Fire prevention early warning method and system for mountain photovoltaic power stations based on visual analysis

Abstract

The present application provides a fire prevention and early warning method and system for mountain photovoltaic power stations based on visual analysis, which relates to the field of intelligent fire protection technology in power stations. The method includes: based on edge nodes in each key monitoring area in the mountain photovoltaic power station, collecting real-time monitoring images and real-time smoke data of the corresponding key monitoring areas; sequentially detecting fire detection results based on lightweight models and high-precision models in edge nodes; in the case of determining that the fire detection result is a fire hazard, triggering the central server to receive historical monitoring images and historical smoke data from edge nodes, and construct monitoring image time series data and smoke time series data; inputting image time series features and smoke time series data into the fire time series analysis model to determine the fire identification result; when there is a fire, executing a fire warning operation. Thus, the edge network and multi-level model are used to implement multi-level and multi-dimensional data analysis to achieve rapid identification and early warning of early fires in power stations.

Description

Technical Field

[0001] The present application relates to the field of intelligent fire protection technology for power stations, and in particular to a fire prevention and warning method and system for mountain photovoltaic power stations based on visual analysis. Background Art

[0002] Mountain photovoltaic power stations refer to photovoltaic power stations built in mountainous or hilly areas. Compared with photovoltaic power stations in plain areas, mountain photovoltaic power stations have the characteristics of complex terrain, changeable climate and dense vegetation. Once a fire occurs in the mountains, it will cause the fire to spread rapidly, making it more difficult to extinguish.

[0003] At present, the fire prevention and early warning of mountain photovoltaic power stations mainly relies on smoke sensing technology and manual inspections. However, on the one hand, in the open and well-ventilated environment of mountain photovoltaic power stations, it is difficult for smoke sensors to quickly detect diluted smoke, resulting in delayed detection of fires. In addition, factors such as dust and fog in the environment may cause false alarms from sensors, affecting the reliability of the system. On the other hand, regular patrol inspections of photovoltaic power stations by manual inspections are difficult to achieve all-weather real-time monitoring due to the limited frequency of manual inspections, and are not conducive to timely detection of fire hazards, which can easily miss the best time to put out fires.

[0004] In response to the above problems, the industry has not yet proposed a better technical solution. Summary of the invention

[0005] The present application provides a fire prevention and warning method, system, storage medium, computer program product and electronic device for a mountain photovoltaic power station based on visual analysis, which is used to at least solve the problems of delayed fire detection and false fire alarms caused by fire detection technology that relies on manual inspections or one-dimensional smoke sensor information in the current related technology.

[0006] In the first aspect, an embodiment of the present application provides a fire prevention and warning method for a mountain photovoltaic power station based on visual analysis, including: based on edge nodes in each key monitoring area in the mountain photovoltaic power station, collecting real-time monitoring images and real-time smoke data of the corresponding key monitoring areas; a lightweight model and a high-precision model are respectively deployed in each of the edge nodes; the lightweight model uses MobileNet as the backbone network, and the high-precision model uses ResNet50 as the backbone network; based on the first edge node, data preprocessing is performed on each of the collected real-time monitoring images to obtain a corresponding standardized real-time monitoring image; based on the lightweight model deployed in the first edge node, identifying whether there is a potential fire pixel block in each of the standardized real-time monitoring images, and cropping and enlarging the image for the potential fire pixel block to determine the corresponding suspected fire area image; based on the suspected fire area image, triggering the call to the deployed in the first edge node a high-precision model so that the high-precision model detects the fire detection result corresponding to the image of the suspected fire area; when it is determined that the fire detection result is that there is a fire hazard, the central server is triggered to perform the following operations: receiving multiple historical monitoring images and multiple historical smoke sensing data for the corresponding first key monitoring area from the first edge node, and constructing monitoring image time series data based on the multiple historical monitoring images and the real-time monitoring images, and constructing smoke sensing time series data based on the multiple historical smoke sensing data and the real-time smoke sensing data; extracting image time series features corresponding to the monitoring image time series data, the image time series features including color time series features, texture time series features and motion time series features; inputting the image time series features and the smoke sensing time series data into the fire time series analysis model to determine the fire recognition result corresponding to the first key monitoring area; when the fire recognition result is that there is a fire, performing a fire warning operation for the first key monitoring area.

[0007] In the second aspect, the embodiment of the present application provides a fire prevention and warning system for a mountain photovoltaic power station based on visual analysis, including: an edge acquisition unit, which is used to collect real-time monitoring images and real-time smoke data of corresponding key monitoring areas based on edge nodes in each key monitoring area in the mountain photovoltaic power station; a lightweight model and a high-precision model are respectively deployed in each of the edge nodes; the lightweight model uses MobileNet as the backbone network, and the high-precision model uses ResNet50 as the backbone network; an edge preprocessing unit, which is used to perform data preprocessing on each of the real-time monitoring images collected based on the first edge node to obtain a corresponding standardized real-time monitoring image; a lightweight edge recognition unit, which is used to identify whether there is a potential fire pixel block in each of the standardized real-time monitoring images based on the lightweight model deployed in the first edge node, and to crop and enlarge the image for the potential fire pixel block to determine the corresponding suspected fire area image; a high-precision edge recognition unit, which is used to Triggering the call of the high-precision model deployed in the first edge node so that the high-precision model detects the fire detection result corresponding to the image of the suspected fire area; the central risk identification unit is used to trigger the central server to perform the following operations when it is determined that the fire detection result is a fire hazard: receiving multiple historical monitoring images and multiple historical smoke data for the corresponding first key monitoring area from the first edge node, and constructing monitoring image time series data based on the multiple historical monitoring images and the real-time monitoring images, and constructing smoke sensing time series data based on the multiple historical smoke sensing data and the real-time smoke sensing data; extracting image time series features corresponding to the monitoring image time series data, the image time series features including color time series features, texture time series features and motion time series features; inputting the image time series features and the smoke sensing time series data into the fire time series analysis model to determine the fire identification result corresponding to the first key monitoring area; when the fire identification result is that there is a fire, performing a fire warning operation for the first key monitoring area.

[0008] According to a third aspect, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the steps of the fire warning method for a mountain photovoltaic power station based on visual analysis according to any embodiment of the present application.

[0009] In a fourth aspect, an embodiment of the present application provides a storage medium on which a computer program is stored, characterized in that when the program is executed by a processor, the steps of the fire prevention and warning method for a mountain photovoltaic power station based on visual analysis of any embodiment of the present application are implemented.

[0010] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the fire prevention and warning method for a mountain photovoltaic power station based on visual analysis of any embodiment of the present application.

[0011] The fire prevention and early warning method for mountain photovoltaic power stations based on visual analysis provided by this application can produce at least the following technical effects:

[0012] (1) By deploying edge nodes in key monitoring areas of mountain photovoltaic power plants, real-time monitoring images and smoke data are collected, and a hierarchical detection method combining lightweight models and high-precision models is used to quickly identify potential fire pixel blocks and further accurately detect fire conditions. This hierarchical detection strategy can not only quickly detect fire conditions under diluted smoke conditions, but also reduce false alarms caused by environmental dust and fog, thereby improving the detection sensitivity and accuracy of the system.

[0013] (2) By adopting the edge computing architecture, lightweight models (with MobileNet as the backbone network) and high-precision models (with ResNet50 as the backbone network) are deployed on edge nodes, so that both preliminary detection and fine detection can be completed on edge nodes, reducing data transmission delays, achieving real-time monitoring and early warning of fire conditions, and avoiding missing the best time to extinguish fires.

[0014] (3) Preliminary screening is performed through a lightweight model, and only when a potential fire pixel block is detected will the high-precision model be called for detailed detection, which greatly reduces the system's computing resource consumption. Compared with directly using a high-precision model or a central server to analyze and process all monitoring images, it saves system resources and effectively improves the system's recognition efficiency.

[0015] (4) After the fire hazard is determined, the central server receives historical monitoring images and smoke sensor data, combines real-time data to construct time series data, and conducts a comprehensive analysis of the time series features through the fire time series analysis model. The image time series features (color time series features, texture time series features, and motion time series features) and smoke sensor time series features are extracted, and the fire identification results are comprehensively judged using multi-dimensional features, avoiding possible misjudgments caused by a single data source and enhancing the reliability of fire identification.

[0016] (5) Through the analysis of the time series characteristics of images and smoke sensor data, the dynamic changes of the fire situation are fully considered, and the early warning capability of mountain photovoltaic power station fires is improved. The image time series characteristics include multi-dimensional information such as color, texture, and motion, and are integrated with the smoke sensor time series data. The fire time series analysis model can learn the characteristic evolution pattern of the dynamic time series of fires, making the fire identification results more comprehensive and accurate.

[0017] (6) When the fire identification result indicates that there is a fire, the central platform server can automatically trigger the early warning operation, promptly notify relevant personnel and take emergency measures. Based on this kind of automated and intelligent rapid early warning response mechanism, it ensures that appropriate emergency measures can be taken at the early stage of a fire, reducing the risk of fire spread and increasing the success rate of fire extinguishing.

[0018] Through this technical solution, multi-level and multi-dimensional data analysis is implemented using edge networks and multi-level models, which has significant advantages in sensitivity, real-time, resource utilization, reliability, multi-dimensional analysis and intelligence, and solves the problems of smoke sensor detection delay, high false alarm rate and low manual inspection efficiency in the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 A flowchart showing an example of a fire prevention and early warning method for a mountain photovoltaic power station based on visual analysis according to an embodiment of the present application is shown;

[0021] Figure 2 A schematic diagram showing a structural connection of an example of a lightweight model according to an embodiment of the present application is shown;

[0022] Figure 3 A schematic diagram showing a structural connection of an example of a high-precision model according to an embodiment of the present application is shown;

[0023] Figure 4 A schematic diagram of structural connections of an example of a fire time sequence analysis model according to an embodiment of the present application is shown;

[0024] Figure 5 A structural block diagram of an example of a fire warning system for a mountain photovoltaic power station based on visual analysis according to an embodiment of the present application is shown;

[0025] Figure 6 It is a schematic structural diagram of an embodiment of an electronic device of the present application. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0027] In the technical solution of this application, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved shall comply with the provisions of relevant laws and regulations and shall not violate public order and good morals.

[0028] Figure 1 A flowchart of an example of a fire prevention and early warning method for a mountain photovoltaic power station based on visual analysis according to an embodiment of the present application is shown.

[0029] Regarding the executor of the method of the embodiment of the present application, it can be any controller or processor with computing or processing capabilities. Through edge networks, multi-model collaboration, combination of visual analysis and smoke data, combination of real-time and historical data and other technical means, the timeliness, accuracy and reliability of fire prevention warnings in mountain photovoltaic power stations are effectively improved.

[0030] like Figure 1 As shown, in step S111, based on the edge nodes in each key monitoring area in the mountain photovoltaic power station, the real-time monitoring images and real-time smoke data of the corresponding key monitoring area are collected.

[0031] In some embodiments, edge nodes are deployed in each key monitoring area of ​​a mountain photovoltaic power station, and these nodes are responsible for collecting real-time monitoring images and smoke data in the area. Lightweight models and high-precision models are deployed in each edge node. The lightweight model in the edge node uses MobileNet as the backbone network, which realizes real-time data collection and preliminary processing with its efficient computing power and low resource consumption. The high-precision model uses ResNet50 as the backbone network, providing higher accuracy and deeper feature extraction capabilities.

[0032] It should be noted that MobileNet is a lightweight deep convolutional neural network designed to achieve efficient image classification and target detection tasks on resource-constrained devices such as mobile devices and embedded systems. In addition, ResNet50 is a deep convolutional neural network with a clever structure and powerful performance, which is part of the ResNet series of networks.

[0033] Here, the key monitoring areas can be diverse and can be adjusted according to operational needs. Exemplarily, the key monitoring areas can include at least one of the following: inverter area, transformer area, photovoltaic module array area, cable line area, power station control room area and energy storage system area.

[0034] In this embodiment, on the one hand, by deploying edge nodes in each key monitoring area, wide coverage and detailed monitoring of the entire mountain photovoltaic power station is achieved. Based on the distributed monitoring method, the real-time situation of each area can be quickly captured, which improves the sensitivity and coverage of the monitoring system. On the other hand, by deploying lightweight and high-precision models at the edge nodes, efficient collection and processing of real-time monitoring data is achieved, the delay of data transmission is reduced, and the response speed and monitoring accuracy of the system are improved.

[0035] In step S112, data preprocessing is performed on each collected real-time monitoring image based on the first edge node to obtain a corresponding standardized real-time monitoring image.

[0036] It should be understood that the first edge node is not specific and may be any one or more edge nodes in the edge network.

[0037] Through standardization processing, the heterogeneity of image data is eliminated, the consistency and standardization of image data are ensured, and a reliable data basis is provided for subsequent model recognition. It can effectively reduce the impact of external factors such as ambient light changes on image recognition accuracy and improve the stability and accuracy of the model.

[0038] In step S113, based on the lightweight model deployed in the first edge node, it is identified whether there are potential fire pixel blocks in each standardized real-time monitoring image, and the potential fire pixel blocks are cropped and enlarged to determine the corresponding suspected fire area image.

[0039] Here, the introduction of lightweight models enables rapid preliminary analysis of large-scale monitoring images, and can efficiently screen out areas where fires may exist. By cropping and enlarging suspected fire areas, the amount of data is reduced, saving computing resources for subsequent in-depth detection of high-precision models and improving processing efficiency.

[0040] In step S114, the high-precision model deployed in the first edge node is triggered to be called based on the suspected fire area image, so that the high-precision model detects the fire detection result corresponding to the suspected fire area image.

[0041] In some embodiments, while generating an image of a suspected fire area, a high-precision model (ResNet50) is triggered to perform further fire detection. The fire detection result can provide more detailed or refined fire detection results, including whether a fire exists and the specific characteristics of the fire.

[0042] Here, by using a high-precision model, the accuracy of fire detection is improved, fires can be identified more accurately, and the false alarm rate can be reduced. Therefore, by using lightweight models and high-precision models in the edge nodes, a combination of fast detection and accurate identification is achieved, significantly improving the overall fire detection performance for the corresponding key monitoring areas.

[0043] When the fire detection result is determined to be a fire hazard, the central server is triggered to perform the following operations:

[0044] In step S121, multiple historical monitoring images and multiple historical smoke sensing data for the corresponding first key monitoring area are received from the first edge node, and monitoring image time series data are constructed based on the multiple historical monitoring images and real-time monitoring images, and smoke sensing time series data are constructed based on the multiple historical smoke sensing data and real-time smoke sensing data.

[0045] Here, after the first edge node determines that the fire detection result is that there is a fire hazard, the central server receives multiple historical monitoring images and historical smoke data of the corresponding monitoring area from the first edge node, and constructs monitoring image time series data and smoke time series data, thereby realizing the construction of complete and continuous time series data by combining historical data and real-time data.

[0046] In step S122, image time series features corresponding to the monitoring image time series data are extracted, and the image time series features include color time series features, texture time series features and motion time series features.

[0047] It should be noted that the color time series features, texture time series features and motion time series features comprehensively reflect the dynamic changes of the fire situation in the monitoring area. The extraction of multiple time series features provides comprehensive data support for the fire time series analysis model, helps to accurately capture the occurrence and development patterns of fires, improves the accuracy and sensitivity of fire identification, and supports more effective distinction between fires and other environmental changes in complex environments.

[0048] In some examples of the embodiments of the present application, in view of the obvious color features of flames and smoke in the image, the monitoring images are converted into HSV color space respectively, and the corresponding color distribution features are extracted at each time step to determine the corresponding color time series features. Thus, extracting the characteristics of the color distribution of the monitoring image changing over time helps to identify the color changes caused by the fire, such as the appearance of smoke and flames. In addition, in view of the fact that the smoke and flames generated by the fire have specific texture features, the local binary pattern algorithm is applied to analyze the change information of the local texture features in the monitoring image over time to determine the corresponding texture time series features. Thus, by analyzing the subtle texture changes in the image, the texture features of the smoke and flames in the image at the initial stage of the fire can be identified, providing a more fine-grained fire detection capability. In addition, in view of the fact that smoke and flames have obvious dynamic change rules when a fire occurs, the motion vector of the pixel points in the monitoring image is calculated by the optical flow method to determine the corresponding motion time series features. Thus, by analyzing the motion of the pixel points in the image, the dynamic changes of flames and smoke can be identified, and the monitoring capability of the fire development process can be improved.

[0049] In step S123, the image time series features and smoke sensor time series data are input into a fire time series analysis model to determine a fire identification result corresponding to the first key monitoring area.

[0050] The time series analysis model can comprehensively analyze the time series changes of image features and smoke sensor data to provide more accurate fire identification results. Through the joint analysis of multi-dimensional time series data, the system can detect fires earlier and reduce false alarms, thereby enhancing the timeliness and accuracy of early warnings.

[0051] In step S124, when the fire identification result is that there is a fire, a fire warning operation is performed for the first key monitoring area.

[0052] In some embodiments, when the fire identification result is that there is a fire, the system automatically performs a fire warning operation, sends a warning message to relevant personnel (for example, the administrator responsible for maintaining the power station or the corresponding key monitoring area), and triggers the corresponding emergency response mechanism. Thus, the timely fire warning operation ensures the rapid response and handling of fire hazards, reducing the losses and harms that may be caused by the fire.

[0053] Through this embodiment, a multi-level detection framework is innovatively proposed, combining lightweight models and high-precision models to achieve the unity of high efficiency and high precision. The color, texture and motion features of image recognition are integrated to improve the comprehensive ability of fire detection and adapt to the changing mountain environment. In addition, by using edge computing for local fast processing and combining the powerful computing and storage capabilities of the cloud platform, a balance between real-time and large-scale data processing is achieved. As a result, not only the accuracy and response speed of fire detection are improved, but also the adaptability of the system in complex environments is enhanced, providing reliable protection for the safe operation of photovoltaic power stations.

[0054] In the practice of this application, combined with the characteristics of the fire monitoring and early warning scenario of mountain photovoltaic power stations, we also carried out a series of enhanced designs on each model in the multi-level detection framework to improve its advantages in this scenario. More details will be expanded in conjunction with other examples below.

[0055] Figure 2 A schematic diagram of the structural connection of an example of a lightweight model according to an embodiment of the present application is shown.

[0056] like Figure 2 As shown, the lightweight model 200 adopts the MobileNet based on the attention mechanism, which includes a cascaded input layer 210, a convolutional layer 220, a CBAM (Convolutional Block Attention Module) module 230, a pooling layer 240 and an output layer 250.

[0057] The convolution layer 220 is used to receive the input feature map corresponding to the standardized real-time monitoring image, and independently perform convolution operations on each channel of the input feature map, and perform convolution operations on the output of the deep convolution. Convolution operation is used to determine the corresponding channel integration features by integrating channel information:

[0058] , Formula (1)

[0059] , Formula (2)

[0060] In the formula, represents the input feature map, represents the depth convolution kernel, Represents the convolution operation; represents the output of the depthwise convolution, represents the point-wise convolution kernel, Indicates channel integration characteristics;

[0061] It should be noted that MobileNet is a lightweight convolutional neural network that uses deep separable convolution to reduce the amount of calculation and parameters. In this embodiment, after the CBAM module is integrated into each convolution block of MobileNet, channel attention adjustment is performed first, and then spatial attention adjustment is performed.

[0062] More specifically, the CBAM module 230 includes a cascaded channel attention module 231 and a spatial attention module 232. The channel attention module 231 is used to generate channel features through global pooling, and generate channel weights through a fully connected layer, and use the channel weights to perform weighted calculations on the feature map output by the convolution layer to obtain the corresponding channel attention feature map:

[0063] , Formula (3)

[0064] , Formula (4)

[0065] , Formula (5)

[0066] , Formula (6)

[0067] In the formula, and Represent the height and width of the feature map respectively; and Represent the results of global average pooling and global maximum pooling respectively; and Represent the weights of the fully connected layer respectively; ReLU activation function. represents the Sigmoid activation function, represents the channel weight, Represents the channel attention feature map.

[0068] Based on the channel attention module, by re-weighting the importance of each channel, it can better highlight key features such as flames and smoke, thereby improving the detection accuracy of fire areas. In addition, it can better adapt to fire detection under different lighting conditions and improve the robustness under complex lighting conditions such as shadows and reflections.

[0069] The spatial attention module 232 is used to generate spatial features for the channel attention feature map in the channel dimension through maximum pooling and average pooling, and then generate spatial weights through convolution operations, and perform weighted calculation on the spatial weights and the channel attention feature map to obtain the CBAM feature map:

[0070] , Formula (7)

[0071] , Formula (8)

[0072] , Formula (9)

[0073] , Formula (10)

[0074] In the formula, and Respectively represent the average pooling results and maximum pooling results in the channel dimension; represents the convolution kernel of the spatial attention module, represents the spatial weight, Represents the CBAM feature map.

[0075] Based on the spatial attention module, by weighting the spatial position of the feature map, the specific area where the fire occurs can be located more accurately, reducing false alarms and missed alarms.

[0076] The pooling layer 240 is used to perform a pooling operation on the CBAM feature map to obtain a corresponding pooled feature map.

[0077] The output layer 250 is used to identify whether there is a potential fire pixel block based on the pooled feature map.

[0078] Through the embodiments of the present application, the depth-separable convolution and point-by-point convolution are integrated into the lightweight MobileNet model, so that the MobileNet model still has a strong feature extraction capability while maintaining efficient computing. In addition, by integrating a relatively lightweight CBAM module, the computational complexity of the model will not be significantly increased, ensuring that the model can run efficiently on resource-constrained embedded devices and is suitable for real-time fire detection needs. Therefore, by integrating the CBAM module in MobileNet, the system can significantly improve the accuracy, robustness and real-time performance of detecting signs of fire in the early stages of a fire without significantly increasing the computational burden, and can ensure the efficient operation of the system in complex environments and changing climate conditions, effectively reduce false alarms, detect fires in a timely manner and issue warnings, and ensure the safe operation of photovoltaic power stations.

[0079] Figure 3 A schematic diagram of the structural connection of an example of a high-precision model according to an embodiment of the present application is shown.

[0080] like Figure 3 As shown, the high-precision model adopts the ResNet50 model based on hybrid convolution, which includes a cascaded input layer 310, a ResNet50 feature extraction module 320, a hybrid convolution module 330 and an output layer 340.

[0081] The input layer 310 is used to receive images of suspected fire areas. .

[0082] The ResNet50 feature extraction module 320 is used to extract multi-level image features corresponding to the suspected fire area image:

[0083] , Formula (11)

[0084] In the formula, Represents the multi-level image features extracted by ResNet50.

[0085] Here, the ResNet50 feature extraction module uses the residual network in ResNet50 to extract multi-level features of images of suspected fire areas, ensuring high-quality feature representation and effectively capturing detailed information in the image, thereby improving the accuracy of fire detection.

[0086] The hybrid convolution module 330 is used to perform hybrid convolution processing on multi-level image features to obtain corresponding hybrid convolution image features:

[0087] , Formula (12)

[0088] , Formula (13)

[0089] , Formula (14)

[0090] , Formula (15)

[0091] In the formula, Indicates The standard convolution output of the layer is at position The value of Indicates The standard convolution kernel of the layer corresponds to the convolution window The weight matrix at Indicates The weights and biases of the standard convolution kernels of the layer; Indicates The layer depthwise convolution output is at position The value of Indicates The weight matrix of the layer depth convolution kernel in channel 𝑐; Indicates The convolution output of the layer is at position The value of Indicates The weight matrix of the layer point convolution kernel in channel 𝑐, Indicates The bias term of the layer point convolution kernel; Indicates The layer mixed convolution output is at position The value of .

[0092] The hybrid convolution module combines standard convolution and depthwise separable convolution, taking full advantage of both to further enhance feature extraction capabilities. Standard convolution can capture global features, while depthwise separable convolution improves computational efficiency and feature extraction accuracy. In addition, the hybrid convolution strategy generates a richer feature representation by adding the standard convolution output and the point convolution output, improving the overall performance of the model.

[0093] The output layer 340 is used to determine the corresponding fire detection result according to the mixed convolution image features:

[0094] , Formula (16)

[0095] , Formula (17)

[0096] , Formula (18)

[0097] In the formula, Represents the mixed convolution image features finally output by the mixed convolution module, represents the flattened eigenvector, and Represent the weight matrix and bias term of the output layer respectively; Represents the linear transformation result of the output layer; Indicates the fire detection result.

[0098] In the embodiment of the present application, through the multi-level features extracted by ResNet50 and the hybrid convolution features processed by the hybrid convolution module, the model can generate a feature vector with stronger expressive power. In addition, the efficient hybrid convolution strategy is combined with the ResNet50 feature extraction module, so that the model can quickly process the input image, generate fire detection results, realize the enhancement of real-time detection capabilities, and significantly reduce the delay in the fire detection process.

[0099] Figure 4 A schematic diagram of structural connections of an example of a fire sequence analysis model according to an embodiment of the present application is shown.

[0100] like Figure 4 As shown, the fire time series analysis model 400 includes an input layer 410, a multi-scale convolutional pooling module 420, a bidirectional LSTM module 430, a multi-head self-attention module 440, a feature fusion layer 450 and an output layer 460.

[0101] The input layer 410 is used to receive image temporal features and smoke sensor time series data ,in Contains color timing features, texture timing features and motion timing features.

[0102] The multi-scale convolution pooling module 420 is used to extract the image features of the input image features at each time step in the image temporal features at different scales, and perform pooling to reduce the feature map size:

[0103] , Formula (19)

[0104] , formula (20)

[0105] , Formula (21)

[0106] , Formula (22)

[0107] , Formula (23)

[0108] , Formula (24)

[0109] In the formula, Represents convolution kernels of different scales, and Respectively The height and width of the convolution kernel; , and Respectively represent Layer for the The convolution kernel of each scale is located at The color features, texture features and motion features output at represents the activation function; , and Respectively represent Tier The convolution kernels of different scales correspond to the convolution windows for color features, texture features, and motion features. The convolution kernel weight matrix; , and They represent the bias terms of the 𝑘th scale convolution kernel of the lth layer for color features, texture features, and motion features respectively; Indicates that in the pooling window Perform the maximum pooling operation inside; , and Respectively represent the time step t The input image features correspond to the Layer input color features, input texture features and input motion features at position The value at , and Respectively represent Tier The convolution kernel of each scale is located at The output eigenvalue at ; , and Respectively expressed in Layer in position Output color features, output texture features, and output motion features after the pooling operation.

[0110] By extracting color features, texture features and motion features from image temporal features through multi-scale convolutional pooling layers, image features can be processed at different scales, the accuracy and diversity of feature extraction can be improved, and image information can be fully captured.

[0111] The bidirectional LSTM module 430 is used to process the smoke sensor time series data to capture the temporal dependencies. The data at each moment is processed by the forward and backward LSTM layers respectively, and then the outputs of the two are connected together to form the final time series features:

[0112] , Formula (25)

[0113] , Formula (26)

[0114] , Formula (27)

[0115] In the formula, Indicates that at time step t Smoke sensor data, Respectively represent the time step t The hidden states of the forward LSTM and backward LSTM; Respectively represent the time step t The cell states of the forward LSTM and backward LSTM; represents the feature concatenation operation, Represents bidirectional LSTM at time step t The final hidden state.

[0116] By processing smoke sensor time series data through the bidirectional LSTM module, the temporal dependencies can be captured, effectively improving the ability to extract time series features of smoke sensor data. It has a significant advantage in capturing the temporal dependencies of time series data, making the model more accurate when processing time series data.

[0117] The multi-head self-attention module 440 is used to use multiple attention heads to focus on the multi-scale image features output by the multi-scale convolutional pooling layer. Perform attention weighted calculation to obtain image attention features:

[0118] , formula (28)

[0119] , Formula (29)

[0120] , formula (30)

[0121] , Formula (31)

[0122] In the formula, The query matrix, key matrix and value matrix represent the image features respectively; represents the attention mechanism function, Represents a multi-head self-attention mechanism; The weight matrices representing query, key, and value respectively; represents the output weight matrix; Image attention features representing the final output of the multi-head attention mechanism.

[0123] The multi-head self-attention module performs weighted calculations on multi-scale image features through multiple attention heads. It can process features in parallel on different attention heads and capture important information in image features, which significantly enhances the model's ability to recognize and analyze image features and improves the overall performance and efficiency of the model.

[0124] The feature fusion layer 450 is used to fuse the smoke features processed by the bidirectional LSTM module and the image features processed by the multi-head self-attention module to obtain multimodal fusion features:

[0125] , Formula (32)

[0126] In the formula, Represents multimodal fusion features.

[0127] The processed smoke features and image features are fused through the feature fusion layer to make full use of the information of multimodal data and improve the accuracy of fire identification.

[0128] The output layer 460 is used to generate fire identification results based on the multimodal fusion features:

[0129] , Formula (33)

[0130] , formula (34)

[0131] , formula (35)

[0132] Where SELU represents the SELU activation function, and is a preset constant; Indicates the fire identification result. Represents the softmax function.

[0133] The output layer uses the SELU activation function, which makes the neural network converge faster during the training process, and generates the probability distribution of fire risk through the softmax function to ensure the reliability and stability of the classification results. Specifically, the SELU activation function automatically normalizes the data to keep the mean of zero and the variance of one, avoiding the problems of gradient vanishing and gradient exploding, and improving the training stability of the deep neural network. In addition, by integrating multimodal data and multiple feature processing modules, the model has stronger generalization ability in different environments and scenarios, and can adapt to different types of fire warning needs.

[0134] Through the above-mentioned enhanced design for the basic timing model, the multi-scale convolutional pooling layer can extract different scale image features of color, texture and motion timing features respectively. The multi-head self-attention mechanism and the bidirectional LSTM layer process these features and capture timing information. The feature fusion layer effectively fuses multimodal features. The output layer is enhanced with the SELU activation function to ensure the efficiency and accuracy of the model in fire prevention and early warning of mountain photovoltaic power stations.

[0135] In some examples of the embodiments of the present application, a novel loss function is provided for the structural design of the above-mentioned fire time series analysis model, which combines the classification loss and the consistency loss of the time series features to improve the accuracy and robustness of the model.

[0136] More specifically, the loss function of the fire time series analysis model is for:

[0137] , Formula (36)

[0138] , Formula (37)

[0139] , formula (38)

[0140] , formula (39)

[0141] In the formula, , and They represent classification loss, temporal consistency loss and multimodal fusion loss respectively. , and Respectively represent the corresponding loss item weight coefficients; Represents the total number of samples in the data sample set. Indicates the total number of fire classification categories, which can be 2 (i.e., whether there is a fire) or other integers with larger values ​​(for example, to distinguish the fire stage or the size of the fire). It can be set or adjusted according to the business application scenario or user needs. Representation sample In category The real label, Represents the sample predicted by the model In category The probability of represents the total number of time steps, Representation sample At time step The hidden state of Representation sample The multimodal fusion features of and Respectively represent samples The corresponding image features and smoke features.

[0142] In the embodiment of the present application, through the classification loss, the model can more accurately identify fire and non-fire situations, so that the model can effectively reduce false alarms and missed alarms in practical applications, thereby improving the reliability of the fire warning system. Because the development of fire is continuous and time-dependent, through the time consistency loss, the model can better handle continuous time series data, constrain the differences between hidden states of consecutive time steps, ensure the smoothness and consistency of the prediction results, and reduce the volatility of the prediction results. Through the multimodal fusion loss, the model can make full use of the information of image features and smoke features, constrain the consistency between the fused feature representation and the separate image features and smoke features, and the effective fusion of multimodal data can enhance the model's adaptability to complex scenes and improve the accuracy and robustness of fire identification. A comprehensive loss function based on the above three losses is adopted, and the weight coefficients in the loss function can be adjusted. , and , the impact of each part of the loss on the total loss can be flexibly controlled to adapt to different application requirements and scenarios. As a result, the model can be optimized in different aspects, with better generalization ability and accuracy, and can provide stable performance in complex fire detection scenarios.

[0143] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of actions combined, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application. In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0144] Figure 5 A structural block diagram of an example of a fire warning system for a mountain photovoltaic power station based on visual analysis according to an embodiment of the present application is shown.

[0145] like Figure 5 As shown, the fire warning system 500 for mountain photovoltaic power stations based on visual analysis includes an edge acquisition unit 510, an edge preprocessing unit 520, a lightweight edge recognition unit 530, a high-precision edge recognition unit 540 and a central risk recognition unit 550.

[0146] The edge acquisition unit 510 is used to collect real-time monitoring images and real-time smoke data of corresponding key monitoring areas based on edge nodes in each key monitoring area in the mountain photovoltaic power station; a lightweight model and a high-precision model are deployed in each of the edge nodes respectively; the lightweight model uses MobileNet as the backbone network, and the high-precision model uses ResNet50 as the backbone network.

[0147] The edge preprocessing unit 520 is used to perform data preprocessing on each of the collected real-time monitoring images based on the first edge node to obtain a corresponding standardized real-time monitoring image.

[0148] The lightweight edge identification unit 530 is used to identify whether there are potential fire pixel blocks in each of the standardized real-time monitoring images based on the lightweight model deployed in the first edge node, and to crop and enlarge the potential fire pixel blocks to determine the corresponding suspected fire area image.

[0149] The high-precision edge recognition unit 540 is used to trigger the call of the high-precision model deployed in the first edge node based on the suspected fire area image, so that the high-precision model detects the fire detection result corresponding to the suspected fire area image.

[0150] The central risk identification unit 550 is used to trigger the central server to perform the following operations when determining that the fire detection result indicates that there is a fire hazard:

[0151] Receiving a plurality of historical monitoring images and a plurality of historical smoke sensing data for a corresponding first key monitoring area from the first edge node, and constructing monitoring image time series data according to the plurality of historical monitoring images and the real-time monitoring images, and constructing smoke sensing time series data according to the plurality of historical smoke sensing data and the real-time smoke sensing data;

[0152] Extracting image time series features corresponding to the monitoring image time series data, wherein the image time series features include color time series features, texture time series features and motion time series features;

[0153] Inputting the image time series features and the smoke sensor time series data into a fire time series analysis model to determine a fire identification result corresponding to the first key monitoring area;

[0154] When the fire identification result is that a fire exists, a fire warning operation is performed for the first key monitoring area.

[0155] In some embodiments, an embodiment of the present application provides a non-volatile computer-readable storage medium, in which one or more programs including execution instructions are stored, and the execution instructions can be read and executed by an electronic device (including but not limited to a computer, a server, or a network device, etc.) to execute the steps of any of the above-mentioned visual analysis-based fire prevention and warning methods for mountain photovoltaic power stations in the present application.

[0156] In some embodiments, the embodiments of the present application also provide a computer program product, which includes a computer program stored on a non-volatile computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer executes any step of the above-mentioned fire prevention and warning method for a mountain photovoltaic power station based on visual analysis.

[0157] In some embodiments, the embodiments of the present application also provide an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the steps of a fire prevention and warning method for a mountain photovoltaic power station based on visual analysis.

[0158] Figure 6 is a schematic diagram of the hardware structure of an electronic device for executing a fire prevention warning method for a mountain photovoltaic power station based on visual analysis provided by another embodiment of the present application, such as Figure 6 As shown, the device includes:

[0159] One or more processors 610 and memory 620, Figure 6 A processor 610 is taken as an example.

[0160] The device for executing the fire prevention warning method for a mountain photovoltaic power station based on visual analysis may further include: an input device 630 and an output device 640 .

[0161] The processor 610, the memory 620, the input device 630 and the output device 640 may be connected via a bus or other means. Figure 6 The example of connecting through bus is taken in the following.

[0162] The memory 620, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer executable programs and modules, such as program instructions / modules corresponding to the fire prevention early warning method for mountain photovoltaic power stations based on visual analysis in the embodiment of the present application. The processor 610 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions and modules stored in the memory 620, that is, realizing the fire prevention early warning method for mountain photovoltaic power stations based on visual analysis in the above method embodiment.

[0163] The memory 620 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 620 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 620 may optionally include a memory remotely arranged relative to the processor 610, and these remote memories may be connected to the electronic device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0164] The input device 630 can receive input digital or character information and generate signals related to user settings and function control of the electronic device. The output device 640 can include a display device such as a display screen.

[0165] The one or more modules are stored in the memory 620, and when executed by the one or more processors 610, the fire prevention and early warning method for a mountain photovoltaic power station based on visual analysis in any of the above method embodiments is executed.

[0166] The above-mentioned product can execute the method provided in the embodiment of the present application, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the method provided in the embodiment of the present application.

[0167] The electronic devices of the embodiments of the present application exist in various forms, including but not limited to:

[0168] (1) Mobile communication equipment: This type of equipment is characterized by having mobile communication functions and its main purpose is to provide voice and data communications. This type of terminal includes: smart phones, multimedia phones, functional phones, and low-end phones.

[0169] (2) Ultra-mobile personal computer devices: These devices belong to the category of personal computers, have computing and processing functions, and generally also have mobile Internet access features. These terminals include: PDA, MID and UMPC devices, etc.

[0170] (3) Portable entertainment devices: These devices can display and play multimedia content. They include audio and video players, handheld game consoles, e-books, smart toys, and portable car navigation devices.

[0171] (4) Other onboard electronic devices with data interaction functions, such as on-board devices installed in vehicles.

[0172] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0173] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a general hardware platform, and of course, by hardware. Based on this understanding, the above technical solution, in essence, or the part that contributes to the relevant technology, can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiment.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A fire prevention early warning method for a mountain photovoltaic power station based on visual analysis, comprising: Based on the edge nodes in each key monitoring area of ​​the mountain photovoltaic power station, real-time monitoring images and real-time smoke data of the corresponding key monitoring areas are collected; A lightweight model and a high-precision model are respectively deployed in each of the edge nodes; the lightweight model uses MobileNet as the backbone network, and the high-precision model uses ResNet50 as the backbone network; the key monitoring area includes at least one of the following: an inverter area, a transformer area, a photovoltaic module array area, a cable line area, a power station control room area, and an energy storage system area in a mountain photovoltaic power station; Performing data preprocessing on each of the collected real-time monitoring images based on the first edge node to obtain a corresponding standardized real-time monitoring image; Based on the lightweight model deployed in the first edge node, identifying whether there are potential fire pixel blocks in each of the standardized real-time monitoring images, and cropping and enlarging the potential fire pixel blocks to determine the corresponding suspected fire area image; Based on the suspected fire area image, triggering the calling of the high-precision model deployed in the first edge node, so that the high-precision model detects the fire detection result corresponding to the suspected fire area image; When it is determined that the fire detection result indicates that there is a potential fire hazard, the central server is triggered to perform the following operations: Receiving a plurality of historical monitoring images and a plurality of historical smoke sensing data for a corresponding first key monitoring area from the first edge node, and constructing monitoring image time series data according to the plurality of historical monitoring images and the real-time monitoring images, and constructing smoke sensing time series data according to the plurality of historical smoke sensing data and the real-time smoke sensing data; Extracting image time series features corresponding to the monitoring image time series data, wherein the image time series features include color time series features, texture time series features and motion time series features; Inputting the image time series features and the smoke sensor time series data into a fire time series analysis model to determine a fire identification result corresponding to the first key monitoring area; When the fire identification result is that there is a fire, performing a fire warning operation for the first key monitoring area; The lightweight model adopts MobileNet based on attention mechanism, which includes a cascaded input layer, a convolutional layer, a CBAM module, a pooling layer and an output layer; The convolution layer is used to receive the input feature map corresponding to the standardized real-time monitoring image, and independently perform convolution operations on each channel of the input feature map, and perform convolution operations on the output of the deep convolution. Convolution operation is used to determine the corresponding channel integration features by integrating channel information: In the formula, represents the input feature map, represents the depth convolution kernel, Represents the convolution operation; represents the output of the depthwise convolution, represents the point-wise convolution kernel, Indicates channel integration characteristics; The CBAM module includes a cascaded channel attention module and a spatial attention module; the channel attention module is used to generate channel features through global pooling, and generate channel weights through a fully connected layer. The channel weights are used to perform weighted calculations on the feature maps output by the convolutional layer to obtain the corresponding channel attention feature maps: In the formula, and Represent the height and width of the feature map respectively; and Represent the results of global average pooling and global maximum pooling respectively; and Represent the weights of the fully connected layer respectively; ReLU activation function. represents the Sigmoid activation function, represents the channel weight, Represents the channel attention feature map; The spatial attention module is used to generate spatial features for the channel attention feature map in the channel dimension through maximum pooling and average pooling, and then generate spatial weights through convolution operations, and perform weighted calculation on the spatial weights and the channel attention feature map to obtain the CBAM feature map: In the formula, and Respectively represent the average pooling results and maximum pooling results in the channel dimension; represents the convolution kernel of the spatial attention module, represents the spatial weight, represents the CBAM feature map; The pooling layer is used to perform a pooling operation on the CBAM feature map to obtain a corresponding pooling feature map; The output layer is used to identify whether there is a potential fire pixel block based on the pooled feature map; The high-precision model adopts a ResNet50 model based on hybrid convolution, which includes a cascaded input layer, a ResNet50 feature extraction module, a hybrid convolution module and an output layer; The input layer is used to receive images of suspected fire areas ; The ResNet50 feature extraction module is used to extract the multi-level image features corresponding to the suspected fire area image: In the formula, Represents the multi-level image features extracted by ResNet50; The hybrid convolution module is used to perform hybrid convolution processing on the multi-level image features to obtain corresponding hybrid convolution image features: In the formula, Indicates The standard convolution output of the layer is at position The value of Indicates The standard convolution kernel of the layer corresponds to the convolution window The weight matrix at Indicates The weights and biases of the standard convolution kernels of the layer; Indicates The layer depthwise convolution output is at position The value of Indicates The weight matrix of the layer depth convolution kernel in channel 𝑐; Indicates The convolution output of the layer is at position The value of Indicates The weight matrix of the layer point convolution kernel in channel 𝑐, Indicates The bias term of the layer point convolution kernel; Indicates The layer mixed convolution output is at position The value of The output layer is used to determine the corresponding fire detection results according to the mixed convolution image features: In the formula, Represents the mixed convolution image features finally output by the mixed convolution module, represents the flattened eigenvector, and Represent the weight matrix and bias term of the output layer respectively; Represents the linear transformation result of the output layer; Indicates the fire detection result.

2. The method according to claim 1, wherein: The extracting the image time series features corresponding to the monitoring image time series data includes: The monitoring images are converted into HSV color space respectively, and the corresponding color distribution features are extracted at each time step to determine the corresponding color temporal features; The local binary pattern algorithm is used to analyze the temporal changes of local texture features in the monitoring image to determine the corresponding texture temporal features. The motion vectors of pixels in the monitoring image are calculated using the optical flow method to determine the corresponding motion timing characteristics.

3. The method according to claim 1, wherein: The fire time series analysis model includes an input layer, a multi-scale convolutional pooling module, a bidirectional LSTM module, a multi-head self-attention module, a feature fusion layer and an output layer; The input layer is used to receive image temporal features and smoke sensor time series data ,in Contains color timing features, texture timing features and motion timing features; The multi-scale convolution pooling module is used to extract the image features of the input image features at each time step in the image temporal features, and pool them at different scales: In the formula, Represents convolution kernels of different scales, and Respectively The height and width of the convolution kernel; , and Respectively represent Layer for the The convolution kernel of each scale is located at The color features, texture features and motion features output at represents the activation function; , and Respectively represent Tier The convolution kernels of different scales correspond to the convolution windows for color features, texture features, and motion features. The convolution kernel weight matrix; , and They represent the bias terms of the 𝑘th scale convolution kernel of the lth layer for color features, texture features, and motion features respectively; Indicates that in the pooling window Perform the maximum pooling operation inside; , and Respectively represent the time step t The input image features correspond to the Layer input color features, input texture features and input motion features at position The value at , and Respectively indicate Tier The convolution kernel of each scale is located at The output eigenvalue at ; , and Respectively expressed in Layer in position Output color features, output texture features, and output motion features after the pooling operation; The bidirectional LSTM module is used to process smoke sensor time series data to capture the temporal dependencies: In the formula, Indicates that at time step t Smoke sensor data, Respectively represent the time step t The hidden states of the forward LSTM and backward LSTM; Respectively represent the time step t The cell states of the forward LSTM and backward LSTM; represents the feature concatenation operation, Represents bidirectional LSTM at time step t The final hidden state of The multi-head self-attention module is used to use multiple attention heads to focus on the multi-scale image features output by the multi-scale convolutional pooling layer. Perform attention weighted calculation to obtain image attention features: In the formula, The query matrix, key matrix and value matrix represent the image features respectively; represents the attention mechanism function, Represents a multi-head self-attention mechanism; weight matrices representing query, key, and value respectively; represents the output weight matrix; Image attention features representing the final output of the multi-head attention mechanism; The feature fusion layer is used to fuse the smoke features processed by the bidirectional LSTM module and the image features processed by the multi-head self-attention module to obtain multimodal fusion features: In the formula, Represents multimodal fusion features; The output layer is used to generate fire identification results based on multimodal fusion features: Where SELU represents the SELU activation function, and is a preset constant; Indicates the fire identification result. Represents the softmax function.

4. The method according to claim 3, wherein: The loss function of the fire time series analysis model is for: In the formula, , and They represent classification loss, temporal consistency loss and multimodal fusion loss respectively. , and Respectively represent the corresponding loss item weight coefficients; represents the total number of samples in the data sample set, Represents the total number of fire classification categories, Representation sample In category The real label, Represents the sample predicted by the model In category probability; represents the total number of time steps, Representation sample At time step The hidden state of Representation sample The multimodal fusion features of and Respectively represent samples The corresponding image features and smoke features.

5. A fire prevention early warning system for a mountain photovoltaic power station based on visual analysis, used to implement the fire prevention early warning method for a mountain photovoltaic power station based on visual analysis as claimed in claim 1; The system includes: The edge acquisition unit is used to collect real-time monitoring images and real-time smoke data of the corresponding key monitoring areas based on the edge nodes in each key monitoring area of ​​the mountain photovoltaic power station; a lightweight model and a high-precision model are respectively deployed in each of the edge nodes; the lightweight model uses MobileNet as the backbone network, and the high-precision model uses ResNet50 as the backbone network; An edge preprocessing unit, configured to perform data preprocessing on each of the real-time monitoring images collected based on the first edge node to obtain a corresponding standardized real-time monitoring image; A lightweight edge recognition unit, configured to identify whether there are potential fire pixel blocks in each of the standardized real-time monitoring images based on the lightweight model deployed in the first edge node, and to crop and amplify the potential fire pixel blocks to determine a corresponding suspected fire area image; A high-precision edge recognition unit, configured to trigger the calling of the high-precision model deployed in the first edge node based on the suspected fire area image, so that the high-precision model detects the fire detection result corresponding to the suspected fire area image; The central risk identification unit is used to trigger the central server to perform the following operations when it is determined that the fire detection result indicates that there is a fire hazard: Receiving a plurality of historical monitoring images and a plurality of historical smoke sensing data for a corresponding first key monitoring area from the first edge node, and constructing monitoring image time series data according to the plurality of historical monitoring images and the real-time monitoring images, and constructing smoke sensing time series data according to the plurality of historical smoke sensing data and the real-time smoke sensing data; Extracting image time series features corresponding to the monitoring image time series data, wherein the image time series features include color time series features, texture time series features and motion time series features; Inputting the image time series features and the smoke sensor time series data into a fire time series analysis model to determine a fire identification result corresponding to the first key monitoring area; When the fire identification result is that a fire exists, a fire warning operation is performed for the first key monitoring area.

Images