An intelligent control method for an acoustic-optic alarm

Abstract

The present invention relates to the field of alarm control. More specifically, the present invention relates to an intelligent control method for an audible and visual alarm. The method includes: obtaining a data set and a first-stage network; inputting a fire monitoring image into the first-stage network to obtain a loss value, calculating the comprehensive importance of each network parameter, and using the loss value to supervise and train the first-stage network to obtain a second-stage network; screening out some redundant network parameters based on the comprehensive importance; continuing to train the second-stage network until the training is completed, pruning the redundant network parameters, and using the pruned fire monitoring network to control the audible and visual alarm; wherein, during the process of training the second-stage network using the fire monitoring image, other network parameters are made to concurrently learn the feature map information extracted by the redundant network parameters. While achieving network lightweight through accurate pruning control, the impact on calculation accuracy can be reduced as small as possible.

Description

Technical Field

[0001] The present invention relates to the field of alarm control. More specifically, the present invention relates to an intelligent control method for an audible and visual alarm device. Background Art

[0002] In case of a fire, an audible and visual alarm device can emit sound and light to give people a prompt so that people can escape from the fire scene in time. Traditional audible and visual alarm devices mostly rely on people to identify fires, and send control instructions to the audible and visual alarm device according to the identification results, and the audible and visual alarm device makes corresponding responses according to the control instructions. Therefore, this control method has low intelligence.

[0003] In order to realize the intelligent control of the audible and visual alarm device, the audible and visual alarm device needs to have a fire recognition function. Traditionally, a neural network is mostly used for fire recognition. Since there are many network parameters in the neural network, the neural network has high computing requirements for hardware, so it is difficult to configure it inside a lightweight audible and visual alarm device. In order to reduce the amount of calculation, it is necessary to perform lightweight processing on the neural network. To achieve the lightweight processing of the neural network, some network parameters need to be pruned, and pruning will inevitably damage the computing accuracy of the neural network. Therefore, how to perform pruning control to minimize the impact on the computing accuracy has become the research focus of this solution. Summary of the Invention

[0004] To solve the problem of how to perform pruning control to minimize the impact on the computing accuracy, the present invention proposes an intelligent control method for an audible and visual alarm device, and the method includes the following steps:

[0005] Obtain a data set, where the data set contains several batches of fire monitoring images;

[0006] Use the data set to perform one round of training on a pre-constructed fire monitoring network to obtain a first-stage network;

[0007] Input a batch of fire monitoring images into the first-stage network in sequence for calculation processing. Randomly select and discard the feature maps corresponding to some network parameters in each convolutional layer, obtain the loss value obtained for each fire monitoring image, calculate the importance of the network parameter according to whether there is a loss value corresponding to the feature map of any network parameter, use the loss value to perform supervised training on the first-stage network, obtain the importance of each network parameter obtained from a preset number of batches of fire monitoring images and calculate the average value to obtain the comprehensive importance of each network parameter; Denote the first-stage network after training with a preset number of batches of fire monitoring images as the second-stage network;

[0008] Select some redundant network parameters based on comprehensive importance; use the fire monitoring images to continue training the second-stage network until the training is completed, prune the redundant network parameters, and use the pruned fire monitoring network to control the audible and visual alarm; among them, during the process of using the fire monitoring images to train the second-stage network, let other network parameters concurrently learn the feature map information extracted by the redundant network parameters.

[0009] In the present invention, by analyzing the importance of each network parameter, redundant network parameters are selected according to the importance for pruning processing. While achieving network lightweight, it can also reduce the impact on the calculation accuracy as small as possible; further, when analyzing the importance of each network parameter, by analyzing the network loss situation when the feature maps of each network parameter exist and do not exist, the importance of each network parameter can be accurately reflected, providing a basis for subsequent accurate network parameter pruning; further, before pruning the network parameters, use other network parameters to concurrently learn the information in the redundant network parameters, thereby further reducing the impact of redundant network parameter pruning on the network detection accuracy.

[0010] Preferably, the step of using the data set to perform one round of training on the pre-constructed fire monitoring network to obtain the first-stage network includes:

[0011] Obtain the pre-constructed fire monitoring network;

[0012] Use the fire monitoring images in the data set to train the fire monitoring network in sequence to obtain the first-stage network.

[0013] In the present invention, by performing one round of training on the network, the network parameters learn certain information, providing a data basis for subsequent accurate analysis of the importance of each network parameter.

[0014] Preferably, the step of randomly selecting and discarding the feature maps corresponding to some network parameters in each convolutional layer, and obtaining the loss value of each fire monitoring image includes:

[0015] Input any one fire monitoring image in a batch of fire monitoring images into the fire monitoring network, and randomly select some network parameters as candidate parameters in the first convolutional layer; discard the feature maps corresponding to the candidate parameters, input the remaining feature maps into the next convolutional layer, participate in the calculation and analysis of the next convolutional layer, randomly select some network parameters as candidate parameters in the second convolutional layer, discard the feature maps corresponding to the candidate parameters, input the remaining feature maps into the next convolutional layer, participate in the calculation and analysis of the next convolutional layer, and so on until all convolutional layers are calculated. Obtain the loss value of this fire monitoring image at the end of the fire monitoring network;

[0016] Obtain the loss value of each fire monitoring image in this batch of fire monitoring images.

[0017] The present invention simulates the network detection accuracy when network parameters are missing by discarding the feature map information corresponding to some network parameters, providing a data basis for subsequent analysis of the influence of network parameters on network detection accuracy.

[0018] Preferably, calculating the importance of the network parameter according to the loss value obtained based on whether there is a feature map corresponding to any network parameter includes:

[0019] Calculating the information gain of each network parameter, denoted as the importance of each network parameter, according to whether the feature maps corresponding to each network parameter are discarded when each fire monitoring image in this batch of fire monitoring images is input and the loss values obtained from each fire monitoring image.

[0020] The present invention accurately reflects the importance of each network parameter through information gain, providing a basis for subsequent accurate network pruning.

[0021] Preferably, screening out some redundant network parameters based on the comprehensive importance includes:

[0022] Marking the network parameters with the normalized value of the comprehensive importance less than the preset importance threshold as redundant network parameters.

[0023] Preferably, pruning the redundant network parameters and using the fire monitoring network after pruning to control the sound and light alarm includes:

[0024] Pruning the redundant network parameters, arranging the fire monitoring network after pruning on the embedded platform of the sound and light alarm, performing fire detection processing on the fire monitoring images collected by the sound and light alarm to obtain a detection result, and controlling the sound and light alarm according to the detection result.

[0025] Preferably, controlling the sound and light alarm according to the detection result includes:

[0026] If the detection result is that there is a fire, controlling the sound and light alarm to emit a sound and light alarm;

[0027] If the detection result is that there is no fire, controlling the sound and light alarm not to emit a sound and light alarm.

[0028] Preferably, during the process of training the second-stage network using the fire monitoring images, enabling other network parameters to concurrently learn the feature map information extracted by the redundant network parameters includes:

[0029] Obtaining the concurrent learning network parameters of each redundant network parameter; obtaining the feature map corresponding to the redundant network parameter and the feature map corresponding to the concurrent learning network parameter;

[0030] Constructing a concurrent learning loss function:

[0031] ;

[0032] wherein, represents the feature map corresponding to the parameters of the merging learning network, represents the feature map corresponding to the redundant network parameters, represents the comprehensive importance of the redundant network parameters, represents the comprehensive importance of the parameters of the merging learning network, represents the L2 norm of the matrix, represents a preset adjustment coefficient;

[0033] The training of the second-stage network is jointly supervised by using the loss value and the loss value obtained from the merging learning loss function.

[0034] In the present invention, the merging learning loss function is used to control the relevant information of other network parameters for merging and learning redundant network parameters, thereby reducing to a certain extent the impact of redundant network parameter pruning on the network detection accuracy. Further, when constructing the merging learning loss function, by introducing the difference between the feature map obtained from the merging learning network parameters and the feature map obtained from the redundant network parameters, as well as the importance of the redundant network parameters, the merging learning network parameters can learn the information in the redundant network parameters more accurately and appropriately, providing a basis for reducing the impact of redundant network parameter pruning on the network detection accuracy.

[0035] Preferably, the obtaining of the merging learning network parameters of each redundant network parameter includes:

[0036] Denote the feature map corresponding to the redundant network parameter as the redundant feature map, and denote the feature map corresponding to other network parameters as the non-redundant feature map; calculate the similarity between the redundant feature map and the non-redundant feature map, and obtain the network parameter corresponding to the non-redundant feature map that is most similar to each redundant feature map as the merging learning network parameter of the redundant network parameter.

[0037] The present invention has the following beneficial effects:

[0038] By analyzing the importance of each network parameter, screening out redundant network parameters for pruning, while realizing network lightweighting, it can also reduce the impact on the calculation accuracy as small as possible;

[0039] Further, when analyzing the importance of each network parameter, by analyzing the network loss situation when the feature map of each network parameter exists and does not exist, the importance of each network parameter can be accurately reflected, providing a basis for subsequent accurate network parameter pruning;

[0040] Further, before pruning the network parameters, other network parameters are used to annex and learn the information in the redundant network parameters, so as to further reduce the impact of pruning the redundant network parameters on the network detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] By referring to the following detailed description with reference to the accompanying drawings, the above and other objects, features, and advantages of the exemplary embodiments of the present invention will become readily understood. In the drawings, several embodiments of the present invention are shown by way of example and not limitation, and identical or corresponding reference numerals represent identical or corresponding parts, wherein:

[0042] Figure 1 is a flowchart of the steps of an intelligent control method for an acoustic-optic alarm in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0044] The following will describe in detail the specific embodiments of the present invention with reference to the accompanying drawings.

[0045] Please refer to Figure 1 , which shows a flowchart of the steps of an intelligent control method for an acoustic-optic alarm provided by an embodiment of the present invention. The method includes the following steps:

[0046] S1: Obtain a data set, where the data set contains several batches of fire monitoring images.

[0047] Specifically, the images of the area to be monitored are collected by a camera and recorded as fire monitoring images;

[0048] The fire monitoring images are manually labeled. The fire monitoring images with fire are labeled as 1, and the fire monitoring images without fire are labeled as 0.

[0049] Taking N fire monitoring images as a batch, several batches of fire monitoring images are obtained. N represents a preset first quantity. In this embodiment, N is taken as 50 for description. Other embodiments can take other values, and this embodiment does not make specific limitations.

[0050] S2: Use the data set to perform one round of training on a pre-constructed fire monitoring network to obtain a first-stage network.

[0051] It should be noted that since the network parameters of the neural network are randomly generated in the initial stage, none of the network parameters of the neural network have the function of fire detection in the initial stage. Therefore, the importance of each network parameter for fire detection cannot be determined by the values of the network parameters in the initial stage. Therefore, the neural network needs to be trained to a certain extent so that the network parameters have the function of fire detection, and then it is possible to determine which network parameters are important for fire detection.

[0052] Preferably, as an example, use the data set to train the pre-constructed fire monitoring network for one round to obtain the first-stage network, including:

[0053] Obtain the pre-constructed fire monitoring network. In this embodiment, the fire monitoring network adds an operation of image downsampling processing before the fully connected layer of the traditional VGG network, and the rest is the same as the traditional VGG network. Since the traditional VGG network is an existing network, the content of the VGG network will not be elaborated here. The following mainly introduces the content of image pooling.

[0054] The implementation process of the image pooling includes: obtaining all the feature maps to be input into the fully connected layer and recording them as the feature maps to be analyzed, and taking any pixel position as the target position; obtaining the pixel values at all target positions in all the feature maps to be analyzed, and recording the sequence composed of all the obtained pixel values as the sequence to be analyzed; using a convolution kernel with a preset size to perform sliding convolution processing on the sequence to be analyzed, and then performing pooling processing on the convolved sequence to be analyzed to obtain the pooled sequence to be analyzed; processing the sequence to be analyzed obtained at each pixel position in the same way to obtain the pooled sequence to be analyzed at each pixel position; restoring the sequence to be analyzed into several feature maps and recording them as the pooled feature maps, and using the pooled feature maps as the input of the fully connected layer. In this embodiment, the preset size is taken as , and other embodiments can take other values, and this embodiment does not make specific restrictions.

[0055] It should be noted that since the number of network parameters in the fully connected layer is fixed, the dimension of the data input into the fully connected layer is fixed. Adding image pooling processing here is to reduce the dimension of the data input into the fully connected layer to prevent the problem of the reduction of the dimension of the data input into the fully connected layer caused by removing some convolutional layer network parameters later.

[0056] It should be further noted that the above convolution and pooling processing are existing knowledge and will not be elaborated here.

[0057] Use the fire monitoring images in the data set to train the fire monitoring network in turn to obtain the first-stage network.

[0058] It should be noted that training the network using images is a prior art and will not be elaborated here.

[0059] S3: Input a batch of fire monitoring images into the first-stage network for calculation and processing. Randomly discard some of the feature maps corresponding to the network parameters in each convolutional layer, obtain the loss value of each fire monitoring image, calculate the importance of the network parameter according to the loss value obtained with or without the feature map corresponding to any network parameter, and use the loss value to perform supervised training on the first-stage network; obtain the importance of each network parameter obtained from a preset number of batches of fire monitoring images and calculate the average value to obtain the comprehensive importance of each network parameter; record the first-stage network trained with a preset number of batches of fire monitoring images as the second-stage network.

[0060] It should be noted that in order to perform pruning on the network, it is necessary to judge the importance of the network parameters. The following mainly introduces the relevant content of judging the importance of network parameters.

[0061] It should be further noted that the information obtained by the convolutional network parameters in the network is mainly presented through feature maps. If the feature maps obtained by the convolutional network parameters are removed, the role of the convolutional network parameters will be truncated. Therefore, the importance of each convolutional network parameter for fire detection can be judged by analyzing the accuracy of fire detection when the feature maps obtained by the convolutional network parameters exist or do not exist.

[0062] S30: Input a batch of fire monitoring images into the first-stage network for calculation and processing. Randomly discard some of the feature maps corresponding to the network parameters in each convolutional layer, obtain the loss value of each fire monitoring image, calculate the importance of the network parameter according to the loss value obtained with or without the feature map corresponding to any network parameter, and use the loss value to perform supervised training on the first-stage network.

[0063] S300: Input a batch of fire monitoring images into the first-stage network for calculation and processing. Randomly discard some of the feature maps corresponding to the network parameters in each convolutional layer, and obtain the loss value of each fire monitoring image.

[0064] Preferably, as an example, input a batch of fire monitoring images into the first-stage network for calculation and processing in sequence. Randomly discard some of the feature maps corresponding to the network parameters in each convolutional layer, and obtain the loss value of each fire monitoring image, including:

[0065] Input any one of a batch of fire monitoring images into the fire monitoring network, and randomly select some network parameters in the first convolutional layer as candidate parameters; discard the feature maps corresponding to the candidate parameters, and input the remaining feature maps into the next convolutional layer to participate in the calculation and analysis of the next convolutional layer. Randomly select some network parameters in the second convolutional layer as candidate parameters, discard the feature maps corresponding to the candidate parameters, and input the remaining feature maps into the next convolutional layer to participate in the calculation and analysis of the next convolutional layer, and so on until all convolutional layers are calculated. Obtain the loss value obtained from this fire monitoring image at the end of the fire monitoring network;

[0066] Obtain the loss value obtained from each fire monitoring image in this batch of fire monitoring images.

[0067] It should be added that after removing the feature maps corresponding to some network parameters, the data dimension input into the fully connected layer will be smaller. Therefore, the pooling degree of the image pooling process before the fully connected layer can be adjusted so that the data dimension input into the fully connected layer is fixed. For example, the dimension of the input data of the fully connected layer is 100. If no feature maps are removed in step S2, 2 pooling processes can be performed at this time to obtain input data with a dimension of 100. However, in this step, after removing some feature maps, the dimension is reduced, and only one pooling process is required to obtain a dimension of 100. Therefore, only one pooling process is required at this time. This example is to reflect that the pooling degree of the pooling process is adjusted according to the data dimension.

[0068] S301: Calculate the importance of the network parameter according to the loss value obtained with or without the feature map corresponding to any network parameter.

[0069] It should be noted that information gain can reflect the influence of the presence or absence of the feature maps corresponding to each network parameter on the fire detection loss. Therefore, the importance of each network parameter can be judged through information gain.

[0070] Preferably, as an example, calculate the importance of the network parameter according to the loss value obtained with or without the feature map corresponding to any network parameter.

[0071] Record the loss values obtained from all fire monitoring images in this batch of fire monitoring images as the loss values to be analyzed. Obtain the value range of all loss values to be analyzed, evenly divide the value range into a preset second number of sub-ranges, and obtain the probability of the loss values to be analyzed existing in each sub-range. In this embodiment, the preset second number is taken as 5 for description, and other values can be taken in other embodiments, and this embodiment does not make specific limitations.

[0072] Calculate the information entropy: , It represents the probability that there is a loss value to be analyzed in the t-th sub-range. log() represents the logarithmic function with base 10. Y represents the number of sub-ranges, and X represents the information entropy of the network parameter.

[0073] Obtain all the loss values to be analyzed corresponding to the feature map of the network parameter and denote them as the first loss values to be analyzed. Obtain all the loss values to be analyzed that do not correspond to the feature map of the network parameter and denote them as the second loss values to be analyzed. Obtain the probability that the first loss values to be analyzed exist among the loss values to be analyzed and the probability that the second loss values to be analyzed exist. Obtain the probability that the first loss values to be analyzed exist in each sub-range and the probability that the second loss values to be analyzed exist in each sub-range.

[0074] Calculate the conditional entropy: , It represents the probability that the first loss values to be analyzed exist among the loss values to be analyzed, It represents the probability that the second loss values to be analyzed exist among the loss values to be analyzed, It represents the probability that the first loss values to be analyzed exist in the t-th sub-range, It represents the probability that the second loss values to be analyzed exist in the t-th sub-range. Y represents the number of sub-ranges, It represents the conditional entropy of the network parameter.

[0075] Subtract the conditional entropy X1 from the information entropy X0 to obtain the information gain of the network parameter, and use the information gain as the importance of the network parameter.

[0076] It should be noted that the information gain is a prior art, which is used to reflect the influence of the determination of each piece of information on the final result. Therefore, it can reflect the influence of each network parameter on fire detection. In other words, it can reflect the importance of each network parameter.

[0077] S302: Use the loss

[0078] value to supervise and train the first-stage network.

[0079] It should be noted that using the loss value to supervise and train the first-stage network is a prior art, and no further elaboration will be provided here.

[0080] S31: Obtain the importance of each network parameter obtained from a preset number of batches of fire monitoring images and calculate the mean value to obtain the comprehensive importance of each network parameter. Denote the first-stage network trained with a preset number of batches of fire monitoring images as the second-stage network.

[0081] It can be understood that the above analysis of the loss situation of the network when each network parameter exists and does not exist is used to judge the importance of each network parameter, providing a basis for the accurate pruning of subsequent network parameters.

[0082] S4: Select some redundant network parameters based on the comprehensive importance; continue to train the second-stage network with the fire monitoring images until the training is completed, perform pruning on the redundant network parameters, and use the pruned network to control the audible and visual alarm; among them, during the process of training the second-stage network with the fire monitoring images, let other network parameters incorporate and learn the feature map information extracted by the redundant network parameters.

[0083] S40: Select some redundant network parameters based on the comprehensive importance.

[0084] Preferably, as an example, selecting some redundant network parameters based on the comprehensive importance includes:

[0085] Perform normalization processing on the comprehensive importance, and record the network parameters with the normalized value of the comprehensive importance less than the preset importance threshold as redundant network parameters.

[0086] It can be understood that a smaller comprehensive importance indicates that the network parameter is not important, and removing this network parameter has a relatively small impact on the network accuracy. Therefore, unimportant redundant network parameters can be selected according to the comprehensive importance.

[0087] S41: Continue to train the second-stage network with the fire monitoring images until the training is completed, and perform pruning on the redundant network parameters.

[0088] S410: Continue to train the second-stage network with the fire monitoring images until the training is completed.

[0089] It should be noted that although the redundant network parameters have a relatively low importance, they will also affect the network accuracy to a certain extent. Therefore, if the information learned by the redundant network parameters can be incorporated and learned by other network parameters, the impact on the network accuracy after removing these redundant network parameters will be relatively reduced. Therefore, a loss function needs to be added here to let the loss function control other network parameters to be able to incorporate and learn the information of the redundant network parameters.

[0090] During the process of training the second-stage network with the fire monitoring images, let other network parameters incorporate and learn the feature map information extracted by the redundant network parameters.

[0091] Preferably, as an example, letting other network parameters incorporate and learn the feature map information extracted by the redundant network parameters includes:

[0092] Denote the feature map corresponding to the redundant network parameter as the redundant feature map, and the feature map corresponding to other network parameters as the non-redundant feature map; calculate the similarity between the redundant feature map and the non-redundant feature map, and obtain the network parameter corresponding to the non-redundant feature map that is most similar to each redundant feature map as the merger learning network parameter of the redundant network parameter; obtain the feature map corresponding to the redundant network parameter and the feature map corresponding to the merger learning network parameter.

[0093] Construct a merger learning loss function:

[0094]

[0095] Wherein, represents the feature map corresponding to the merger learning network parameter, represents the feature map corresponding to the redundant network parameter, represents the comprehensive importance of the redundant network parameter, represents the comprehensive importance of the merger learning network parameter, represents the L2 norm of the matrix, represents a preset adjustment coefficient, which is used to adjust the learning degree of the merger learning network parameter for the feature map information of the redundant network parameter. In this embodiment, taking 0.1 as an example for description. Other embodiments can take other values, and this embodiment does not make specific limitations. represents the merger learning loss function.

[0096] It can be understood that reflects the difference between the feature map learned by the merger learning network parameter and the feature map learned by the redundant network parameter. If the merger learning network parameter is to learn the information related to the redundant network parameter, it is necessary to make the difference between the feature map learned by the merger learning network parameter and the feature map learned by the redundant network parameter smaller; reflects the importance of the redundant network parameter compared to the merger learning network parameter. This value is used to adjust the learning degree of the merger learning network parameter for the feature map information of the redundant network parameter. The larger this value is, the more important the redundant network parameter is, and the merger learning network parameter should learn more feature map information of the redundant network parameter.

[0097] Use the loss value and the loss value obtained by the merger learning loss function to jointly supervise the training of the second-stage network.

[0098] It should be noted that using the loss value and the loss value obtained by the merger learning loss function to jointly supervise the training of the second-stage network is the same as the method of using multiple loss values to supervise network training in the prior art, so it will not be elaborated here.

[0099] S411: Prune the redundant network parameters.

[0100] Preferably, as an example, pruning the redundant network parameters includes:

[0101] Prune the redundant network parameters and arrange the pruned fire monitoring network on the embedded platform of the sound and light alarm.

[0102] S42: Control the sound and light alarm using the pruned fire monitoring network.

[0103] Preferably, as an example, controlling the sound and light alarm using the pruned fire monitoring network includes:

[0104] Perform fire detection processing on the fire monitoring image collected by the sound and light alarm to obtain a detection result. If the detection result indicates a fire, control the sound and light alarm to emit a sound and light alarm; if the detection result indicates no fire, control the sound and light alarm not to emit a sound and light alarm.

[0105] Thus, this embodiment is completed.

[0106] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. An intelligent control method for an audible and visual alarm, characterized in that: include: Acquire a data set, wherein the data set includes a plurality of batches of fire monitoring images; The pre-built fire monitoring network is trained using the dataset to obtain the first-stage network. A batch of fire monitoring images are sequentially input into the first-stage network for computational processing, and the feature maps corresponding to some network parameters are randomly extracted and discarded in each convolution layer to obtain the loss value obtained for each fire monitoring image. The importance of the network parameter is calculated according to the loss value obtained by the feature map corresponding to any network parameter. The first-stage network is supervised and trained using the loss value to obtain the importance of each network parameter obtained by a preset number of batches of fire monitoring images and the average value is calculated to obtain the comprehensive importance of each network parameter. The first stage network training after training with a preset number of batches of fire monitoring images is recorded as the second stage network; Filter out some redundant network parameters based on comprehensive importance; The second-stage network is continuously trained with fire monitoring images until the training is completed, and the redundant network parameters are pruned. The pruned fire monitoring network is used to control the sound and light alarm. In the process of training the second-stage network with fire monitoring images, other network parameters are allowed to merge and learn the feature map information extracted by the redundant network parameters.

2. The intelligent control method of an audible and visual alarm according to claim 1, characterized in that: The first-stage network is obtained by performing a round of training on the pre-built fire monitoring network using the data set, including: Get pre-built fire monitoring networks; The fire monitoring network is trained in sequence using the fire monitoring images in the dataset to obtain the first stage network.

3. The intelligent control method of an audible and visual alarm according to claim 1, characterized in that: The feature maps corresponding to some network parameters are randomly extracted and discarded in each convolution layer to obtain the loss value of each fire monitoring image, including: Input any fire monitoring image from a batch of fire monitoring images into the fire monitoring network, randomly extract some network parameters in the first convolution layer as candidate parameters; discard the feature maps corresponding to the candidate parameters, input the remaining feature maps into the next convolution layer, participate in the calculation and analysis of the next convolution layer, randomly extract some network parameters as candidate parameters in the second convolution layer, discard the feature maps corresponding to the candidate parameters, input the remaining feature maps into the next convolution layer, participate in the calculation and analysis of the next convolution layer, and so on, until all convolution layers are calculated, and finally obtain the loss value of the fire monitoring image at the end of the fire monitoring network; Obtain a loss value for each fire monitoring image in the batch of fire monitoring images.

4. The intelligent control method of the sound and light alarm according to claim 3 is characterized in that: The calculating the importance of the network parameter according to the loss value obtained by whether any network parameter corresponds to the feature graph includes: According to whether the feature map corresponding to each network parameter is discarded when each fire monitoring image in the batch of fire monitoring images is input and the loss value obtained for each fire monitoring image, the information gain of each network parameter is calculated and recorded as the importance of each network parameter.

5. The intelligent control method of an audible and visual alarm according to claim 1, characterized in that: The method of screening out some redundant network parameters based on comprehensive importance includes: The network parameters whose normalized values ​​of comprehensive importance are less than the preset importance threshold are recorded as redundant network parameters.

6. The intelligent control method of the sound and light alarm according to claim 1 is characterized in that: The method of pruning redundant network parameters and controlling the sound and light alarm using the pruned fire monitoring network includes: The redundant network parameters are pruned, and the pruned fire monitoring network is arranged on the embedded platform of the sound and light alarm. The fire monitoring images collected by the sound and light alarm are processed for fire detection to obtain the detection results, and the sound and light alarm is controlled according to the detection results.

7. The intelligent control method of the sound and light alarm according to claim 6 is characterized in that: The control of the sound and light alarm according to the detection result includes: If the detection result shows that there is a fire, the sound and light alarm is controlled to sound a sound and light alarm; If the detection result is that there is no fire, the sound and light alarm will be controlled not to sound the sound and light alarm.

8. The intelligent control method of the sound and light alarm according to claim 1 is characterized in that: In the process of training the second-stage network using the fire monitoring image, other network parameters are allowed to learn the feature map information extracted by the redundant network parameters, including: Obtaining the merged learning network parameters of each redundant network parameter; obtaining the characteristic graph corresponding to the redundant network parameter and the characteristic graph corresponding to the merged learning network parameter; Construct the merged learning loss function: ； in, Represents the feature map corresponding to the merged learning network parameters, Represents the feature graph corresponding to the redundant network parameters, represents the comprehensive importance of redundant network parameters, represents the comprehensive importance of the combined learning network parameters, represents the L2 norm of the matrix, Indicates the preset adjustment factor; The loss value and the loss value obtained by merging the learning loss function are used to jointly supervise the training of the second stage network.

9. The intelligent control method of the sound and light alarm according to claim 8, characterized in that: The step of acquiring the combined learning network parameters of the redundant network parameters includes: The feature maps corresponding to the redundant network parameters are recorded as redundant feature maps, and the feature maps corresponding to other network parameters are recorded as non-redundant feature maps; the similarity between the redundant feature maps and the non-redundant feature maps is calculated, and the network parameters corresponding to the non-redundant feature maps that are most similar to each redundant feature map are obtained as the merged learning network parameters of the redundant network parameters.

Images