Intelligent fire detection method based on acceleration signal

Abstract

The application discloses an intelligent fire detection method based on acceleration signals, static acceleration data, vibration acceleration data and combustion acceleration data are collected by an acceleration sensor in a fireproof material, and data preprocessing is performed, then data feature analysis is performed on the preprocessed acceleration signals by using signal processing technology including time domain analysis and wavelet packet decomposition technology, and reasonable feature extraction is performed on the signals, after data standardization is performed on the extracted feature data, limit gradient boosting algorithm training feature data is established, and parameter optimization is simultaneously performed to obtain a classification model with optimal accuracy, compared with a traditional method, the intelligent fire detection method can more timely distinguish a fire, can maximally reduce losses caused by the fire, and has low implementation cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fire safety early warning, and in particular to an intelligent fire detection method based on acceleration signals. BACKGROUND

[0002] Fire detection technology is an important research content in the field of fire safety. The goal of fire detection technology is to quickly and accurately detect the presence of fire and alert people to safety so that they can take appropriate action to prevent harm to themselves and others. Traditional fire detection technology usually detects the presence of fire through a single type of sensor. The problem with this method is that the sensor often needs to be close to the fire source to work, and the fire burning process needs to reach a certain level to exceed the sensor's threshold point to produce an alarm condition. This detection method provides a long time window for the spread of fire. In recent years, important breakthroughs have been made in the research of fire detection technology. With the development of artificial intelligence technology, fire detection technology has become more intelligent, automated and efficient. At present, image processing technology based on smoke recognition has a wide application in modern fire detection technology. The technology of using neural network model to recognize smoke image to detect fire (Wildland Forest Fire Smoke Detection Based on Faster R-CNN using Synthetic Smoke Images, Qi-xing ZHANG, Gao-hua LIN*, Yong-ming ZHANG, Gao XU, Jin-jun WANG, State Key Laboratory of Fire Science, University of Science and Technology of China Hefei 230026, China) has a high recognition rate, but this technology has a common defect, which is the need for a large amount of smoke in the fire scene. The production of a large amount of smoke basically indicates that the fire has reached a certain stage and cannot identify early fire signals, which may cause irreparable loss. At the same time, the implementation cost of fire detection technology based on image recognition technology is high, and the recognition accuracy will be affected in the condition of low environmental visibility. The existing method of detecting fire based on infrared thermal radiation and ultraviolet rays detects fire in real time by arranging differential thermal infrared sensors and ultraviolet sensors. This technology can detect the presence of fire in the early stage by detecting the thermal radiation produced by the fire, but due to the complexity of the technology and high implementation cost, it is not conducive to large-scale promotion.The current fire detection technology is basically divided into two directions of realizing the identification and monitoring of fire from the chemical property changes such as smoke generated by fire and the heat changes generated by fire. There is no technology for detecting fire from the change amount of physical vibration. The method for detecting fire from the physical vibration of materials has multiple advantages. First, due to the transmission of vibration, only a few acceleration sensors need to be arranged in the material to detect the vibration changes in a large range. Second, the material can generate vibration signals as soon as it contacts the heat source, which can greatly shorten the response time of the fire signal and realize the detection of the early stage of fire. Finally, the monitoring sensor used is an acceleration sensor, which is low in cost and conducive to the implementation and popularization of the technology. SUMMARY

[0003] In order to solve the problem that the existing technology cannot detect early fire in the field of fire safety warning technology, and reduce the implementation cost of fire detection technology, the present application provides an intelligent fire detection method based on acceleration signal. The acceleration data of fireproof material is collected by acceleration sensor, and then the data is preprocessed and feature extracted. Then, the acceleration data is identified and judged whether a fire occurs by using machine learning algorithm limit gradient boosting tree algorithm.

[0004] The present application is realized by at least one of the following technical solutions.

[0005] An intelligent fire detection method based on acceleration signal, comprising the following steps:

[0006] Collecting static acceleration data, vibration acceleration data and combustion acceleration data by using acceleration sensor implanted in fireproof material;

[0007] Data preprocessing is performed on the three types of acceleration data, and the preprocessing includes low-pass filtering and sliding operation to filter the collected original acceleration data;

[0008] Feature extraction is performed on the preprocessed acceleration data, and the feature extraction includes time domain feature extraction and energy feature extraction;

[0009] Training the extracted feature data by using machine learning algorithm model: normalizing the data, training the divided feature data by using limit gradient boosting tree classification model after normalizing the feature data, performing five-fold cross-validation after grid search on the parameters of the limit gradient boosting tree classification model, and outputting the confusion matrix and evaluation index column chart of the model, and selecting the fire acceleration data classification model with the best classification effect in the confusion matrix and evaluation index column chart.

[0010] Further, the static acceleration data is acceleration data collected by the acceleration sensor when the fireproof material is in a static state without external interference; the vibration acceleration data is acceleration data collected by the acceleration sensor when the fireproof material is subjected to vibration caused by human or environment; and the combustion acceleration data is acceleration data collected by the acceleration sensor when the fireproof material is subjected to vibration caused by combustion expansion.

[0011] Further, the low-pass filtering refers to low-pass filtering of the acceleration data by using a Butterworth low-pass filter, and an amplitude square function expression thereof is as follows:

[0012]

[0013] wherein, , is an order of the filter, is an angular frequency of the filter, is a cutoff frequency;

[0014] The sliding operation adopts a sliding average algorithm, and a formula thereof is as follows:

[0015]

[0016] wherein, is an actual value of data numbered , is a calculated value of data numbered , is a period of the sliding average; since the algorithm loses data for the previous -1 units, the data for the previous -1 units are offset by 1 unit, the data for the previous -1 units are taken as actual values.

[0017] Further, seven time-domain indexes are adopted to extract time-domain characteristics of the three types of preprocessed acceleration data, and the seven time-domain characteristic indexes are respectively variance, mean square value, waveform value, kurtosis value, peak-to-peak value, mean difference, and mean crossing frequency; the acceleration data is subjected to eight-layer wavelet packet decomposition, and energy values of the first eight frequency bands of the 256 frequency bands after the eight-layer decomposition are taken as energy characteristics of the acceleration data.

[0018] Further, the time-domain characteristics are extracted by taking 1000 data points as a group, and five time-domain indexes, namely, variance, mean square value, waveform value, kurtosis value, and peak-to-peak value, are adopted, wherein the variance is an average of squares of differences between each data value and an average of all data values, and a calculation formula thereof is as follows:

[0019]

[0020] wherein The total number of data, The data number, For each independent data, The average value of data;

[0021] Mean square value Also called effective value; the mean square value is obtained by squaring all data values and then taking the average of the sum , The formula is as follows:

[0022]

[0023] Waveform value Describe the degree of fluctuation of the signal, which is the quotient of the mean square value and the absolute mean value, the formula is as follows:

[0024]

[0025] Kurtosis value It is a dimensionless parameter, defined as the normalized 4th order central moment of data, the formula is as follows:

[0026]

[0027] Peak to peak value It refers to the difference between the highest value and the lowest value of a signal in a period, the formula is as follows:

[0028]

[0029] Where, The maximum value of data, The minimum value of data;

[0030] Mean difference It is obtained by calculating the difference between the mean values of the data in the two time periods of a cycle, so the absolute value of the difference between the mean values of the data in the first half of the cycle and the mean values of the data in the second half of the cycle is calculated, which can describe the mean value change of the fire signal in the continuous time period, the formula is as follows:

[0031]

[0032] The number of data in the first half of the cycle, The number of data in the second half of the cycle, The data value of the first half of the cycle with data number , The data value of the first half of the cycle with data number ;

[0033] Mean crossing times ​The formula for calculating the distribution of the fire signal with respect to the mean value in a period is as follows:

[0034]

[0035] wherein is a symbol function:

[0036] .

[0037] Further, the energy feature extraction is to take the energy value of the data after wavelet packet transformation as the feature data.

[0038] The formula for wavelet packet decomposition is as follows:

[0039]

[0040] wherein, is the wavelet packet coefficient of the original signal at the layer and the wavelet packet base; is the low-frequency wavelet packet coefficient at the layer; is the high-frequency wavelet packet coefficient at the layer; and are the low-frequency and high-frequency wavelet packet bases at the layer and the wavelet packet base, is the frequency band number of each layer of wavelet packet coefficient;

[0041] The decomposition process starts from the first layer, and the original signal is passed through low-pass and high-pass filters to obtain the low-frequency signal and the high-frequency signal Then, the low-frequency signal is decomposed by low-pass and high-pass filters to obtain the second layer low-frequency signal and the high-frequency signal This process is repeated until the wavelet packet decomposition coefficient of the signal in different frequency bands is obtained at the last layer. The energy of each frequency band at the last layer is calculated by the following formula:

[0042]

[0043] wherein is the wavelet packet coefficient of the original signal after wavelet packet decomposition at the last layer, with the frequency band number is the corresponding energy value calculated from

[0044] ​​Further, the three types of acceleration data are decomposed by db2 wavelet base for 8-layer wavelet packet decomposition, and a 256xT dimensional energy distribution matrix is obtained at the 8th layer of the wavelet packet decomposition tree, where 256 represents the frequency band number of the bottom layer node of the 8-layer wavelet packet tree, and T represents the data amount; the energy values of the first eight frequency bands in the 256 frequency bands are taken as the final energy feature data, and finally fifteen-dimensional feature data of the three types of acceleration data are obtained, including seven time domain indexes and eight frequency band energy values.

[0045] Further, the data standardization is to normalize the feature data of the three types of acceleration data de-meaning scaling, so that the data obeys a normal distribution with a mean of 0 and a variance of 1, and the calculation formula is as follows:

[0046]

[0047] wherein is the normalized data, is the average value of the data, is the standard deviation of the data.

[0048] Further, the grid search adjustment parameter refers to dividing the value list of the number of weak evaluators n_estimators, the data proportion subsample, and the learning rate learning_rate in the extreme gradient boosting tree classification model into several groups, then randomly selecting three values from the list in the three groups as the input values of the three parameters of the extreme gradient boosting tree classification model, then inputting the fifteen-dimensional feature vector training set into the model for training, and then performing five-fold cross-validation, i.e., randomly dividing the training data into five parts, and sequentially selecting one part as a validation set for model evaluation, and calculating the average value of the accuracy of the five validation results as the model evaluation index under the parameter combination, and finally traversing all the values in the three groups to obtain the best combination of the model classification effect.

[0049] Further, the fire acceleration data classification model is obtained by analyzing the confusion matrix and four evaluation indexes of the extreme gradient boosting tree classification model, and the four evaluation indexes are accuracy, precision, recall, and F1 score, where T (True) represents correct, F (False) represents error, P (Positive) represents classification result 1, N (Negative) represents classification result 0, represents the amount of data classified correctly as result 1, represents the amount of data classified correctly as result 0, represents the amount of data classified incorrectly as result 1, represents the amount of data classified incorrectly as result 0; the accuracy To predict the percentage of correct results out of the total data, the formula is as follows:

[0050]

[0051] Precision is the probability of actual positive data among all data predicted to be positive , the formula is as follows:

[0052]

[0053] Recall is the probability of data predicted to be positive among actual positive data , the formula is as follows:

[0054]

[0055] The F1 Score is a comprehensive index for evaluating the balance between the two , the formula is as follows:

[0056] .

[0057] Further, the acceleration sensor is MPU6050, which is implanted in the fireproof material to convert the acceleration signal of the fireproof material into an electrical signal, and communicates with the single-chip microcomputer through the IIC communication protocol. The single-chip microcomputer converts the electrical signal of the acceleration sensor into an acceleration value and saves the acceleration data through a computer. The sampling frequency of the acceleration sensor is set to 500Hz.

[0058] Further, the fire detection method uses a single-chip microcomputer development board with model STM32F103RCT6. The development board is equipped with an acceleration acquisition sensor, a buzzer, and a red LED lamp. The acceleration acquisition sensor is used to collect acceleration data of the fireproof material in real time, and the buzzer and the red LED lamp are used as alarm devices.

[0059] The specific process of the fire detection method is as follows:

[0060] First step: With 1000 data points as a detection period, the acceleration sensor in the single-chip microcomputer collects 1000 acceleration data points after each detection period, and then processes the acceleration data;

[0061] Second step: The single-chip microcomputer chip pre-processes and extracts features from the data;

[0062] Third step: After standardizing the feature data, input the data into a classification algorithm model;

[0063] Fourth step: when the classification model identifies as non-fire acceleration data, enter the next detection period; when the model identifies as fire acceleration data, start the buzzer and red LED light for fire alarm and enter the next detection period at the same time until a certain detection period classification model judges as non-fire signal.

[0064] Compared with the prior art, the present application has the beneficial effects that:

[0065] The present application realizes the detection of fire signal by collecting vibration signal of fireproof material through acceleration sensor, so that the acceleration signal change caused by the expansion of fireproof material due to the emergence of fire signal can be recognized by the intelligent fire detection system. Since the physical vibration change of fireproof material is monitored, the reaction time of fire detection is greatly shortened, so that there is sufficient time to maximize the protection of property and personal safety. At the same time, since the acceleration sensor and the composition hardware of the integrated intelligent fire detection system relied on by the present application are low in price and easy to obtain, the present application has great application value and popularization prospect. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 A flow chart of the intelligent fire detection method based on acceleration signal according to the present application;

[0067] Figure 2 A confusion matrix diagram of the extreme gradient boosting tree classification model according to the present application;

[0068] Figure 3 A four-type evaluation index columnar diagram of the extreme gradient boosting tree classification model according to the present application;

[0069] Figure 4 A fire detection system experimental effect diagram according to the present application. DETAILED DESCRIPTION

[0070] In order to enable the personnel in the art to better understand the present application scheme, the present application will be further described in detail below in combination with the drawings and specific embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by the person skilled in the art without creative labor fall within the scope of protection of the present application.

[0071] As shown in a kind of intelligent fire detection method based on acceleration signal according to the present application, it includes the following steps: Figure 1

[0072] MPU6050 acceleration sensor is implanted into fireproof material to collect static acceleration data, vibration acceleration data and combustion acceleration data;

[0073] ​The aforementioned accelerometer is an MPU6050, which is embedded inside the fireproof material. It converts the acceleration signal of the fireproof material into an electrical signal and communicates with the microcontroller via the IIC communication protocol. The microcontroller converts the electrical signal of the accelerometer into an acceleration value and saves the acceleration data through a connection with a computer. The sampling frequency of the accelerometer is set to 500Hz.

[0074] The static data refers to the acceleration data collected by the accelerometer when the fireproof material is in a static state without external interference; the vibration data refers to the acceleration data collected by the accelerometer when the material is subjected to vibration caused by human or environmental factors; and the combustion data refers to the acceleration data collected by the accelerometer when the fireproof material vibrates due to the expansion caused by combustion.

[0075] The three types of acceleration data collected were preprocessed, including low-pass filtering and moving average filtering.

[0076] Step 1: Low-pass filtering. The acceleration data is low-pass filtered using a Butterworth low-pass filter. The expression for its squared amplitude function is as follows:

[0077]

[0078] in, , Let the order be the filter order. The angular frequency of the filter. The cutoff frequency is the frequency at which the amplitude decreases. The frequency of time. As a preferred embodiment, setting... =100Hz, order selected as 5th;

[0079] Step 2: Moving average filtering. The moving average algorithm formula is as follows:

[0080]

[0081] in, The data is numbered as The actual value of the data. The data is numbered as Calculated data values This refers to the number of periods in the moving average; because this algorithm will lose previous periods... -1 data point, generating data -1 unit offset, therefore the previous -1 data point is taken as the actual value of the data. As a preferred embodiment, this is set... = 20.

[0082] Next, feature extraction is performed on the preprocessed acceleration data, including time-domain feature extraction and energy feature extraction.

[0083]

[0084] where is the total number of data, is the data number, is each independent data, is the average value of data.

[0085] The mean square value is also called the effective value; the mean square value is obtained by squaring all data values, summing them up, and then taking the mean value , and the calculation formula is as follows:

[0086]

[0087] Waveform value The waveform value describes the fluctuation degree of the signal, which is the quotient of the mean square value and the absolute mean value, and the calculation formula is as follows:

[0088]

[0089] Kurtosis value The kurtosis value is a dimensionless parameter, which is defined as the normalized 4th central moment of data, and the calculation formula is as follows:

[0090]

[0091] Peak-to-peak value The peak-to-peak value refers to the difference between the highest value and the lowest value of the signal in a period, and the calculation formula is as follows:

[0092]

[0093] is the maximum value of data, is the minimum value of data.

[0094] Mean value difference The mean value difference is obtained by calculating the difference between the mean values of the data in the two time periods before and after a period, so the absolute value of the difference between the mean value of the data in the first half of the period and the mean value of the data in the second half of the period is calculated, which can describe the mean value change of the fire signal in the continuous time period, and the calculation formula is as follows:

[0095]

[0096] is the data quantity in the first half of the period, is the data quantity in the second half of the period, is the data value of the data numbered in the first half of the period, is the data value of the data numbered in the first half of the period.

[0097] Mean crossing number The formula for calculating the mean distribution of fire signal in a cycle is as follows:

[0098]

[0099] Wherein is a symbol function:

[0100]

[0101] Energy feature extraction, the energy value of the acceleration data after wavelet packet transformation is used as the feature data, the three types of acceleration data are decomposed by 8 layers of wavelet packet with "db2" wavelet basis, a 256xT dimensional energy distribution matrix is obtained at the 8th layer of the wavelet packet decomposition tree, 256 represents the frequency band number of the bottom layer node of the 8-layer wavelet packet tree, and T represents the data amount. The energy of the first eight frequency bands in 256 frequency bands is used as the final energy feature data.

[0102] The formula for wavelet packet decomposition is:

[0103]

[0104] Wherein, is the wavelet packet coefficient of the original signal at the layer, the th wavelet packet basis; is the low-frequency wavelet packet coefficient at the layer; is the high-frequency wavelet packet coefficient at the layer; and are the low-frequency and high-frequency wavelet packet bases at the layer, the th wavelet packet basis, is the frequency band number of each layer of wavelet packet coefficient.

[0105] The decomposition process starts from the first layer, and the original signal is passed through low-pass and high-pass filters to obtain low-frequency signal and high-frequency signal . Then the low-frequency signal is decomposed by low-pass and high-pass filters to obtain the second layer low-frequency signal and high-frequency signal . Repeat this process until the last layer of wavelet packet decomposition coefficient of the signal in different frequency bands is obtained. Taking three-layer wavelet packet as an example, the energy of each frequency band in the last layer is calculated by the following calculation formula:

[0106]

[0107] the wavelet packet coefficient of the frequency band number obtained after the last layer of wavelet packet decomposition of the original signal, the corresponding energy value calculated by .

[0108] Finally, fifteen-dimensional feature data of three types of acceleration data can be obtained, including seven time domain indicators and eight energy values of frequency bands.

[0109] Then, the feature data is trained by a machine learning algorithm model.

[0110] S1, data standardization: data standardization is a kind of data dimensionless, the feature data of three types of acceleration data is scaled by the standard deviation after mean removal, so that the data obeys the normal distribution with mean 0 and variance 1 (standard state distribution), and the calculation formula is as follows:

[0111]

[0112] wherein is the standardized data, is the average value of the data, is the standard deviation of the data

[0113] S2, model training: first, adjust the parameter value list of the number of weak estimators n_estimators, the proportion of randomly sampled data subsample, and the learning rate learning_rate in the extreme gradient boosting tree classification algorithm, and divide it into several groups, then randomly select three values from the list in the three groups as the input values of the three parameters of the extreme gradient boosting tree classification model, then input the fifteen-dimensional feature vector training set into the model for training, and then randomly divide the thirteen-dimensional feature data into five parts, and select one part as the validation set in turn, until all the feature data is taken five-fold cross-validation, and the average value of the confusion matrix and evaluation index bar chart of the model under each group of model parameters is output, as shown in Figure 3 .

[0114] The model training is to train the feature data by using the extreme gradient boosting tree classification algorithm model; the extreme gradient boosting tree classification algorithm model is built by using the extreme gradient boosting tree classification algorithm model “XGBClassifier” method in the “xgboost” library in Python, and the five-fold cross-validation is realized by using the “cross_val_score” method in the “sklearn.model_selection” library in Python.

[0115] S3, select the fire acceleration data classification model with the best classification effect in the confusion matrix and evaluation index column chart;

[0116] The best classification effect model is obtained by analyzing the confusion matrix and four evaluation indexes of the model. The four evaluation indexes are accuracy, precision, recall and F1 score, T (True) represents correct, F (False) represents error, P (Positive) represents classification result 1, and N (Negative) represents classification result 0. representing the amount of data classified correctly for result 1, representing the amount of data classified correctly for result 0, representing the amount of data classified incorrectly for result 1, representing the amount of data classified incorrectly for result 0; accuracy The accuracy is the percentage of correct prediction results in the total data amount, and the calculation formula is as follows:

[0117]

[0118] The precision is the probability of actual positive data in all predicted positive data, and the calculation formula is as follows:

[0119]

[0120] The recall is the probability of being predicted as positive data in the actual positive data, and the calculation formula is as follows:

[0121]

[0122] The F1 score is a comprehensive index for considering the balance of the two, and the calculation formula is as follows:

[0123] .

[0124] The system for realizing the intelligent fire detection method based on acceleration signal comprises a single-chip microcomputer development board, an acceleration acquisition sensor, a buzzer and a red LED lamp arranged on the development board. The acceleration sensor is used to acquire real-time acceleration data of fireproof materials, the buzzer and the red LED lamp are used as alarm devices, and the data preprocessing code, the data feature extraction code and the data standardization C language code are written and burned into the single-chip microcomputer together with the best classification model code to detect the actual acceleration data.

[0125] As a preferred embodiment, the single-chip microcomputer is an STM32F103RCT6 single-chip microcomputer development board.

[0126] The specific step flow of the fire detection method comprises:

[0127] (1) take 1000 data points as a detection cycle, the acceleration sensor in the single-chip microcomputer collects 1000 acceleration data points after each detection cycle, and processes the acceleration data;

[0128] (2) the single-chip microcomputer chip pre-processes and extracts features from data;

[0129] (3) the feature data is standardized and input into a fire acceleration data classification model;

[0130] (4) when the classification model identifies non-fire acceleration data, enter the next detection cycle; when the model identifies fire acceleration data, start the buzzer and red LED light for fire alarm and simultaneously enter the next detection cycle until the classification model judges non-fire signal in a certain detection cycle, and the experimental result graph is shown in Figure 4

[0131] The preferred embodiments of the present application disclosed above are only used to help explain the present application. The preferred embodiments do not describe all the details and limit the present application to the specific embodiments described. Obviously, according to the content of the present application, many modifications and changes can be made. The present application selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present application, so that those skilled in the art can well understand and use the present application. The present application is limited by the claims and their entire scope and equivalents.​

Claims

1. A smart fire detection method based on acceleration signals, characterized in that, The method comprises the following steps: Collect static acceleration data, vibration acceleration data and combustion acceleration data by using an acceleration sensor implanted with fireproof material; the static acceleration data is acceleration data collected by the acceleration sensor when the fireproof material is in a static state without external interference; the vibration acceleration data is acceleration data collected by the acceleration sensor when the fireproof material is subjected to vibration caused by human or environment; the combustion acceleration data is acceleration data collected by the acceleration sensor when the fireproof material is subjected to vibration caused by combustion expansion; Perform 8-layer wavelet packet decomposition on the three types of acceleration data by using a db2 wavelet basis, and obtain a 256xT-dimensional energy distribution matrix at the 8th layer of the wavelet packet decomposition tree, wherein 256 represents the frequency band number of the bottom layer node of the 8-layer wavelet packet tree, and T represents the data amount; take the energy values of the first eight frequency bands in the 256 frequency bands as the final energy feature data, and finally obtain fifteen-dimensional feature data of the three types of acceleration data, including seven time domain indexes and the energy values of the eight frequency bands; Perform time domain feature extraction on the preprocessed three types of acceleration data by using seven time domain indexes, and the seven time domain feature indexes are variance, mean square value, waveform value, kurtosis value, peak-to-peak value, mean difference and mean crossing frequency; perform eight-layer wavelet packet decomposition on the acceleration data, and take the energy values of the first eight frequency bands of the 256 frequency bands after eight-layer decomposition as the energy features of the acceleration data; Perform data preprocessing on the three types of acceleration data, and the preprocessing comprises low-pass filtering and sliding operation; Perform feature extraction on the preprocessed acceleration data, and the feature extraction comprises time domain feature extraction and energy feature extraction; Train a machine learning algorithm model by using the extracted feature data: normalize the data, train the limit gradient boosting tree classification model after normalizing the feature data, perform grid search on the parameters of the limit gradient boosting tree classification model, and then perform five-fold cross-validation, output the confusion matrix and evaluation index column chart of the model, and select the fire acceleration data classification model with the best classification effect in the confusion matrix and evaluation index column chart.

2. The intelligent fire detection method based on acceleration signals according to claim 1, wherein, The low-pass filtering refers to low-pass filtering the acceleration data by using a Butterworth low-pass filter, and the amplitude square function expression is as follows: wherein , is the order of the filter, is the corner frequency of the filter, is the cutoff frequency; The sliding operation adopts a sliding average algorithm, and the formula is as follows: Wherein, is the data actual value of data numbered , is the data calculated value of data numbered , is the period of the moving average; since the algorithm loses the previous -1 data, there is a -1 unit offset to the data, so the previous -1 data takes the data actual value.

3. The intelligent fire detection method based on acceleration signal according to claim 1, characterized in that, The energy feature extraction takes the energy values of the data after wavelet packet transformation as the feature data; The formula of the wavelet packet decomposition is as follows: wherein, is the wavelet packet coefficient of the original signal at the layer, the th wavelet packet basis; is the low frequency wavelet packet coefficient at the layer; is the high frequency wavelet packet coefficient at the layer; and are the low frequency and high frequency wavelet packet basis at the layer, the th wavelet packet basis, respectively, is the frequency band number of each layer wavelet packet coefficient. The decomposition process starts from the first layer, the original signal is passed through low-pass and high-pass filters to obtain low-frequency signal and high-frequency signal Then the low-frequency signal is decomposed by low-pass and high-pass filters to obtain second layer low-frequency signal and high-frequency signal This process is repeated until the last layer is decomposed to obtain wavelet packet decomposition coefficients of the signal in different frequency bands. The energy of each frequency band in the last layer is calculated by the following formula: wherein is the frequency band number obtained after the last wavelet packet decomposition of the original signal, is the wavelet packet coefficient of the frequency band, is the corresponding energy value calculated from .

4. The intelligent fire detection method based on acceleration signal according to claim 1, characterized in that, The data standardization is to scale the feature data of the three types of acceleration data After mean removal, standard deviation Scaling, making the data conform to the normal distribution with mean 0 and variance 1, the calculation formula is as follows: wherein is the standardized data, is the average value of the data, is the standard deviation of the data.

Images