An indoor fire detection method based on deep neural network multi-source domain transfer learning

Abstract

The application provides an indoor fire detection method based on deep neural network multi-source domain transfer learning, belongs to the technical field of fire monitoring, and is used for monitoring the occurrence of indoor fire. The method comprises the following steps: 1, collecting different source fire dataset samples as a neural network training set; 2, adopting a special neural network structure, so that the neural network can generate different weight parameters according to different sample inputs; 3, training the neural network by taking the training set as input; 4, inputting the fire dataset test sample into the neural network, and outputting the classification result of the neural network. The method provided by the application can effectively utilize different source datasets to train the neural network, avoids the embarrassing situation that the neural network cannot be trained due to a small number of samples, and has certain practicability.

Description

Technical Field

[0001] This invention belongs to the field of fire detection technology, and in particular relates to an indoor fire detection method based on deep neural network multi-source domain transfer learning. Background Technology

[0002] Traditional methods of monitoring fires using various sensors such as smoke sensors, temperature sensors, and air quality sensors have the advantages of being easy to deploy and low in cost, but they also have the following drawbacks: 1. Such sensors cannot be deployed in relatively spacious areas; 2. Their fire detection is delayed and cannot achieve real-time fire monitoring.

[0003] In modern society, camera deployment is extremely common, making vision-based fire monitoring possible. Vision-based fire monitoring does not require the deployment of additional sensors besides cameras, which effectively saves on the deployment cost of fire monitoring systems. Secondly, the monitoring accuracy of the deployed system will not decrease due to the environment being spacious, and it can achieve real-time detection and alarm of fires, thus solving the two major shortcomings of traditional monitoring systems mentioned above.

[0004] Before the rise of deep learning, vision-based fire detection made full use of features such as flame color, texture, and shape. The performance of these methods mainly depended on manually designing flame characteristics. However, in reality, the color, shape, and brightness of flames vary depending on the actual situation, making traditional vision-based fire detection methods less than ideal.

[0005] With the rapid development of deep learning today, it has become the preferred method for solving various vision-based problems. The success of deep learning is due to the use of massive amounts of data to train neural networks, enabling the network parameters to converge to the optimal solution. However, in the problem of indoor fire detection, the number of indoor fire samples is relatively small, and their sources are also different, making it difficult to train a good network model for indoor fire monitoring. Summary of the Invention

[0006] The purpose of this invention is to provide an indoor fire detection method based on multi-source domain transfer learning of deep neural networks, characterized by the following steps:

[0007] S1: Collect fire data samples from different sources as the training set for the neural network;

[0008] S2: Design a neural network structure that can generate different weight parameters based on different samples and effectively classify samples from different sources;

[0009] S3: Use the training set of S1 as input to train the neural network of S2, so that the neural network converges to a suitable point;

[0010] S4: Input the test sample into the trained neural network. The neural network classifies the test sample and determines whether a fire has occurred in the test sample.

[0011] Furthermore, in S1, different fire data samples simultaneously have both positive and negative examples.

[0012] Furthermore, in S2, the neural network is a convolutional neural network. By adding an increment related to the network parameters to the original network parameters, the parameters of the convolutional neural network can vary with different sample inputs. The formula is defined as follows:

[0013] ;

[0014] in, These are the parameters of the neural network before training; These are the parameters of the trained neural network; This is the parameter increment related to the input.

[0015] Furthermore, the generation of increments is directly related to the input samples, and only the convolutional layers of the processed neural network are changed.

[0016] Furthermore, the increment is specifically generated by: setting the parameters of a single convolutional layer to... The convolution kernel is defined as The parameters of the convolutional layer after training are:

[0017] ；

[0018] middle, and The dimensions are the same as the input dimension of the next layer and the output dimension of the previous layer, respectively; in the parameters of the trained convolutional layer, This is the input to the convolutional layer; For parameter increments related to the input; With input It changes with the changes.

[0019] Furthermore, It is generated through the following steps:

[0020] SA1: Input The neural network sequentially passes through an average pooling layer, a fully connected layer, a ReLU layer, a fully connected layer, and a sigmoid layer, outputting a... square array Then the increment The first part is represented as:

[0021] ;

[0022] SA2: Input The neural network sequentially passes through an average pooling layer, a fully connected layer, a ReLU layer, a fully connected layer, and a softmax layer, outputting a... square array ,and Composed of random numbers within a specified range The matrix, then the increment The second part is represented as:

[0023] ;

[0024] Overall Increment In the first and second parts, the final parameters of the convolution kernel of this additional convolutional layer are:

[0025] .

[0026] Furthermore, the range is limited to (-10~10). k The value is 1.

[0027] Furthermore, during the training of S3, the cross-entropy loss function is used, expressed as:

[0028]

[0029] in, The number of input samples; For the input of the first One sample; For the first The labels of each sample; during training, the loss... The training of the neural network stops when the current is gradually reduced until it converges relatively.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows: the indoor fire detection method based on deep neural network multi-source domain transfer learning proposed in this invention can effectively utilize datasets from different sources to train the neural network, avoiding the embarrassing situation of not being able to train the neural network due to a small number of samples. The method has certain practicality. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the neural network structure used in this invention;

[0032] Figure 2 This is a schematic flowchart of the method of the present invention. Detailed Implementation

[0033] The following will describe in more detail an indoor fire detection method based on deep neural network multi-source domain transfer learning according to the present invention, with reference to the schematic diagrams, which illustrate preferred embodiments of the present invention. It should be understood that those skilled in the art can modify the present invention described herein while still achieving the advantageous effects of the present invention. Therefore, the following description should be understood as being of general knowledge to those skilled in the art and is not intended to limit the present invention.

[0034] like Figure 2 As shown, an indoor fire detection method based on multi-source domain transfer learning of deep neural networks is presented. The main steps of the method include:

[0035] Step 1: Collect fire dataset samples from different sources as the training set for the neural network. The fire dataset samples contain both positive and negative examples.

[0036] Step 2: Design the structure of the neural network. Unlike other neural networks, this network generates different weight parameters based on different sample inputs. This enables it to effectively classify samples from different sources.

[0037] Step 3: Use the training set as input to train the neural network until it converges to a suitable point.

[0038] Step 4: Input the test samples into the neural network, and the network classifies the test samples.

[0039] The fire dataset samples in step 1 must be fire-related datasets and not arbitrarily selected datasets. Many research institutions have published fire datasets, most of which are outdoor fire datasets, which can be used as training sets for neural networks in the method of this invention.

[0040] For the network structure design in step 2, this embodiment uses ResNet50 as the backbone network of the neural network, and adds an additional convolutional layer before each convolutional layer of ResNet50. The convolutional kernel of this additional convolutional layer is...

[0041] The convolutional kernel, the structure of this additional convolutional layer is as follows Figure 1 As shown,

[0042] k =1, where and The dimensions are the same as the input dimension of the next layer and the output dimension of the previous layer, respectively.

[0043] Let the convolution kernel be... ,enter After such Figure 1As shown in Module 1, the input of Module 1 is After passing through an average pooling layer, a fully connected layer, a ReLU layer, a fully connected layer, and a sigmoid layer, a result is output. square array Then the increment :

[0044] ；

[0045] Module 1 output The neural network sequentially passes through an average pooling layer, a fully connected layer, a ReLU layer, a fully connected layer, and a softmax layer, outputting a... square array ,and Composed of random numbers within a limited range (-10 to 10) The matrix, then the increment :

[0046] ；

[0047] The final parameters of the convolution kernel of this additional convolutional layer are:

[0048] ；

[0049] In step 3 of the training process, the commonly used cross-entropy loss function is employed, as shown below:

[0050] ；

[0051] in, The number of input samples. For the input of the first One sample, For the first The labels of each sample. During training, the loss... Gradually decrease the value until it relatively converges, at which point training stops.

[0052] In step 4, the test sample is input into the trained neural network, and the neural network outputs the classification result. Based on the result, it can be determined whether a fire has occurred in the test sample. The test sample can be a photo or a video.

[0053] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. An indoor fire detection method based on multi-source domain transfer learning of deep neural networks, characterized in that, Includes the following steps: S1: Collect fire data samples from different sources as the training set for the neural network; S2: Design a neural network structure that has the ability to generate different weight parameters based on different samples and to classify samples from different sources; S3: Use the training set of S1 as input to train the neural network of S2, so that the neural network converges to a suitable point; S4: Input the test sample into the trained neural network, which classifies the test sample and determines whether a fire has occurred in the test sample; In S2, the neural network is a convolutional neural network. By adding an increment related to the network parameters to the original network parameters, the parameters of the convolutional neural network can be made to vary with different sample inputs. The increment is generated specifically by adjusting the parameters of a single convolutional layer to... The convolution kernel is defined as The parameters of the convolutional layer after training are: ； middle, and The dimensions are the same as the input dimension of the next layer and the output dimension of the previous layer, respectively. In the parameters of the convolutional layer after training, This is the input to the convolutional layer; For parameter increments related to the input; With input It changes with the changes; It is generated through the following steps: SA1: The input The neural network sequentially passes through an average pooling layer, a fully connected layer, a ReLU layer, a fully connected layer, and a sigmoid layer, outputting a... square array Then the increment The first part is represented as: ； SA2: The input The neural network sequentially passes through an average pooling layer, a fully connected layer, a ReLU layer, a fully connected layer, and a softmax layer, outputting a... square array ,and Composed of random numbers within a specified range The matrix, then the increment The second part is represented as: ； In summary, the increments The first part and the second part, then: .

2. The indoor fire detection method based on deep neural network multi-source domain transfer learning according to claim 1, characterized in that, In S1, different fire data samples simultaneously have positive and negative examples.

3. The indoor fire detection method based on deep neural network multi-source domain transfer learning according to claim 1, characterized in that, By adding an increment related to the original network parameters, the parameters of the convolutional neural network can be made to vary with different sample inputs. The formula is defined as follows: ； in, These are the parameters of the neural network before training; These are the parameters of the trained neural network; This is the parameter increment related to the input.

4. The indoor fire detection method based on deep neural network multi-source domain transfer learning according to claim 3, characterized in that, The generation of increments is directly related to the input samples, and only the convolutional layers of the processed neural network are changed.

5. The indoor fire detection method based on deep neural network multi-source domain transfer learning according to claim 1, characterized in that, The defined range is (-10~10). k The value is 1.

6. The indoor fire detection method based on deep neural network multi-source domain transfer learning according to claim 1, characterized in that, During the training of S3, the cross-entropy loss function is used, which is expressed as: ； in, The number of input samples; For the input of the first One sample; For the first The labels of each sample; during training, the loss... The training of the neural network stops when the current is gradually reduced until it converges relatively.

Images