Infrared multi-band intelligent fire identification method and system based on time sequence classification model

Abstract

The invention relates to an infrared multi-band intelligent fire behavior identification method and system based on a time sequence classification model, belongs to the technical field of fire behavior identification, and solves the problems that in the prior art, the identification distance is short, the anti-interference capability is weak, misjudgment is easily caused, and reliable identification cannot be realized when radiation is strong. Comprising the steps of collecting environmental radiation intensity data and performing preprocessing; inputting the data obtained after preprocessing into a time sequence classification model to obtain an identification result; the model comprises a flame feature extraction module which is used for calculating a flame matching coefficient by using data of an optimal wave band; the multi-band information fusion module is used for calculating the probability of fire behavior estimation by using the flame matching coefficient, the absolute radiation intensity mean value of the main band and the relative radiation intensity mean value of the auxiliary band; the saturation judgment module is used for judging whether the main wave band is saturated; and the threshold judgment module is used for judging whether a fire occurs or not. And the identification distance is long, the anti-interference capability is strong, and the reliable identification can still be realized when the radiation is strong.

Description

Technical Field

[0001] The present invention relates to the technical field of fire identification, and in particular to an infrared multi-band intelligent fire identification method and system based on a time series classification model. Background Art

[0002] Traditional infrared multi-band optical detectors collect infrared sensor response values ​​in multiple bands according to a fixed sampling period, and then design rule-based expert algorithms, such as slope / threshold discrimination, different band response ratio discrimination, and FFT frequency analysis algorithm to perform flame detection.

[0003] Since the traditional method only uses signal strength information, when the fire source is far away, the infrared radiation power attenuates severely, resulting in unclear signal changes and the recognition distance is usually limited to within 2 meters. The traditional threshold comparison method cannot effectively distinguish flames from other interference sources, such as flickering lights, welding sparks, direct or reflected sunlight, and can easily cause misjudgment. When the sensor is subjected to strong radiation, the main band is saturated and the features required for fire identification are lost, making it impossible to reliably identify the fire. Summary of the invention

[0004] In view of the above analysis, an embodiment of the present invention aims to provide an infrared multi-band intelligent fire identification method based on a time series classification model to solve the problems of the prior art such as short identification distance, weak anti-interference ability, easy misjudgment, and inability to reliably identify when the radiation is strong.

[0005] On the one hand, an embodiment of the present invention provides an infrared multi-band intelligent fire identification method based on a time series classification model, comprising the following steps:

[0006] Collect and pre-process the ambient radiation intensity data of the main band and the auxiliary band to obtain the relative radiation intensity mean of the auxiliary band, the absolute radiation intensity mean of the main band, and the optimal band;

[0007] The data obtained after preprocessing is input into the time series classification model to obtain the recognition result; wherein the time series classification model includes: a flame feature extraction module, a multi-band information fusion module, a saturation judgment module and a threshold judgment module; wherein the flame feature extraction module is used to calculate the flame matching coefficient using the data of the optimal band; the multi-band information fusion module is used to calculate the probability of fire prediction using the flame matching coefficient, the absolute radiation intensity mean of the main band and the relative radiation intensity mean of the auxiliary band; the saturation judgment module is used to judge whether the main band is saturated; the threshold judgment module is used to select the flame matching coefficient or the probability of fire prediction to judge whether a fire occurs using the result output by the saturation judgment module.

[0008] Based on the further improvement of the above method, the flame feature extraction module includes a fixed multi-scale convolution kernel, a pooling layer and a Ridge binary classification module; wherein,

[0009] A fixed multi-scale convolution kernel is used to extract features from the data of the optimal band;

[0010] The pooling layer is used to generate high-dimensional feature vectors using the extracted feature data;

[0011] The Ridge binary classification module is used to calculate the weight coefficient of each feature using the high-dimensional feature vector, and perform weighted summation of each feature and the corresponding weight coefficient to obtain a flame matching coefficient.

[0012] Based on the further improvement of the above method, the fixed multi-scale convolution kernel includes: the length of the convolution kernel is fixed and remains unchanged during the training and use stages of the module.

[0013] Based on the further improvement of the above method, the multi-band information fusion module includes a multi-layer perceptron network and a softmax layer; the flame matching coefficient, the absolute radiation intensity mean of the main band and the relative radiation intensity mean of the auxiliary band obtained by the flame feature extraction module are used as the input of the multi-layer perceptron network to calculate the classification of fire and no fire; the classification of fire and no fire is input into the softmax layer to calculate the probability of fire estimation.

[0014] Based on the further improvement of the above method, the saturation judgment module is used to judge whether the main band is saturated. If the main band is saturated, the flame matching coefficient obtained by the flame feature extraction module is directly used to judge whether there is a fire; otherwise, the probability of fire estimation calculated by the multi-band information fusion module is used to judge whether there is a fire.

[0015] Based on the further improvement of the above method, the threshold judgment module includes: a flame matching coefficient threshold module and a flame probability threshold module:

[0016] The flame matching coefficient threshold module is used to compare with the flame matching coefficient. If the flame matching coefficient is greater than the flame matching coefficient threshold, it is considered that there is a fire;

[0017] The flame probability threshold module is used to compare with the estimated probability of fire. If the estimated probability of fire is greater than the flame probability threshold, it is considered that there is a fire.

[0018] Based on the further improvement of the above method, the optimal band is obtained by using the following steps:

[0019] The saturation state of each band is calculated respectively, and the first band whose corresponding time series data is not saturated is found in the order of radiation intensity from large to small, which is regarded as the optimal band; wherein,

[0020] Calculating the saturation state for a band includes: calculating the median of time series data of a band collected by the sensor in a recent time period, and if the median is greater than 90% of the sensor range, the band is considered to be saturated.

[0021] Based on the further improvement of the above method, the relative radiation intensity mean of the auxiliary band is obtained by using the following steps:

[0022] The absolute radiation intensity average of each auxiliary band is calculated;

[0023] The absolute radiation intensity mean of the main band is compared with 0.1, the larger one is taken as the denominator, the absolute radiation intensity mean of the auxiliary band is taken as the numerator, and the relative radiation intensity mean of each auxiliary band is calculated respectively.

[0024] Based on the further improvement of the above method, the preprocessing includes: down sampling, background removal difference processing, exponential moving average processing and saturation state statistics.

[0025] On the other hand, an embodiment of the present invention provides an infrared multi-band intelligent fire identification system based on a time series classification model, comprising a data acquisition module, a data preprocessing module and a time series classification module, wherein:

[0026] A data acquisition module, used to collect the environmental radiation intensity data of the main band and the auxiliary band;

[0027] The data preprocessing module is used to preprocess the collected data to obtain the relative radiation intensity mean of the auxiliary band, the absolute radiation intensity mean of the main band, and the optimal band;

[0028] The time series classification module includes a flame feature extraction module, a multi-band information fusion module, a saturation judgment module and a threshold judgment module; wherein the flame feature extraction module is used to calculate the flame matching coefficient using the data of the optimal band; the multi-band information fusion module is used to calculate the probability of fire prediction using the flame matching coefficient, the absolute radiation intensity mean of the main band and the relative radiation intensity mean of the auxiliary band; the saturation judgment module is used to judge whether the main band is saturated; the threshold judgment module is used to select the flame matching coefficient or the probability of fire prediction to judge whether a fire occurs using the result output by the saturation judgment module.

[0029] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0030] 1. The recognition method of the present invention designs a multi-scale time series classification model to improve the detection distance and expand the practical scope in view of the flickering phenomenon of the radiation intensity caused by the turbulence effect during flame combustion;

[0031] 2. In view of the different intensity distribution characteristics of combustion radiation sources and interference radiation sources in the infrared spectrum, a multi-layer perceptron network is used to interpret information in different bands, which significantly improves the anti-false alarm performance, improves the accuracy and reliability of judgment, and reduces the misjudgment rate;

[0032] 3. Through the saturation judgment module, different bands are adaptively selected for fire identification, which enhances the system's ability to cope with high-intensity radiation interference. When the radiation is strong and the main band is saturated, it can still be reliably identified;

[0033] 4. The use of fixed-length convolution kernels makes model training simple and fast.

[0034] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0036] Figure 1 A flowchart of an infrared multi-band intelligent fire identification method based on a time series classification model shown in one embodiment of the present invention;

[0037] Figure 2 A structural diagram of a flame feature extraction module shown in one embodiment of the present invention;

[0038] Figure 3 The block diagram of an infrared multi-band intelligent fire identification system based on a time series classification model is shown as an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0040] A specific embodiment of the present invention discloses an infrared multi-band intelligent fire identification method based on a time series classification model, such as Figure 1 As shown, the following steps are included:

[0041] S11: Collect and pre-process the ambient radiation intensity data of the main band and the auxiliary band to obtain the relative radiation intensity mean of the auxiliary band, the absolute radiation intensity mean of the main band, and the optimal band.

[0042] Specifically, a multi-band infrared sensor is used to collect the ambient infrared radiation intensity of the four bands of 4.5um, 3.9um, 5.0um and 2.7um; the collected main band and auxiliary band data are converted into digital form to obtain the time series data of each band. The main band refers to the band with higher radiation intensity and greater influence on the flame characteristics; the auxiliary band refers to the band with relatively low radiation intensity in the flame spectrum, but it can complement the main band and enhance the fire situation discrimination ability. For example, for oil and gas fires, the 4.5um band is the band with the highest flame radiation intensity during combustion, which is mainly used for pattern matching of combustion flicker characteristics and is regarded as the main band; the combustion radiation intensity of the other three bands is relatively weak, which is mainly used for interference source identification and is regarded as the auxiliary band.

[0043] The obtained time series data of each band are preprocessed, including: downsampling, background removal difference processing, exponential moving average processing and saturation state statistics for the main band and the auxiliary band respectively, and structured data that can be directly input into the subsequent module are calculated; the structured data include: the time series data of each band after preprocessing, the absolute radiation intensity mean of each band, the relative radiation intensity mean of each auxiliary band relative to the main band, and the saturation state of each band.

[0044] Specifically, since the sampling frequency of common optical sensors is above 2KHz, and the time series data features that this embodiment focuses on are caused by the turbulence effect during flame combustion, and the characteristic components are mainly concentrated between 3Hz and 25Hz, it is necessary to downsample the time series data of each band. Exemplarily, this embodiment performs an exponential moving average process of N=4 on data with a length of 4ms, and then performs a downsampling operation with a period of 16ms. The sampling frequency corresponding to the period is 62.5Hz, which complies with the Nyquist sampling theorem;

[0045] Next, the downsampled data is obtained and background difference processing is performed to reduce background interference and improve data accuracy; illustratively, the time series data of each band after background elimination is obtained by first-order difference calculation.

[0046] Furthermore, in order to quantify the signal fluctuations of the main band and the auxiliary band in the local time, the time series data of each band after eliminating the background is modulo, and the obtained modulo is processed by exponential moving average; it should be noted that because the period of the downsampling operation is 16ms, in order to obtain the estimated value of the fluctuation size of the signal of each band at the 1s time level, the equivalent length N of the exponential moving average is 64, and the obtained estimated values ​​include: the mean absolute radiation intensity of the main band and the auxiliary band. For example, the mean absolute radiation intensity of the 4.5um band is VR_IR45, the mean absolute radiation intensity of the 3.9um band is VR_IR39, the mean absolute radiation intensity of the 5.0um band is VR_IR50, and the mean absolute radiation intensity of the 2.7um band is VR_IR27.

[0047] Furthermore, since the response distribution of different infrared radiation sources in the four bands is very different, one band is selected as the main band for flame detection, and the absolute radiation intensity mean of the main band is compared with 0.1, the larger one is taken as the denominator, and the absolute radiation intensity mean of the auxiliary band is taken as the numerator, and the relative radiation intensity mean of each auxiliary band is calculated using the following formula:

[0048]

[0049] For example, for gasoline flames, 4.5um is selected as the main band for flame detection. In this embodiment, the absolute radiation intensity mean value VR_IR45 of the 4.5um band is used as the reference value, wherein the relative radiation intensity mean value of the 3.9um band is recorded as The average relative radiation intensity in the 5.0um band is recorded as The average relative radiation intensity in the 2.7um band is recorded as

[0050] Furthermore, the saturation state of each band is calculated respectively to obtain the unsaturated optimal band, including: finding the first band whose corresponding time series data is unsaturated in the order of 4.5um, 3.9um, 5.0um, and 2.7um, and considering it as the optimal band; wherein, calculating the saturation state of a band includes: calculating the median of the time series data of the corresponding band collected by the sensor in the most recent time period, exemplarily, the time period is 1 second, if the median is greater than 90% of the sensor range, the band is considered to be saturated; otherwise, it is considered to be unsaturated.

[0051] For example, for oil and gas fires, the infrared radiation intensities of the four bands are arranged from large to small as follows: 4.5um>3.9um>5.0um>2.7um. Therefore, the band sequence is used to find the optimal band.

[0052] It should be noted that when the sampling data of a band is saturated, the time series data corresponding to the band will tend to a straight line, the turbulent fluctuation information of the flame combustion will be lost, and the sensor will be blinded to the band. Therefore, it is meaningless to perform time series analysis on the saturated band. This embodiment selects the optimal unsaturated band. When one or more bands are saturated, that is, blinded, flame identification can still be performed, and it has a high resistance to blinding. It should be noted that in the process of flame combustion, there is no situation where all bands are saturated. If all bands are saturated, it is considered a false alarm.

[0053] S12: Input the data obtained after preprocessing into the time series classification model to obtain the recognition result; wherein, the time series classification model includes: a flame feature extraction module, a multi-band information fusion module, a saturation judgment module and a threshold judgment module; wherein, the flame feature extraction module is used to calculate the flame matching coefficient using the data of the optimal band; the multi-band information fusion module is used to calculate the probability of fire prediction using the flame matching coefficient, the absolute radiation intensity mean of the main band and the relative radiation intensity mean of the auxiliary band; the saturation judgment module is used to judge whether the main band is saturated; the threshold judgment module is used to use the result output by the saturation judgment module to select the flame matching coefficient or the probability of fire prediction to judge whether a fire has occurred.

[0054] Specifically, the flame feature extraction module includes a fixed multi-scale convolution kernel, a pooling layer and a Ridge binary classification module, such as Figure 2 As shown; among them,

[0055] A fixed multi-scale convolution kernel is used to extract features from the data of the optimal band;

[0056] The pooling layer is used to generate high-dimensional feature vectors using the extracted feature data;

[0057] The Ridge binary classification module is used to calculate the weight coefficient of each feature using the high-dimensional feature vector, and perform weighted summation of each feature and the corresponding weight coefficient to obtain a flame matching coefficient.

[0058] The fixed multi-scale convolution kernel includes a variety of different convolution kernels, and the weights of the convolution kernels only include two numbers, -1 and 2, and are fixed during the training and use phases of the module to adapt to various change modes of the flame waveform. , where 9 represents the convolution kernel length, of which there are 6 -1s and 3 2s; this method can cover almost all situations where fire sources and false alarm sources occur, satisfying the adaptability.

[0059] Specifically, the time series data x = [x1, x2, x3, ..., xn], where n represents the length of the sequence. For example, the sensor samples once every 16 ms, and the length of the sequence n of 1 second is 64; the convolution kernel w = [w1, w2, w3, ..., wm], where m represents the length of the convolution kernel. For example, m is 9;

[0060] The one-dimensional convolution of time series data x and convolution kernel w can be expressed as:

[0061]

[0062] Among them, t = 1, 2, ..., n-m+1; n-m+1 represents the sequence length of the convolution output.

[0063] By using convolution kernels of different lengths, the module can extract features of data sequences at different time scales. Shorter convolution kernels are responsible for extracting local features on the time axis, while longer convolution kernels can capture features in a wider range. This multi-scale convolution mechanism enables the model to fully explore the connections and differences between data at different time spans.

[0064] Furthermore, this embodiment uses the expansion factor to stretch and transform the convolution kernel on the time axis, changing the time scale of the data calculated with the convolution kernel in the convolution operation. If the expansion coefficient is i, the convolution kernel is stretched to i times the original length, and each convolution kernel template corresponds to a different number of features under each expansion coefficient. Specifically, the smaller the expansion coefficient i, the more features Mi corresponding to the expansion coefficient, and the better the effect after multi-time scale fusion. Exemplarily, this embodiment uses 8 expansion factors, considering the model performance and the processing power of the machine, and sets the total feature number of the model to 84×(M1+M2+…+M8)=9996; wherein each feature has a corresponding bias coefficient, so the dimension of the bias coefficient is 9996, and the convolution operation is performed on the time series of length 64, and the obtained feature map can be regarded as a 9996×64 matrix, and the bias coefficient will be superimposed on the feature map in a broadcast manner, and the specific bias coefficient is obtained through module training.

[0065] Furthermore, the convolved feature map is input into the pooling layer for processing, including condensing the 1×64 sequence into a scalar value, which comprehensively reflects the matching degree of the convolution kernel in the entire time span; the pooling layer outputs a high-dimensional feature vector with a dimension of 9996, and the high-dimensional feature vector encodes the waveform feature information of the corresponding input time series data.

[0066] Next, the high-dimensional feature vector is subjected to classification training, including: considering that the length of the high-dimensional feature vector is relatively large, the trained Ridge binary classification module is combined with the label information to calculate the weight coefficient of each feature, wherein the label information includes: 0 represents no fire, and 1 represents fire. Compared with the linear regression method, the Ridge binary classification method is less prone to overfitting.

[0067] It should be noted that the Ridge binary classification module is a linear regression model, which adds an L2 regularization term to the loss function to effectively prevent overfitting. The output of the trained Ridge binary classification module is the weighted sum of each feature and the corresponding weight coefficient, that is, the flame matching coefficient γ, which reflects the degree of match between the change pattern of the input time series data and the change pattern of the radiation intensity when the flame is burning.

[0068] Specifically, if the flame matching coefficient is greater than the flame matching coefficient threshold γ thr , it is considered as fire; otherwise, it is considered as no fire; the larger the γ, the closer the sensor is to the flame. In addition, due to the different emphasis on the module detection sensitivity and anti-false alarm capability according to the actual mission scenario, γ can be set as needed. thr The size of ; among them, set γ greater than 0 thr It means reducing the detection sensitivity and improving the ability to prevent false alarms. Set γ to less than 0. thr It means increasing the detection sensitivity and reducing the false alarm prevention capability.

[0069] It should be noted that, in the training process of the flame feature extraction module, since the 84 convolution kernel templates are fixed values ​​and do not require training, the expansion rules of each convolution kernel and the feature number after expansion are also generated according to the rules and do not require training, so there is no need to input a large amount of data and use the back propagation algorithm to back-calculate the weight gradient; what needs to be trained with real data is the bias coefficient, and the dimension of the bias coefficient is consistent with the dimension of the feature, and the training of the bias coefficient includes: taking the data of a certain time period obtained by random collection as input, performing a convolution operation to obtain a high-dimensional result, and then performing a random quantile calculation on the high-dimensional result, and the quantile value is uniformly randomly selected between 0% and 100% to obtain the bias coefficient; illustratively, the length of the time period is 1 second, and a high-dimensional result of 9996x64 dimensions is obtained, each feature is a time series with a length of 64, and a random quantile calculation is performed on these 64 numbers, and the quantile value is uniformly randomly selected between 0% and 100% to obtain the bias coefficient.

[0070] Next, forward inference is performed on all samples in the data set to obtain high-dimensional results, and the results are divided into training set and validation set at a ratio of 70% and 30%, and the training set and labels are input into the Ridge binary classification module for training, minimizing the cross entropy loss with L2 regularization, obtaining the weight parameters and intercept parameters of the Ridge binary classification module, and completing the training. The training operation of the flame feature extraction module is simple and time-saving.

[0071] It should be noted that the saturation judgment module is used to judge whether the main band is saturated. If the main band is saturated, the flame matching coefficient obtained by the flame feature extraction module is directly used to judge whether there is a fire; otherwise, the probability of fire estimation calculated by the multi-band information fusion module is used to judge whether there is a fire. Through the judgment of the saturation detection module, the model can adaptively select different band combinations, enhance the system's ability to cope with high-intensity radiation interference, improve the accuracy of the model, and expand the applicable scenarios of the model for fire identification.

[0072] The multi-band information fusion module includes a multi-layer perceptron network and a softmax layer; the flame matching coefficient, the absolute radiation intensity mean of the main band and the relative radiation intensity mean of the auxiliary band obtained by the flame feature extraction module are used as inputs of the multi-layer perceptron network to calculate the classification of fire and no fire; the classification of fire and no fire is input into the softmax layer to calculate the probability of fire situation estimation.

[0073] The multi-layer perceptron network structure includes: an input layer with 5 nodes, a first hidden layer with 8 nodes, a second hidden layer with 12 nodes, a third hidden layer with 8 nodes, and an output layer with 2 nodes, wherein the output layers respectively represent the classification of fire and no fire. The data output by the multi-layer perceptron is input into the softmax layer to calculate the probability of fire prediction.

[0074] In addition to receiving the results from the flame feature extraction module, the multi-band information fusion module also takes the mean absolute radiation intensity of the main band and the mean relative radiation intensity of the three auxiliary bands as input for analysis, and marks the data whose timing characteristics are very close to the flame but whose radiation intensity does not match the flame as an interference source, thereby reducing the probability of false alarms in the module.

[0075] The training of the multi-layer perceptron network includes: using dropout technology to prevent overfitting, dividing the data set into a training set and a validation set at a ratio of 70% and 30%; using the SGD optimizer to minimize the cross entropy loss function, iteratively updating the weights of the multi-layer perceptron network, and completing the training.

[0076] Specifically, the threshold judgment module includes: a flame matching coefficient threshold module and a flame probability threshold module:

[0077] The flame matching coefficient threshold module is used to compare with the flame matching coefficient. If the flame matching coefficient is greater than the flame matching coefficient threshold γ thr , it is considered as a fire;

[0078] The flame probability threshold module is used to compare with the fire prediction probability. If the fire prediction probability is greater than the flame probability threshold γ th , it is considered as a fire.

[0079] If the main band is saturated, the flame matching coefficient obtained by the flame feature extraction module is input into the flame matching coefficient threshold module to determine whether there is a fire; otherwise, the fire estimated probability obtained by the multi-band information fusion module is input into the flame probability threshold module to determine whether there is a fire.

[0080] By modifying the threshold, it is possible to adapt to different usage scenarios and meet different requirements for false positive rate and false negative rate. thr The default setting is 0, th The default setting is 50%; to increase the alarm threshold, you can thr Set to twice the standard deviation, γ th Set to 75% to reduce the false positive rate of the model but increase the false negative rate of the model.

[0081] In terms of recognition accuracy, the accuracy of this method on the self-made brazier fire / interference source data set reached 99.5%; in terms of detection distance, in the normal temperature brazier fire test, the detection distance of this method can reach more than 3 times the detection distance of the traditional detection method, greatly improving the detection distance without changing the performance of the sensor device.

[0082] Compared with the prior art, the recognition method provided in this embodiment targets the flickering phenomenon of the radiation intensity caused by the turbulence effect during flame combustion, and designs a multi-scale time series classification model to distinguish the similarity between the flickering pattern of the sampled data and the flickering pattern of the standard flame radiation, thereby improving the detection distance and expanding the practical scope; in view of the different intensity distribution characteristics of the combustion radiation source and the interference radiation source in the infrared spectrum, a multi-layer perceptron network is used to realize the interpretation of information in different bands, which significantly improves the anti-false alarm performance, improves the accuracy and reliability of the judgment, and reduces the misjudgment rate; through the saturation judgment module, different bands are adaptively selected for fire identification, which enhances the system's ability to cope with high-intensity radiation interference, and can still reliably identify when the radiation is strong and the main band is saturated; the use of a fixed-length convolution kernel makes model training simple and fast.

[0083] A specific embodiment of the present invention discloses an infrared multi-band intelligent fire identification system based on a time series classification model, such as Figure 3As shown, it is characterized by comprising a data acquisition module 31, a data preprocessing module 32 and a time series classification module 33, wherein:

[0084] The data acquisition module 31 is used to collect the environmental radiation intensity data of the main band and the auxiliary band;

[0085] The data preprocessing module 32 is used to preprocess the collected data to obtain the relative radiation intensity mean value of the auxiliary band, the absolute radiation intensity mean value of the main band, and the optimal band;

[0086] The time series classification module 33 includes a flame feature extraction module 331, a multi-band information fusion module 332, a saturation judgment module 333 and a threshold judgment module 334; wherein the flame feature extraction module 331 is used to calculate the flame matching coefficient using the data of the optimal band; the multi-band information fusion module 332 is used to calculate the probability of fire prediction using the flame matching coefficient, the absolute radiation intensity mean of the main band and the relative radiation intensity mean of the auxiliary band; the saturation judgment module 333 is used to judge whether the main band is saturated; the threshold judgment module 334 is used to select the flame matching coefficient or the probability of fire prediction to judge whether a fire has occurred using the result output by the saturation judgment module.

[0087] It should be noted that the system and method of the present invention are based on the same inventive concept, and the specific contents can be referenced from each other and will not be described repeatedly here; including data acquisition process, preprocessing process, etc.

[0088] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0089] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. An infrared multi-band intelligent fire identification method based on a time series classification model, characterized in that: The steps include: Collect and pre-process the ambient radiation intensity data of the main band and the auxiliary band to obtain the relative radiation intensity mean of the auxiliary band, the absolute radiation intensity mean of the main band, and the optimal band; The data obtained after preprocessing is input into the time series classification model to obtain the recognition result; wherein the time series classification model includes: a flame feature extraction module, a multi-band information fusion module, a saturation judgment module and a threshold judgment module; wherein the flame feature extraction module is used to calculate the flame matching coefficient using the data of the optimal band; the multi-band information fusion module is used to calculate the probability of fire prediction using the flame matching coefficient, the absolute radiation intensity mean of the main band and the relative radiation intensity mean of the auxiliary band; the saturation judgment module is used to judge whether the main band is saturated; the threshold judgment module is used to select the flame matching coefficient or the probability of fire prediction to judge whether a fire occurs using the result output by the saturation judgment module.

2. The infrared multi-band intelligent fire identification method based on time series classification model according to claim 1 is characterized in that: The flame feature extraction module includes a fixed multi-scale convolution kernel, a pooling layer and a Ridge binary classification module; wherein, A fixed multi-scale convolution kernel is used to extract features from the data of the optimal band; The pooling layer is used to generate high-dimensional feature vectors using the extracted feature data; The Ridge binary classification module is used to calculate the weight coefficient of each feature using the high-dimensional feature vector, and perform weighted summation of each feature and the corresponding weight coefficient to obtain a flame matching coefficient.

3. The infrared multi-band intelligent fire identification method based on time series classification model according to claim 2 is characterized in that: The fixed multi-scale convolution kernel includes: the length of the convolution kernel is fixed and remains unchanged during the training and use phases of the module.

4. According to the infrared multi-band intelligent fire identification method based on time series classification model according to claim 1, it is characterized in that: The multi-band information fusion module includes a multi-layer perceptron network and a softmax layer; the flame matching coefficient, the absolute radiation intensity mean of the main band and the relative radiation intensity mean of the auxiliary band obtained by the flame feature extraction module are used as inputs of the multi-layer perceptron network to calculate the classification of fire and no fire; the classification of fire and no fire is input into the softmax layer to calculate the probability of fire situation estimation.

5. The infrared multi-band intelligent fire identification method based on time series classification model according to claim 1 is characterized in that: The saturation judgment module is used to judge whether the main band is saturated. If the main band is saturated, the flame matching coefficient obtained by the flame feature extraction module is directly used to judge whether there is a fire; otherwise, the probability of fire estimation calculated by the multi-band information fusion module is used to judge whether there is a fire.

6. The infrared multi-band intelligent fire identification method based on time series classification model according to claim 1 is characterized in that: The threshold judgment module includes: a flame matching coefficient threshold module and a flame probability threshold module: The flame matching coefficient threshold module is used to compare with the flame matching coefficient. If the flame matching coefficient is greater than the flame matching coefficient threshold, it is considered that there is a fire; The flame probability threshold module is used to compare with the estimated probability of fire. If the estimated probability of fire is greater than the flame probability threshold, it is considered that there is a fire.

7. The infrared multi-band intelligent fire identification method based on time series classification model according to claim 1 or 2 is characterized in that: The optimal band is obtained by using the following steps: The saturation state of each band is calculated respectively, and the first band whose corresponding time series data is not saturated is found in the order of radiation intensity from large to small, which is regarded as the optimal band; wherein, Calculating the saturation state for a band includes: calculating the median of time series data of a band collected by the sensor in a recent time period, and if the median is greater than 90% of the sensor range, the band is considered to be saturated.

8. The infrared multi-band intelligent fire identification method based on time series classification model according to claim 1 or 4 is characterized in that: The relative radiation intensity mean value of the auxiliary band is obtained by using the following steps: The absolute radiation intensity average of each auxiliary band is calculated; The absolute radiation intensity mean of the main band is compared with 0.1, the larger one is taken as the denominator, the absolute radiation intensity mean of the auxiliary band is taken as the numerator, and the relative radiation intensity mean of each auxiliary band is calculated respectively.

9. The infrared multi-band intelligent fire identification method based on time series classification model according to claim 1 is characterized in that: The preprocessing includes: down sampling, background removal difference processing, exponential moving average processing and saturation state statistics.

10. An infrared multi-band intelligent fire identification system based on a time series classification model, characterized in that: It includes data acquisition module, data preprocessing module and time series classification module, among which: A data acquisition module, used to collect the environmental radiation intensity data of the main band and the auxiliary band; The data preprocessing module is used to preprocess the collected data to obtain the relative radiation intensity mean of the auxiliary band, the absolute radiation intensity mean of the main band, and the optimal band; The time series classification module includes a flame feature extraction module, a multi-band information fusion module, a saturation judgment module and a threshold judgment module; wherein the flame feature extraction module is used to calculate the flame matching coefficient using the data of the optimal band; the multi-band information fusion module is used to calculate the probability of fire prediction using the flame matching coefficient, the absolute radiation intensity mean of the main band and the relative radiation intensity mean of the auxiliary band; the saturation judgment module is used to judge whether the main band is saturated; the threshold judgment module is used to select the flame matching coefficient or the probability of fire prediction to judge whether a fire occurs using the result output by the saturation judgment module.

Images