Intelligent control method of audible and visual alarm

Abstract

The invention relates to the field of alarm control, in particular to an intelligent control method for an audible and visual alarm, and the method comprises the steps: obtaining a data set and a first-stage network; inputting the fire monitoring image into the first-stage network to obtain a loss value, calculating the comprehensive importance of each network parameter, and performing supervision training on the first-stage network by using the loss value to obtain a second-stage network; screening out part of redundant network parameters based on comprehensive importance; the second-stage network continues to be trained until training is completed, redundant network parameters are pruned, and the pruned fire monitoring network is used for audible and visual alarm control; wherein in the process of training the second-stage network by using the fire monitoring image, other network parameters are merged to learn feature map information extracted by redundant network parameters. The network lightweight is realized through accurate pruning control, and meanwhile, the influence on the calculation precision can be reduced as much as possible.

Description

Technical Field

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

[0002] When a fire occurs, the sound and light alarm can emit sound and light to alert people and allow them to escape from the fire scene in time. Traditional sound and light alarms mostly rely on people to identify the fire, and then send control instructions to the sound and light alarm based on the identification results. The sound and light alarm responds accordingly, so this control method is less intelligent.

[0003] In order to realize the intelligent control of the sound and light alarm, the sound and light alarm needs to have the fire identification function. Traditionally, neural networks are used for fire identification. Since there are many network parameters in the neural network, the neural network has high requirements for hardware calculations, so it is difficult to configure it inside the lightweight sound and light alarm. In order to reduce the amount of calculation, the neural network needs to be lightweight. To achieve lightweight processing of the neural network, some network parameters need to be pruned, and pruning will inevitably damage the calculation accuracy of the neural network. Therefore, how to perform pruning control to minimize the impact on the calculation accuracy has become the research focus of this program. Summary of the invention

[0004] In order to solve the problem of how to perform pruning control to minimize the impact on calculation accuracy, the present invention proposes an intelligent control method for an audible and visual alarm, which comprises the following steps: 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, feature maps corresponding to some network parameters are randomly extracted and discarded in each convolution layer, and the loss value obtained for each fire monitoring image is obtained. The importance of the network parameter is calculated according to the loss value obtained by the feature map corresponding to any network parameter, and 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 calculate the average value to obtain the comprehensive importance of each network parameter; the first-stage network training after the preset number of batches of fire monitoring images is recorded as the second-stage network; Based on the comprehensive importance, some redundant network parameters are screened out; the second-stage network is continuously trained with the fire monitoring images until the training is completed, the redundant network parameters are pruned, and the pruned fire monitoring network is used to control the sound and light alarm; wherein, in the process of training the second-stage network with the fire monitoring images, other network parameters are allowed to merge and learn the feature map information extracted by the redundant network parameters.

[0005] The present invention analyzes the importance of each network parameter, selects redundant network parameters according to the importance and performs pruning processing, thereby achieving network lightweighting while minimizing the impact on calculation accuracy; further, when analyzing the importance of each network parameter, the importance of each network parameter is accurately reflected by analyzing the network loss situation when the characteristic graph of each network parameter exists and does not exist, providing a basis for subsequent accurate network parameter pruning; further, before pruning the network parameters, other network parameters are used to merge and learn the information in the redundant network parameters, thereby further reducing the impact of the redundant network parameter pruning on the network detection accuracy.

[0006] Preferably, the method of using the data set to perform a round of training on the pre-built fire monitoring network to obtain the first-stage network includes: 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.

[0007] The present invention allows the network parameters to learn certain information by performing a round of training on the network, thereby providing a data basis for the subsequent accurate analysis of the importance of each network parameter.

[0008] Preferably, randomly extracting and discarding feature maps corresponding to some network parameters in each convolution layer to obtain a loss value for each fire monitoring image includes: 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.

[0009] The present invention simulates the network detection accuracy when network parameters are missing by discarding the characteristic graph information corresponding to some network parameters, providing a data basis for subsequent analysis of the impact of network parameters on network detection accuracy.

[0010] Preferably, 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.

[0011] The present invention accurately reflects the important conditions of each network parameter through information gain, providing a basis for subsequent accurate network pruning.

[0012] Preferably, the 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.

[0013] Preferably, the pruning of redundant network parameters and using the pruned fire monitoring network to control the sound and light alarms include: 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.

[0014] Preferably, the controlling 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.

[0015] Preferably, in the process of training the second-stage network using the fire monitoring image, allowing other network parameters 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.

[0016] The present invention controls other network parameters by merging learning loss functions to merge and learn related information of redundant network parameters, thereby reducing the influence of redundant network parameter pruning on network detection accuracy to a certain extent. Furthermore, when constructing the merged learning loss function, by introducing the difference between the feature graph obtained by merging learning network parameters and the feature graph obtained by redundant network parameters and the importance of redundant network parameters, the merged learning network parameters can be more accurately and appropriately learned from the information in the redundant network parameters, providing a basis for reducing the influence of redundant network parameter pruning on network detection accuracy.

[0017] Preferably, 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.

[0018] The present invention has the following beneficial effects: The present invention analyzes the importance of each network parameter, selects redundant network parameters according to the importance, and performs pruning processing, thereby achieving network lightweighting while minimizing the impact on calculation accuracy; Furthermore, when analyzing the importance of each network parameter, the importance of each network parameter is accurately reflected by analyzing the network loss when the characteristic graph of each network parameter exists and does not exist, providing a basis for subsequent accurate network parameter pruning; Furthermore, before pruning the network parameters, other network parameters are used to merge and learn the information in the redundant network parameters, thereby further reducing the impact of the redundant network parameter pruning on the network detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] By reading 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 accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein: Figure 1The present invention is a flowchart of an intelligent control method for an audible and visual alarm according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0021] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0022] See also Figure 1 , which shows a flowchart of a method for intelligently controlling an audible and visual alarm provided by an embodiment of the present invention, the method comprising the following steps: S1: Obtain a data set, wherein the data set includes several batches of fire monitoring images.

[0023] Specifically, the image of the area to be monitored is collected by using a camera and recorded as a fire monitoring image; The fire monitoring images are manually labeled, and the fire monitoring images with fire are labeled as 1, and the fire monitoring images without fire are labeled as 0.

[0024] N fire monitoring images are taken as a batch to obtain several batches of fire monitoring images. N represents a preset first number. This embodiment takes N as 50 as an example for description. Other embodiments may take other values, which are not specifically limited in this embodiment.

[0025] S2: Use the dataset to perform a round of training on the pre-built fire monitoring network to obtain the first-stage network.

[0026] It should be noted that since the network parameters of the neural network are randomly generated in the initial stage, the network parameters of the neural network have no fire detection function in the initial stage; therefore, the importance of each network parameter to 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 fire detection function, so as to determine which network parameters are important for fire detection.

[0027] Preferably, as an example, a pre-built fire monitoring network is trained for one round using a data set to obtain a first-stage network, including: A pre-built fire monitoring network is obtained. The fire monitoring network of this embodiment adds an inter-image downsampling operation before the fully connected layer of the traditional VGG network. 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 repeated here. The following mainly introduces the content of inter-image pooling.

[0028] The implementation process of the inter-image pooling includes: obtaining all feature maps to be input to the fully connected layer as feature maps to be analyzed, and taking any pixel position as the target position; obtaining pixel values ​​at all target positions in all feature maps to be analyzed, and recording the sequence composed of all obtained pixel values ​​as the sequence to be analyzed; using a convolution kernel of a preset size to perform sliding convolution on the sequence to be analyzed, and then performing pooling on the sequence to be analyzed after convolution 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 recorded as 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 , other embodiments may take other values, and this embodiment does not make any specific limitations.

[0029] It should be noted that since the number of network parameters in the fully connected layer is fixed, the dimension of the data input to the fully connected layer is fixed. The inter-image pooling process is added here to reduce the dimension of the data input to the fully connected layer, in order to prevent the problem of reducing the dimension of the data input to the fully connected layer due to the subsequent removal of some convolutional layer network parameters.

[0030] It should be further explained that the above convolution and pooling processing are existing knowledge and will not be repeated here.

[0031] The fire monitoring network is trained in sequence using the fire monitoring images in the dataset to obtain the first stage network.

[0032] It should be noted that using images to train a network is an existing technology and will not be described in detail here.

[0033] S3: Input a batch of fire monitoring images into the first-stage network for computational processing, randomly extract and discard the feature maps corresponding to some 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 by whether there is any network parameter corresponding to the feature map, and use the loss value to supervise the first-stage network training; obtain the importance of each network parameter obtained from a preset number of batches of fire monitoring images and calculate the average to obtain the comprehensive importance of each network parameter; record the first-stage network training after training the preset number of batches of fire monitoring images as the second-stage network.

[0034] It should be noted that in order to prune the network, it is necessary to judge the importance of network parameters. The following mainly introduces the relevant content of judging the importance of network parameters.

[0035] It should be further explained that the information obtained by the convolutional network parameters in the network is mainly presented through the feature map. If the feature map obtained by the convolutional network parameters is removed, the effect of the convolutional network parameters will be truncated. Therefore, the importance of each convolutional network parameter to fire detection can be judged by analyzing the accuracy of fire detection when the feature map obtained by the convolutional network parameters exists or does not exist.

[0036] S30: Input a batch of fire monitoring images into the first-stage network for computational processing, randomly extract feature maps corresponding to some network parameters in each convolutional layer and discard them, obtain the loss value of each fire monitoring image, calculate the importance of the network parameter according to the loss value obtained based on whether there is any network parameter corresponding to the feature map, and use the loss value to supervise the training of the first-stage network.

[0037] S300: A batch of fire monitoring images are input into the first-stage network for computational processing, and feature maps corresponding to some network parameters are randomly extracted and discarded in each convolutional layer to obtain the loss value of each fire monitoring image.

[0038] Preferably, as an example, a batch of fire monitoring images are sequentially input into the first-stage network for computational processing, and 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.

[0039] It should be added that after the feature maps corresponding to some network parameters are removed, the dimension of the data input to the fully connected layer will be smaller. Therefore, the pooling degree of the inter-image pooling process before the fully connected layer can be adjusted to fix the dimension of the data input to the fully connected layer. For example, the dimension of the input data of the fully connected layer is 100, and the feature map is not removed in step S2. At this time, two pooling processes can be performed to obtain input data with a dimension of 100. After removing some feature maps in this step, 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 dimension of the data.

[0040] S301: Calculate the importance of the network parameter according to the loss value obtained by whether there is any network parameter corresponding to the feature map.

[0041] It should be noted that information gain can reflect the influence of the existence or non-existence of the characteristic graph corresponding to each network parameter on the fire detection loss, so the importance of each network parameter can be judged by information gain.

[0042] Preferably, as an example, the importance of the network parameter is calculated based on the loss value obtained based on whether there is any network parameter corresponding to the feature graph.

[0043] The loss values ​​obtained from all fire monitoring images in the batch of fire monitoring images are recorded as loss values ​​to be analyzed, the value ranges of all loss values ​​to be analyzed are obtained, the value ranges are evenly divided into a preset second number of sub-ranges, and the probability that the loss value to be analyzed exists in each sub-range is obtained. This embodiment is described by taking the preset second number of 5 as an example, and other embodiments may take other values, which are not specifically limited in this embodiment.

[0044] Calculate information entropy: , represents the probability that there is a loss value to be analyzed in the tth 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 parameters.

[0045] All the loss values ​​to be analyzed that exist in the characteristic graph corresponding to the network parameter are recorded as the first loss values ​​to be analyzed, and all the loss values ​​to be analyzed that do not exist in the characteristic graph corresponding to the network parameter are recorded as the second loss values ​​to be analyzed. The probability of the existence of the first loss value to be analyzed and the probability of the existence of the second loss value to be analyzed in the loss values ​​to be analyzed are obtained; the probability of the existence of the first loss value to be analyzed in each sub-range and the probability of the existence of the second loss value to be analyzed in each sub-range are obtained.

[0046] Calculate the conditional entropy: , represents the probability that the first loss value to be analyzed exists in the loss values ​​to be analyzed, represents the probability that there is a second loss value to be analyzed in the loss value to be analyzed, represents the probability that the first loss value to be analyzed exists in the t-th sub-range, represents the probability that the second loss value to be analyzed exists in the t-th sub-range, Y represents the number of sub-ranges, Represents the conditional entropy of the network parameters.

[0047] The information gain of the network parameter is obtained by subtracting the information entropy X0 from the conditional entropy X1, and the information gain is used as the importance of the network parameter.

[0048] It should be noted that information gain is an existing technology, which is used to reflect the influence of the determination of each information on the final result, so it can reflect the influence of each network parameter on fire detection, in other words, it can reflect the importance of each network parameter.

[0049] S302: Utilization loss The first stage network is trained with supervised values.

[0050] It should be noted that using loss values ​​to supervise the training of the first-stage network is an existing technology and will not be described in detail here.

[0051] S31: Obtain the importance of each network parameter obtained from a preset number of batches of fire monitoring images and calculate the average to obtain the comprehensive importance of each network parameter; record the first stage network training after the preset number of batches of fire monitoring images as the second stage network.

[0052] It can be understood that the importance of each network parameter is judged by analyzing the network loss when each network parameter exists and does not exist, which provides a basis for the subsequent accurate pruning of network parameters.

[0053] S4: Filter out some redundant network parameters based on comprehensive importance; continue to train the second-stage network with fire monitoring images until the training is completed, prune the redundant network parameters, and use the pruned network to control the sound and light alarm; in the process of training the second-stage network with fire monitoring images, let other network parameters merge and learn the feature map information extracted by the redundant network parameters.

[0054] S40: Filter out some redundant network parameters based on comprehensive importance.

[0055] Preferably, as an example, some redundant network parameters are screened out based on comprehensive importance, including: The comprehensive importance is normalized, and the network parameters whose normalized values ​​of comprehensive importance are less than the preset importance threshold are recorded as redundant network parameters.

[0056] It can be understood that a smaller comprehensive importance indicates that the network parameter is not important, and the removal of the network parameter has little effect on the network accuracy. Therefore, unimportant redundant network parameters can be screened out according to the comprehensive importance.

[0057] S41: Continue to train the second-stage network using the fire monitoring image until the training is completed, and prune the redundant network parameters.

[0058] S410: Continue training the second-stage network using the fire monitoring images until the training is completed.

[0059] It should be noted that although redundant network parameters are less important, they will also affect network accuracy to a certain extent. Therefore, if the information learned by the redundant network parameters is merged and learned by other network parameters, the impact of these redundant network parameters on network accuracy will be relatively reduced after they are removed. Therefore, a loss function needs to be added here to allow the loss function to control other network parameters to merge and learn the information of redundant network parameters.

[0060] In the process of training the second-stage network using fire monitoring images, other network parameters are allowed to merge and learn the feature map information extracted by redundant network parameters.

[0061] Preferably, as an example, other network parameters are allowed to merge and learn the feature map information extracted by the redundant network parameters, including: 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; the feature maps corresponding to the redundant network parameters and the feature maps corresponding to the merged learning network parameters are obtained.

[0062] Construct the merged learning loss function:

[0063] 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, Represents a preset adjustment coefficient, which is used to adjust the learning degree of the merged learning network parameters to the feature map information of the redundant network parameters. Take 0.1 as an example for description, other embodiments may take other values, and this embodiment does not specifically limit it. Represents the merged learning loss function.

[0064] Understandably, It reflects the difference between the feature map learned by the merged learning network parameters and the feature map learned by the redundant network parameters. If the merged learning network parameters are allowed to learn the information related to the redundant network parameters, the difference between the feature map learned by the merged learning network parameters and the feature map learned by the redundant network parameters should be small. It reflects the importance of redundant network parameters compared to merged learning network parameters. This value is used to adjust the degree to which merged learning network parameters learn the feature graph information of redundant network parameters. The larger the value, the more important the redundant network parameters are. The merged learning network parameters should learn more feature graph information of redundant network parameters.

[0065] 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.

[0066] It should be noted that the use of loss values ​​and the loss values ​​obtained by merging the 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 repeated here.

[0067] S411: Pruning redundant network parameters.

[0068] Preferably, as an example, pruning redundant network parameters 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.

[0069] S42: Use the post-pruning fire monitoring network to control the sound and light alarms.

[0070] Preferably, as an example, using the post-pruning fire monitoring network to control the sound and light alarm includes: Fire detection processing is performed on the fire monitoring image collected by the sound and light alarm to obtain a detection result. If the detection result is 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 is controlled not to sound a sound and light alarm.

[0071] At this point, this embodiment is completed.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in 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 an audible and visual alarm according to claim 1, 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