A point-type smoke detection method and device for precise classification of fire smoke

Abstract

The application discloses a point type smoke detection method and device for accurate classification of fire smoke, and belongs to the technical field of multi-dimensional perception and intelligent identification of initial fire information. The method comprises the following steps: S1, preprocessing of acquired detection data; S2, two-dimensional recombination of N period preprocessing data corresponding to the number N of detector channels, and line convolution and column convolution of the recombined two-dimensional data respectively, so as to extract a time feature vector and a space feature vector of the detection data; S3, cross multiplication of the two vectors to obtain a two-dimensional time-space interlaced feature; input of the two-dimensional time-space interlaced feature into a multi-layer convolutional neural network, and output of a one-dimensional expansion to obtain a classification feature vector, and then input of the classification feature vector into a fully connected network for classification. In this way, the application can accurately classify fire smoke under fewer channels, and can divide fire smoke and non-fire smoke.

Description

Technical Field

[0001] This invention belongs to the field of multi-dimensional perception and intelligent identification technology of early fire information, and more specifically, relates to a point-type smoke detection method and device for accurate classification of fire smoke. Background Technology

[0002] Point-type smoke detectors are an important component of fire alarm systems. They primarily detect smoke and provide early warning of fires by responding to solid particles generated by combustion or heat. They are suitable for locations where fires are in the early stages of smoldering and play a vital role in protecting people's lives and property. Based on different response principles, point-type smoke detectors can be further classified into photoelectric, ionization, and pyroelectric particle types. Different types of detectors have different application scenarios, among which optical scattering-based point-type smoke detectors are the most widely used and common.

[0003] In the area of ​​multi-dimensional sensing and intelligent identification technology for early-stage fire information, the billions of point-type fire detectors currently widely used in large complexes and high-rise / super high-rise buildings primarily employ light scattering smoke particle detection technology. This technology boasts advantages such as low cost, small size, and light weight. However, existing point-type smoke detectors are highly susceptible to triggering false alarms when interfered with by non-fire aerosols (such as dust, water vapor, and cooking fumes). This is because existing detectors mainly focus on detecting the presence of fire smoke particles and cannot accurately classify the types of smoke particles. Therefore, there is an urgent need to develop a new generation of anti-interference fire detection technology to improve the accuracy and reliability of point-type smoke detection devices. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a point-type smoke detection method and device for accurate classification of fire smoke, aiming to solve the problem of high false alarm rate of current point-type smoke detectors based on optical scattering.

[0005] To achieve the above objectives, according to a first aspect of the present invention, a point-type smoke detection method for accurate classification of fire smoke is provided, comprising:

[0006] S1, preprocess the acquired detection data;

[0007] S2, take N-cycle preprocessed data corresponding to the number of detector channels N, perform two-dimensional reconstruction, and perform row convolution and column convolution on the reconstructed two-dimensional data respectively to extract the temporal feature vector and spatial feature vector of the detection data.

[0008] S3. Perform a cross product on the two vectors to obtain a two-dimensional spatiotemporal interleaved feature. Input the two-dimensional spatiotemporal interleaved feature into a multi-layer convolutional neural network, and expand the output into one dimension to obtain a classification feature vector. Then, input the classification feature vector into a fully connected network for classification.

[0009] Furthermore, the method also includes:

[0010] S4, calculate the cosine similarity between the classification feature vector and the specimen vector. If the cosine similarity is greater than a preset threshold, the classification result is determined to be valid; otherwise, the corresponding probe data is labeled as unlabeled data. The specimen vector is the mean of the classification feature vectors of each category when correctly inferring from the training data of all known categories under the current network parameters.

[0011] Furthermore, in step S1, the detection data is obtained in the following manner:

[0012] M laser emitters are fixed on a circular plane. With the center of the circular plane as the origin, N photodetectors are installed at N different angles within a 180° clockwise or counterclockwise range. A housing with smoke inlet is also made.

[0013] During measurement, the MCU controls M laser emitters to turn on sequentially, the photodetector converts the collected light signal into an electrical signal output, the analog-to-digital converter converts the electrical signal into a digital value within the measurement range, and the microcontroller controls the multiplexer to acquire the data of each channel in sequence and send it to the data receiving end via serial port.

[0014] Furthermore, in S1, the preprocessing includes: mean filtering, background removal, and setting a threshold.

[0015] Further, S2 includes:

[0016] Take N preprocessed data corresponding to the number of detector channels N, and reassemble them into a two-dimensional array according to the time × space dimension combination method. The reassembled data dimension is [N,N].

[0017] Construct two types of convolutional kernels with sizes [1,N] and [N,1], and perform row convolution and column convolution on the two-dimensional reconstructed data with a stride of 1 to obtain the time feature vectors t = [T1 T2 … T N ], and spatial feature vectors s = [S1 S2 …S N ].

[0018] Furthermore, in S3, the two-dimensional spatiotemporal interleaving feature is represented as follows:

[0019]

[0020] Among them, TS NN T represents N ×S N .

[0021] Furthermore, in S3, the multi-layer convolutional neural network is a two-layer convolutional neural network or a three-layer convolutional neural network; wherein, the single-layer convolutional neural network includes a convolutional layer, a normalization layer and an activation layer, and the activation function adopts a linear rectified function.

[0022] Secondly, the present invention provides a point-type smoke detection device for accurate classification of fire smoke, comprising:

[0023] The preprocessing module is used to preprocess the acquired detection data;

[0024] The two-dimensional reconstruction module is used to take N-cycle preprocessed data corresponding to the number of detector channels N, reconstruct them in two dimensions, and perform row convolution and column convolution on the reconstructed two-dimensional data to extract the temporal feature vector and spatial feature vector of the detection data.

[0025] The classification module is used to perform a cross product of two vectors to obtain two-dimensional spatiotemporal interleaved features. The two-dimensional spatiotemporal interleaved features are input into a multi-layer convolutional neural network, and the output is expanded into one dimension to obtain a classification feature vector. The classification feature vector is then input into a fully connected network for classification.

[0026] Furthermore, the device also includes:

[0027] The verification module is used to calculate the cosine similarity between the classification feature vector and the specimen vector. If the cosine similarity is greater than a preset threshold, the classification result is determined to be valid; otherwise, the corresponding probe data is labeled as unlabeled data. The specimen vector is the mean of the classification feature vectors of each category when correctly inferring from the training data of all known categories under the current network parameters.

[0028] Thirdly, the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method as described in the first aspect.

[0029] Fourthly, the present invention provides a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the method as described in the first aspect.

[0030] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0031] (1) This invention is aimed at point-type smoke detectors based on the principle of optical scattering, which can achieve high classification accuracy with fewer channels and save hardware costs.

[0032] (2) The convolutional neural network used has few parameters, low computational cost, easy model quantification, simple and practical method, easy to deploy on embedded systems with low computing power, and easy to modify according to actual scenarios;

[0033] (3) Applying artificial intelligence methods to point-type smoke detection systems can effectively solve the problem of false alarm rate that is common in current point-type smoke detectors, and it also has strong versatility for other point-type measurement devices. Attached Figure Description

[0034] Figure 1 This is a flowchart of a point-type smoke detection method for accurate classification of fire smoke provided by the present invention;

[0035] Figure 2 This is a neural network structure diagram for fire smoke classification provided by the present invention;

[0036] Figure 3 This is a schematic diagram of a reasoning structure for distinguishing between labeled and unlabeled samples provided by the present invention;

[0037] Figure 4 This is a plan view of a point-type smoke detector for accurate classification of fire smoke provided by the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0039] In this invention, the terms "first," "second," etc. (if present) in the invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0040] Considering that smoke particles produced by the combustion of different substances have different physical properties, their optical scattering characteristics also vary, which provides the possibility to distinguish these smoke particles. Combined with the wide application of artificial intelligence nowadays, artificial intelligence algorithms have strong advantages in obtaining data differences and distinguishing sample categories. Therefore, applying artificial intelligence algorithms to fire smoke detection scenarios undoubtedly has broad development space and application prospects. Based on this, the present invention provides a point-type smoke detection method for accurate classification of fire smoke. Different types of smoke data are obtained by burning different substances, and after processing the data, a convolutional neural network is built to achieve accurate classification.

[0041] Example 1:

[0042] This embodiment provides a point-type smoke detection method for accurate classification of fire smoke, as shown in Figure 1 , the method comprises: (S1) performing preprocessing operations such as mean filtering, background removal, and threshold setting on detector data; (S2) performing two-dimensional reorganization on N-period data corresponding to the number N of detector channels, and performing row convolution and column convolution respectively on the reorganized two-dimensional data to extract time feature vectors and spatial feature vectors of the data; (S3) performing cross product on the two types of vectors to obtain two-dimensional spatiotemporal interleaved features, constructing a multi-layer convolutional neural network and a one-layer fully connected network, inputting the above two-dimensional spatiotemporal interleaved features, obtaining classification feature vectors after expanding the output one-dimensionally, and then sending the vectors to the fully connected layer for classification; (S4) processing data and training the network according to the above method, quantizing parameters after the network is fitted; (S5) during prediction, on one hand, the classification feature vector is transmitted to the fully connected layer, and the final prediction result is obtained after normalization processing; on the other hand, the cosine similarity between the classification feature vector and the sample vector is calculated to assist in determining the prediction result.

[0043] (S1) Specifically, the data obtained by the detector contains noise and invalid data (natural state), which needs to be processed before actual classification, such as mean filtering, background removal, threshold setting and other preprocessing operations. The specific method is as follows:

[0044] For mean filtering, assuming that the data collected by a single sensor is d i , i=1,2,3, ……n, the number of filtering points is selected as 2s+1, s=1,2,3...... When i<s+1 or i>n-s, no processing is performed on the data; when s+1≤i≤n-s, the mean value of this data and s data before and after it is calculated, and each data is smoothed successively. The mean filtering formula is as follows:

[0045]

[0046] The background value refers to the detector's measurement of ambient air when no smoke is generated. When smoke appears, the detector's measurement value gradually increases. Based on this, the average value of the detector's measurements of natural air over a certain time T is calculated, while a smoke threshold is set. When the measurement value exceeds the threshold, smoke is considered to have occurred. The change data is obtained by subtracting the background value from the measured data. Assuming smoke occurs at point t (t>20), the calculation formula for this part is as follows:

[0047]

[0048] For example, five substances—cotton rope, beech wood, polyurethane, n-heptane, and pork chops (cooking fumes)—were burned in a confined environment. Data was collected using a four-channel smoke detector based on optical scattering, and the data was subjected to mean filtering and background removal. Because the particle sizes formed after the combustion of different substances are different, the scattering intensity varies at different angles, and the light intensity values ​​captured by the photodetector are different, thus enabling the differentiation of different substances.

[0049] (S2) Specifically, assuming the detector has N channels, meaning there are N photoelectric sensors arranged in space according to certain rules, there are spatial characteristics among the N sets of data. For a single channel, the sensor data also changes with the air environment in the discrete time dimension, thus single-channel data has temporal characteristics. To achieve adaptive extraction of spatial and temporal characteristics, the following method is adopted:

[0050] Take N periods of data with the same number of channels, and reassemble them into a two-dimensional array according to the time × space dimension combination. The reassembled data dimension is [N,N].

[0051] Construct two types of convolution kernels with sizes [1,N] and [N,1], and perform row convolution and column convolution on the above two-dimensional reconstructed data with a stride of 1 to obtain the time feature vectors t=[T1T2…T N ], and spatial eigenvectors s = [S1S2…S N Because features are extracted using convolutional kernels, parameters can be automatically fitted during subsequent training.

[0052] In particular, when performing two-dimensional data reconstruction, the length of the selected time dimension does not necessarily have to match the number of channels; it can be arranged according to the actual measurement cycle and computing resources. When the measurement cycle is short or computing resources are sufficient, it can be appropriately extended, and vice versa.

[0053] (S3) Specifically, after obtaining the temporal and spatial feature vectors, the following method is used to further fuse and extract features for classification:

[0054] Performing a cross product on the two feature vectors above yields a spatiotemporally interleaved feature with dimensions [N, N], as shown in the following formula:

[0055]

[0056] This feature re-integrates the spatiotemporal characteristics of the detector data, containing deeper information compared to the original data.

[0057] Subsequently, a multi-layer convolutional neural network and a single fully connected network are constructed. Each convolutional neural network consists of a convolutional layer, a normalization layer, and a ReLU activation layer. The parameters of the convolutional layer are as follows: Figure 2 As shown. Due to the small feature size, it is better to build a neural network with two to three layers, and the kernel size can be [3,3] or [2,2]. After obtaining the output of the neural network, the output is expanded in one dimension to obtain the classification feature vector, and this vector is fed into a fully connected network for classification. The data is processed and the network is trained in the order of (S1) to (S3) above. After the network is fitted, the subsequent steps are performed.

[0058] (S4) Specifically, to achieve deployment and accelerate network computing speed, the network is quantized here, and the specific method is as follows:

[0059] A single-layer convolutional neural network consists of a Conv layer, a BatchNorm layer, and a ReLU layer, and the calculation formulas for these three layers can be fused. Assuming the input is x, the calculation formula for the Conv layer is as follows:

[0060] y = w conv x+b conv

[0061] Among them, w conv That is, the convolution kernel parameter, b conv This indicates the bias. After the convolution operation, the output is fed into the BactchNorm layer, calculated as follows:

[0062]

[0063] Where, μ bn and σ bn 2 ε and β represent the mean and variance of the data, respectively; ε is extremely small to prevent the denominator from being zero; β represents the scaling factor; and B represents the bias. Decomposing the above formula yields the following formula:

[0064] z = w bn y+b bn

[0065] in,

[0066] From the above formula, we can see that the calculation of the BatchNorm layer is equivalent to performing a convolution on the input with a 1*1 kernel and adding a bias. Combining this with the ReLU calculation formula: f(x)=max(0,x), we finally obtain the fusion formula:

[0067] f(x) = max(0, w fu x+b fu )

[0068] Among them, w fu =w conv *w bn b fu =w bn *b conv +b bn

[0069] Quantization refers to converting the original 32-bit floating-point arithmetic operations into 8-bit operations. Static quantization is used here, and the calculation formula is as follows:

[0070]

[0071] Here, Input represents the input data, Point represents the distance between data centers, and Scale represents the quantization scale. After the above fusion is completed, the model parameters are statically quantized using the above formula, and a small amount of data and labels are fed in for training to determine the values ​​of the center points. Furthermore, the model input data is quantized, and dequantization is performed after output to prevent data type mismatch errors. This significantly reduces the model size and improves computation speed.

[0072] After quantization, a forward inference is performed on the existing dataset to retain the classification feature vectors of all correctly predicted samples and calculate the mean of the classification feature vectors of each category, resulting in 5 sets of sample vectors for subsequent similarity matching.

[0073] (S5) Specifically, after obtaining the classification feature vector, the classification feature vector can be fed into the classifier for prediction, such as... Figure 3 As shown, the specific method is as follows:

[0074] On one hand, the classification feature vector is fed into a fully connected network, and the output value is converted into a probability distribution of 0 to 1 using the Softmax normalization function. The Softmax calculation formula is as follows:

[0075]

[0076] On the other hand, when unlabeled samples appear, the classifier cannot determine their specific category. In this case, the cosine similarity between the classification feature vector of the sample and the sample vector is calculated, and a decision threshold is set to assist in determining the category of the sample data. The formula for calculating cosine similarity is as follows:

[0077]

[0078] The cosine similarity value ranges from -1 to 1. Assuming the calculated result is C and the decision threshold is Y, the above prediction result is valid when C > Y; otherwise, the result is invalid, and the sample is labeled as unlabeled data.

[0079] For a sample vector, it refers to the mean of the classification feature vectors of each category when correctly inferring from training data of all known categories under the current network parameters. When the number of training categories in the convolutional neural network is K, after the network completes the fitting, a forward prediction is performed on all samples, the classification feature vectors of the correctly predicted samples are retained, and then the mean of the classification feature vectors of each category is calculated. A total of K sets of vectors are used as samples.

[0080] Example 2:

[0081] This invention also provides a point-type smoke detector for accurate classification of fire smoke. The detector mainly consists of M laser emitters, N photoelectric detectors, N analog-to-digital converters, multiple multiplexers, and an MCU.

[0082] The laser emitter is fixed on a circular plane. With the center of the plane as the origin (0), N different angles are selected within a 180° clockwise or counterclockwise range (clockwise and counterclockwise angles are equivalent) to install photodetectors. A black casing with smoke inlets is fabricated to enclose the measurement system and reduce ambient light interference. For example,... Figure 4 As shown, the detector mainly consists of a laser emitter, four photodetectors, four analog-to-digital converters, a multiplexer, and an MCU. The photodetectors are installed at four positions: 15° and 30° clockwise, and 45° and 60° counterclockwise.

[0083] During measurement, the MCU controls the laser emitters to turn on sequentially. At this point, one laser emitter, one photodetector, and one analog-to-digital converter constitute a simplified channel, resulting in a total of M*N channels. The photodetector converts the acquired optical signal into an electrical signal for output, and the analog-to-digital converter converts this electrical signal into a digital value within the measurement range. Subsequently, the MCU controls a multiplexer to sequentially acquire data from each channel and transmits it to the data receiving end via serial port. The data receiving end can be programmed with a corresponding serial port receiving program to save the measurement data for each measurement, used for data processing and model training in Embodiment 1 of this invention.

[0084] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A point-type smoke detection method for accurate classification of fire smoke, characterized in that, include: S1, preprocess the acquired detection data; S2, taken as the number of detector channels Corresponding The periodically preprocessed data is reconstructed in two dimensions, and row convolution and column convolution are performed on the reconstructed two-dimensional data respectively to extract the temporal feature vector and spatial feature vector of the probe data. Wherein, S2 includes: Take the number of detector channels Corresponding The preprocessed data is reassembled into a two-dimensional array according to the time-space dimension combination. The reassembled data has the following dimensions: ; Build size is and Two convolutional kernels were used to perform row and column convolutions on the two-dimensional reconstructed data, with a stride of 1, to obtain temporal feature vectors respectively. and spatial feature vectors ; S3, perform a cross product on the two vectors to obtain a two-dimensional spatiotemporal interleaved feature; input the two-dimensional spatiotemporal interleaved feature into a multi-layer convolutional neural network, and expand the output into one dimension to obtain a classification feature vector, and then input the classification feature vector into a fully connected network for classification; The two-dimensional spatiotemporal interleaving feature is represented as follows: in, express .

2. The point-type smoke detection method for accurate classification of fire smoke according to claim 1, characterized in that, The method further includes: S4, calculate the cosine similarity between the classification feature vector and the specimen vector. If the cosine similarity is greater than a preset threshold, the classification result is determined to be valid; otherwise, the corresponding probe data is labeled as unlabeled data. The specimen vector is the mean of the classification feature vectors of each category when correctly inferring from the training data of all known categories under the current network parameters.

3. The point-type smoke detection method for accurate classification of fire smoke according to claim 1 or 2, characterized in that, In step S1, the detection data is obtained in the following manner: Fixed on a circular plane A laser emitter is selected within a 180° clockwise or counterclockwise range, with the center of the circular plane as the origin. Install at different angles A photoelectric detector was manufactured, and a housing with a smoke inlet was made. During measurement, the MCU controls... The laser emitters are turned on in sequence, the photodetector converts the collected light signal into an electrical signal and outputs it, the analog-to-digital converter converts the electrical signal into a digital value within the range, and then the microcontroller controls the multiplexer to acquire the data of each channel in sequence and send it to the data receiving end via serial port.

4. The point-type smoke detection method for accurate classification of fire smoke according to claim 1, characterized in that, In S1, the preprocessing includes: mean filtering, background removal, and setting a threshold.

5. The point-type smoke detection method for accurate classification of fire smoke according to claim 1, characterized in that, In S3, the multi-layer convolutional neural network is a two-layer convolutional neural network or a three-layer convolutional neural network; wherein, the single-layer convolutional neural network includes a convolutional layer, a normalization layer and an activation layer, and the activation function adopts a linear rectified function.

6. A point-type smoke detection device for precise classification of fire smoke, characterized in that, include: The preprocessing module is used to preprocess the acquired detection data; The two-dimensional reconstruction module is used to obtain the number of detector channels. Corresponding The periodically preprocessed data is reconstructed in two dimensions, and row convolution and column convolution are performed on the reconstructed two-dimensional data to extract the temporal feature vector and spatial feature vector of the probe data. The two-dimensional recombination module is used to perform the following steps: Take the number of detector channels Corresponding The preprocessed data is reassembled into a two-dimensional array according to the time-space dimension combination. The reassembled data has the following dimensions: ; Build size is and Two convolutional kernels were used to perform row and column convolutions on the two-dimensional reconstructed data, with a stride of 1, to obtain temporal feature vectors respectively. and spatial feature vectors ; The classification module is used to perform a cross product of two vectors to obtain two-dimensional spatiotemporal interleaved features; the two-dimensional spatiotemporal interleaved features are input into a multi-layer convolutional neural network, and the output is expanded into one dimension to obtain a classification feature vector, which is then input into a fully connected network for classification; The two-dimensional spatiotemporal interleaving feature is represented as follows: in, express .

7. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method as claimed in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the method as described in any one of claims 1-5.

Images