to be the first fire mode, the first causal verification chain is activated to verify whether the deviation information of the volatile organic compound sensor and the ultrafine particulate matter sensor simultaneously shows a continuous upward trend, whether they are highly synchronized in time, whether they last for a preset duration, and to confirm whether the deviation information of the temperature sensor and the carbon monoxide sensor remains below a preset level within the same time window. If the current situation is determined to belong to the second fire mode, the second causal verification chain is activated to verify whether the deviation information simultaneously shows a consistent upward trend, and to check whether the consistent upward trend has a reasonable order of occurrence in physical terms. The consensus threshold is dynamically adjusted based on the uncertainty estimate of the current situation; Weights are assigned based on the verification results of the enabled causal verification chain, and it is determined whether the dynamically adjusted consensus threshold has been reached.

[0052] Specifically, analyzing deviation information to identify the main source of the current abnormal signal involves comprehensively analyzing the amplitude, rate of change, and duration of deviation information from different types of sensors, combined with a pre-set knowledge base or machine learning model, to infer the main physical or chemical processes causing the anomaly. For example, a significant increase in volatile organic compounds (VOCs) and ultrafine particulate matter (UFP) may indicate a smoldering fire, while a rapid increase in temperature and carbon monoxide may indicate an open flame fire. Based on the identified main source, the system categorizes the current situation into a first fire mode or a second fire mode. The first fire mode typically corresponds to a smoldering or initial fire characterized by VOCs and UFP, while the second fire mode corresponds to a rapidly developing fire characterized by temperature and carbon monoxide.

[0053] If the current situation is determined to be a first-mode fire, a first causal verification chain is activated. This verification chain aims to confirm the typical characteristics of smoldering fires. Specifically, it verifies whether the deviation information of the volatile organic compound (VOC) sensor and the ultrafine particulate matter sensor simultaneously exhibits a continuous upward trend, indicating continuous organic matter decomposition and particulate matter generation. It also checks whether these upward trends are highly synchronized in time to rule out accidental, inconsistent fluctuations. Furthermore, these trends need to be maintained for a preset duration to avoid instantaneous interference. As a key exclusion criterion, it also confirms whether the deviation information of the temperature sensor and the carbon monoxide sensor remains consistently below preset levels within the same time window, because smoldering fires typically do not initially involve significant increases in temperature and carbon monoxide.

[0054] In practical applications, if the current situation is determined to belong to the second fire mode, a second causal verification chain is activated. This verification chain primarily targets the characteristics of open flame fires. Specifically, it verifies whether the deviation information of all sensors (such as temperature sensors, smoke sensors, carbon monoxide sensors, etc.) simultaneously exhibits a consistent upward trend, reflecting the rapid development of the fire. Furthermore, it checks whether this consistent upward trend has a physically plausible order of occurrence. For example, smoke sensors typically respond earlier than temperature sensors, while the increase in carbon monoxide may follow closely behind; this temporal relationship is an important manifestation of fire dynamics.

[0055] Furthermore, the consensus threshold is dynamically adjusted based on the uncertainty estimate of the current situation. Uncertainty estimate can be understood as the confidence information attached to the output anomaly probability by the recurrent neural network model, such as that obtained through Monte Carlo dropout sampling. When the model's judgment of the current situation is highly uncertain, the consensus threshold is appropriately increased, requiring stronger evidence to trigger an alarm, thus reducing false alarms; conversely, when the uncertainty is low, the consensus threshold can be appropriately decreased to improve the sensitivity of early detection.

[0056] Finally, weights are assigned based on the verification results of the enabled causal verification chains, and it is determined whether a dynamically adjusted consensus threshold has been reached. Specifically, the verification results of each verification chain (first or second) are assigned different weights based on their importance or reliability. For example, if all conditions of the first causal verification chain are met, it will receive a higher score. These weighted verification results are then combined and compared with the dynamically adjusted consensus threshold. Only when the combined score reaches or exceeds the threshold is an anomalous trend pointing to a fire finally confirmed.

[0057] This application's solution effectively addresses the limitations of traditional consistency judgment in distinguishing between real fires and non-fire interference in complex environments by introducing refined fire mode judgment and causal verification chains. Through the above technical solution, this application can significantly improve the accuracy and reliability of early fire detection.

[0058] The following is a specific example to illustrate this. Suppose that in a data center server room, temperature sensors, smoke sensors, carbon monoxide sensors, volatile organic compound sensors, and ultrafine particulate matter sensors are deployed.

[0059] At a certain moment, the anomaly probability output by the anomaly probability assessment module exceeded the preset threshold, triggering the consistency judgment process.

[0060] The system first analyzes the current deviation information. It assumes that the reading deviations of the volatile organic compound (VOC) sensor and the ultrafine particulate matter sensor have shown a synchronous, slow upward trend over the past 10 minutes, while the reading deviations of the temperature sensor and the carbon monoxide sensor have remained below normal levels. Based on this characteristic, the system identifies the main source of the current abnormal signal as the slow decomposition of organic matter and the generation of particulate matter, thus determining that the current situation belongs to the first fire mode (smoldering fire mode).

[0061] The system then activates the first causal verification chain. This chain checks: 1. Do the deviation information of the volatile organic compound sensor and the ultrafine particulate matter sensor both show a continuous upward trend? Yes, both show a continuous upward trend.

[0062] 2. Are they highly synchronized in time? Yes, the upward trends start and continue almost simultaneously.

[0063] 3. Whether to continue for the preset duration (e.g., 5 minutes): Yes, it has been 10 minutes.

[0064] 4. Whether the deviation information of the temperature sensor and the carbon monoxide sensor remains below the preset level within the same time window: Yes, neither of them showed a significant increase.

[0065] Meanwhile, based on the uncertainty estimate output by the recurrent neural network model, the system finds that the uncertainty of the current situation is low, and therefore dynamically adjusts the consensus threshold to a relatively low value.

[0066] Since all conditions in the first causal verification chain are met, its verification result is given high weight. The combined weighted verification result far exceeds the dynamically adjusted consensus threshold. Therefore, the consensus judgment module ultimately confirms the existence of an abnormal trend pointing to a fire and triggers a fire alarm. This example demonstrates how this solution can accurately identify initial smoldering fires by recognizing specific fire patterns and activating corresponding causal verification chains, providing early warning even before temperatures and carbon monoxide levels have significantly increased. This avoids the possibility of delayed alarms due to a lack of comprehensive evidence, a situation that traditional methods might encounter.

[0067] This application further proposes the following steps for adjusting the update parameters or consistency judgment parameters of the behavioral pattern baseline based on feedback information from fire alarms: Obtain feedback information from the fire alarm, including feedback on actual fires, false alarms, and missed alarms; Each type of feedback information is encoded into a vector representation and input into the meta-learning network to obtain parameter adjustment amounts; A confidence score is assigned to each type of feedback information, and the confidence score indicates the reliability of each type of feedback information and participates in the output adjustment of the meta-learning network; Based on the parameter adjustment amount and the confidence score, adjust the update parameters of the behavioral pattern baseline or the parameters of the consistency judgment; When conflict feedback information is received, the conflict feedback information is integrated through an arbitration process. The arbitration process determines the direction and magnitude of parameter adjustment based on the authenticity of historical alarm events, long-term operating data of each sensor, and the changing trend of the environmental background.

[0068] Specifically, fire alarm feedback information refers to the information obtained after a fire alarm is triggered, either manually or through other verification mechanisms, regarding the authenticity of the alarm event. True fire feedback indicates that the alarm accurately identified a real fire event; false alarm feedback indicates that the alarm was triggered incorrectly, and no fire actually occurred; missed alarm feedback indicates that a fire occurred but the system failed to issue an alarm in a timely manner. This feedback information is crucial for the system's self-learning and optimization.

[0069] Furthermore, encoding each type of feedback information into a vectorized representation means converting different types of feedback (such as real fires, false alarms, and missed alarms) into numerical vectors to facilitate processing by machine learning models. For example, one-hot encoding or vector embedding can be used. This vectorized feedback information is then fed into a meta-learning network. A meta-learning network is a model that learns how to learn; its purpose is to generate adjustments to parameters for updating the behavioral pattern baseline or for consistency judgments based on the input feedback information. These adjustments can be parameter increments, decrements, or new parameter values.

[0070] The assignment of a confidence score to each type of feedback information can be understood as a quantitative assessment of the reliability or accuracy of the feedback information. For example, real fire feedback confirmed by experienced operators may have a high confidence score, while feedback from uncertain sources may have a low confidence score. This confidence score indicates the reliability of each type of feedback information and participates in the output regulation of the meta-learning network. Specifically, the confidence score can serve as a weighting factor, giving greater influence to high-confidence feedback information and less influence to low-confidence feedback information when the meta-learning network calculates parameter adjustments, thereby ensuring the accuracy and stability of parameter adjustments.

[0071] In practical applications, adjusting the updated parameters or consistency judgment parameters of the behavioral pattern baseline based on parameter adjustment amounts and confidence scores means combining the parameter adjustment amounts output by the meta-learning network with the confidence scores of the feedback information to jointly determine the final parameter adjustment strategy. For example, if the meta-learning network suggests increasing a certain parameter, and this suggestion is based on high-confidence real fire feedback, then that parameter may be significantly adjusted; conversely, if the suggestion is based on low-confidence false alarm feedback, the adjustment may be reduced or ignored.

[0072] Furthermore, when conflicting feedback is received—for example, when the system receives both false alarms and actual fire reports within the same time period, or when feedback from different areas contradicts each other—an arbitration process is required to reconcile the conflicting feedback. This arbitration process aims to resolve inconsistencies between feedback information to determine the most reasonable direction and magnitude of parameter adjustments.

[0073] This application's solution effectively addresses the limitations of traditional methods in parameter adjustment by introducing a multi-dimensional feedback information processing mechanism. Through the aforementioned technical solution, this application can significantly improve the intelligence and robustness of early fire detection systems in terms of adaptive parameter adjustment.

[0074] In some preferred embodiments, a specific example is given below. Suppose a data center has deployed the early fire detection system described in this application. During a period of operation, the system triggered three alarms: First Alarm: The system issues a fire alarm, which is confirmed as a real fire after on-site verification (e.g., a short circuit and fire inside a server rack). At this point, the system receives feedback confirming the fire. This feedback is encoded as a vector and assigned a high confidence score (e.g., 0.9). Based on this feedback and the confidence score, the meta-learning network calculates the adjustment amount needed to modify the baseline update parameters of the behavioral pattern, making the system more sensitive to anomalies in similar features.

[0075] Second alarm: The system issues a fire alarm, but after on-site verification, it is found that the problem is a localized temperature increase caused by an air conditioning system malfunction, not a fire. At this point, the system receives a false alarm feedback. This feedback is encoded as a vector and assigned a moderate confidence score (e.g., 0.7, because air conditioning malfunctions can also have similar characteristics to fire). Based on this feedback and the confidence score, the meta-learning network calculates the adjustment amount needed to adjust the consistency judgment parameters to improve the ability to distinguish non-fire anomalies.

[0076] The third alarm: The system issues a fire alarm in one area, but almost simultaneously, sensor data in another area also shows anomalies and is manually flagged as a false alarm (e.g., cleaning agent evaporation causing an increase in volatile organic compound (VOC) sensor readings). At this point, the system receives conflict feedback. An arbitration process is triggered, and the system comprehensively analyzes: historical alarm records for the area (e.g., a high false alarm rate in the past), long-term operating data from each sensor (e.g., recent drift trends in VOC sensors), and environmental background change trends (e.g., recent cleaning work). Based on this information, the arbitration process assigns greater weight to false alarm feedback and adjusts parameters accordingly, for example, appropriately increasing the threshold for VOC sensor deviation information in consistency assessment to avoid similar false alarms.

[0077] Through the above process, the system can continuously optimize and adaptively adjust the parameters for updating the behavior pattern baseline or the parameters for consistency judgment based on feedback information of different types and reliability, as well as intelligent arbitration when facing conflicting information, thereby continuously improving the accuracy and robustness of early fire detection.

[0078] The steps described above for establishing and updating the baseline of the behavioral patterns of each type of sensor in the current environment based on sensor readings of each type of sensor include: Obtain information on the current operating mode of the data center, including the computing load of the servers, the operating status of the cooling system, and environmental control parameters. Based on the operating mode information, select a preset behavior pattern library that matches the current operating mode; In the preset behavior pattern library, multiple behavior pattern baselines are maintained for each type of sensor, and each behavior pattern baseline corresponds to a certain operating mode; Based on the historical data of the sensor, the behavior pattern baseline of the sensor in the current operating mode is established and updated in the behavior pattern baseline corresponding to the current operating mode.

[0079] Specifically, when establishing and updating behavioral pattern baselines, it is first necessary to obtain information on the current operating mode of the data center. This operating mode information is a key indicator reflecting the current working status of the data center, and may include server computing load, cooling system operating status, and environmental control parameters. For example, server computing load can indicate the intensity of current data processing, the cooling system operating status can reflect heat dissipation requirements and energy consumption, and environmental control parameters are directly related to the set values ​​of environmental indicators such as temperature and humidity. This information collectively depicts the operational profile of the data center at a specific point in time.

[0080] Furthermore, based on the acquired operating mode information, the system selects a set of baselines matching the current operating mode from a preset behavior pattern library. This preset behavior pattern library is a pre-built knowledge base that stores sensor behavior characteristics of the data center under different typical operating modes. For example, the library may include behavior pattern data corresponding to various operating modes such as "peak load mode," "off-peak load mode," and "maintenance mode." By matching the current operating mode, it ensures that subsequent baseline establishment and updates are based on the current actual operating conditions.

[0081] In the preset behavior pattern library, multiple behavior pattern baselines are maintained for each type of sensor. This means that for each type of sensor, such as temperature sensors, smoke sensors, and carbon monoxide sensors, there may be multiple behavior pattern baselines, each corresponding to a specific operating mode. For example, the normal temperature range and fluctuation characteristics of a temperature sensor under high-load server operating mode are different from those under low-load operating mode, so corresponding behavior pattern baselines need to be established separately.

[0082] Therefore, based on the historical data of the sensor, a behavioral pattern baseline for the sensor under the current operating mode is established and updated within the behavioral pattern baseline corresponding to the current operating mode. This means that the system utilizes the historical data of the sensor under the current specific operating mode to learn and construct the normal behavioral pattern baseline under that mode. As time goes by and new data continues to flow in, this baseline will be continuously updated to adapt to subtle changes and long-term drifts in the data center's operating mode, ensuring the real-time performance and accuracy of the baseline.

[0083] This application's solution integrates the establishment and updating of behavioral pattern baselines with actual operating conditions by incorporating data center operation mode information. When the data center operation mode changes, such as switching from low load to high load, the system can identify this mode shift and automatically switch to or update the behavioral pattern baseline corresponding to the high load mode. This mechanism allows the baseline to dynamically reflect the true "normal" state of the current environment, rather than a generalized, potentially inaccurate average state. Through the above technical solution, this application can significantly improve the accuracy and robustness of early fire detection systems in complex dynamic environments.

[0084] In some preferred embodiments, a specific example is given below. Assume a data center operates in two main modes: a high-load mode during the day and a low-load mode at night. During the day, the server's computing load is high, the cooling system typically operates at high power, and the ambient temperature may be maintained at 22-24 degrees Celsius. At night, the computing load decreases, the cooling system power is correspondingly reduced, and the ambient temperature may be maintained at 20-22 degrees Celsius.

[0085] The proposed solution first obtains information about the current operating mode of the data center. For example, by monitoring parameters such as server CPU utilization, memory usage, and cooling fan speed, the system can determine whether it is currently in a high-load or low-load mode.

[0086] When the system detects that it is currently in a high-load mode, it selects a set of behavior pattern baselines from a preset behavior pattern library that match the "high-load mode". Within this set, temperature sensors will have a specific baseline for high-load mode, which may reflect normal temperature fluctuations within the range of 22-24 degrees Celsius. Similarly, smoke sensors, carbon monoxide sensors, and others will also have their own behavior pattern baselines for high-load mode.

[0087] Subsequently, the system continuously establishes and updates the baseline of temperature behavior patterns under high-load mode based on historical data from the temperature sensor. For example, if historical data shows that the temperature typically fluctuates around 23 degrees Celsius under high-load mode, but suddenly rises to 26 degrees Celsius, this represents a significant deviation from the baseline for that mode, and the system will identify it as a potential anomaly.

[0088] Conversely, when the system detects that it is currently in a low-load mode, it switches to a baseline set of behavioral patterns that matches the "low-load mode." In this case, the temperature sensor baseline might reflect normal temperature fluctuations within the 20-22 degree Celsius range. If the temperature rises to 24 degrees Celsius at this point, although this temperature might be considered normal in high-load mode, in low-load mode it will deviate significantly from the current mode's baseline, thus being identified as abnormal by the system and triggering subsequent judgment processes.

[0089] In this way, the solution proposed in this application can ensure that the behavior pattern baseline can accurately reflect the current normal state under any operating mode, thereby effectively avoiding false alarms or missed alarms caused by changes in operating mode, and significantly improving the accuracy and reliability of early fire detection.

[0090] This application further proposes a method for collaboratively training the recurrent neural network model using a hierarchical federated learning framework across multiple data centers to overcome the aforementioned limitations. The method also includes: The recurrent neural network model is trained collaboratively using a hierarchical federated learning framework across multiple data centers; wherein, the hierarchical federated learning framework includes horizontal federation within data centers, vertical federation between data centers, and cross-regional hierarchical aggregation. The cross-regional hierarchical aggregation adopts a differential privacy-weighted aggregation strategy based on the historical accuracy of alarms; the differential privacy-weighted aggregation strategy adds Gaussian noise to the model gradient, and the noise level is adjusted inversely according to the historical accuracy of alarms.

[0091] Specifically, the hierarchical federated learning framework aims to achieve distributed and collaborative model training, enabling various data centers to jointly improve the performance of recurrent neural network models without sharing raw data. The framework is designed with a multi-level structure to adapt to data distributions and privacy requirements of varying granularities and scales. Within a data center, horizontal federation refers to the local aggregation of model parameters among different sensor groups, areas, or devices within a single data center. For example, a data center may include multiple server rooms; sensor data from each server room can independently train local models, and then the updates of these local models are aggregated to form a model at the data center level. This approach helps to fully utilize heterogeneous data within the data center and reduces the burden of internal data transmission.

[0092] Furthermore, vertical federation between data centers refers to the collaborative training of models between different data centers when they share common users or entities but possess different feature sets. For example, two data centers may both monitor the same batch of servers, but one data center primarily collects data from temperature and smoke sensors, while the other collects data from carbon monoxide and volatile organic compound sensors. Through vertical federation, a more comprehensive model can be jointly trained while protecting the privacy of each data source.

[0093] Furthermore, cross-regional hierarchical aggregation refers to the summarization and fusion of model updates from different geographical regions or data centers. At this level, to ensure the effectiveness and security of the aggregation, a differential privacy-weighted aggregation strategy based on historical alarm accuracy is adopted. The core of this strategy is to allocate the weight of model updates according to the historical alarm accuracy of each data center or region; that is, model updates with higher accuracy are given greater weight, thereby guiding the global model towards more reliable optimization.

[0094] This application's solution effectively addresses issues such as data privacy, data silos, and communication efficiency in traditional centralized training models by introducing a hierarchical federated learning framework. Through this technical solution, the application can significantly improve the training efficiency and generalization ability of early fire detection recurrent neural network models, while effectively ensuring data privacy and security.

[0095] As a specific implementation method, assume there are three geographically distinct data centers, A, B, and C, each deploying various types of sensors for early fire detection. Data center A primarily serves the North China region, data center B serves the East China region, and data center C serves the South China region.

[0096] First, within data center A, sensor data from its multiple server rooms is used to train their respective local recurrent neural network models. These local models are then aggregated within data center A through lateral federated learning to form a regional model for data center A. Similarly, data centers B and C also form their respective regional models through internal lateral federated learning.

[0097] Secondly, suppose data centers A and B share some common samples from monitoring certain servers, but collect different types of sensors. In this case, data centers A and B can perform vertical federated learning to collaboratively train a more comprehensive model to capture richer fire features without sharing the original data.

[0098] Finally, the regional models (or the models after vertical federation) of data centers A, B, and C need to be aggregated across geographical regions to form a global early fire detection model applicable to all regions. During the aggregation process, a differential privacy-weighted aggregation strategy based on historical alarm accuracy is employed. For example, if data center A has the highest historical alarm accuracy, followed by data center B, and then data center C has the lowest, then when aggregating their model gradients, the gradient of data center A will be assigned the largest weight, followed by data center B, and the smallest weight for data center C. Simultaneously, to protect privacy, Gaussian noise is added to the gradient of each data center inversely proportional to its historical alarm accuracy before it is uploaded. For example, due to its high accuracy, data center A will have a relatively small amount of noise added to its gradient to preserve its high-quality contribution; while data center C, with its lower accuracy, will have a relatively large amount of noise added to its gradient to further enhance privacy protection and prevent potentially low-quality updates from negatively impacting the global model. In this way, the final global recurrent neural network model will possess high accuracy, strong generalization ability, and strict privacy protection.

[0099] This application further proposes an online incremental update for the recurrent neural network model, specifically as follows: Monitor the degree of change in the distribution of the deviation information; When the degree of change in the distribution exceeds a preset drift threshold, an incremental learning process is triggered; A representative sample pool is constructed by sampling historical samples; The recurrent neural network model is updated using an elastic weight solidification method, which imposes constraints on important parameters to avoid catastrophic forgetting. Representative samples in the experience replay pool are jointly trained with the current incremental samples to mitigate catastrophic forgetting.

[0100] Specifically, online incremental updates refer to continuously training and adjusting the model based on new data after deployment, rather than retraining the entire model. The goal is to enable the model to adapt to new data distributions and environmental changes while maintaining its detection performance. Monitoring the degree of change in the distribution of the bias information can be understood as continuously evaluating the data distribution characteristics of the bias information input to the recurrent neural network model using statistical methods or machine learning techniques, such as mean, variance, covariance, and eigenvalues, and comparing them with historical distributions. For example, KL divergence, JS divergence, or distance-based metrics can be used to quantify the difference between the current bias information distribution and the baseline distribution. The aim is to promptly detect potential drifts in the data distribution, providing triggering conditions for incremental updates to the model.

[0101] In practical applications, when the degree of distribution change exceeds a preset drift threshold, it indicates a significant difference between the current data distribution and the data distribution during model training, triggering an incremental learning process. The preset drift threshold can be set based on the actual application scenario and the acceptable level of performance degradation, for example, through cross-validation or expert experience. The incremental learning process involves initiating a series of operations, including data collection, sample selection, model training, and parameter updates, to enable the model to learn new data patterns.

[0102] Furthermore, constructing an experience replay pool by sampling representative samples from historical data refers to selecting representative samples from past data and storing them in a buffer (i.e., the experience replay pool). These representative samples can be selected through clustering, importance sampling, or methods based on diversity measures to ensure they reflect the main characteristics of the historical data. The purpose of the experience replay pool is to provide historical data context during incremental learning, preventing the model from forgetting old knowledge while learning new knowledge.

[0103] The use of elastic weight consolidation (EWC) to update the parameters of the recurrent neural network model refers to imposing specific constraints on the updating of model parameters during incremental learning. EWC is a commonly used incremental learning technique. Its core idea is to identify model parameters that are important for previous tasks and penalize the updates of these important parameters when learning new tasks, thereby mitigating their changes.

[0104] Specifically, the Elastic Weight Solidification (EWC) method imposes constraints on important parameters to avoid catastrophic forgetting. Catastrophic forgetting refers to the phenomenon where a neural network's performance on previously learned tasks drops sharply due to parameter updates when learning a new task. EWC calculates the importance of parameters to previous tasks (e.g., through the Fisher information matrix) and applies regularization terms to the updates of important parameters, thereby retaining as much knowledge of old tasks as possible while learning new tasks. Furthermore, joint training with representative samples from the experience replay pool and current incremental samples further mitigates catastrophic forgetting. By training the model together with newly collected incremental samples, the model not only learns new data patterns but also consolidates its understanding of old patterns, effectively mitigating the catastrophic forgetting problem and ensuring the model maintains stable and high performance during continuous learning.

[0105] This application's solution effectively addresses the potential performance degradation issue of recurrent neural network models during long-term operation by introducing an online incremental update mechanism. Through the above technical solution, this application enables recurrent neural network models in early fire detection systems to possess online adaptive capabilities.

[0106] refer to Figure 2 , Figure 2 This is a schematic diagram of an early fire detection system based on a recurrent neural network provided in an embodiment of the present invention, which includes: The data acquisition and preprocessing module is used to acquire sensor readings for each type of sensor and perform noise reduction and normalization processing on the sensor readings for each type of sensor; each type of sensor includes one or more of the following: temperature sensor, smoke sensor, carbon monoxide sensor, volatile organic compound sensor, and ultrafine particulate matter sensor. The behavior pattern baseline establishment and update module is used to establish and update the behavior pattern baseline of each type of sensor in the current environment based on the sensor readings of each type of sensor; the behavior pattern baseline includes the statistical characteristics and temporal characteristics of the sensor readings; The deviation information calculation module is used to calculate the deviation information of each type of sensor based on the real-time readings of each type of sensor and the behavior pattern baseline. An anomaly probability assessment module is used to input the deviation information into a trained recurrent neural network model. The multi-scale temporal feature extraction layer of the recurrent neural network model adaptively adjusts the sampling weights of multiple sliding windows of different lengths based on the rate of change of the deviation information to sample the deviation information and obtain multi-scale temporal features. The bidirectional long short-term memory encoding layer of the recurrent neural network model performs forward and backward encoding on the multi-scale temporal features to obtain cross-scale interactive forward and backward encodings. The temporal attention fusion layer of the recurrent neural network model performs weighted fusion of the forward and backward encodings based on fire dynamics priors to output the anomaly probability of the area where each type of sensor is located. The consistency judgment module is used to perform a consistency judgment based on the deviation information when the anomaly probability exceeds a preset threshold, to confirm whether there is an abnormal trend pointing to a fire; the consistency judgment includes analyzing the changing trend of the deviation information and verifying whether there is a physical correlation between the deviation information. The alarm triggering module is used to trigger a fire alarm based on the consistency judgment result; The parameter adjustment module is used to adjust the update parameters of the behavior pattern baseline or the parameters of the consistency judgment based on the feedback information of the fire alarm.

[0107] This application proposes an early fire detection system based on recurrent neural networks. Through modular design and intelligent data processing, it aims to effectively address the declining accuracy and reliability of existing early fire detection systems in the face of continuous environmental changes and sensor aging and drift. The system employs a data acquisition and preprocessing module to collect and optimize multi-source sensor data in real time; a behavior pattern baseline establishment and update module to dynamically construct and adapt normal behavior patterns to environmental changes; and a deviation information calculation module to accurately quantify abnormal deviations. The core anomaly probability assessment module utilizes an advanced recurrent neural network model to deeply mine multi-scale temporal features and integrate fire dynamics priors, outputting high-confidence anomaly probabilities. Subsequently, a consistency judgment module further reduces false alarms through physical correlation verification, and an alarm triggering module issues timely warnings. Finally, a parameter adjustment module adaptively optimizes system parameters based on feedback information, forming a closed-loop learning mechanism to ensure the system maintains high accuracy and robustness in complex and changing environments, thereby achieving timely and reliable early fire detection.

[0108] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. An early fire detection method based on a recurrent neural network, executed by a computer system, characterized in that, include: Acquire sensor readings for each type of sensor, and perform noise reduction and normalization processing on the sensor readings for each type of sensor; each type of sensor includes one or more of the following: temperature sensor, smoke sensor, carbon monoxide sensor, volatile organic compound sensor, and ultrafine particulate matter sensor; Based on the sensor readings of each type of sensor, a baseline of the behavioral pattern of each type of sensor in the current environment is established and updated; the baseline of the behavioral pattern includes the statistical characteristics and temporal characteristics of the sensor readings. Based on the real-time readings of each type of sensor and the baseline of the behavior pattern, the deviation information of each type of sensor is calculated; The deviation information is input into a trained recurrent neural network model. The multi-scale temporal feature extraction layer of the recurrent neural network model adaptively adjusts the sampling weights of multiple sliding windows of different lengths according to the rate of change of the deviation information, and samples the deviation information to obtain multi-scale temporal features. The multi-scale temporal features are encoded forward and backward by the bidirectional long short-term memory encoding layer of the recurrent neural network model to obtain cross-scale interactive forward and backward encodings; The temporal attention fusion layer of the recurrent neural network model performs weighted fusion of the forward encoding and the backward encoding based on fire dynamics priors, and outputs the anomaly probability of the area where each type of sensor is located; When the anomaly probability exceeds a preset threshold, a consistency judgment is performed based on the deviation information to confirm whether there is an abnormal trend pointing to a fire; the consistency judgment includes analyzing the changing trend of the deviation information and verifying whether there is a physical correlation between the deviation information. A fire alarm is triggered based on the result of the consistency assessment. Based on the feedback information from the fire alarm, adjust the update parameters of the behavior pattern baseline or the parameters of the consistency judgment.

2. The early fire detection method based on recurrent neural networks according to claim 1, characterized in that, The recurrent neural network model outputs the anomaly probability, uncertainty estimate, anomaly type classification, and anomaly time window prediction. The uncertainty estimate is obtained by the recurrent neural network model through Monte Carlo dropout sampling and includes cognitive uncertainty and accidental uncertainty. The consistency judgment is based on the anomaly type classification, the uncertainty estimate, and the anomaly time window prediction, and uses Bayesian inference to perform multidimensional evidence fusion. The posterior probability after fusion is used to determine whether there is an abnormal trend pointing to a fire.

3. The early fire detection method based on recurrent neural networks according to claim 1, characterized in that, The method further includes: A fire propagation physical information embedding layer is constructed, in which the fire heat release rate, smoke spread speed, and toxic gas diffusion coefficient are embedded into the feature space of the recurrent neural network model; The fire propagation physical information embedding layer is used to dynamically adjust the embedding parameters based on the real-time readings of each type of sensor; the recurrent neural network model is used to adjust the network parameters in reverse through physical constraint loss terms so that the network output satisfies the fire dynamics conservation relationship.

4. The early fire detection method based on recurrent neural networks according to claim 1, characterized in that, The method further includes training the recurrent neural network model, specifically: Construct a training dataset for each type of sensor, including normal operating conditions, early fire operating conditions, and interference operating conditions. Establish the behavior pattern baseline and calculate the deviation information for the sensor readings in each type of sensor training dataset to obtain training samples of deviation information from the same source. The bias information training samples are segmented by a multi-scale sliding window to generate fixed-length time-series samples that cover short-term mutations and long-term drifts; Data augmentation is performed on the fixed-length time series samples based on the physical fire propagation model to generate synthetic samples that conform to the laws of fire dynamics. The main task is trained using focus loss, and the class weights of focus loss are dynamically adjusted according to the output of the physical consistency regularization term so that the network output satisfies the laws of fire dynamics. The recurrent neural network model is trained together to obtain a trained recurrent neural network model.

5. The early fire detection method based on recurrent neural networks according to claim 1, characterized in that, The step of performing a consistency judgment based on the deviation information to confirm whether there is an abnormal trend pointing to a fire when the anomaly probability exceeds a preset threshold includes: Analyze the deviation information to identify the main source of the current abnormal signal, and determine whether the current situation belongs to the first fire mode or the second fire mode based on the main source. If the current situation is determined to be the first fire mode, the first causal verification chain is activated to verify whether the deviation information of the volatile organic compound sensor and the ultrafine particulate matter sensor simultaneously shows a continuous upward trend, whether they are highly synchronized in time, whether they last for a preset duration, and to confirm whether the deviation information of the temperature sensor and the carbon monoxide sensor remains below a preset level within the same time window. If the current situation is determined to belong to the second fire mode, the second causal verification chain is activated to verify whether the deviation information simultaneously shows a consistent upward trend, and to check whether the consistent upward trend has a reasonable order of occurrence in physical terms. The consensus threshold is dynamically adjusted based on the uncertainty estimate of the current situation. Weights are assigned based on the verification results of the enabled causal verification chain, and it is determined whether the dynamically adjusted consensus threshold has been reached.

6. The early fire detection method based on recurrent neural networks according to claim 1, characterized in that, The step of adjusting the update parameters of the behavior pattern baseline or the parameters of the consistency judgment based on the feedback information of the fire alarm includes: Obtain feedback information from the fire alarm, including feedback on actual fires, false alarms, and missed alarms; Each type of feedback information is encoded into a vector representation and input into the meta-learning network to obtain parameter adjustment amounts; A confidence score is assigned to each type of feedback information, and the confidence score indicates the reliability of each type of feedback information and participates in the output adjustment of the meta-learning network; Based on the parameter adjustment amount and the confidence score, adjust the update parameters of the behavioral pattern baseline or the parameters of the consistency judgment; When conflict feedback information is received, the conflict feedback information is integrated through an arbitration process. The arbitration process determines the direction and magnitude of parameter adjustment based on the authenticity of historical alarm events, long-term operating data of each sensor, and the changing trend of the environmental background.

7. The early fire detection method based on recurrent neural networks according to claim 1, characterized in that, The step of establishing and updating the baseline of the behavioral patterns of each type of sensor in the current environment based on the sensor readings of each type of sensor includes: Obtain information on the current operating mode of the data center, including the computing load of the servers, the operating status of the cooling system, and environmental control parameters. Based on the operating mode information, select a preset behavior pattern library that matches the current operating mode; In the preset behavior pattern library, multiple behavior pattern baselines are maintained for each type of sensor, and each behavior pattern baseline corresponds to a certain operating mode; Based on the historical data of each type of sensor, establish and update the baseline of the behavior pattern of each type of sensor in the current operating mode.

8. The early fire detection method based on recurrent neural networks according to claim 1, characterized in that, The method further includes: The recurrent neural network model is trained collaboratively using a hierarchical federated learning framework across multiple data centers. The hierarchical federated learning framework includes horizontal federation within a data center, vertical federation between data centers, and cross-regional hierarchical aggregation. The cross-regional hierarchical aggregation adopts a differential privacy-weighted aggregation strategy based on alarm history accuracy. The differential privacy-weighted aggregation strategy adds Gaussian noise to the model gradient, and the noise level is adjusted inversely according to the alarm history accuracy.

9. The early fire detection method based on recurrent neural networks according to claim 1, characterized in that, The method further includes: incrementally updating the recurrent neural network model online, specifically: Monitor the degree of change in the distribution of the deviation information; When the degree of change in the distribution exceeds a preset drift threshold, an incremental learning process is triggered; A representative sample pool is constructed by sampling historical samples; The recurrent neural network model is updated using an elastic weight solidification method, which imposes constraints on important parameters to avoid catastrophic forgetting. Representative samples in the experience replay pool are jointly trained with the current incremental samples to mitigate catastrophic forgetting.

10. An early fire detection system based on a recurrent neural network, characterized in that, include: The data acquisition and preprocessing module is used to acquire sensor readings for each type of sensor and perform noise reduction and normalization processing on the sensor readings for each type of sensor; each type of sensor includes one or more of the following: temperature sensor, smoke sensor, carbon monoxide sensor, volatile organic compound sensor, and ultrafine particulate matter sensor. The behavior pattern baseline establishment and update module is used to establish and update the behavior pattern baseline of each type of sensor in the current environment based on the sensor readings of each type of sensor; the behavior pattern baseline includes the statistical characteristics and temporal characteristics of the sensor readings; The deviation information calculation module is used to calculate the deviation information of each type of sensor based on the real-time readings of each type of sensor and the behavior pattern baseline. An anomaly probability assessment module is used to input the deviation information into a trained recurrent neural network model. The multi-scale temporal feature extraction layer of the recurrent neural network model adaptively adjusts the sampling weights of multiple sliding windows of different lengths according to the rate of change of the deviation information and then samples the deviation information to obtain multi-scale temporal features. The multi-scale temporal features are encoded forward and backward by the bidirectional long short-term memory encoding layer of the recurrent neural network model to obtain cross-scale interactive forward and backward encodings; The temporal attention fusion layer of the recurrent neural network model performs weighted fusion of the forward encoding and the backward encoding based on fire dynamics priors, and outputs the anomaly probability of the area where each type of sensor is located; The consistency judgment module is used to perform a consistency judgment based on the deviation information when the anomaly probability exceeds a preset threshold, to confirm whether there is an abnormal trend pointing to a fire; the consistency judgment includes analyzing the changing trend of the deviation information and verifying whether there is a physical correlation between the deviation information. The alarm triggering module is used to trigger a fire alarm based on the result of the consistency judgment. The parameter adjustment module is used to adjust the update parameters of the behavior pattern baseline or the parameters of the consistency judgment based on the feedback information of the fire alarm.

Images