A method and system for early fire detection based on anomaly detection

Abstract

The application discloses a fire early detection method and system based on abnormality detection, and relates to the technical field of fire early detection. The method comprises the following steps: S1, synchronously collecting temperature signals, specific gas concentration signals and infrared radiation intensity signals in an environment to obtain multidimensional environment collection data; and S2, based on the multidimensional environment collection data, constructing an environment adaptive dynamic reference library, and dynamically updating the environment adaptive dynamic reference library to obtain a dynamic reference curve of the normal fluctuation range of the environment. The application synchronously collects three types of signals, i.e. temperature signals, specific gas concentration signals and infrared radiation intensity signals, and realizes accurate alignment in the time dimension, combines a sliding time window and a Gaussian mixture model to construct an environment adaptive dynamic reference library, dynamically adjusts the normal fluctuation range, and then separates fire early signals and environmental interference signals through frequency-time domain joint processing, and verifies and confirms the signal effectiveness through double-dimensional feature verification.

Description

Technical Field

[0001] This invention relates to the field of early fire detection technology, specifically to a method and system for early fire detection based on anomaly detection. Background Technology

[0002] In the early stages of a fire, obvious smoke and flames don't immediately appear. Instead, subtle changes in environmental signals first emerge. These early anomalies include weak fluctuations in temperature, gas composition, or energy radiation. Timely detection and identification of these signals are crucial to preventing the fire from spreading and minimizing losses. Anomaly detection technology, with its ability to capture signals in the environment that deviate from normal conditions, is widely used in the field of early fire detection. Its core logic is to continuously monitor environmental signals and compare them with a baseline signal under normal conditions to identify abnormal patterns that may indicate the occurrence of a fire.

[0003] However, existing fire early detection technologies based on anomaly detection face the problem of signal differentiation in complex environments. Specifically, the weak characteristic signals generated in the early stages of a fire are highly similar in feature to naturally occurring interference signals in the environment. Existing technologies lack effective means to accurately separate these two types of signals, making it impossible to accurately extract the unique anomaly features of the early stages of a fire. Consequently, the early warning results are easily affected by interference signals, making it difficult to achieve stable and reliable early detection in the nascent stage of a fire. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for early fire detection based on anomaly detection, which solves the problems of existing technologies in distinguishing weak early fire signals from environmental interference signals in complex environments, the susceptibility of early warnings to interference, and the difficulty in reliably achieving early fire detection.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for early fire detection based on anomaly detection, comprising: S1. Synchronously collect temperature signals, specific gas concentration signals, and infrared radiation intensity signals in the environment to obtain multi-dimensional environmental data. S2. Based on multi-dimensional environmental data collection, construct an environmental adaptive dynamic benchmark library, and dynamically update the environmental adaptive dynamic benchmark library to obtain a dynamic benchmark curve of the normal fluctuation range of the environment. S3. Perform frequency-time domain joint feature enhancement processing on multi-dimensional environmental acquisition data to separate early fire signals from environmental interference signals and obtain enhanced signal features; S4. Perform two-dimensional feature verification on the enhanced signal features to confirm the effectiveness of the signal features; S5. When both the dual-dimensional feature verifications meet the preset conditions, a fire classification early warning is triggered.

[0006] Preferably, step S1 includes: Temperature signals are acquired using a high-precision thermistor, with the sampling frequency set to 10Hz. The gas sensor collects specific gas concentration signals and specifically detects changes in carbon monoxide and formaldehyde concentrations. Infrared radiation intensity signals are collected by an infrared radiation sensor, which can also directionally collect energy radiation signals in specific wavelength bands. Temperature signals, specific gas concentration signals, and infrared radiation intensity signals are transmitted to the data processing unit through a synchronous triggering mechanism to ensure precise alignment of the three types of signals in the time dimension.

[0007] Preferably, step S2 includes: Using a sliding time window algorithm, we continuously extract multi-dimensional environmental data collected in the past 24 hours as analysis samples. A Gaussian mixture model is used to perform cluster analysis on the analysis samples, automatically delineating the boundaries of normal environmental fluctuations and forming a dynamic baseline curve; When the environment undergoes long-term, slow changes, the dynamic benchmark library gradually adjusts the range of the dynamic benchmark curve with an update factor of 0.05. Set a mutation suppression threshold to avoid abnormal fluctuations in the dynamic baseline curve caused by short-term drastic disturbances.

[0008] Preferably, the step of performing cluster analysis on the analysis samples using a Gaussian mixture model includes: Based on the analysis samples ,in A data point for collecting data from a multi-dimensional environment. The number of samples; The Gaussian mixture model is parameter estimated using the expectation-maximization algorithm. The probability density function of the Gaussian mixture model is expressed as: ; in, The number of Gaussian components. For the first The weights of the Gaussian components, For the first A multivariate Gaussian distribution with Gaussian components, the mean of which is The covariance matrix is , This is the parameter set of the Gaussian mixture model; The range of the dynamic baseline curve is determined based on the parameter set.

[0009] Preferably, the frequency-time domain joint feature enhancement processing of the multi-dimensional environmental acquisition data includes: At the frequency domain level, multi-dimensional environmental acquisition data is converted into frequency domain data through Fast Fourier Transform; The formula for Fast Fourier Transform is: ; in, It is a time-domain signal. It is a frequency domain component. It is the number of signal sampling points. It is a frequency index; By using an adaptive filtering algorithm, specific frequency band components corresponding to early fire signals in the frequency domain data are retained, while interference frequency bands corresponding to interference signals are filtered out.

[0010] Preferably, the frequency-time domain joint feature enhancement processing of the multi-dimensional environmental acquisition data further includes: At the time domain level, trend features and duration features of multi-dimensional environmental data collection are extracted; The trend slope of the trend feature is calculated using a trend fitting algorithm. : ; in, The number of data points within the time window. For time indexing, The signal value; By filtering by duration, we can determine the duration of a trend feature that exceeds a preset threshold.

[0011] Preferably, the two-dimensional feature verification of the enhanced signal features includes: Perform feature matching degree verification on the enhanced signal feature vector. Compared with the preset early fire feature template vector The two are compared and their similarity scores are calculated. ; Similarity score The formula is: ; in, The dimension of the feature vector. The first eigenvector of the enhanced signal One portion, The first feature template vector of the early stage of fire One component; When the similarity score reaches 85% or higher, the time series consistency verification stage begins.

[0012] Preferably, the step of performing two-dimensional feature verification on the enhanced signal features further includes: Perform time series consistency verification and continuously monitor the enhanced signal characteristics; The enhanced signal characteristics were confirmed to remain stable over three consecutive sampling periods.

[0013] Preferably, the method further includes: The system self-calibration process is initiated every 72 hours, sending analog signals with known parameters to the sensor array through a built-in standard signal source; By comparing the deviation between the sensor's acquisition results and the standard signal, the sensor's sensitivity parameters are automatically adjusted. When a single sensor detects an anomaly, cross-validation of data from multiple sensors automatically uses collaborative data from two other types of sensors to supplement the judgment.

[0014] The present invention also provides a fire early detection system based on anomaly detection, comprising: The multi-dimensional signal synchronous acquisition module is used to synchronously acquire temperature signals, specific gas concentration signals, and infrared radiation intensity signals in the environment to obtain multi-dimensional environmental acquisition data. The environment adaptive dynamic benchmark update unit is used to construct an environment adaptive dynamic benchmark library based on multi-dimensional environmental data, and to dynamically update the environment adaptive dynamic benchmark library to obtain a dynamic benchmark curve of the normal fluctuation range of the environment. The frequency-time domain joint feature enhancement module is used to perform frequency-time domain joint feature enhancement processing on multi-dimensional environmental acquisition data to separate early fire signals from environmental interference signals and obtain enhanced signal features. The dual-dimensional feature verification and early warning triggering unit is used to perform dual-dimensional feature verification on the enhanced signal features, confirm the validity of the signal features, and trigger a fire classification early warning when both dual-dimensional feature verifications meet the preset conditions.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention simultaneously collects three types of signals—temperature, specific gas concentration, and infrared radiation intensity—and achieves precise alignment in the time dimension. It combines a sliding time window and a Gaussian mixture model to construct an environmentally adaptive dynamic benchmark library, dynamically adjusting the normal fluctuation range. Then, through joint frequency-time domain processing, it separates early fire signals from environmental interference signals. The validity of the signals is confirmed through dual-dimensional feature verification. Coupled with periodic self-calibration and multi-sensor cross-verification mechanisms, this invention effectively solves the problem of existing technologies' difficulty in accurately distinguishing weak early fire signals from environmental interference signals. It can accurately extract unique abnormal features in the early stages of a fire, significantly reducing the probability of false alarms and achieving stable and reliable early detection in the nascent stage of a fire. This buys time to prevent the fire from spreading and reduce losses. Simultaneously, it improves the system's adaptability in complex environments and its long-term operational stability, allowing it to continue operating normally even when a single sensor fails. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system structure diagram of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This invention provides a method for early fire detection based on anomaly detection, comprising: S1. Synchronously collect temperature signals, specific gas concentration signals, and infrared radiation intensity signals in the environment to obtain multi-dimensional environmental data. S2. Based on multi-dimensional environmental data collection, construct an environmental adaptive dynamic benchmark library, and dynamically update the environmental adaptive dynamic benchmark library to obtain a dynamic benchmark curve of the normal fluctuation range of the environment. S3. Perform frequency-time domain joint feature enhancement processing on multi-dimensional environmental acquisition data to separate early fire signals from environmental interference signals and obtain enhanced signal features; S4. Perform two-dimensional feature verification on the enhanced signal features to confirm the effectiveness of the signal features; S5. When both the dual-dimensional feature verifications meet the preset conditions, a fire classification early warning is triggered.

[0019] Specifically, firstly, three signals strongly correlated with early-stage fire characteristics—temperature, specific gas concentration, and infrared radiation intensity—were selected. These three signals reflect environmental changes in the early stages of a fire from different physical dimensions, and their coordinated acquisition can compensate for the susceptibility of individual signals to interference. Temperature signals are acquired using high-precision thermistors at a sampling frequency of 10Hz. This frequency ensures the capture of subtle temperature fluctuations without generating excessive redundant data. Specific gas concentration signals focus on carbon monoxide and formaldehyde, typical products of incomplete combustion of organic matter in the early stages of a fire, and are specifically detected using dedicated gas sensors. Infrared radiation intensity signals are acquired directionally using infrared radiation sensors, focusing on capturing specific wavelengths of energy radiation released in the early stages of a fire, such as the mid-infrared band, which is more easily distinguishable from environmental interference radiation in the early stages. To ensure precise temporal alignment of the three signals, a synchronous triggering mechanism is used to transmit data to the data processing unit. Specifically, a unified synchronous clock signal controls the start and stop of sensor acquisition, keeping the sampling timestamp error of the three signals within the microsecond range, laying a data foundation for subsequent joint analysis.

[0020] The core of the dynamic benchmark library is to dynamically adjust the normal fluctuation range of the environment according to the actual situation, avoiding the problem that a fixed benchmark is difficult to adapt to environmental changes. A sliding time window algorithm is used to continuously extract multi-dimensional environmental data collected in the most recent 24 hours as analysis samples. The 24-hour time span can cover the periodic environmental changes in most scenarios. Since the three types of signals have different dimensions, normalization processing is required before cluster analysis. Temperature, gas concentration, and infrared radiation intensity data are uniformly mapped to the [0,1] interval to ensure dimension consistency. Then, a Gaussian mixture model is used to perform cluster analysis on the normalized analysis samples, automatically defining the boundaries of the normal fluctuation range of the environment and forming a dynamic benchmark curve. When the environment undergoes long-term slow changes, such as gradual temperature changes due to seasonal changes, the dynamic benchmark library gradually adjusts the benchmark curve range with an update coefficient of 0.05. This rate can adapt to environmental changes in a timely manner without causing benchmark instability due to excessively rapid updates. Simultaneously, a mutation suppression threshold is set. For example, if the signal value at a certain moment exceeds 30% of the current benchmark range and the duration is less than 1 second, it is judged as a short-term severe disturbance, and the benchmark curve adjustment is not triggered to avoid the benchmark being affected by instantaneous interference.

[0021] At the frequency domain level, multi-dimensional environmental data is converted into frequency domain data using Fast Fourier Transform (FFT). The frequency distribution bands of early-stage fire signals and interference signals differ, and this characteristic is used for initial separation. At the time domain level, trend and duration characteristics of the signals are extracted. Early-stage fire signals exhibit continuous and stable trends, while interference signals are mostly transient fluctuations. These two characteristics are used to further filter for valid signals. The joint processing of the frequency and time domains significantly improves the accuracy of signal separation, solving the problem that single-domain processing is insufficient to completely distinguish between the two types of signals.

[0022] Next, a two-dimensional feature verification is performed to ensure that the enhanced signal features truly correspond to early fire phenomena. The first dimension is feature matching verification, which compares the enhanced signal feature vector with the preset early fire feature template vector, and calculates a similarity score to determine the degree of feature matching. The second dimension is time series consistency verification, which continuously monitors the enhanced signal features to ensure that their trend remains stable over multiple consecutive sampling periods, avoiding misjudging instantaneous false features as fire signals. This two-dimensional verification provides double protection, further reducing the probability of false alarms.

[0023] When both dimensions of feature verification meet the preset conditions, the warning level is divided according to the feature matching score and the signal trend strength. For example, it is divided into alert warning, attention warning and emergency warning. Different levels correspond to different alarm methods and response procedures.

[0024] In this embodiment, step S1 includes: Temperature signals are acquired using a high-precision thermistor, with the sampling frequency set to 10Hz. The gas sensor collects specific gas concentration signals and specifically detects changes in carbon monoxide and formaldehyde concentrations. Infrared radiation intensity signals are collected by an infrared radiation sensor, which can also directionally collect energy radiation signals in specific wavelength bands. Temperature signals, specific gas concentration signals, and infrared radiation intensity signals are transmitted to the data processing unit through a synchronous triggering mechanism to ensure precise alignment of the three types of signals in the time dimension.

[0025] Specifically, a high-precision thermistor with an accuracy of ±0.1℃ is used for temperature signal acquisition. A sampling frequency of 10Hz ensures that a temperature data point is acquired every 0.1 seconds, capturing minute temperature fluctuations within 0.5℃ in the early stages of a fire. The gas sensor is an electrochemical sensor, with a detection range of 0-1000ppm for carbon monoxide concentration changes and 0-10ppm for formaldehyde concentration changes. These concentration ranges cover the typical gas release range in the early stages of a fire, and the sensor's response time is less than 2 seconds, allowing for rapid capture of subtle changes in gas concentration. The infrared radiation sensor uses a focal plane array sensor to directionally acquire mid-wave infrared energy radiation signals in the 3-5μm band. This band is the main distribution band of thermal radiation in the early stages of a fire and is less affected by ambient light interference. The synchronization triggering mechanism is achieved through the synchronization clock signal output by the data processing unit. After receiving the synchronization clock signal, the sensor array starts acquisition simultaneously. After acquisition, the data is transmitted to the data processing unit along with the timestamp. The data processing unit aligns the three types of signals through the timestamp to ensure that the three types of data at the same time correspond one-to-one, providing an accurate time reference for subsequent joint analysis.

[0026] In this embodiment, step S2 includes: Using a sliding time window algorithm, we continuously extract multi-dimensional environmental data collected in the past 24 hours as analysis samples. A Gaussian mixture model is used to perform cluster analysis on the analyzed samples to automatically delineate the boundaries of the normal fluctuation range of the environment and form a dynamic baseline curve; that is, a Gaussian mixture model is used to estimate the probability density of the analyzed samples and delineate the normal fluctuation range. When the environment undergoes long-term, slow changes, the dynamic benchmark library gradually adjusts the range of the dynamic benchmark curve with an update factor of 0.05. Set a mutation suppression threshold to avoid abnormal fluctuations in the dynamic baseline curve caused by short-term drastic disturbances.

[0027] Specifically, the sliding time window has a fixed length of 24 hours, and slides forward in 1-minute increments. After each slide, older data exceeding 24 hours is automatically discarded, retaining only the latest 24-hour data as the analysis sample to ensure that the sample always reflects the recent environmental conditions. When using a Gaussian mixture model for cluster analysis, the multi-dimensional normalized data within the past 24 hours is preprocessed to remove obvious outliers before being input into the model. The number of Gaussian components K in the Gaussian mixture model is adaptively determined based on the distribution complexity of the samples, typically ranging from 3 to 5, which can cover environmental fluctuation patterns in most scenarios. When the environment undergoes long-term, slow changes, such as an indoor temperature gradually rising from 25°C to 30°C in summer, the dynamic benchmark library adjusts the benchmark curve with an update coefficient of 0.05. That is, each update sets the weight of new environmental data to 0.05 and the weight of historical benchmark data to 0.95, gradually adjusting the benchmark range through a weighted average to avoid abrupt benchmark changes. The mutation suppression threshold is determined based on historical interference data. For example, it is set to 1.3 times the upper limit and 0.7 times the lower limit of the current reference range. When the signal value exceeds the threshold but the duration is less than 1 second, it is judged as a short-term severe interference. The reference curve is not adjusted, but only the abnormal event is recorded to ensure the stability of the reference curve.

[0028] In this embodiment, a Gaussian mixture model is used to perform cluster analysis on the analysis samples, including: Based on the analysis samples ,in A data point for collecting data from a multi-dimensional environment. The number of samples; The Gaussian mixture model is parameter estimated using the expectation-maximization algorithm. The probability density function of the Gaussian mixture model is expressed as: ; in, The number of Gaussian components. For the first The weights of the Gaussian components, For the first A multivariate Gaussian distribution with Gaussian components, the mean of which is The covariance matrix is , This is the parameter set of the Gaussian mixture model; The range of the dynamic baseline curve is determined based on the parameter set.

[0029] Specifically, the analysis sample It consists of multi-dimensional environmental data collected over the past 24 hours, where each data point All are three-dimensional vectors, corresponding to the normalized temperature signal value, gas concentration signal value, and infrared radiation intensity signal value, respectively. The total number of samples within 24 hours, calculated at a sampling frequency of 10Hz, is N = 24 × 3600 × 10 = 864000. The expectation-maximization algorithm is used to estimate the parameters of the Gaussian mixture model. The specific steps are: first, initialize the parameter set. Including the weights of each Gaussian component mean Covariance Matrix ,in The initial value is set to 1 / K. The initial values ​​were obtained by clustering the samples using the K-means algorithm. The initial value is set to the identity matrix. Then, the expectation step and the maximization step are performed iteratively until the parameters converge. The expectation step calculates the posterior probability of each sample belonging to each Gaussian component, and the maximization step updates the parameter set based on the posterior probabilities. The probability density function of the Gaussian mixture model is: in, The number of Gaussian components. The weight of the k-th Gaussian component and satisfying , The k-th Gaussian component is a multivariate Gaussian distribution. This is the mean vector of this component, with dimensions equal to the data points. Consistent, This is the covariance matrix of the component, used to describe the degree of dispersion of the data distribution within the component. This is the complete set of parameters for the Gaussian mixture model. After the parameters converge, based on the distribution range of each Gaussian component, the interval covering 95% of the samples is taken as the normal fluctuation range of the environment, thus forming a dynamic baseline curve.

[0030] In this embodiment, frequency-time domain joint feature enhancement processing is performed on multi-dimensional environmental acquisition data, including: At the frequency domain level, multi-dimensional environmental acquisition data is converted into frequency domain data through Fast Fourier Transform; The formula for Fast Fourier Transform is: ; in, It is a time-domain signal. It is a frequency domain component. It is the number of signal sampling points. It is a frequency index; By using an adaptive filtering algorithm, specific frequency band components corresponding to early fire signals in the frequency domain data are retained, while interference frequency bands corresponding to interference signals are filtered out.

[0031] Specifically, the frequency domain processing first involves performing a Fast Fourier Transform (FFT) on each type of signal in the multi-dimensional environmental acquisition data. The formula for the Fast Fourier Transform is: ; in It is the original signal value in the time domain. These are the transformed frequency domain components. It is the number of sampling points for a single signal. It is a frequency index, with values ​​ranging from 0 to N-1. The unit is the imaginary unit. This transformation converts two types of signals that are difficult to distinguish in the time domain into components of different frequency bands in the frequency domain. Based on extensive experimental data, the specific frequency band of early-stage fire signals in the frequency domain is 0.1-1Hz, while the interference frequency band of environmental interference signals is mostly concentrated above 5Hz. Therefore, the filtering parameters of the adaptive filtering algorithm are set to retain the 0.1-1Hz frequency band components and filter out the interference frequency band above 5Hz. The adaptive filtering algorithm uses the minimum mean square error algorithm, adjusting the filtering coefficients in real time to minimize the mean square error between the output signal and the desired signal (i.e., the ideal early-stage fire frequency domain signal), ensuring accurate preservation of the target frequency band components under different environments and achieving signal separation at the frequency domain level.

[0032] In this embodiment, the frequency-time domain joint feature enhancement processing of multi-dimensional environmental acquisition data also includes: At the time domain level, trend features and duration features of multi-dimensional environmental data collection are extracted; The trend slope of the trend feature is calculated using a trend fitting algorithm. : ; in, The number of data points within the time window. For time indexing, The signal value; By filtering by duration, we can determine the duration of a trend feature that exceeds a preset threshold.

[0033] Specifically, time-domain processing first determines the time window length, for example, set to 10 seconds, corresponding to 100 sampling points, with a sampling frequency of 10Hz, i.e., L=100. Within each time window, trend features and duration features of the signal are extracted. The trend features are then used to calculate the trend slope using a trend fitting algorithm. The formula is: ; in The number of data points within the time window. This is a time index, with values ​​from 1 to L. This is the normalized signal value corresponding to this time index. Trend slope. The positive or negative sign of the value reflects the direction of signal change, while the absolute value reflects the rate of change. The trend slope of early fire signals is mostly positive and the absolute value gradually increases. The duration screening step sets a preset threshold, such as 3 seconds. That is, when the duration of the trend feature exceeds 3 seconds, it is judged as a potential early fire signal trend. If the duration is less than 3 seconds, it is judged as transient interference. This screening can further eliminate short-term fluctuation interference signals.

[0034] In this embodiment, the enhanced signal features are verified using a two-dimensional feature model, including: Perform feature matching degree verification on the enhanced signal feature vector. Compared with the preset early fire feature template vector The two are compared and their similarity scores are calculated. ; Similarity score The formula is: ; in, The dimension of the feature vector. The first eigenvector of the enhanced signal One portion, The first feature template vector of the early stage of fire One component; When the similarity score reaches 85% or higher, the time series consistency verification stage begins.

[0035] Specifically, the feature matching verification uses a cosine similarity algorithm to calculate the similarity score between the enhanced signal feature vector A and the preset early fire feature template vector B. The enhanced signal feature vector A is a three-dimensional vector, corresponding to the feature values ​​of temperature, gas concentration, and infrared radiation intensity after joint processing in the frequency and time domains, and has been normalized. The preset early fire feature template vector B is extracted from early data from a large number of fire simulation experiments and is also a three-dimensional normalized vector. Similarity score. The calculation formula is: ; in M represents the dimension of the feature vector, where M=3. For the i-th component of the enhanced signal feature vector A, Let be the i-th component of the early fire feature template vector B. The similarity score calculated by this formula ranges from [0,1], with a higher score indicating greater similarity between the two sets of features. When the similarity score reaches 85% or higher, it indicates that the current signal features highly match the early fire features, and the process proceeds to the subsequent time series consistency verification stage; if the score is below 85%, it is determined to be a non-fire signal, and data is collected again.

[0036] In this embodiment, the two-dimensional feature verification of the enhanced signal features also includes: Perform time series consistency verification and continuously monitor the enhanced signal characteristics; The enhanced signal characteristics were confirmed to remain stable over three consecutive sampling periods.

[0037] Specifically, time series consistency verification is used to confirm the stability of signal characteristics and avoid misjudging interference signals that only occasionally match. The sampling period is consistent with the signal acquisition frequency, i.e., 0.1 seconds, continuously monitoring the enhanced signal characteristics. The verification standard is to confirm that the trend of the signal characteristics remains stable within three consecutive sampling periods. Specific criteria include that the change in the trend slope does not exceed 10%, and the fluctuation range of each component of the feature vector does not exceed 5%. For example, if the trend slopes in three consecutive sampling periods are 0.2, 0.21, and 0.205, with changes all within 10% and fluctuations of each feature component not exceeding 5%, then the time series consistency verification is considered successful. If the trend in any of the sampling periods shows a significant reversal or the fluctuation exceeds the threshold, the verification fails, and the signal is identified as interference.

[0038] In this embodiment, the method further includes: The system self-calibration process is initiated every 72 hours, sending analog signals with known parameters to the sensor array through a built-in standard signal source; By comparing the deviation between the sensor's acquisition results and the standard signal, the sensor's sensitivity parameters are automatically adjusted. When a single sensor detects an anomaly, cross-validation of data from multiple sensors automatically uses collaborative data from two other types of sensors to supplement the judgment.

[0039] Specifically, the system's self-calibration process automatically starts every 72 hours to ensure the stability and data accuracy of the sensors during long-term operation. A built-in standard signal source can generate simulated signals with known parameters, such as a 0.3°C increase in temperature, a 5ppm increase in carbon monoxide concentration, and a 0.2 ppm increase in infrared radiation intensity. These parameters correspond to typical subtle changes in the early stages of a fire. After receiving the simulated signals, the sensor array collects them. The data processing unit compares the collected results with the standard signals. If the deviation is within the allowable range, such as ±5%, the current sensor sensitivity parameters are maintained. If the deviation exceeds the allowable range, the sensor's amplification factor, threshold, and other sensitivity parameters are automatically adjusted until the deviation between the collected results and the standard signals meets the requirements. When a single sensor exhibits abnormal data, such as temperature sensor data consistently exceeding the dynamic reference range and inconsistent with the trends of data from the other two types of sensors, the system automatically initiates multi-sensor data cross-validation. It uses collaborative data from the gas concentration sensor and infrared radiation sensor for supplementary judgment. If the data from the latter two types of sensors are not abnormal, the temperature sensor is determined to be faulty, a sensor fault warning is issued, and analysis is performed only based on normal sensor data, ensuring the system continues to function normally even when a single sensor fails.

[0040] Please see Figure 2 The present invention also provides a fire early detection system based on anomaly detection, comprising: The multi-dimensional signal synchronous acquisition module is used to synchronously acquire temperature signals, specific gas concentration signals, and infrared radiation intensity signals in the environment to obtain multi-dimensional environmental acquisition data. The environment adaptive dynamic benchmark update unit is used to construct an environment adaptive dynamic benchmark library based on multi-dimensional environmental data, and to dynamically update the environment adaptive dynamic benchmark library to obtain a dynamic benchmark curve of the normal fluctuation range of the environment. The frequency-time domain joint feature enhancement module is used to perform frequency-time domain joint feature enhancement processing on multi-dimensional environmental acquisition data to separate early fire signals from environmental interference signals and obtain enhanced signal features. The dual-dimensional feature verification and early warning triggering unit is used to perform dual-dimensional feature verification on the enhanced signal features, confirm the validity of the signal features, and trigger a fire classification early warning when both dual-dimensional feature verifications meet the preset conditions.

[0041] Specifically, the multi-dimensional signal synchronous acquisition module consists of a sensor array and a synchronization trigger circuit. The sensor array includes a high-precision thermistor, an electrochemical gas sensor, and a focal plane array infrared radiation sensor, responsible for acquiring temperature, specific gas concentration, and infrared radiation intensity signals, respectively. The synchronization trigger circuit is connected to the data processing unit, receives the synchronization clock signal output by the data processing unit, and controls the three sensors to start and stop acquisition simultaneously, ensuring data time alignment. The acquired multi-dimensional environmental data is sent in real-time to the environmental adaptive dynamic benchmark update unit via a data transmission interface.

[0042] The environmental adaptive dynamic benchmark update unit consists of a data storage module and a benchmark calculation module. The data storage module uses a circular buffer to continuously store multi-dimensional environmental data collected in the last 24 hours, automatically overwriting expired data. The benchmark calculation module incorporates a sliding time window algorithm and a Gaussian mixture model algorithm. It first normalizes the stored data stream, then calculates the normal fluctuation range of the environment using the algorithm, generates a dynamic benchmark curve, and dynamically adjusts the benchmark range according to environmental changes. Simultaneously, it executes mutation suppression logic to ensure benchmark stability.

[0043] The frequency-time domain joint feature enhancement module comprises a frequency domain processing unit and a time domain processing unit. The frequency domain processing unit incorporates a Fast Fourier Transform (FFT) algorithm and an adaptive filtering algorithm to convert multi-dimensional environmental data into frequency domain data, filter out interfering frequency bands, and retain the target frequency band components. The time domain processing unit incorporates a trend fitting algorithm and duration filtering logic to extract signal trend and duration features. These features are then fused with the frequency domain processing results to obtain enhanced signal features, achieving effective separation of early fire signals from interfering signals.

[0044] The dual-dimensional feature verification and early warning triggering unit consists of a feature matching unit, a time-series verification unit, and an early warning output unit. The feature matching unit calculates the similarity score between the enhanced signal features and the template features using a cosine similarity algorithm. The time-series verification unit monitors the temporal consistency of the signal features. Only when both verification steps meet the requirements can the early warning output unit trigger a fire classification early warning. The early warning output unit can output early warning information through various methods such as audible and visual alarms, SMS notifications, and platform push notifications. Different early warning levels correspond to different alarm intensities and notification ranges to ensure timely response from relevant personnel.

[0045] In summary, this invention simultaneously collects three types of signals—temperature, specific gas concentration, and infrared radiation intensity—and achieves precise alignment in the time dimension. It combines a sliding time window and a Gaussian mixture model to construct an environmentally adaptive dynamic benchmark library, dynamically adjusting the normal fluctuation range. Then, through joint frequency-time domain processing, it separates early fire signals from environmental interference signals. Signal validity is confirmed through dual-dimensional feature verification. Coupled with periodic self-calibration and multi-sensor cross-validation mechanisms, this invention effectively solves the problem of existing technologies' difficulty in accurately distinguishing weak early fire signals from environmental interference signals. It can accurately extract unique abnormal features in the early stages of a fire, significantly reducing the probability of false alarms and achieving stable and reliable early detection in the nascent stage of a fire. This buys time to prevent fire spread and reduce losses, while also improving the system's adaptability and long-term operational stability in complex environments. It can still operate normally even when a single sensor fails.

[0046] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0047] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for early fire detection based on anomaly detection, characterized in that, include: S1. Synchronously collect temperature signals, specific gas concentration signals, and infrared radiation intensity signals in the environment to obtain multi-dimensional environmental data. S2. Based on multi-dimensional environmental data collection, construct an environmental adaptive dynamic benchmark library, and dynamically update the environmental adaptive dynamic benchmark library to obtain a dynamic benchmark curve of the normal fluctuation range of the environment. S3. Perform frequency-time domain joint feature enhancement processing on multi-dimensional environmental acquisition data to separate early fire signals from environmental interference signals and obtain enhanced signal features; S4. Perform two-dimensional feature verification on the enhanced signal features to confirm the effectiveness of the signal features; S5. When both the dual-dimensional feature verifications meet the preset conditions, a fire classification early warning is triggered.

2. The fire early detection method based on anomaly detection according to claim 1, characterized in that, Step S1 includes: Temperature signals are acquired using a high-precision thermistor, with the sampling frequency set to 10Hz. The gas sensor collects specific gas concentration signals and specifically detects changes in carbon monoxide and formaldehyde concentrations. Infrared radiation intensity signals are collected by an infrared radiation sensor, which can also directionally collect energy radiation signals in specific wavelength bands. Temperature signals, specific gas concentration signals, and infrared radiation intensity signals are transmitted to the data processing unit through a synchronous triggering mechanism to ensure precise alignment of the three types of signals in the time dimension.

3. The fire early detection method based on anomaly detection according to claim 1, characterized in that, Step S2 includes: Using a sliding time window algorithm, we continuously extract multi-dimensional environmental data collected in the past 24 hours as analysis samples. A Gaussian mixture model is used to perform cluster analysis on the analysis samples, automatically delineating the boundaries of normal environmental fluctuations and forming a dynamic baseline curve; When the environment undergoes long-term, slow changes, the dynamic benchmark library gradually adjusts the range of the dynamic benchmark curve with an update factor of 0.

05. Set a mutation suppression threshold to avoid abnormal fluctuations in the dynamic baseline curve caused by short-term drastic disturbances.

4. The fire early detection method based on anomaly detection according to claim 3, characterized in that, The cluster analysis of the analysis samples using a Gaussian mixture model includes: Based on the analysis samples ,in A data point for collecting data from a multi-dimensional environment. The number of samples; The Gaussian mixture model is parameter estimated using the expectation-maximization algorithm. The probability density function of the Gaussian mixture model is expressed as: ； in, The number of Gaussian components. For the first The weights of the Gaussian components, For the first A multivariate Gaussian distribution with Gaussian components, the mean of which is The covariance matrix is , This is the set of parameters for the Gaussian mixture model; The range of the dynamic baseline curve is determined based on the parameter set.

5. The fire early detection method based on anomaly detection according to claim 1, characterized in that, The frequency-time domain joint feature enhancement processing of multi-dimensional environmental data includes: At the frequency domain level, multi-dimensional environmental acquisition data is converted into frequency domain data through Fast Fourier Transform; The formula for Fast Fourier Transform is: ； in, It is a time-domain signal. It is a frequency domain component. It is the number of signal sampling points. It is a frequency index; By using an adaptive filtering algorithm, specific frequency band components corresponding to early fire signals in the frequency domain data are retained, while interference frequency bands corresponding to interference signals are filtered out.

6. The fire early detection method based on anomaly detection according to claim 1, characterized in that, The frequency-time domain joint feature enhancement processing of multi-dimensional environmental data also includes: At the time domain level, trend features and duration features of multi-dimensional environmental data collection are extracted; The trend slope of the trend feature is calculated using a trend fitting algorithm. : ； in, The number of data points within the time window. For time indexing, The signal value; By filtering by duration, we can determine the duration of a trend feature that exceeds a preset threshold.

7. The fire early detection method based on anomaly detection according to claim 1, characterized in that, The two-dimensional feature verification of the enhanced signal features includes: Perform feature matching degree verification on the enhanced signal feature vector. Compared with the preset early fire feature template vector The two are compared and their similarity scores are calculated. ; Similarity score The formula is: ； in, The dimension of the feature vector. The first eigenvector of the enhanced signal One portion, The first feature template vector of the early stage of fire One component; When the similarity score reaches 85% or higher, the time series consistency verification stage begins.

8. The fire early detection method based on anomaly detection according to claim 1, characterized in that, The two-dimensional feature verification of the enhanced signal features also includes: Perform time series consistency verification and continuously monitor the enhanced signal characteristics; The enhanced signal characteristics were confirmed to remain stable over three consecutive sampling periods.

9. The method for early fire detection based on anomaly detection according to claim 1, characterized in that, Also includes: The system self-calibration process is initiated every 72 hours, sending analog signals with known parameters to the sensor array through a built-in standard signal source; By comparing the deviation between the sensor's acquisition results and the standard signal, the sensor's sensitivity parameters are automatically adjusted. When a single sensor detects an anomaly, cross-validation of data from multiple sensors automatically uses collaborative data from two other types of sensors to supplement the judgment.

10. A fire early detection system based on anomaly detection, applied to the fire early detection method based on anomaly detection as described in claims 1-9, characterized in that, include: The multi-dimensional signal synchronous acquisition module is used to synchronously acquire temperature signals, specific gas concentration signals, and infrared radiation intensity signals in the environment to obtain multi-dimensional environmental acquisition data. The environment adaptive dynamic benchmark update unit is used to construct an environment adaptive dynamic benchmark library based on multi-dimensional environmental data, and to dynamically update the environment adaptive dynamic benchmark library to obtain a dynamic benchmark curve of the normal fluctuation range of the environment. The frequency-time domain joint feature enhancement module is used to perform frequency-time domain joint feature enhancement processing on multi-dimensional environmental acquisition data to separate early fire signals from environmental interference signals and obtain enhanced signal features. The dual-dimensional feature verification and early warning triggering unit is used to perform dual-dimensional feature verification on the enhanced signal features, confirm the validity of the signal features, and trigger a fire classification early warning when both dual-dimensional feature verifications meet the preset conditions.

Images