An infrared flame detection method with anti-interference and model learning

Abstract

The present invention proposes an infrared flame detection method with anti-interference and model learning, the steps are: collecting detection signals of multiple cycles in a unit time, Fourier transforming the detection signals and storing them in an array; convolving the data in one cycle of two channels with a discrete data method to obtain a feature set of a burning material; collecting data after combustion of multiple burning materials, obtaining the feature set of each burning material through data analysis, and establishing a basic model; collecting real flame signals, Fourier transforming the real flame signals to obtain detection data, comparing the detection data with the stored multiple sets of feature sets, giving an alarm and adding the detection data to the basic model, and updating the weight parameters of the basic model. The present invention uses convolution operation to classify and calculate the collected data, saves the calculation results as flame features to form a self-learning model, which can effectively reduce false alarms.

Description

Technical Field

[0001] The invention relates to the technical field of infrared flame detection, and in particular to an infrared flame detection method with anti-interference and model learning capabilities. Background Art

[0002] Fire detectors are one of the most important components of fire detection systems. The most common ones are smoke, temperature, infrared and image. Smoke fire detectors mainly identify the smoke produced in the early smoldering stage of a fire, thereby realizing an alarm. Temperature fire detectors mainly judge the fire condition by the temperature at a specific location. Infrared fire detectors generally use multi-band judgments, using the changes in sensor signals in a specific band to judge the fire condition. Image fire detectors generally use infrared thermal imaging and image analysis as the principle, and have monitoring functions at the same time, but they will be affected by high-temperature objects and cause false alarms, such as high-temperature water vapor, moving red objects, red clothes, sunlight, fluttering red flags, etc. The application scenarios of infrared pyroelectric flame detectors are very wide, including aerospace industry, chemical industry, highway tunnels, ammunition warehouses, paint factories, petroleum and petrochemicals, pharmaceutical factories, power plants, warehouses and other flammable and explosive places.

[0003] In recent years, explosions and gas leaks have occurred frequently in various industrial and commercial places, eventually causing fires and major accidents, and people's safety awareness has become increasingly strong. While accurately detecting open flames, it is also necessary to be able to withstand complex industrial scenes. Therefore, the demand for flame detectors with high stability is also growing. How to highly match the industrial site environment, accurately detect without false alarms, is a new test for flame products in the current form.

[0004] At present, infrared flame detectors use Fourier transform technology, and generally use multi-channel acquisition method combined with threshold method to judge fire conditions. However, during the period of interference from interference sources, data in some frequency bands have certain flame characteristics. During data collection, more commonly used filtering algorithms are generally used. Although false alarms can be prevented to a certain extent, they cannot be reliably and effectively suppressed. Monotonous algorithms will increase uncertainty and fail to bring out the characteristics that the product should have.

[0005] Flame detectors detect open flames. The important feature of flames is that burning materials will radiate electromagnetic waves when burning, mainly concentrated in ultraviolet and mid-infrared. Another important feature is the flicker effect, whose frequency is affected by the surrounding environment such as wind, but is basically within the range of 0.5-30Hz. The infrared radiation spectrum emitted by heating objects is not used for the radiation spectrum emitted by flames, nor does it have the flickering characteristics of flames. Under ideal conditions, the above two important characteristics can be used to distinguish between real and effective flames and interference sources. In actual use, the impact of interference sources cannot be ignored.

[0006] The invention patent with application number 202111214941.1 discloses a flame detection method based on FFT-GMM-K nearest neighbors, including: sample recording, wherein the types of samples include flames and interference sources, and each sample is recorded in combination at different distances and angles from the flame detector; the recorded samples are cropped to obtain a sample set of fixed length; infrared feature extraction and ultraviolet feature extraction are performed on the cropped sample set; GMM fitting is performed according to the results of infrared feature extraction and ultraviolet feature extraction, wherein each GMM output represents the similarity between the predicted sample and the interference source or flame, and probability is used as a measure of similarity. The greater the probability, the greater the possibility that the predicted sample belongs to the interference source or flame, so as to finally select the most similar K neighbors, calculate the sum of the neighbor probabilities that the predicted sample belongs to the flame and the sum of the neighbor probabilities that the predicted sample belongs to the interference source; according to the size of the sum of the neighbor probabilities that the predicted sample belongs to the flame and the sum of the neighbor probabilities that the predicted sample belongs to the interference source, the type to which the predicted sample belongs is judged. The above invention patent application better utilizes the advantages of the combination of red and ultraviolet sensors, and the detection accuracy of flames is significantly improved. However, in the above invention, the step of cutting the recorded samples to obtain a sample set of fixed length points out that a sample matrix is ​​X 4×1280 , the duration is 1280 / Fs=10 seconds; that is, a sampling period is 10 seconds. If there is a fire on site, the response time is too long. Some specific scenarios that require rapid alarms cannot be met, such as hazardous chemical production lines and explosive production lines. In addition, the Gaussian mixture model has a large amount of calculations and is prone to errors. It does not work well for high-dimensional data because there is a lot of redundant information, which will lead to a decrease in model effectiveness. At the same time, the requirements for the CPU are relatively high, the cost cannot be controlled, and it cannot be commercialized well. Summary of the invention

[0007] In view of the technical problems that the existing infrared flame detection methods cannot reliably and effectively suppress interference signals and have low detection accuracy, the present invention proposes an infrared detection method with anti-interference and model learning capabilities, establishes an anti-interference and model learning mechanism, has the ability of anti-interference and model learning, and increases product reliability.

[0008] In order to achieve the above object, the technical solution of the present invention is implemented as follows: an infrared flame detection method with anti-interference and model learning, the steps of which are as follows:

[0009] Step 1: Collect detection signals of multiple cycles per unit time through the pyroelectric sensor, perform Fourier transform on the detection signals and store them in an array; perform discrete data convolution on the data of one cycle of two channels of the pyroelectric sensor to obtain a feature set of a burning substance;

[0010] Step 2: Collect data after combustion of various combustion substances, analyze the data as the feature set of each combustion substance, and establish a basic model based on the feature set;

[0011] Step 3: Use the pyroelectric sensor to collect the real flame signal, perform Fourier transform on the real flame signal to obtain detection data, compare the detection data with the stored multiple feature sets, and when it meets the basic model, issue an alarm and add the detection data to the basic model, while updating the weight parameters of the basic model.

[0012] Preferably, in the flame detection device, the pyroelectric infrared sensor is connected to the signal conversion circuit through the signal processing circuit, the signal conversion circuit is connected to the data processing unit, the data processing unit is a microprocessor, the microprocessor is respectively connected to the signal output circuit, the indicator light circuit, and the relay output circuit, the signal output circuit is used to output the processed signal, the indicator light circuit is used to indicate the working state of the detector, the relay circuit is used to output the signal to the outside, and in the alarm state, the normally open end of the relay is closed, and a passive switch signal is output to the outside;

[0013] The signal processing circuit is an amplifier circuit, the signal conversion circuit is an AD conversion circuit, and the signal output circuit is a 4-20mA output circuit.

[0014] Preferably, the pyroelectric sensor includes a 3.8um sensor, a 4.3um sensor and a 5.0um sensor respectively, and judgment is made by detecting infrared rays of a specific band emitted by the flame, wherein the 4.3um sensor works at a wavelength that reflects flame information as a detection channel, and the 3.8um sensor and the 5.0um sensor are used to monitor other infrared radiation in the environment as reference channels, and the actual fire condition is judged in combination with the flickering characteristics of flame burning.

[0015] Preferably, several cycles of data are collected in a unit time, and the sampling frequency is assumed to be Fs, the number of sampling points is N, and the resolution is Fs / N. After Fourier transform, the results of 2N points are obtained, and the results are stored in an array. The first N data are taken for judgment. The inherent flicker spectrum of the flame 1-30Hz is selected for data analysis, and then the characteristics of the detection channel and the reference channel data are combined for judgment. The data with the detection channel energy exceeding a certain value and the reference channel energy below a certain value are considered to be real flames.

[0016] Preferably, convolution is performed according to the actual situation of flame combustion, and N data of one period of two channels are convolved, and another reference channel and detection channel are convolved to obtain a feature set of the burning material; when there are two channel data inputs, the reference channel function is first inverted, and then slid, multiplied at each frequency value, and N values ​​are obtained, and then the characteristic values ​​are selected as the reference feature set according to the distribution of the calculated values.

[0017] Preferably, the convolution uses a discrete method to perform data calculation:

[0018]

[0019] Among them, f and g represent a detection channel and a reference channel respectively, s represents the convolution data correspondence between a detection channel and a reference channel, n represents the center frequency of flame combustion, Index variable representing the convolution summation.

[0020] Preferably, the method for establishing the basic model is: according to the different combustion materials, combustion data is collected, each type of data is calculated separately, and then the calculated data is classified and sorted, and a feature set is obtained by analyzing the results of different combustion materials to establish a model learning mechanism.

[0021] Preferably, the combustion data processing method is: the signal is represented by S, N points are collected in a unit time, and the original data collected by the three channels are:

[0022] 3.8S[0], 3.8S[1], 3.8S[2], 3.8S[3]…………3.8S[N]

[0023] 4.3S[0], 4.3S[1], 4.3S[2], 4.3S[3]…………4.3S[N]

[0024] 5.0S[0], 5.0S[1], 5.0S[2], 5.0S[3]…………5.0S[N]

[0025] After FFT calculation, the output result is:

[0026] 3.8SFFT[0], 3.8SFFT[1], 3.8SFFT[2], 3.8SFFT[3]…………3.8SFFT[N]

[0027] 4.3SFFT[0], 4.3SFFT[1], 4.3SFFT[2], 4.3SFFT[3]…………4.3SFFT[N]

[0028] 5.0SFFT[0], 5.0SFFT[1], 5.0SFFT[2], 5.0SFFT[3]…………5.0SFFT[N]

[0029] And it is the spectrum information, the frequency information represented by the nth point is Fn = (n-1) * Fs / N, and the amplitude of each frequency point is FFT [n] / N / 2; combined with the flame detection principle, according to the flickering characteristics of flame combustion, FFT [1] to FFT

[30] are selected as the main calculation data, and two groups of data are obtained, A [1], A [2], A [3] ... ... A

[30] and B [1], B [2], B [3] ... ... B

[30]

[0030] express;

[0031] Use a sample feature set to represent the combustion characteristics of a substance:

[0032]

[0033] Preferably, the combustion substances of the basic model include: n-heptane, alcohol, gasoline, diesel, charcoal, kerosene, alkane gas, candle, paper and coal; the basic model does not include interference sources, which include sunlight, human body, electric soldering iron, halogen lamp, photographic lamp, high-power LED, electric furnace wire and incandescent lamp;

[0034] The received infrared radiation energy exceeding a certain value and having a certain proportional relationship between the detection channel and the reference channel is regarded as a fire source signal; the 2-10Hz low-frequency information in the basic model represents the fire source signal; and a threshold range is obtained based on multiple sets of feature sets.

[0035] Preferably, the method for updating the weight parameters of the basic model is: the data in the basic model satisfies the normal distribution as the data input of the Gaussian fuzzy algorithm, and the weight of each integer frequency of 2-10Hz is obtained with 6Hz as the center; when calculating the average value, the center point is taken as the origin, and the other points are obtained according to the position on the normal curve to obtain a weighted average value;

[0036] Gaussian function of the Gaussian blur algorithm Among them, x represents N data in a period, μ is the mean value of x, is the variance of x.

[0037] Compared with the prior art, the present invention has the following beneficial effects: a large amount of test data is accumulated in the early stage, a large amount of interference data is collected according to the characteristics of the pyroelectric sensor, the data are classified, and the calculation results are obtained, and the calculation results are stored in the processor as a data training set. Later, in the process of working in the industrial scene, the calculation results stored in the early stage are compared with the data collected on site. If there is similar interference data, the test results are saved and a detector learning model is established at the same time; the present invention uses convolution operations to classify and calculate the collected data, thereby realizing the classification of fire signals and interference signals, and the calculation results are saved as flame characteristics to form a self-learning model, which can effectively reduce false alarms.

[0038] The present invention extracts real and effective flame signals through early test data combined with later data analysis of the on-site working environment, and classifies the signals into the current feature set to generate a new feature set, thereby increasing the judgment criteria and making the product more reliable. In addition, through the early collection and processing of interference source signals, signals different from real flames are avoided, interference sources are eliminated, and false alarms of detectors can be reliably prevented. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0040] Figure 1 A flow chart is established for the early basic model of the present invention.

[0041] Figure 2 It is a flow chart of the later stage self-learning of the present invention.

[0042] Figure 3 It is a schematic diagram of the normal curve of the present invention. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in 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 creative work are within the scope of protection of the present invention.

[0044] like Figure 1 As shown, an infrared detection method with anti-interference and model learning, the steps are as follows:

[0045] Step 1: Collect detection signals of multiple cycles in unit time through the pyroelectric sensor, perform Fourier transform on the detection signals and store them in an array; perform discrete data convolution on the data in one cycle of two channels of the pyroelectric sensor to obtain a feature set of a burning substance.

[0046] The signal is collected in real time, and is continuously detected regardless of whether there is a fire or not. The unit time is a period of time for the CPU to collect data, which can also be 1 second. If 1 second is taken, the data collected within 1 second is the data of one cycle.

[0047] The flame detection device includes: a pyroelectric infrared sensor, a signal processing circuit, a signal conversion circuit, a data processing unit, a signal output circuit and an indicator light circuit. The pyroelectric infrared sensor is connected to the signal conversion circuit through the signal processing circuit. The signal processing circuit is an amplifier circuit. The signal conversion circuit is an AD conversion circuit. The signal conversion circuit is connected to the data processing unit. The data processing unit is a microprocessor. The microprocessor is respectively connected to the signal output circuit, the power supply circuit, the indicator light circuit, and the relay output circuit. The signal output circuit is a 4-20mA output circuit for outputting processed signals. The power supply circuit provides the CPU with working voltage, sensor power supply, and signal processing circuit power supply. The indicator light circuit is used to indicate the working state of the detector, normal detection state or alarm state. The relay circuit is used to output signals to the outside. In the alarm state, the normally open end of the relay is closed, and a passive switch signal is output to the outside.

[0048] Pyroelectric sensors include 3.8um sensors, 4.3um sensors and 5.0um sensors, which detect flames by detecting infrared rays of a specific band emitted by the flame. Among them, the 4.3um sensor works at a wavelength that reflects flame information, and the 3.8um sensor and 5.0um sensor are used to monitor other infrared radiation in the environment. Together with the signal acquisition unit and the data processing unit, the signals collected per unit time are analyzed, and the actual fire condition is judged by combining the flickering characteristics of flame combustion. The signal acquisition unit is an AD analog acquisition module inside the CPU, which can continuously collect voltage.

[0049] The process of establishing the feature set: collect interference source signals in a unit time, analyze the interference source signals, obtain the results, and store the feature information according to the signal characteristics. First, collect several cycles of data in a unit time, set the sampling frequency to Fs, the number of sampling points to N (generally N is an integer power of 2), and the resolution to Fs / N. After Fourier transform, 2N points of results are obtained. This result will be stored in an array. The first bit in the array is the DC component, which is generally not involved in the calculation. Since the FFT output result is symmetrical, the first N data are taken for judgment.

[0050] The CPU processor needs to collect 1 cycle of data within 1 second of unit time, assuming that the sampling frequency is Fs (the present invention adopts 256Hz), the number of sampling points is N (generally N is an integer power of 2, and here a 256-point FFT transformation is selected, so N is also 256), and the resolution is Fs / N=1Hz. After Fourier transformation, 2N (512) points of results will be obtained, and this result will be stored in an array, which is an array output after using the CPU's internal DSP processing unit. The subscript of the array is the spectrum information. The first bit in the array is the DC component, which is generally not involved in the calculation. Since the fft output result is symmetrical, the first N data are taken for judgment. This calculation process transforms the time domain signal into a frequency domain signal, selects the inherent flicker spectrum of the flame 1-30Hz for data analysis, and then combines the characteristics of the detection channel and the reference channel data for judgment, and considers the data with the detection channel energy exceeding a certain value and the reference channel energy below a certain value as a real flame.

[0051] Using discrete method convolution for data calculation

[0052]

[0053] Among them, f and g represent a detection channel and a reference channel respectively, s represents the convolution data correspondence between a detection channel and a reference channel, n represents the center frequency of flame combustion, Represents the index variable of the convolution sum. Convolution is performed according to the actual situation of flame combustion. For different signals, N data of one cycle (0 to N) of the two channels are convolved. The other reference channel and the detection channel are convolved in the same way to obtain a characteristic set of a burning substance. When there are two channel data inputs, first invert the reference channel function, then slide, multiply at each frequency value, and obtain N values. Then, according to the distribution of the calculated values, select the characteristic value. For example, if the maximum calculated value appears at 6Hz, then we select 4-8Hz data as the reference characteristic set. And so on. The three-channel detector must calculate the detection channel and the two reference channels before it can be used as a basis for judgment. Three infrared selects 3.8um, 4.3um, and 5.0um, of which the 4.3um channel is used as the detection channel, and 3.8um and 5.0um are used as reference channels.

[0054] Step 2: Collect data after combustion of various combustion substances, analyze the data as the feature set of each combustion substance, and establish a basic model based on the feature set.

[0055] Model establishment: According to the different combustion materials, their combustion data are collected, each type of data is calculated separately, and then the calculated data is classified and sorted. By analyzing the results of different combustion materials, a model learning mechanism is established. The signal is represented by S, because N points need to be collected in a unit time. The original data collected by the three channels is

[0056] 3.8S[0], 3.8S[1], 3.8S[2], 3.8S[3]…………3.8S[N]

[0057] 4.3S[0], 4.3S[1], 4.3S[2], 4.3S[3]…………4.3S[N]

[0058] 5.0S[0], 5.0S[1], 5.0S[2], 5.0S[3]…………5.0S[N]

[0059] After FFT calculation, the output result is:

[0060] 3.8SFFT[0], 3.8SFFT[1], 3.8SFFT[2], 3.8SFFT[3]…………3.8SFFT[N]

[0061] 4.3SFFT[0], 4.3SFFT[1], 4.3SFFT[2], 4.3SFFT[3]…………4.3SFFT[N]

[0062] 5.0SFFT[0], 5.0SFFT[1], 5.0SFFT[2], 5.0SFFT[3]…………5.0SFFT[N]

[0063] And is the spectrum information, the frequency information represented by the nth point is Fn = (n-1) * Fs / N, and the amplitude of each frequency point is FFT [n] / N / 2. Combined with the flame detection principle, the 4.3um sensor channel is the detection channel, the 3.8um and 5.0um sensors are the reference channels, and according to the flickering characteristics of flame combustion, FFT [1] to FFT

[30] are selected as the main calculation data, and two sets of data can be obtained.

[0064] A[1], A[2], A[3]…………A

[30]

[0065] B[1], B[2], B[3]…………B

[30]

[0066] express.

[0067] Therefore, a sample feature set can be used to represent the combustion characteristics of a certain substance:

[0068]

[0069] Different sample feature sets can be obtained for different combustion materials. According to different combustion materials, the following are listed and experimental data are collected to obtain multiple feature sets. Using the above method, different combustion materials have different performances and calculation results. In the early stage, ten sets of results are recorded and stored in the CPU as the early combustion model.

[0070] Table 1 Multiple feature sets

[0071] n-Heptane S1 kerosene S6 Alcohol S2 Alkane gas S7 gasoline S3 Candle S8 diesel fuel S4 Paper S9 charcoal S5 coal S10

[0072] The feature set is saved as the flame criterion in the processor, and combined with the distribution of the ten sets of data, it is determined whether there is a fire at present. The data of the ten kinds of burning substances can be used as a basic model.

[0073] Step 3: Use the pyroelectric sensor to collect the real flame signal, perform Fourier transform on the real flame signal to obtain detection data, compare the detection data with the stored multiple feature sets, and when it meets the basic model, issue an alarm and add the detection data to the basic model, while updating the weight parameters of the basic model.

[0074] like Figure 2 As shown in the figure, the received infrared radiation energy exceeds a certain value, and the detection channel and the reference channel have a certain proportional relationship, which is regarded as the fire source signal. Because the flickering frequency is concentrated in the low frequency band when the flame is burning, the low-frequency information of the above 10 types of data can represent the fire source signal. The 2-10Hz data is mainly analyzed. The fire of some burning objects is large and the flickering is slow, so the frequency information is concentrated in 2-6Hz. The fire of some burning objects is small and the flickering is fast, so the frequency information is concentrated in 5-10Hz. Therefore, these 10 results can be used as a basic model. According to the above multiple sets of feature sets, a threshold range is obtained. Data that meets this feature set are regarded as real flames. At the same time, the feature is included and saved as new data to generate a new feature set, thereby establishing a model learning mechanism.

[0075] It is known that the data results, that is, the 10 basic models satisfy the normal distribution, and are used as the data input of the Gaussian fuzzy algorithm. Based on the data of the above ten fire sources, the weight of each integer frequency of 2-10Hz is obtained with 6Hz as the center, as follows: Figure 3 As shown in the figure, the normal distribution is a bell-shaped curve. The closer to the center, the larger the value, and the farther away from the center, the smaller the value. When calculating the average value, you only need to use the "center point" as the origin, and the other points can get a weighted average value according to their positions on the normal curve. According to the Gaussian function,

[0076]

[0077] Among them, x represents N data in a period, μ is the mean value of x, is the variance of x. Based on this function, the weight of each point can be calculated. The data of each frequency point, for example, the data of the flame signal at 2Hz is 20 units, at 3Hz is 60 units, at 4Hz is 40, and at 5Hz is 20 units. The proportion of each frequency subscript can be calculated. The larger the proportion, the closer it is to the flame signal. The self-learning process is realized according to this proportion. The self-learning process can be realized according to the weight coefficient. Therefore, in the actual work of the detector in the later stage, any data that meets the conditions can be used as a real flame signal.

[0078] For interference signals, use the same method in step 1 to test multiple interference sources, and record them as follows:

[0079] sunlight X1 High Power LED X6 human body X2 Electric furnace wire X7 Electric soldering iron X3 Incandescent lamp X8 Halogen lamp X4 Photography Light X5

[0080] Because the information of these signals after Fourier transformation no longer meets the data in the basic model. After analyzing the above steps, it is concluded that the data collected from high-temperature heating objects and low-temperature objects are quite different from the real flame signal and do not meet the functional relationship. Therefore, similar signals can be excluded as interference sources, and there will be no false alarms, which can effectively solve the problem of false alarms in existing detectors. Such signals are quite different from flame signals, the flickering information does not meet the requirements, and the data deviation from the above basic model is large.

[0081] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An infrared flame detection method with anti-interference and model learning, characterized in that: The steps are as follows: Step 1: Collect detection signals of multiple cycles per unit time through the pyroelectric sensor, perform Fourier transform on the detection signals and store them in an array; perform discrete data convolution on the data of one cycle of two channels of the pyroelectric sensor to obtain a feature set of a burning substance; Collect several cycles of data in a unit time, set the sampling frequency to Fs, the number of sampling points to N, and the resolution to Fs / N. After Fourier transform, get the results of 2N points, store the results in an array, and take the first N data for judgment; select the inherent flicker spectrum of the flame 1-30Hz for data analysis, and then combine the characteristics of the detection channel and reference channel data for judgment, and consider the data with the detection channel energy exceeding a certain value and the reference channel energy below a certain value as real flame; According to the actual situation of flame combustion, convolution is performed, N data of one cycle of two channels are convolved, and another reference channel and detection channel are convolved to obtain a characteristic set of a burning substance; when there are two channel data inputs, the reference channel function is first inverted, then slid, and multiplied at each frequency value to obtain N values, and then the characteristic value is selected as the reference characteristic set according to the distribution of the calculated value; The convolution uses a discrete method to calculate data: Among them, f and g represent a detection channel and a reference channel respectively, s represents the convolution data correspondence between a detection channel and a reference channel, n represents the center frequency of flame combustion, The index variable representing the convolution summation; Step 2: Collect data after combustion of various combustion substances, analyze the data as the feature set of each combustion substance, and establish a basic model based on the feature set; The method for establishing the basic model is as follows: according to different combustion materials, combustion data is collected, each type of data is calculated separately, and then the calculated data is classified and sorted, and a feature set is obtained by analyzing the results of different combustion materials, and a model learning mechanism is established; The combustion data processing method is as follows: the signal is represented by S, N points are collected in a unit time, and the original data collected by the three channels are: 3.8S[0], 3.8S[1], 3.8S[2], 3.8S[3]…………3.8S[N] 4.3S[0], 4.3S[1], 4.3S[2], 4.3S[3]…………4.3S[N] 5.0S[0], 5.0S[1], 5.0S[2], 5.0S[3]…………5.0S[N] After FFT calculation, the output result is: 3.8SFFT[0], 3.8SFFT[1], 3.8SFFT[2], 3.8SFFT[3]…………3.8SFFT[N] 4.3SFFT[0], 4.3SFFT[1], 4.3SFFT[2], 4.3SFFT[3]…………4.3SFFT[N] 5.0SFFT[0], 5.0SFFT[1], 5.0SFFT[2], 5.0SFFT[3]…………5.0SFFT[N] and are spectrum information, the frequency information represented by the nth point is Fn=(n-1)*Fs / N, and the amplitude of each frequency point is FFT[n] / N / 2; combined with the flame detection principle and according to the flickering characteristics of flame combustion, FFT[1] to FFT[30] are selected as the main calculation data, and two groups of data are obtained, A[1], A[2], A[3]…………A[30] and B[1], B[2], B[3]…………B[30] express; Use a sample feature set to represent the combustion characteristics of a substance: Step 3: Use the pyroelectric sensor to collect the real flame signal, perform Fourier transform on the real flame signal to obtain detection data, compare the detection data with the stored multiple feature sets, and when it meets the basic model, issue an alarm and add the detection data to the basic model, while updating the weight parameters of the basic model.

2. The infrared flame detection method with anti-interference and model learning according to claim 1 is characterized in that: In the flame detection device, the pyroelectric infrared sensor is connected to the signal conversion circuit through the signal processing circuit, the signal conversion circuit is connected to the data processing unit, the data processing unit is a microprocessor, the microprocessor is respectively connected to the signal output circuit, the indicator light circuit, and the relay output circuit, the signal output circuit is used to output the processed signal, the indicator light circuit is used to indicate the working state of the detector, and the relay circuit is used to output the signal to the outside. In the alarm state, the normally open end of the relay is closed, and a passive switch signal is output to the outside; The signal processing circuit is an amplifier circuit, the signal conversion circuit is an AD conversion circuit, and the signal output circuit is a 4-20mA output circuit.

3. The infrared flame detection method with anti-interference and model learning according to claim 1 or 2, characterized in that: The pyroelectric sensors include 3.8um sensor, 4.3um sensor and 5.0um sensor, which make judgments by detecting infrared rays of a specific band emitted by the flame. The 4.3um sensor works at a wavelength that reflects flame information as a detection channel, and the 3.8um sensor and 5.0um sensor are used to monitor other infrared radiation in the environment as reference channels, and the actual fire situation is judged in combination with the flickering characteristics of flame burning.

4. The infrared flame detection method with anti-interference and model learning according to claim 3 is characterized in that: The combustion substances of the basic model include: n-heptane, alcohol, gasoline, diesel, charcoal, kerosene, alkane gas, candle, paper and coal; the basic model does not include interference sources, which include sunlight, human body, electric soldering iron, halogen lamp, photographic lamp, high-power LED, electric furnace wire and incandescent lamp; The received infrared radiation energy exceeding a certain value and having a certain proportional relationship between the detection channel and the reference channel is regarded as a fire source signal; the 2-10Hz low-frequency information in the basic model represents the fire source signal; and a threshold range is obtained based on multiple sets of feature sets.

5. The infrared flame detection method with anti-interference and model learning according to claim 4 is characterized in that: The method for updating the weight parameters of the basic model is as follows: the data in the basic model satisfying the normal distribution is used as the data input of the Gaussian fuzzy algorithm, and the weight of each integer frequency of 2-10Hz is obtained with 6Hz as the center; when calculating the average value, the center point is used as the origin, and the other points are obtained according to the position on the normal curve to obtain a weighted average value; Gaussian function of the Gaussian blur algorithm Among them, x represents N data in a period, μ is the mean value of x, is the variance of x.

Images