Explosive and powder thermal runaway detection method and signal processing algorithm

Abstract

The invention relates to an explosive thermal runaway detection method and a signal processing algorithm, and the method comprises the steps: based on the molecular characteristics of a nitrogen-containing solid propellant combustion product, selecting an infrared absorption peak of a nitrogen oxide (NOx) in the nitrogen-containing solid propellant combustion product as a main wave band, and taking an infrared absorption peak of a hydrocarbon (CHxOy) as a background wave band, synchronously acquiring spectral signals of a target area through three infrared sensors; sequentially performing Kalman filtering noise reduction, dynamic background modeling, explosive and powder specificity spectrum extraction, time domain rapid detection, frequency domain feature analysis and multi-dimensional feature fusion on the acquired signals; and finally, judging according to the comprehensive decision score D, outputting a state flag bit TrigFlagD, and transmitting a signal to a control host through a CAN (Controller Area Network) bus so as to control a fire extinguishing device and a sound-light alarm system to act. According to the algorithm, short-time energy detection, fundamental frequency analysis and harmonic energy ratio judgment are integrated, high-performance hardware and a real-time operation system are combined, millisecond-level explosive flame detection is achieved, the method has the advantages of being high in sensitivity, low in false alarm rate and rapid in response, and the safety of explosive operators can be effectively guaranteed.

Description

Technical Field

[0001] This invention patent relates to the field of fire prevention and control technology for explosives, specifically a method and signal processing algorithm for detecting thermal runaway of explosives. It is particularly suitable for high-sensitivity, low-false-alarm fire detection of explosives such as nitrogen-containing solid propellants during combustion, and can achieve rapid response and accurate identification of thermal runaway of explosives. Background Technology

[0002] In the explosives industry (such as solid propellant and pyrotechnics manufacturing), fire prevention is a core safety issue. Currently, domestic explosives fire detection technology faces the following problems:

[0003] (1) Insufficient response speed: Traditional smoke detectors cannot achieve millisecond-level fast response, and ultraviolet detectors are easily interfered with by dense smoke, resulting in false alarms or delayed alarms.

[0004] (2) Weak anti-interference capability: Traditional infrared flame detectors are based on the 4.3μm CO2 radiation peak, which is easily affected by sunlight, lamplight and heat radiation, and cannot effectively identify nitrogen oxides (NOx) produced by the combustion of nitrogen-containing solid propellants (such as double-base, AP-based composite, and azide-based propellants). x Characteristic spectrum.

[0005] (3) Lack of specificity recognition: Existing technologies do not specifically identify NO in the combustion products of explosives. x and hydrocarbons (CH) x O y Differentiated detection schemes based on spectral design are insufficient to meet the precision requirements of active firefighting equipment.

[0006] The aforementioned problems pose significant hidden dangers to the prevention and control of fires involving high-burning-rate solid propellants, necessitating a highly sensitive detection method and algorithm with low false alarm rates. Summary of the Invention

[0007] This invention provides a method and signal processing algorithm for detecting thermal runaway of explosives, aiming to solve the problems of slow response speed, high false alarm rate and insufficient specific spectrum identification in the detection of explosive fires. Through composite band screening, Kalman filtering, dynamic background modeling, time and frequency domain analysis and multi-dimensional feature fusion, millisecond-level fire detection and active protection can be achieved.

[0008] This invention proposes a method for detecting thermal runaway of explosives, characterized by comprising the following steps:

[0009] Step S1: Composite band spectral screening, focusing on nitrogen oxides (NOx) in the combustion products of nitrogen-containing solid propellants. x The specific infrared absorption band of ) is used as the main detection band, with hydrocarbons (CH4) as the primary detection band. x O yThe infrared absorption band of NO2 at 6.2 μm and the infrared absorption peak of NO at 5.3 μm are selected as the main band; the infrared absorption peaks of hydrocarbons (CHx Oy) at 5 μm and 6.5 μm are selected as the background band; the infrared sensors of the background bands 1 and 2 are single-channel structures with center wavelengths of 5 μm and 6.5 μm, respectively; the infrared sensors of the main bands are dual-channel structures, with each channel of the main band infrared sensor equipped with a narrowband filter, and center wavelengths of 5.3 μm and 6.2 μm, respectively.

[0010] Step S2: Signal acquisition. The radiation spectrum of the explosive operation area is synchronously and continuously acquired in real time through the background band 1 and 2 infrared sensors and the main band infrared sensor at the working sampling rate. The acquisition time is at least 1 minute, and the optical signal is converted into an analog electrical signal.

[0011] Step S3: Signal processing. Based on the algorithm preset in the microcontroller (MCU), the above analog electrical signal is preprocessed, the explosive-specific spectrum is extracted, and the flame feature is identified to generate the corresponding output signal.

[0012] Step S4: Signal output and response. The above output signal is transmitted to the control host through the Controller Area Network (CAN) bus. The output current of the control host varies depending on the output signal. The magnitude of the output current determines whether the fire extinguishing device and the audible and visual alarm system are triggered.

[0013] Furthermore, the signal preprocessing step includes: adjusting the gain of the analog electrical signal through an infrared signal conditioning circuit to obtain an amplified signal; discretizing the amplified signal into a digital signal through an analog-to-digital converter (ADC); and executing a thermal runaway signal processing algorithm for explosives through an MCU to obtain an output signal.

[0014] Furthermore, the control logic of the fire extinguishing device is as follows: the above output signal is transmitted to the control host through the Controller Area Network (CAN) bus. The control host will output different currents according to the output signal. The control host determines whether to trigger the fire extinguishing device and the audible and visual alarm system based on the current magnitude. The current output range is 0 to 20mA. Different current output results represent different states of the control host, as shown in Table 1.

[0015] Table 1 Definition of 0-20mA Current Output

[0016] Serial Number Output signal Circuit characteristics Control host status Control host indicator light status 1 00 1mA Normal monitoring status Green light 2 11 15mA Fire alarm output status Red light on 3 01、10 20mA Equipment Fault Status Yellow light on

[0017] The present invention proposes an algorithm for processing thermal runaway signals of explosives, characterized by the following processing steps:

[0018] Step P1: Kalman filtering. The aim is to continuously refine the state estimate using prior estimates and new measurements, reducing noise and improving signal accuracy. Kalman filtering is applied to the three channels of the background band 1, 2, and main band digital signal to obtain the filtered signal.

[0019] Step P2: Dynamic Background Modeling. This aims to estimate and subtract environmental background interference in real time to accurately extract flame feature signals. The filtered signal x... est Using [n] as input, a background noise model B is obtained through dynamic background modeling. model [t].

[0020] Step P3: Specific Spectral Extraction. The filtered signal S... main_est [t] and background noise model B model [t] is used as input. The effective signal is calculated by subtracting background noise from the main band:

[0021] S eff [t] = S main_est [t]-B model [t]

[0022] Among them, S eff [t]: Effective flame signal (pyrotechnic-specific spectrum), used as the output signal for dynamic background modeling; S main_est [t]: The estimated value of the main band at time t; B model [t]: The estimated value of the background model (dynamic background signal) at time t.

[0023] Step P4: Fast Time-Domain Detection. This aims to quickly determine the presence of a flame signal by analyzing the short-time energy characteristics of the signal. The obtained valid flame signal S... eff Using [t] as input, the short-time energy E[n] is obtained through short-time energy calculation, and then compared with the dynamic threshold T[n] to output the time-domain trigger flag Trig_Flag_t.

[0024] Step P5: Frequency Domain Feature Analysis. This aims to identify the flame-specific frequency components (such as the fundamental frequency of 4–20 Hz and harmonics) by converting the time-domain signal to the frequency domain. The time-domain trigger flag Trig_Flag_t and the valid flame signal S are then... eff Using [t] as input, the fundamental frequency f is calculated through Fast Fourier Transform (FFT). peak Harmonic energy ratio R harmonic The frequency domain trigger flag Trig_Flag_f is output based on the trigger threshold.

[0025] Step P6: Multi-dimensional feature fusion. This aims to construct a highly robust flame detection model by integrating information from multiple sources, including the time domain, frequency domain, and spatial domain, effectively reducing the false alarm rate. The band signal S... bg1_est S main_est S bg2_est Short-time energy E[n], frequency domain trigger flag Trig_Flag_f, base frequency f peak Harmonic energy ratio R harmonic As input, the presence of a flame signal is determined by the comprehensive decision score D, and the multi-dimensional trigger flag Trig_Flag_D is output.

[0026] Furthermore, Kalman filtering is suitable for dynamic systems, effectively handling noise and performing real-time filtering of infrared signals to improve the signal-to-noise ratio. The core idea of ​​Kalman filtering is prediction and updating; it combines the system model and observation data to progressively optimize the state estimate. The steps are as follows:

[0027] (1) System initialization: Configure the key noise covariance parameters for the Kalman filter algorithm, namely the measurement noise covariance matrix R and the process noise covariance matrix Q.

[0028] In a stable environment free from flames and strong interference, the raw output voltage signals of three infrared sensors are continuously collected, amplified by a conditioning circuit, and converted by an ADC to obtain the digital signal V. bg1 [n],V main [n],V bg2 [n](n=1,2,...,n init n init (This is the initial number of sampling points). Then, the variance of the data from the three channels is calculated separately. To simplify and allow the use of dimensionless numerical values ​​in algorithms, signals are typically normalized during the initialization phase.

[0029] σ 2 norm =σ 2 / (V ref ) 2

[0030] Among them, V ref It is a reference voltage, such as the range of an ADC or the full-scale value of a signal; σ 2 σ represents the variance of the three channel data. 2 norm This represents the normalized variance of the three channel data. The normalized variance values ​​are used as the diagonal elements of the measurement noise covariance matrix R, where R = diag([σ...). 2 norm_bg1 ,σ 2 norm_main ,σ 2norm_bg2 ]).

[0031] The process noise covariance matrix Q is used to model the intensity of the process noise w[n] in the state equation, reflecting the drastic changes in the system state (i.e., the actual radiated signal). Since the future flame state cannot be predicted, the Q matrix is ​​usually set based on prior knowledge under typical conditions. Assuming the signal changes slowly and the flame fluctuations are small, it is initially set to a small diagonal matrix; if the fluctuations are large, the Q value is increased according to the actual situation.

[0032] (2) Kalman filter input: The original output voltage signals of three infrared sensors are continuously acquired, amplified by the conditioning circuit, and converted by the ADC to obtain the digital signal V. bg1 [n],V main [n],V bg2 [n], representing the system's state vector at time n, x[n] = [S] bg1 ,S main ,S bg2 ] T .

[0033] (3) Kalman filtering process: mainly includes two parts—prediction and update.

[0034] 1) State prediction

[0035] x pred =F*x est [n-1]

[0036] Where, x pred The predicted state value at time n is based on the state estimate x at time n-1. est [n-1] is calculated from the state transition matrix F; F is the state transition matrix, and assuming the signal changes slowly, the state transition matrix F = 1 is set; x est [0] = [S bg1 [0],S main [0],S bg2 [0]] T .

[0037] Error covariance prediction:

[0038] P pred =F*P est [n-1]*F T +Q

[0039] Among them, P pred The prediction error covariance matrix at time n, representing the predicted state value x. pred Uncertainty of P est [n-1]: The estimation error covariance matrix at time n-1, representing the state estimate x at time n-1.est [n-1] uncertainty; P est [0] can be set to a small diagonal matrix, indicating a high degree of confidence in the initial state estimate.

[0040] 2) Status Update

[0041] Kalman gain calculation:

[0042] K = P pred *H T *(H*P pred *H T +R) -1

[0043] Where K: Kalman gain matrix, representing the trade-off between confidence in prediction and measurement: if R >> Q, then confidence in prediction (K → 0); if Q >> R, then confidence in measurement (K → H). -1 );

[0044] Status Update:

[0045] x est [n] = x pred +K*(z[n]-H*x pred )

[0046] Where, x est [n] serves as the filtered output signal, used for subsequent dynamic background modeling, effective signal extraction, and time-frequency analysis. This is achieved through the residual (z[n]-H*x). pred Correct the predicted value; z[n]-H*x pred The innovation or residual represents the difference between the actual observed value z[n] and the expected observed value H*x based on the prediction. pred The differences between them are: I: is the identity matrix; z[n] is the observation value of the system at the current moment, z[n] = H*x[n] + v[n]; H is the observation matrix, and the state vector x[n] directly corresponds to z[n], so H is usually a scalar, and H = 1 is set; v[n] is the observation noise, which follows a Gaussian distribution v ~ N(0,R); the measurement noise R reflects the sensor noise.

[0047] Covariance update:

[0048] P est [n] = (I – K * H) * P pred

[0049] Among them, P est [n]: The updated estimation error covariance matrix at time n. The updated covariance reflects the corrected state estimate x. est [n] Uncertainty.

[0050] (4) Kalman filter output: xest [n] represents the filtered output signal.

[0051] Furthermore, during the initialization phase of system operation, the data after Kalman filtering noise reduction automatically learns the environmental background radiation characteristics and establishes a dynamic background noise model. The steps are as follows:

[0052] 1) Signal input: The state estimation vector of the output signal after Kalman filtering. As input for dynamic background modeling;

[0053] 2) Background blending: Utilizing two background bands S bg1_est [t] and S bg2_est The estimated value of [t] is used for background signal fusion.

[0054]

[0055] Among them, B raw [t]: The fused value of the two background band signals at time t; S bg1_est [t] and S bg2_est [t]: The estimated value of the background band at time t.

[0056] 3) Initial background model B model [0] Settings:

[0057] B model [0] Sets all B values ​​during the entire initialization period. raw The arithmetic mean of [t].

[0058]

[0059] Where, n init This is the number of sampling points during the initialization phase.

[0060] 4) Standard deviation σ B Initial value settings:

[0061] Calculate B using the standard statistical variance formula. raw [t] Historical standard deviation of the sequence σ B initial value

[0062]

[0063] Where, σ B It is background signal B raw The historical standard deviation of [t].

[0064] 5) Dynamically Updated Background Model: The background model is dynamically adjusted using a smoothing factor to adapt to environmental changes. The mathematical model is as follows:

[0065] B model [t]=α·B model [t-1]+(1-α)·B raw [t]

[0066] Among them, B model [t]: The estimated value of the background model (dynamic background signal) at time t; α: Smoothing factor (0 < α < 1), controlling the update rate, preferably 0.85 to 0.95. When α → 1: the background update is slow, with strong noise resistance, suitable for stable environments. When α → 0: the background update is fast, with rapid response, suitable for dynamic environments.

[0067] 6) Anomaly detection and freezing mechanism: detects background mutations to avoid temporary interference that could contaminate the model.

[0068] If |B raw [t]-B model [t-1]∣>k·σ B Then the background model will pause updating.

[0069] Where, σ B It is background signal B raw [t] is the historical standard deviation; k is the sensitivity coefficient (usually taken as 3, based on the 3σ criterion).

[0070] Furthermore, based on the temporal characteristics of flames: flame combustion causes a rapid rise in infrared radiation signals and non-stationary bands, manifested as a sudden increase in temporal energy. Therefore, the rapid temporal detection steps are as follows:

[0071] 1) Signal Input: Input the valid flame signal S described in step P3. eff [t] is used as the input signal.

[0072] 2) Bandpass filtering: This filters the input signal S... eff [t], high-frequency noise and low-frequency drift (such as temperature drift) are eliminated by a bandpass filter (0.1Hz~1kHz) to obtain the filtered signal y[n].

[0073] 3) Short-time energy calculation: The preprocessed signal y[n] is then quantized to show the rapid rise of the flame within a short time window.

[0074] E[n] = E[n-1] + y[n] 2 -y[nN] 2

[0075] Where, E[n]: the short-time energy calculated at sampling point n (the sum of signal energy within the current sliding window); E[n-1]: the short-time energy calculated at the previous sampling point n-1; y[n]: the effective signal S at sampling point n after preprocessing (high-pass / low-pass filtering). eff The value of [t]. N: Length of the sliding window (number of sampling points included); y[nN]: Signal value N sampling points ago (i.e., the sampling point that just moved out of the current window).

[0076] 4) Exponentially weighted moving average:

[0077] μ B [n] = 0.99μ B [n-1]+0.01E[n]

[0078]

[0079] Where, μ B [n]: The exponentially weighted moving average (mean estimate) of the background noise energy at sampling point n; μ B [n-1]: The mean estimate at the previous sampling point n-1; At sampling point n, the exponentially weighted moving variance (variance estimate) of the background noise energy; The variance estimate of the previous sampling point n-1.

[0080] 5) Dynamic threshold: The trigger threshold is adaptively adjusted according to the background noise to reduce false alarms caused by environmental interference.

[0081]

[0082] Where, T[n]: the dynamic trigger threshold set at sampling point n; The standard deviation estimate of background noise energy (σ) B [n]).

[0083] 6) Trigger condition judgment:

[0084]

[0085] Wherein, Trig_Flag_t: a time-domain trigger flag, representing a binary flag indicating whether the trigger condition is met (1 = met, 0 = not met); E[ni]>T[ni]: determines whether the short-time energy E[ni] at sampling point ni is greater than the corresponding dynamic threshold T[ni]. The result is 1 (true) or 0 (false). Count the number of points that satisfy E>T among M-1 consecutive points back from the current sampling point n (a total of M consecutive points). M: the number of sampling points that require consecutive exceedance of the threshold (M=3), that is, the binary flag of the trigger condition needs to be 1 for 3 consecutive exceedances; i: the index of the backtracking sampling point.

[0086] Furthermore, based on the frequency domain characteristics of flames: during combustion, due to turbulence and thermal convection, the infrared radiation signal exhibits fundamental frequency fluctuations and harmonic components in the range of 4Hz to 20Hz. Therefore, the frequency domain characteristic analysis steps are as follows:

[0087] 1) Signal Input: Input the valid flame signal S eff [t] and the time-domain trigger flag Trig_Flag_t are used as input signals. If Trig_Flag_t = 1, then frequency domain feature analysis is performed; if Trig_Flag_t = 0, then the process returns to valid flame signal extraction.

[0088] 2) Add a window: For S eff [t] Windowing reduces spectral leakage and removes the DC component from the signal, resulting in x. f [n]. Hanning window function:

[0089]

[0090] Where w[n] is the Hanning window function value at time index n; n: time index of the sampling point within the window (n = 0, 1, ..., N-1); N: window length (number of FFT points). X[k]: frequency domain complex output, where k is the frequency index.

[0091] 3) FFT: Converts the time-domain signal x f [n] is converted to the complex frequency domain X f [k] represents.

[0092]

[0093] 4) Amplitude spectrum calculation: Take the modulus of the FFT result to obtain the amplitude energy of each frequency component, obtain A[k], and search for the frequency with the largest amplitude in the target frequency band (4~20Hz) as the fundamental frequency f. peak .

[0094]

[0095] 5) Harmonic energy ratio:

[0096]

[0097] Among them, R harmonic Harmonic energy ratio. Represents the fundamental frequency (f). peak The second harmonic of ) (2f peak) and third harmonic (3f peak The ratio of the sum of energies to the fundamental frequency energy; k peak : The frequency index corresponding to the base frequency.

[0098] 6) Trigger condition judgment:

[0099]

[0100] Among them, Trig_Flag_f: frequency domain trigger flag, which is a binary flag indicating whether the trigger condition is met (1 = met, 0 = not met).

[0101] Furthermore, the multi-dimensional feature fusion process includes the following steps:

[0102] 1) Signal Input: Input the band signal S described in step P1. bg1_est S main_est S bg2_est (used to calculate band proportions), the short-time energy E[n] of claim 8, the frequency domain trigger flag Trig_Flag_f, and the fundamental frequency f peak Harmonic energy ratio R harmonic As the input signal. If Trig_Flag_f = 1, then perform multi-dimensional feature fusion; if Trig_Flag_f = 0, then return to step P3.

[0103] Band ratio verification:

[0104]

[0105] Among them, the intensity of the flame in the main band is significantly higher than that in the background band, and the normal scene R band <2, Fire Scene R band >3.

[0106] 2) Feature normalization: Scaling features of different dimensions to a uniform range (e.g., [0,1]) to avoid the influence of numerical differences on weights.

[0107]

[0108] Among them, F i F represents the original eigenvalues; min / F max This indicates the historical minimum / maximum of a feature (sliding window statistics).

[0109] 3) Feature-weighted fusion: Weights are assigned based on feature importance, and a comprehensive score is calculated. A decision threshold is designed to determine the presence of a flame. Linear weighted decision function:

[0110]

[0111] Where D represents the overall decision score; w i Represents the feature weights (∑w) i =1). w1=0.3, It is the normalized short-time energy; w2 = 0.4. It is the normalized fundamental frequency; w3 = 0.2. It is the normalized harmonic ratio; w4 = 0.1. It is the normalized band ratio.

[0112] 4) Determination of grading thresholds:

[0113]

[0114] Among them, Trig_Flag_D: a multi-dimensional trigger flag used to determine the alarm status. When Trig_Flag_D = 00, the control host output current is 1mA. At this time, the control host is in normal monitoring state, and the fire extinguishing device and audible and visual alarm do not activate. When Trig_Flag_D = 11, the control host output current is 15mA. At this time, the control host is in fire alarm output state, and the fire extinguishing device and audible and visual alarm are activated. When Trig_Flag_D = 01 or 10, the control host output current is 20mA. At this time, the control host is in equipment fault state, and the fire extinguishing device and audible and visual alarm do not activate.

[0115] The Kalman filter described in this invention enables the flame detection system to suppress high-frequency noise, track signal trends, and improve the signal-to-noise ratio. In actual deployment, the optimal values ​​of Q and R need to be determined experimentally, taking into account specific sensor characteristics and environmental conditions (such as temperature drift), to enhance anti-interference capabilities.

[0116] The dynamic background modeling described in this invention significantly improves the anti-interference capability of flame detection by real-time fusion of multi-band data and adaptive filtering. Its core features include: a multi-band redundancy design that utilizes the strong correlation of non-flame bands to cancel environmental noise; an exponentially weighted moving average dynamic update that balances historical data with real-time observations and adapts to gradual environmental changes; and an established anomaly freezing mechanism to prevent transient interference from contaminating the background model. The code implementation requires optimization of computational efficiency based on embedded system characteristics and parameter calibration through on-site testing to ensure reliable operation in complex industrial environments.

[0117] The multi-dimensional feature fusion described in this invention significantly improves the reliability of flame detection systems by integrating time-domain, frequency-domain, and spatial-domain information. Its core lies in: cross-validating the physical properties of features from different dimensions; adapting to environmental changes through normalization and weight adjustment; and optimizing the code to ensure real-time operation on embedded platforms. In actual deployment, parameters need to be adjusted according to specific application scenarios, and the system's stability and anti-interference capabilities need to be verified through long-term operational testing.

[0118] This invention can detect the specific spectrum of explosives and the flashing frequency of explosive flames within milliseconds. Through multi-dimensional feature data fusion, it can quickly and reliably trigger the fire extinguishing system, effectively protecting the safety of personnel handling explosives manually. This solution combines high-performance hardware, detection algorithms, and a real-time operating system to ensure the system's rapid response and high reliability. Attached Figure Description

[0119] Figure 1 Schematic diagram of fire detection methods for flammable and explosive materials;

[0120] Figure 2 A schematic diagram of the signal processing algorithm for flammable explosive fires. Detailed Implementation

[0121] The embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0122] One embodiment of the present invention, such as Figure 1 and Figure 2 As stated above.

[0123] The hardware platform setup in this embodiment, such as Figure 1 The module mainly includes a signal acquisition module, a signal processing module, and a signal output module, each of which includes the following components:

[0124] Background band 1 infrared sensor: A thermopile infrared sensor (such as Heimann Sensor HTPA 5x5L10.00) is selected and configured with a narrowband filter with a center wavelength of 5μm (such as Thorlabs FB5500-500) to acquire 5μm background band signals.

[0125] Background band 2 infrared sensor: Select the same type of thermopile sensor and configure a narrowband filter with a center wavelength of 6.5μm (such as Thorlabs FB6500-500) to collect 6.5μm background band signals.

[0126] Main band infrared sensor: A dual-channel thermopile infrared sensor (such as Heimann Sensor HTPA8x8d L1.0 / 0.7) is selected. Channel 1 is equipped with a narrowband filter (custom-made, FWHM≤100nm) with a center wavelength of 5.3μm, used to acquire NO characteristic signals. Channel 2 is equipped with a narrowband filter (custom-made, FWHM≤100nm) with a center wavelength of 6.2μm, used to acquire NO2 characteristic signals.

[0127] Infrared signal conditioning circuit: A non-inverting amplifier circuit is built using a low-noise operational amplifier (such as TIOPA188), with an adjustable gain range of 60-80dB and a bandwidth of DC-10kHz. The analog voltage signal output from each sensor (typically in the μV to mV range) is amplified and impedance matched.

[0128] A / D conversion circuit: Select a 24-bit high-precision Σ-Δ ADC (such as TIADS1256), set the sampling rate to 1kHz, and convert the conditioned analog signal into a digital signal.

[0129] Microcontroller (MCU): A high-performance ARM Cortex-M4 core MCU (such as STMicroelectronics STM32F407VGT6) with a main frequency of 168MHz is selected. It has a built-in FPU unit and runs a real-time operating system (such as FreeRTOS) to ensure the real-time performance of the algorithm. The MCU has a built-in algorithm for processing thermal runaway signals of explosives.

[0130] Communication interface: Integrated CAN bus controller (such as built into the MCU), using CAN 2.0B protocol, baud rate set to 500kbps, used to transmit processing results (status flags, fault information) to the control host.

[0131] The control host (i.e., the controller) receives status codes (00 / 01 / 10 / 11) sent via the CAN bus and outputs corresponding 0-20mA current signals according to the logic defined in Table 1 to drive the fire extinguishing device (i.e., the fire extinguishing system) and the audible and visual alarm. The fire extinguishing system includes fire extinguishers (such as high-pressure fine water mist nozzles).

[0132] The algorithm flow of this embodiment is as follows: Figure 2 The scenario is set as follows: a solid propellant loading workshop with an ambient temperature of 30℃, and interference from lighting and heat radiation from distant equipment. The specific operating procedure is as follows:

[0133] 1. System initialization:

[0134] Noise Measurement: In a stable environment free from flames and strong interference, three infrared sensors are synchronously and continuously acquired at a working sampling rate (e.g., 1kHz). The original output voltage signal V is then obtained after amplification and discretization. bg1 [n],V main [n],V bg2 [n], with a duration of at least 1 minute (i.e., 60,000 points). Then, calculate the variance of the data from the three channels separately. To simplify and allow the use of dimensionless numerical values ​​in algorithms, signals are typically normalized during the initialization phase.

[0135] σ 2 norm=σ 2 / (V ref ) 2

[0136] Among them, V ref It is a reference voltage, such as the range of an ADC or the full-scale value of a signal; σ 2 σ represents the variance of the three channel data. 2 norm This represents the normalized variance of the three channel data.

[0137] The normalized variance is used as the diagonal element of the measurement noise covariance matrix R, R = diag([σ 2 norm_bg1 ,σ 2 norm_main ,σ 2 norm_bg2 The typical value is R = diag([0.3, 0.4, 0.3]).

[0138] Assuming the signal changes slowly and the flame fluctuations are small, the initial setting is a small diagonal matrix: Q = diag([0.1, 0.1, 0.1]).

[0139] Assuming a high confidence level in the initial state estimate, P is set as a small diagonal matrix. est [0] = diag([0.1,0.1,0.1]).

[0140] Dynamic background model initialization: After the system is powered on, it runs for 5 minutes in a "clean" environment where no operators have entered and no equipment is running. During this period, the dynamic background modeling algorithm continuously collects background band data and calculates the initial background model B. model [0] and smoothing factor α (fixed to 0.9 in this embodiment), while estimating the background signal standard deviation σ. B The initial value.

[0141] 2. Normal monitoring status: All sensors operate continuously. Background band 1 and 2 infrared sensor outputs are stable (S bg1_est ≈1.2V,S bg2_est ≈1.1V). The output of the main band infrared sensor fluctuates slightly (S). main_est ≈1.3V). Kalman filtering effectively suppresses high-frequency noise (the standard deviation of observed noise is reduced by approximately 60%). Dynamic background model B model Smooth updates. The time-domain energy E[n] remains at μ. BWith a value approximately 50 (au), the dynamic threshold T[n] ≈ 80 (au), and Trig_Flag_t remains at 0. Frequency domain feature analysis and multi-dimensional feature fusion are not triggered. The MCU sends a fire alarm status code 00 to the control host via the CAN bus. The control host outputs a 1mA current, the green light illuminates, and the fire extinguishing device and audible and visual alarm do not activate.

[0142] 3. Fire Alarm Trigger - Simulated Propellant Ignition: Within the field of view of the infrared sensor (3 meters away), a 10g AP / HTPB composite propellant sample was ignited using a small burner. Combustion produced significant NO / NO2 characteristic radiation.

[0143] (1) Signal change (starting from t=0ms): Main band signal S main_est Within 5ms, the voltage surged from 1.3V to 4.8V (NO channel) and 5.2V (NO2 channel). Background band signal S bg1_est ,S bg2_est It only rose slightly to 1.5V (mainly due to the overall thermal radiation of the flame).

[0144] (2) Specific spectral extraction (t = 10 ms): Effective signal S eff [t] = S main_est [t]-B model [t] increased significantly (>3.0V), successfully separating the characteristic spectrum of the explosive.

[0145] (3) Fast detection in the time domain (t = 15 ms): After bandpass filtering, the energy of signal y[n] increases sharply. The short-time energy E[n] values ​​at three consecutive sampling points (t = 15, 16, 17 ms) are 210, 350, and 520 (au), respectively, all of which far exceed the dynamic threshold T[n] ≈ 80 (au), satisfying the condition of M = 3 consecutive over-threshold, and Trig_Flag_t = 1.

[0146] (4) Frequency domain characteristic analysis (t = 20 ms): For S eff A 512-point FFT was performed on the signal after applying a Hanning window. The peak frequency f was identified within the 4–20 Hz frequency band. peak =12.5Hz (typical pulsating fundamental frequency of a flame). Calculate the harmonic energy ratio R. harmonic = (A

[25] + A[37.5]) / A[12.5] ≈ 0.65 (A[k] is the amplitude corresponding to point k). This satisfies condition f. peak ∈[4,20] and R harmonic If the value is greater than 0.3, set Trig_Flag_f = 1.

[0147] (5) Multi-dimensional feature fusion (t = 25ms):

[0148] 1) Calculate the band ratio: Rband =S main_est / ((S bg1_est +S bg2_est ) / 2)≈4.1>3.

[0149] 2) Normalization characteristics:

[0150]

[0151] 3) Weighted score:

[0152]

[0153] 4) Set Trig_Flag_D = 11.

[0154] (6) System Response (t = 30ms): The MCU sends fire alarm status code 11 to the control host via the CAN bus. The control host outputs 15mA current, triggering the audible and visual alarm (red light illuminates, buzzer sounds) and activating the fire extinguishing device (high-pressure fine water mist nozzles activate). The total delay from the appearance of flames to the activation of fire extinguishing is approximately 30ms.

[0155] 4. Anti-interference verification - false alarm test:

[0156] Scenario 1 (Strong Sunlight Interference): Noon sunlight directly hits the sensor window (illuminance > 100 klux). Both background and main band signals increase, but R... band ≈1.2<2. The time-domain energy E[n] fluctuates, but does not exceed the threshold for three consecutive points. Trig_Flag_t is not set, so subsequent processes will not start. No alarms.

[0157] Scenario 2 (Welding Arc Light Interference): Arc welding is performed 5 meters away. The sensor detects transient intense light. The dynamic background model detects |B raw [t]-B model [t-1]|>3*σ B The freeze mechanism is triggered to pause updates and prevent model contamination. A single spike appears in the time-domain energy E[n] but does not persist; Trig_Flag_t = 0. No alarms are triggered.

[0158] Scenario 3 (Heat Source Interference - Heating Furnace Door Open): A nearby heating furnace door is open, increasing heat radiation. The main band signal rises, R... band ≈2.5. The time-domain energy E[n] rises slowly (rise time > 100ms). Trig_Flag_t = 1 triggers frequency domain characteristic analysis. FFT results show that the energy is concentrated in the low frequency <2Hz, with no significant 4-20Hz fundamental frequency and harmonics (R). harmonic <0.2). Trig_Flag_f=0, multi-dimensional feature fusion is not performed. No alarms.

[0159] The above implementation method and operation process utilize the characteristic absorption of NOx at 5.3μm / 6.2μm, combined with dynamic background subtraction and band ratio verification, to significantly improve the specific identification capability of explosive flames, effectively distinguishing them from environmental heat sources and sunlight interference. Millisecond-level rapid time-domain detection combined with high-performance hardware ensures extremely early detection of high-burning-rate explosive fires (average <35ms). The multi-dimensional feature fusion mechanism integrates time-domain (energy mutation), frequency-domain (fundamental frequency + harmonics), and spatial-domain (band ratio) information, and through strict threshold judgment (such as continuous exceeding of threshold, harmonic ratio >0.3, comprehensive score D>0.6), it can quickly and reliably trigger the fire extinguishing system in strong interference environments, effectively protecting the safety of personnel handling explosives manually. This solution, combined with high-performance hardware, detection algorithms, and a real-time operating system, ensures the system's rapid response and high reliability.

Claims

1. A method for detecting thermal runaway of explosives, characterized in that, Includes the following steps: Step S1: Composite band spectral screening, focusing on nitrogen oxides (NOx) in the combustion products of nitrogen-containing solid propellants. x The infrared absorption peak of ) is used as the main band, with hydrocarbons (CH) as the main band. x O y The infrared absorption peak of ) is used as the background band; Step S2: Signal acquisition. The radiation spectrum of the explosive operation area is acquired in real time by background band 1 and 2 infrared sensors and main band infrared sensor. Each sensor converts the acquired light signal into an analog electrical signal. Step S3: Signal processing. The analog electrical signal is amplified by adjusting the gain of the infrared signal conditioning circuit. The amplified signal is discretized into a digital signal by an analog-to-digital converter (ADC). The digital signal is then subjected to Kalman filtering, dynamic background modeling, explosive-specific spectrum extraction, and flame feature recognition to generate an output signal. Step S4: Signal output and response. The above output signal is transmitted to the control host through the Controller Area Network (CAN) bus. The output current of the control host varies depending on the output signal. The magnitude of the output current determines whether the fire extinguishing device and the audible and visual alarm system are triggered.

2. A method for detecting thermal runaway of explosives, characterized in that, Nitrogen oxides (NOx) in the combustion products of nitrogen-containing solid propellants x The infrared absorption peak of ) is used as the main band, with hydrocarbons (CH) as the main band. x O y The infrared absorption peak of ) is used as the background band; Includes the following steps: 1) Noise Measurement The radiation spectrum of the explosive operation area is collected and processed in real time by background band 1 and 2 infrared sensors and main band infrared sensor to obtain three channels of data: background band 1, 2 and main band digital signals. 2) Filtering The background bands 1 and 2 and the main band digital signals obtained in step 1) are filtered using the Kalman filter algorithm to obtain the filtered output signal x. est [n], S bg1_est [t] and S bg2_est [t] represents the estimated values ​​of the two background bands at time t, S main_est [t] represents the estimated value of the main band at time t; 3) Calculate the fusion value B of the filtered background bands 1 and 2. raw [t], Establish background noise model B model [t], B model [t]=α·B model [t-1]+(1-a)·B raw [t] Among them, B model [t]: Background model estimate at time t; α: Smoothing factor, 0 < α < 1; 4) Extract the valid flame signal S eff [t], S eff [t] = S main_est [t]-B model [t]; 5) Fast time-domain detection According to the valid flame signal S eff [t] Calculate the short-time energy and count the number of points that satisfy the condition that the short-time energy is greater than the dynamic threshold among M-1 points back from the current sampling point n. If the dynamic threshold is exceeded for M consecutive times, output the time-domain trigger flag Trig_Flag_t = 1; otherwise, output the time-domain trigger flag Trig_Flag_t = 0. If Trig_Flag_t = 1, then perform frequency domain feature analysis; if Trig_Flag_t = 0, then repeat steps 4) and 5). 6) Frequency domain feature analysis valid flame signal S eff [t] is converted to the frequency domain to identify whether there are flame-specific frequency components. The frequency domain trigger flag Trig_Flag_f is output based on the trigger threshold. If Trig_Flag_f = 1, the flame-specific frequency components are identified, and a flame determination model is constructed to determine the alarm status. If Trig_Flag_f = 0, steps 4), 5), and 6 are repeated.

3. The method for detecting thermal runaway of explosives according to claim 2, characterized in that, The method for constructing the flame determination model in step 6) includes the following steps: 61) Verify the band ratio Among them, the intensity of the flame in the main band is significantly higher than that in the background band, and the normal scene R band <2, Fire Scene R band >3; 62) Feature normalization The short-time energy E[n] and the fundamental frequency f peak Harmonic energy ratio R harmonic Band R band Feature normalization is performed to obtain the normalized short-time energy. Normalized fundamental frequency Normalized harmonic ratio Normalized band ratio 63) Feature-weighted fusion Linear weighted decision function: Where D represents the overall decision score; w i Represents the feature weights, ∑w i =1, w1=0.3, It is the normalized short-time energy; w2 = 0.

4. It is the normalized fundamental frequency; w3 = 0.

2. It is the normalized harmonic ratio; w4 = 0.

1. It is the normalized band ratio; 64) Grading threshold determination: Among them, Trig_Flag_D: a multi-dimensional trigger flag used to determine the alarm status; when Trig_Flag_D=00, the control host output current is 1mA, at which time the control host is in normal monitoring state, and the fire extinguishing device and audible and visual alarm do not activate; when Trig_Flag_D=11, the control host output current is 15mA, at which time the control host is in fire alarm output state, and the fire extinguishing device and audible and visual alarm are activated; when Trig_Flag_D=01 or 10, the control host output current is 20mA, at which time the control host is in equipment fault state, and the fire extinguishing device and audible and visual alarm do not activate.

4. The method for detecting thermal runaway of explosives according to claim 2, characterized in that, In step 1) noise measurement: Under a stable environment without flames or strong interference, the radiation spectrum of the explosive operation area is collected and processed in real time to obtain three channels of data in the initialization stage. Based on the three channel data from the initialization phase, configure the measurement noise covariance matrix R and the process noise covariance matrix Q, and calculate the initial background model B. model [0] and smoothing factor α, and simultaneously estimate the fusion value B of background bands 1 and 2. raw Historical standard deviation σ of [t] B .

5. The method for detecting thermal runaway of explosives according to claim 4, characterized in that, Step 3) also includes establishing an anomaly detection and freezing mechanism; if |B raw [t]-B model [t-1]∣>k·σ B If so, the background noise model will pause updating; Where k is the sensitivity coefficient.

6. An algorithm for processing thermal runaway signals of explosives, characterized in that, The following processing steps are included: Step P1: Kalman filtering: Real-time filtering is performed on the three channels of the background band 1, 2 and main band digital signal to obtain the filtered signal. Step P2: Dynamic Background Modeling: The filtered signal x from step P1... est Using [n] as input, a background noise model B is obtained through dynamic background modeling. model [t]; Step P3: Specific Spectrum Extraction: Extract the filtered signal S from step P1. main_est [t] and B in step P2 model Using [t] as input, the effective signal is calculated by subtracting background noise from the main band: S eff [t]=S main_est [t]-B model [t] Among them, S eff [t]: Valid flame signal, used as the output signal for dynamic background modeling; S main_est [t]: The estimated value of the main band at time t; B model [t]: The estimated value of the background model at time t; Step P4: Fast Time-Domain Detection: The valid flame signal S obtained in step P3... eff Using [t] as input, the short-time energy E[n] is obtained through short-time energy calculation, and then compared with the dynamic threshold T[n] to output the time-domain trigger flag Trig_Flag_t; Step P5: Frequency Domain Feature Analysis: Combine the time-domain trigger flag Trig_Flag_t obtained in Step P4 with the valid flame signal S obtained in Step P3. eff Using [t] as input, the fundamental frequency f is calculated through Fast Fourier Transform (FFT). peak Harmonic energy ratio R harmonic The frequency domain trigger flag Trig_Flag_f is output based on the trigger threshold. Step P6: Multi-dimensional feature fusion: Combine the band signal S from step P1 bg1_est S main_est S bg2_est The short-time energy E[n] in step P4, and the frequency domain trigger flag Trig_Flag_f and base frequency f in step P5. peak Harmonic energy ratio R harmonic As input, the presence of a flame signal is determined by the comprehensive decision score D, and the multi-dimensional trigger flag Trig_Flag_D is output.

7. The thermal runaway signal processing algorithm for explosives according to claim 6, characterized in that, The Kalman filter includes: (1) System initialization: Configure key noise covariance parameters for the Kalman filter algorithm, namely the measurement noise covariance matrix R and the process noise covariance matrix Q; Under a stable environment free from flame and strong interference, the raw output voltage signals V from three infrared sensors were collected. bg1 [n],V main [n],V bg2 [n], and then calculate the variance of the three channel data respectively. Normalize: s 2 norm =s 2 / (V ref ) 2 Among them, V ref It is a reference voltage, such as the range of an ADC or the full-scale value of a signal; σ 2 σ represents the variance of the three channel data. 2 norm This represents the normalized variance of the three channel data. The normalized variance values ​​are used as the diagonal elements of the measurement noise covariance matrix R, where R = diag([σ 2 norm_bg1 ,σ 2 norm_main ,σ 2 norm_bg2 ]); The process noise covariance matrix Q is used to model the intensity of process noise w[n] in the state equation, Q=diag([0.1,0.1,0.1]); (2) Kalman filter input: The raw output voltage signals V from the three infrared sensors are processed. bg1 [n],V main [n],V bg2 [n], and after being amplified by a conditioning circuit and converted by an ADC, a digital signal is obtained, which serves as the system's state vector x[n] = [S] at time n. bg1 ,S main ,S bg2 ] T ; (3) Kalman filtering process: 1) State prediction x pred =F*x est [n-1] Where, x pred The predicted state value at time n is based on the state estimate x at time n-1. est [n-1] is calculated from the state transition matrix F; F is the state transition matrix, and assuming the signal changes slowly, the state transition matrix F = 1 is set; x est [0] = [S bg1 [0],S main [0],S bg2 [0]] T ; Error covariance prediction: P pred =F*P est [n-1]*F T +Q Among them, P pred The prediction error covariance matrix at time n, representing the predicted state value x. pred Uncertainty of P est [n-1]: The estimation error covariance matrix at time n-1, representing the state estimate x at time n-1. est [n-1] uncertainty; P est [0] is set to a small diagonal matrix, indicating a high degree of confidence in the initial state estimate; 2) Status Update Kalman gain calculation: K=P pred *H T *(H*P pred *H T +R) -1 Where K is the Kalman gain matrix, which balances the confidence levels of prediction and measurement: if R >> Q, then the prediction is trusted and K → 0; if Q >> R, then the measurement is trusted and K → H. -1 ; Status Update: x est [n]=x pred +K*(z[n]-H*x pred ) Where, x est [n] serves as the filtered output signal, used for subsequent dynamic background modeling, effective signal extraction, and time-frequency analysis. This is achieved through the residual (z[n]-H*x). pred Correct the predicted value; z[n]-H*x pred The innovation or residual represents the difference between the actual observed value z[n] and the expected observed value H*x based on the prediction. pred The differences between them are: I: is the identity matrix; z[n] is the observation value of the system at the current moment, z[n] = H*x[n] + v[n]; H is the observation matrix, the state vector x[n] directly corresponds to z[n], so H is usually a scalar, set H = 1; v[n] is the observation noise, which follows a Gaussian distribution v ~ N(0,R); the measurement noise R reflects the sensor noise; Covariance update: P est [n]=(I–K*H)*P pred Among them, P est [n]: The updated estimation error covariance matrix at time n. The updated covariance reflects the corrected state estimate x. est [n] Uncertainty. (4) Kalman filter output: x est [n] represents the filtered output signal.

8. The thermal runaway signal processing algorithm for explosives according to claim 6, characterized in that, The dynamic background modeling includes: 1) Signal input: The state estimation vector of the output signal after Kalman filtering. As input for dynamic background modeling; 2) Background fusion: Using the estimated values ​​S of the two background bands bg1_est [t] and S bg2_est [t] performs background signal fusion; Among them, B raw [t]: The fused value of the two background band signals at time t; S bg1_est [t] and S bg2_est [t]: The estimated value of the background band at time t; 3) Dynamically update the background model: The background model is dynamically adjusted through a smoothing factor to adapt to environmental changes; the mathematical model is as follows: B model [t]=α·B model [t-1]+(1-a)·B raw [t] Among them, B model [t]: At time t, the estimated value of the background model (dynamic background signal); α: Smoothing factor, 0 < α < 1; When α → 1: the background updates slowly, with strong noise resistance, suitable for stable environments; When α → 0: the background updates quickly, with rapid response, suitable for dynamic environments; 4) Anomaly detection and freezing mechanism: Detects sudden background changes to avoid temporary interference that could contaminate the model; If |B raw [t]-B model [t-1]∣>k·σ B Then the background model will pause updating. Where, σ B It is background signal B raw [t] is the historical standard deviation; k is the sensitivity coefficient.

9. The thermal runaway signal processing algorithm for explosives according to claim 6, characterized in that, The time-domain fast detection includes: 1) Signal Input: Input the valid flame signal S described in step P3. eff [t] is used as the input signal. 2) Bandpass filtering: This filters the input signal S... eff [t], high-frequency noise and low-frequency drift (such as temperature drift) are eliminated by a bandpass filter (0.1Hz~1kHz) to obtain the filtered signal y[n]; 3) Short-time energy calculation: The preprocessed signal y[n] is then quantized to show the rapid rise of the flame within a short time window by quantizing the intensity change of the signal. E[n]=E[n-1]+y[n] 2 -y[nN] 2 Where, E[n]: the short-time energy calculated at sampling point n (the sum of signal energy within the current sliding window); E[n-1]: the short-time energy calculated at the previous sampling point n-1; y[n]: the effective signal S at sampling point n after high-pass / low-pass filtering preprocessing. eff The value of [t]; N: length of the sliding window; y[nN]: signal value before N sampling points; 4) Exponentially weighted moving average: m B [n]=0.99μ B [n-1]+0.01E[n] Where, μ B [n]: The exponentially weighted moving average (mean estimate) of the background noise energy at sampling point n; μ B [n-1]: The mean estimate at the previous sampling point n-1; At sampling point n, the exponentially weighted moving variance (variance estimate) of the background noise energy; Variance estimation at the previous sampling point n-1; 5) Dynamic threshold: The trigger threshold is adaptively adjusted according to background noise to reduce false alarms caused by environmental interference; Where, T[n]: the dynamic trigger threshold set at sampling point n; The standard deviation estimate of background noise energy (σ) B [n]); 6) Trigger condition judgment: Wherein, Trig_Flag_t: time-domain trigger flag, representing a binary flag indicating whether the trigger condition is met, 1 = met, 0 = not met; E[ni]>T[ni]: determines whether the short-time energy E[ni] at sampling point ni is greater than the corresponding dynamic threshold T[ni]; the result is 1 or 0; Count the number of points that satisfy E[ni]>T[ni] among M-1 points back from the current sampling point n; M: the number of sampling points that need to exceed the threshold consecutively, i.e., the binary flag bit of the trigger condition needs to be 1 for M consecutive times to exceed the threshold; i: the index of the backtracking sampling point; The frequency domain feature analysis includes: 1) Signal input: The effective flame signal S mentioned in step P3 eff [t] and the time-domain trigger flag Trig_Flag_t are used as input signals; if Trig_Flag_t = 1, then frequency domain feature analysis is performed; if Trig_Flag_t = 0, then return to step P3; 2) Add a window: For S eff [t] Windowing reduces spectral leakage and removes the DC component from the signal, resulting in x. f [n]; Hanning window function: Where w[n] is the Hanning window function value at time index n; n: time index of the sampling point within the window (n = 0, 1, ..., N-1); N: window length (number of FFT points); X[k]: frequency domain complex output, k is the frequency index; 3) FFT: Converts the time-domain signal x f [n] is converted to the complex frequency domain X f [k] indicates; 4) Amplitude spectrum calculation: Take the modulus of the FFT result to obtain the amplitude energy of each frequency component, obtain A[k], and search for the frequency with the largest amplitude in the target frequency band (4~20Hz) as the fundamental frequency f. peak ; 5) Harmonic energy ratio: Among them, R harmonic Harmonic energy ratio, representing the fundamental frequency (f) peak The second harmonic of ) (2f peak ) and third harmonic (3f peak The ratio of the sum of energies to the fundamental frequency energy; k peak : The frequency index corresponding to the base frequency; 6) Trigger condition judgment: Wherein, Trig_Flag_f: frequency domain trigger flag, a binary flag indicating whether the trigger condition is met, 1 = met, 0 = not met.

10. The thermal runaway signal processing algorithm for explosives according to claim 6, characterized in that, The multi-dimensional feature fusion includes: 1) Signal Input: Input the band signal S described in step P1. bg1_est S main_est S bg2_est (used to calculate band proportions), the short-time energy E[n], the frequency domain trigger flag Trig_Flag_f, and the fundamental frequency f peak Harmonic energy ratio R harmonic As the input signal; if Trig_Flag_f = 1, then perform multi-dimensional feature fusion; if Trig_Flag_f = 0, then return to step P3; Band ratio verification: Among them, the intensity of the flame in the main band is significantly higher than that in the background band, and the normal scene R band <2, Fire Scene R band >3; 2) Feature normalization: Scaling features of different dimensions to a uniform range to avoid the influence of numerical differences on weights; Among them, F i F represents the original eigenvalues; min / F max Indicates the historical minimum / maximum value of the feature; 3) Feature-weighted fusion: Assign weights based on feature importance and calculate a comprehensive score; design a decision threshold to determine the presence of flames; use a linear weighted decision function. Where D represents the overall decision score; w i Represents the feature weights (∑w) i =1); w1=0.3, It is the normalized short-time energy; w2 = 0.

4. It is the normalized fundamental frequency; w3 = 0.

2. It is the normalized harmonic ratio; w4 = 0.

1. It is the normalized band ratio; 4) Determination of grading thresholds: Among them, Trig_Flag_D: a multi-dimensional trigger flag used to determine the alarm status.

Images