Method for preventing false alarm of fire smoke detector in high-frequency electromagnetic field environment

Abstract

The invention relates to the technical field of fire alarm, in particular to a method for preventing false alarm of a fire smoke detector in a high-frequency electromagnetic field environment. The method comprises the following steps: S1, acquiring a smoke scattered light intensity signal, an environment high-frequency electromagnetic field intensity signal and a temperature and humidity signal in real time through a multi-sensor synchronous acquisition module; s2, constructing an electromagnetic coupling feature set; s3, inputting the basic smoke characteristics and the electromagnetic coupling characteristics into a pre-trained random forest model; s4, aggregating the output probabilities of all decision trees by adopting a soft voting mechanism; s5, dynamically adjusting an alarm threshold value according to the real-time electromagnetic field intensity; and S6, judging whether fire alarm is triggered or not. According to the method, a diversified and robust decision tree set is constructed by using the random forest, and prediction of the trees is aggregated through soft voting, so that the signal identification accuracy of the fire smoke detector can be remarkably improved. According to the invention, the variance and deviation are effectively reduced, and the robustness of the model to noise and complex interference modes is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of fire alarm technology, and in particular to a method for preventing false alarms of fire smoke detectors in a high-frequency electromagnetic field environment. Background Art

[0002] Traditional smoke detectors face severe false alarm challenges in high-frequency electromagnetic field-intensive environments (such as substations above 110kV, 5G communication base stations, arc welding workshops, etc.). Actual measured data show that their false alarm rate is as high as 15%-30%, far exceeding the <3% standard for civilian scenarios. The fundamental contradiction lies in the fact that electromagnetic interference (EMI) signals and real fire smoke signals are highly similar in the time and frequency domains, specifically manifested as follows: Time domain feature confusion: The rising edge time of a strong electromagnetic pulse (0.2-5μs) overlaps with the initial smoke signal of a smoldering fire (0.5-10μs) Frequency domain feature coupling: The spectral energy distribution of the EMI resonant frequency point (typically 2-10MHz) partially overlaps with the modulation frequency band of the smoke particle scattering signal Envelope correlation: Both exhibit random fluctuation characteristics in pulse amplitude (coefficient of variation >0.35). Improvement plans proposed in recent years still have significant bottlenecks: rule-based feature screening methods rely on prior knowledge By setting the interference feature threshold, the misjudgment rate increases by 2.7 times in a broadband random electromagnetic noise environment (such as inverter radiation); in the prior art, Chinese patent publication number CN114495408A discloses a monitoring and early warning system based on fire big data. It processes data with the help of a big data analysis platform and uses a random forest algorithm for modeling and analysis. Compared with traditional electrical fire alarm systems, the platform is more intelligent and responds more quickly to fire hazard incidents. It can detect fires in the early stages of a fire. However, a single machine learning model has insufficient learning ability for high-dimensional time-frequency mixed features, and the generalization error is large under unknown interference patterns. Summary of the Invention

[0003] The technical problem to be solved by the present invention is: to overcome the shortcomings of the existing technology and provide a method for preventing false alarms of fire smoke detectors in a high-frequency electromagnetic field environment. A diverse and robust set of decision trees is constructed using random forests, and the predictions of these trees are aggregated through soft voting, which can significantly improve the accuracy of fire smoke detector signal recognition.

[0004] The technical solution of the present invention is:

[0005] A method for preventing false alarms of fire smoke detectors in a high-frequency electromagnetic field environment comprises the following steps:

[0006] S1, real-time acquisition of smoke scattered light intensity signal S through multi-sensor synchronous acquisition module smoke (t), ambient high-frequency electromagnetic field strength signal E field (t) and temperature and humidity signals;

[0007] S2. Construct an electromagnetic coupling feature set, including:

[0008] Time domain coupling characteristics:

[0009]

[0010] Frequency domain coupling characteristics:

[0011]

[0012] in represents Fourier transform;

[0013] S3. Input the basic smoke features and electromagnetic coupling features into the pre-trained random forest model, where each decision tree selects from the feature space through the feature randomization strategy. Features are trained, n is the total number of features;

[0014] S4. Aggregate the output probabilities of all decision trees using a soft voting mechanism:

[0015]

[0016] S5. Dynamically adjust the alarm threshold θ according to the real-time electromagnetic field strength:

[0017] θ=0.6-0.025×min(E field ,10)

[0018] S6, when P fire ≥θ, a fire alarm is triggered.

[0019] Preferably, the electromagnetic coupling feature set in step S2 further includes:

[0020] Pulse interference correlation characteristics:

[0021]

[0022] Resonant frequency characteristics:

[0023]

[0024] Among them, E wave The energy distribution of the third layer detail coefficients of wavelet packet decomposition.

[0025] Preferably, the random forest training in step S3 adopts an electromagnetic sensitivity feature weighting strategy, and assigns a weight factor ω to the electromagnetic coupling feature. k ∈[1.5,2], the Gini impurity is calculated as:

[0026]

[0027] Among them, p i is the proportion of samples of category i in the node, and C is the total number of categories.

[0028] Preferably, the parameter configuration of the random forest satisfies:

[0029] Number of trees n estimators ∈[100,500];

[0030] Maximum depth max_depth = 20 ± 5;

[0031] The minimum number of node samples min_samples_split = 5;

[0032] The feature random selection ratio max_features = 0.3 ± 0.1.

[0033] Preferably, the dynamic threshold adjustment in step S5 is provided with a safety margin:

[0034] When E field <0.5V / m,θ=0.65;

[0035] When E field >10V / m,θ=0.35.

[0036] Preferably, multi-sensor collaborative verification is added before the alarm determination in step S6:

[0037] If 0.4≤P<θ for three consecutive sampling periods, the pyroelectric infrared sensor is activated to detect human activity;

[0038] If human activity is detected and the temperature rise rate is less than 1°C / min, the alarm will be suppressed.

[0039] Preferably, the update mechanism of the random forest model includes:

[0040] Incremental learning is automatically performed every morning, using verified non-false positive data from the past 24 hours;

[0041] Emergency model retraining is triggered when three consecutive false positives occur;

[0042] When the pyroelectric sensor suppresses the alarm, the data in this period is automatically marked as a negative sample for incremental learning of the model.

[0043] Preferably, step S4 includes the following sub-steps:

[0044] S41. Each decision tree calculates the fire probability through the leaf node sample distribution based on the feature input;

[0045] S42. Use arithmetic mean to fuse the probability outputs of all trees.

[0046] Preferably, the method is deployed in an edge computing unit and includes:

[0047] Preprocessing module: signal noise reduction at a sampling rate of 10kHz;

[0048] Feature calculation module: This module uses FPGA to parallelly calculate time-frequency features. It employs a parallelized wavelet packet decomposition algorithm and uses a pre-generated DB4 wavelet coefficient lookup table to extract the third-layer detail coefficients in real time. The module transfers the pre-processed feature vectors to the GPU memory of the model inference module via the PCIe interface.

[0049] Model inference module: Accelerate random forest prediction through TensorRT.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] The present invention achieves a significant reduction in the false alarm rate of fire smoke detectors in high-frequency electromagnetic field environments through the triple synergistic mechanism of multi-source signal coupling analysis, random forest dynamic decision-making, and electromagnetic adaptive threshold. The random forest electromagnetic sensitivity weighting strategy strengthens the anti-interference feature effect, the soft voting probability fusion eliminates the prediction bias of a single tree, and the incremental learning mechanism enables the model to continuously adapt to new interference patterns, significantly improving the accuracy of fire smoke detector signal recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0053] Figure 1 It is a schematic flow chart of the method of the present invention. DETAILED DESCRIPTION

[0054] In order to enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0055] Example

[0056] In high-frequency electromagnetic environments, such as those found in 110kV substations, traditional fire and smoke detectors are susceptible to electromagnetic interference, resulting in a high false alarm rate, which seriously impacts the normal operation and safety management of the substation. To address this issue, this embodiment employs the proposed false alarm prevention method, aiming to significantly reduce the false alarm and missed alarm rates, thereby improving fire detection accuracy.

[0057] This example was conducted in a 110kV substation, which has a complex high-frequency electromagnetic field environment and numerous electromagnetic interference sources, including transformers, switchgear, etc. Furthermore, the temperature and humidity in the substation fluctuate significantly, placing high demands on the detector's performance.

[0058] Implementation equipment

[0059] Multi-sensor synchronous acquisition module:

[0060] It uses a high-precision, multi-channel data acquisition card with a sampling rate of 10kHz, which can synchronously collect the smoke scattered light intensity signal S in real time. smoke (t), ambient high-frequency electromagnetic field strength signal E fueld (t) and temperature and humidity signals. The acquisition card has a built-in signal conditioning circuit to perform preliminary amplification and filtering on the collected signals to improve signal quality.

[0061] Edge computing unit:

[0062] The edge computing unit includes a preprocessing module, a feature calculation module, and a model inference module.

[0063] Pre-processing module: A high-performance digital signal processor (DSP) is used to perform noise reduction on the collected signals. A wavelet threshold noise reduction algorithm is used to adaptively select the wavelet basis function and threshold according to the frequency characteristics and noise level of the signal, effectively removing high-frequency noise and random interference from the signal.

[0064] Feature calculation module: A field-programmable gate array (FPGA) is used to perform parallel computation of time-frequency features. The FPGA has multiple built-in computation units capable of simultaneously calculating time-domain features (such as signal rise time tr and pulse duration td), frequency-domain features (such as wavelet packet energy entropy He), and joint time-frequency features (such as the S-transform modulus at t = 50 ms and f = 1 MHz | S(τ,f) |). Through hardware acceleration, the speed and efficiency of feature calculation are greatly improved.

[0065] Model Inference Module: This module uses the NVIDIA Jetson AGX Xavier embedded development board and TensorRT to accelerate the prediction process of the random forest model. TensorRT optimizes the random forest model, including model quantization and layer fusion, to reduce the model's computational workload and memory usage, and improve inference speed.

[0066] Pyroelectric infrared sensors are used to detect human activity under specific conditions. They offer high sensitivity and fast response, enabling them to accurately detect changes in infrared radiation generated by human activity.

[0067] Implementation steps

[0068] Data collection and preprocessing:

[0069] The multi-sensor synchronous acquisition module collects the smoke scattered light intensity signal S in real time smoke (t), ambient high-frequency electromagnetic field strength signal E field (t) and temperature and humidity signals, with a sampling rate of 10kHz.

[0070] The pre-processing module performs noise reduction processing on the collected signals to remove noise and interference in the signals and obtain high-quality signal data.

[0071] Feature calculation:

[0072] Basic smoke feature calculation: Based on the smoke scattered light intensity signal S after noise reduction smoke (t), calculate the signal rise time t r (10%-90% amplitude), pulse duration t d Equal time domain features, and wavelet packet energy entropy H e Equal frequency domain features.

[0073] Electromagnetic coupling characteristic calculation:

[0074] The electromagnetic coupling characteristics include: time domain coupling characteristics F td , frequency domain coupling characteristics F fd , pulse interference correlation feature F pulse , and resonant frequency characteristics F res .

[0075] Time Domain Coupling Features: Calculate F td =max(dS smoke / dt×|E field |), which reflects the time-domain coupling relationship between the smoke signal change rate and the electromagnetic field intensity.

[0076] Frequency domain coupling characteristics: for S smoke (t) and E field (t) respectively perform fast Fourier transform (FFT) to obtain FFT (S smoke ) and FFT(E field ), then calculate It reflects the degree of coupling between the two in the frequency domain.

[0077] Pulse interference correlation characteristics: Calculate the ratio of the integrated area of ​​the smoke signal within 5ms after the rising edge of the electromagnetic pulse to the background noise area to evaluate the degree of interference of the electromagnetic pulse on the smoke signal.

[0078] Resonance frequency characteristics: Wavelet packet decomposition is performed on the smoke signal to extract the energy proportion of the third-layer detail coefficient in the [5MHz, 8MHz] frequency band, reflecting the resonance characteristics of this frequency band.

[0079] Time-frequency joint feature calculation: Calculate the modulus value |S(τ,f)| of the S transform at t = 50ms and f = 1MHz to comprehensively reflect the characteristics of the signal in the time-frequency domain.

[0080] Model training:

[0081] Collect a large amount of smoke signals, electromagnetic field signals, and temperature and humidity data in the substation environment, and mark whether they are fire signals.

[0082] The data is divided into training set and test set. The electromagnetic sensitivity feature weighting strategy is adopted to give the electromagnetic coupling feature a feature importance weight of 1.5-2 times. The calculation of Gini impurity is modified to achieve the following: where ω k is the weight factor of electromagnetic coupling characteristics, p i is the category probability.

[0083] The random forest model is trained using the training set. Each decision tree selects n features from the feature space for training through a feature randomization strategy (n is the total number of features). The parameter configuration of the random forest satisfies: the number of trees n estimators ∈[100,500], maximum depth max_depth=20±5, minimum number of node samples min_samples_split=5, feature random selection ratio max_features=0.3±0.1.

[0084] Use the test set to evaluate the performance of the model and adjust the model parameters according to the evaluation results until the model achieves satisfactory performance indicators.

[0085] Deploy the trained random forest model to the model inference module of the edge computing unit and accelerate and optimize it through TensorRT.

[0086] Dynamic threshold adjustment: According to the real-time electromagnetic field strength E field Dynamically adjust the alarm threshold θ:

[0087] When E field >10 / m,θ=0.35,

[0088] When E field <0.5V / m,θ=0.65,

[0089] When 0.5V / m≤E field ≤10V / m,θ=0.6-0.025×min(E field ,10),

[0090] Alarm judgment:

[0091] The basic smoke features and electromagnetic coupling features are input into the pre-trained random forest model, and the output probabilities of all decision trees are aggregated using a soft voting mechanism:

[0092] If P fire ≥θ, a fire alarm is triggered.

[0093] If 0.4≤P for three consecutive sampling periods fire <θ, the pyroelectric infrared sensor is activated to detect human activity. If human activity is detected and the temperature rise rate is <1℃ / min, the alarm is suppressed.

[0094] Model Updates:

[0095] Daily incremental learning: Incremental learning is automatically performed every morning, using verified non-false positive data from the past 24 hours to update the model, enabling the model to adapt to changes in the environment.

[0096] Emergency retraining: When three consecutive false positives occur, emergency model retraining is triggered to collect data and retrain the model to improve the accuracy and reliability of the model.

[0097] Implementation effect:

[0098] The present invention was field-tested for six months in three typical high-frequency electromagnetic field environments: a 110kV substation, a 5G communication base station, and an arc welding workshop. The results demonstrated that the method achieved breakthrough improvements in multiple dimensions of performance:

[0099]

[0100] The electromagnetic coupling feature set's accuracy in identifying pulsed EMI has been increased to 95.4%, and the false positive rate has been reduced by 4.3 times compared to traditional feature screening methods.

[0101] Dynamic threshold adjustment reduces the risk of missed reports to 1 / 5 of the traditional fixed threshold solution under strong electromagnetic fields;

[0102] Multi-sensor collaborative verification effectively suppresses non-fire interference such as cooking smoke, with a false alarm suppression rate of 89.6%.

[0103] Working principle: This application constructs an electromagnetic coupling feature set by synchronously collecting smoke scattered light, electromagnetic field intensity, and temperature and humidity signals, quantifies the coupling relationship between electromagnetic interference and smoke signals, inputs basic smoke features and electromagnetic coupling features into a weighted random forest model, improves anti-interference capability through an electromagnetic sensitive feature weighting strategy, dynamically lowers the alarm threshold according to the real-time electromagnetic field intensity to reduce missed reports, is deployed in an edge computing unit, supports incremental learning to update the model, and adapts to new interference patterns.

[0104] Although the present invention has been described in detail with reference to the accompanying drawings and in conjunction with preferred embodiments, the present invention is not limited thereto. Without departing from the spirit and essence of the present invention, a person of ordinary skill in the art may make various equivalent modifications or substitutions to the embodiments of the present invention, and such modifications or substitutions shall be within the scope of the present invention. Any person skilled in the art who can easily conceive of changes or substitutions within the technical scope disclosed in the present invention shall be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be based on the scope of protection of the claims.

Claims

1. A method for preventing false alarms of fire smoke detectors in a high-frequency electromagnetic field environment, characterized in that: The steps include: S1, real-time acquisition of smoke scattered light intensity signal S through multi-sensor synchronous acquisition module smoke (t), ambient high-frequency electromagnetic field strength signal E field (t) and temperature and humidity signals; S2. Construct an electromagnetic coupling feature set, including: Time domain coupling characteristics: Frequency domain coupling characteristics: in represents Fourier transform; S3. Input the basic smoke features and electromagnetic coupling features into the pre-trained random forest model, where each decision tree selects from the feature space through the feature randomization strategy. Features are trained, n is the total number of features; S4. Aggregate the output probabilities of all decision trees using a soft voting mechanism: S5. Dynamically adjust the alarm threshold θ according to the real-time electromagnetic field strength: θ=0.6-0.025×min(E field ,10) S6, when P fire ≥θ, a fire alarm is triggered.

2. The method for preventing false alarms of fire smoke detectors in a high-frequency electromagnetic field environment according to claim 1, characterized in that: The electromagnetic coupling feature set in step S2 further includes: Pulse interference correlation characteristics: Resonant frequency characteristics: Among them, E wave The energy distribution of the third layer detail coefficients of wavelet packet decomposition.

3. The method for preventing false alarms of fire smoke detectors in a high-frequency electromagnetic field environment according to claim 1, characterized in that: The random forest training in step S3 adopts an electromagnetic sensitivity feature weighting strategy to assign a weight factor ω to the electromagnetic coupling feature. k ∈[1.5,2], the Gini impurity is calculated as: Among them, p i is the proportion of samples of category i in the node, and C is the total number of categories.

4. The method for preventing false alarms of fire smoke detectors in a high-frequency electromagnetic field environment according to claim 1, characterized in that: The parameter configuration of the random forest satisfies: Number of trees n estimators ∈[100,500]; Maximum depth max_depth = 20 ± 5; The minimum number of node samples min_samples_split = 5; The feature random selection ratio max_features = 0.3 ± 0.

1.

5. The method for preventing false alarms of fire smoke detectors in a high-frequency electromagnetic field environment according to claim 1, characterized in that: The dynamic threshold adjustment in step S5 is provided with a safety margin: When E field <0.5V / m,θ=0.65; When E field > 10 V / m, θ = 0.

35.

6. The method for preventing false alarms of fire smoke detectors in a high-frequency electromagnetic field environment according to claim 1, characterized in that: The multi-sensor collaborative verification is added before the alarm determination in step S6: If 0.4≤P<θ for three consecutive sampling periods, the pyroelectric infrared sensor is activated to detect human activity; If human activity is detected and the temperature rise rate is less than 1°C / min, the alarm will be suppressed.

7. The method for preventing false alarms of fire smoke detectors in a high-frequency electromagnetic field environment according to claim 1, characterized in that: The update mechanism of the random forest model includes: Incremental learning is automatically performed every morning, using verified non-false positive data from the past 24 hours; Emergency model retraining is triggered when three consecutive false positives occur; When the pyroelectric sensor suppresses the alarm and the ambient temperature rise rate continues to be lower than 0.5℃ / min for more than 5 minutes, the data in this period is automatically marked as a negative sample for incremental learning of the model.

8. The method for preventing false alarms of fire smoke detectors in a high-frequency electromagnetic field environment according to claim 1, characterized in that: The step S4 includes the following sub-steps: S41. Each decision tree calculates the fire probability through the leaf node sample distribution based on the feature input; S42. Use arithmetic mean to fuse the probability outputs of all trees.

9. The method for preventing false alarms of fire smoke detectors in a high-frequency electromagnetic field environment according to claim 1, characterized in that: The method is deployed on an edge computing unit and includes: Preprocessing module: signal noise reduction at a sampling rate of 10kHz; Feature calculation module: FPGA is used to parallelize the calculation of time-frequency features. The feature calculation module adopts a parallelized wavelet packet decomposition algorithm and realizes the real-time extraction of the third layer detail coefficients through the pre-generated DB4 wavelet coefficient lookup table; Model inference module: Accelerate random forest prediction through TensorRT.

Images