1. A transformer substation inspection method based on a cloud side system and video intelligent analysis is characterized by comprising the following steps:

step a), establishing a cloud edge cooperative system framework for substation inspection, which is composed of an inspection host, an inspection edge block and an inspection edge node, wherein the inspection host is used for processing and analyzing large image data of the substation, the inspection edge block is used for gathering, storing, processing and intelligent application of front-end inspection node sensing data and is independent and autonomous, the inspection edge node is used for acquiring and intelligently processing multi-dimensional video and audio sensing data, and the inspection edge node sends the inspection data to an adjacent node and the inspection host in a certain period;

b), the transformer substation patrol cloud side system configures intelligent task scheduling between the patrol host and the patrol edge node according to the real-time occupation condition of system resources;

step c), the inspection host establishes a historical data model through a historical inspection database of the transformer substation, identifies the monitoring target of the transformer substation by adopting a target intelligent identification method based on a DRFCN network for the current received inspection stream data, calculates the difference between the monitoring target and the historical normal model, and sends the calculation result to the inspection edge node in real time;

d), broadcasting dynamic information to neighborhood nodes in the edge block by the patrol edge node, periodically sending a part of collected data to the patrol host, identifying the transformer substation target by the patrol edge node by adopting a target intelligent identification method based on a lightweight deep network, calculating an abnormal value, and sending a calculation result to the patrol host in real time if the patrol edge node is calculated to be abnormal;

and e), the inspection host performs cooperative processing on the data of the cloud and the inspection edge nodes, compares the data with an abnormal threshold value predefined by the system, calculates a comprehensive abnormal index A, obtains an abnormal state of the inspection target, further judges the possibly occurred abnormality and updates the system scheduling in time.

2. The substation patrol method based on the cloud-side system and the video intelligent analysis as claimed in claim 1, wherein the step b) comprises the following specific steps:

step b1), the substation patrol cloud side system adopts a DAG task model, which comprises a strong real-time hot spot code analysis mechanism, an application program analysis process and mapping, and a cloud + side oriented patrol task scheduling model, wherein the strong real-time hot spot code analysis mechanism comprises two modules, namely a branch identification module and a path detection module;

step b2), adopting DAG task scheduling under the transformer substation distributed heterogeneous computing system, wherein the DAG task scheduling model under the transformer substation distributed heterogeneous computing system represents how to schedule task nodes in the DAG task graph to computing nodes in the hardware resource topological graph to execute so as to optimize the total operation completion time of the DAG tasks in the heterogeneous system.

3. The substation patrol method based on the cloud-side system and the video intelligent analysis as claimed in claim 1, wherein the step c) comprises the following specific steps:

step c1), establishing a transformer substation data model through a large amount of transformer substation historical inspection data;

step c2), processing the currently received substation patrol data: each convolution module is densely connected with all convolution modules on the upper layer in a mode of densely connecting the convolution modules, and the characteristics of each layer in the depth network model are multiplexed;

step c3), dense convolutional neural network based area sampling: on the basis of an RPN model structure, acquiring a higher-quality sampling area of the substation equipment based on a DRPN method of a dense convolutional neural network, inputting substation equipment images with any size into a DRPN model, and outputting the substation equipment images into a plurality of sampling areas corresponding to each class;

step c4), acquiring a target sampling area of the transformer substation: after transformer substation equipment feature pictures of an input channel are subjected to dense convolutional network layer transformation, generating a plurality of feature maps with W x H size, wherein each pixel of the feature maps has a wide receptive field, generating W x H x k anchor point frames for one W x H feature map, a sampling area comprises a large number of foreground areas containing transformer substation targets and background areas not containing the transformer substation targets, and selecting the anchor point frame which can represent the most sample features from W x H x k samples;

step c5), training DRPN by adopting a joint cost function, and realizing end-to-end training and testing: after cost functions of classification and regional sampling algorithms are respectively set, a combined cost function shown in a formula (1), L, is designed by jointly calculating class LOSS (LOSS) and position LOSS of a sampling region_clsAnd L_regRespectively representing the cost function of classification and anchor box regression, where N_clsNumber of anchor boxes selected in one training, N_regFeature map size representing the selection anchor box:

step c6), rapidly classifying the transformer substation targets based on the dense convolutional neural network;

step c7), feature conversion: performing convolution, pooling, normalization, linear or nonlinear conversion operation on each layer of feature map;

step c8), gradient diffusion and gradient expansion;

step c9), ROIP normalization transformation, which normalizes the feature matrixes of different dimensions to the same dimension;

step c10), joint cost function: DFCN has two sibling outputs: the category information of K +1 targets and the coordinate position of each category, and the joint cost function for training the DFCN are shown as the following formula (2):

L(p,u,t^u,v)＝L_cls(p,u)+λ[u≥1]L_loc(t^u,v)

classified ROIs are obtained by DRPN, each ROI has a class and position coordinate label, denoted u and v, t^uRepresenting the position coordinates to be predicted, in the form of tuplesSetting the background class as class 0, setting lambda as 1, and calculating the position coordinate without regression for the background ROI;

and c11), calculating the difference between the substation patrol target and the substation historical model, and if the calculation result is abnormal, transmitting the calculation result to the patrol edge node in real time.

4. The substation patrol method based on the cloud-side system and the video intelligent analysis as claimed in claim 1, wherein the step d) comprises the following specific steps:

step d1), the patrol edge node collects the video and sound multi-dimensional dynamic information in the edge block and periodically sends a part of collected data to the patrol host through the network;

step d2), using a Se-dressnet lightweight classification network: the convolution operation is utilized to compress the input characteristic channel, so that the calculation amount required in the subsequent module operation is reduced;

step d3), inputting the signals into 1 × 1 and 3 × 3 convolution kernel channels respectively, and realizing diversity of the sizes of the receptive fields through convolution kernels with different sizes, so that the characterization capability of extracting the patrol target characteristics of the transformer substation is more effective;

step d4), performing weighting operation on the characteristic diagram by using a channel weighting method provided in the SeNet network for reference, so that the characteristic diagram with excellent performance can be utilized to the maximum extent;

step d5), lightweight detection network: extracting the target characteristics of the transformer substation equipment by using a Se-DResNet lightweight classification network module;

step d6), selecting a lightweight SSD deep network as a basic framework, and improving the problem of poor SSD accuracy;

step d7), adjusting the positive and negative proportion of the input classifier sample by using a mode of reshaping cross entropy loss, and simultaneously, reasonably distinguishing the difficult samples from the easy-to-separate samples in the sample;

and d8), calculating the difference between the substation patrol target and a preset model of the patrol edge node, and if the patrol edge node calculates abnormal data, sending the calculation result to the patrol host in real time.

5. The substation patrol method based on the cloud-side system and the video intelligent analysis as claimed in claim 1, wherein the step e) comprises the following specific steps:

step e1), the inspection host machine adopts an evaluation mode combining cloud end and edge end detection, and two abnormal detection results are combined in a weighted mode to obtain a comprehensive abnormal index A:

A＝W_N*A_N+W_C*A_C

A_Nrepresenting the anomaly index, W, calculated at the edge_NRepresenting the proportion of the anomaly index calculated at the edge, A_CRepresenting an anomaly index, W, calculated by the cloud_cThe proportion of the abnormal index obtained by cloud computing is represented, and the W can be adjusted in real time according to the data volume collected by the edge end and the communication condition_cAnd W_NA value of (d);

and e2), further judging the possible transformer substation equipment abnormity according to the calculated comprehensive abnormity index A, and updating the system scheduling in time.

Background

In a traditional transformer substation monitoring system, data collection is carried out at a patrol host server/cloud end, and requests and responses are transmitted through a data link after being processed. The transformer substation is concerned with the safety and stability of a large power grid, and has the advantages of full voltage grade coverage, dense and complex equipment types, and high requirements on the feedback speed of the safe and stable operation of the equipment on defects and faults. The traditional patrol mode and patrol host adopt centralized data processing, the operation model is based on a historical database, operation is carried out on the patrol host, a large amount of real-time data generated by an edge end cannot be processed in time, and the dynamic change situation of a target is not well grasped, so that the traditional patrol mode and patrol host are difficult to adapt to new requirements.

The inventor of the present application has found through research in the process of implementing the present invention that: in the unmanned patrol system of the transformer substation, the edge computing technology is adopted, the front-end sensing equipment carries out distributed data processing on the spot and then uploads the processing result, the centralized processing amount of the platform can be reduced, and the response speed of the platform is improved. Therefore, the invention provides a transformer substation patrol method based on a cloud side system and video intelligent analysis, namely, an unmanned patrol function is realized by combining an image intelligent identification algorithm and an online anomaly detection algorithm through the cooperation between a patrol cloud side and a patrol edge front end; the monitoring target is identified and subjected to abnormal calculation through a target intelligent identification method based on a DRFCN (distributed resource control network) under the patrol host cloud computing environment; the patrol edge node adopts a target intelligent identification method based on a lightweight deep network to identify a target and calculates an abnormal value; and finally, the system integrates the data of the cloud end and the patrol edge node and the calculation result, and performs cooperative processing, so that a high-precision low-delay unmanned patrol mode more suitable for the application of the transformer substation is established.

Disclosure of Invention

Aiming at the technical problem that the traditional patrol mode cannot meet the new development requirements of a future power grid and energy Internet, the invention adopts a cloud-edge cooperative system framework under the condition of high real-time requirement and realizes the real-time unmanned patrol mode of the transformer substation based on a mode of a target intelligent identification algorithm combined by different depth networks on the cloud edge; the invention combines the focusing artificial intelligence technology, the edge computing technology and the power transformation patrol service, constructs an unmanned patrol mode of combined auxiliary patrol based on a cloud-edge system and centered by an intelligent camera, realizes that the daily patrol of primary and secondary equipment and auxiliary equipment of a transformer substation is completely replaced by 'machines', innovates the power transformation patrol mode, promotes the quality and efficiency improvement of the power transformation operation and maintenance service, and has important significance for guaranteeing the safe operation of electric power.

A transformer substation inspection method based on a cloud side system and video intelligent analysis comprises the following steps:

step a), establishing a cloud edge cooperative system framework for substation inspection, which is composed of an inspection host, an inspection edge block and an inspection edge node, wherein the inspection host is used for processing and analyzing large image data of the substation, the inspection edge block is used for gathering, storing, processing and intelligent application of front-end inspection node sensing data and is independent and autonomous, the inspection edge node is used for collecting and intelligently processing multi-dimensional sensing data such as videos and audios, and the inspection edge node sends the inspection data to adjacent nodes and the inspection host in a certain period;

b), the transformer substation patrol cloud side system configures intelligent task scheduling between the patrol host and the patrol edge node according to the real-time occupation condition of system resources;

step c), the inspection host establishes a historical data model through a historical inspection database of the transformer substation, identifies the monitoring target of the transformer substation by adopting a target intelligent identification method based on a DRFCN network for the current received inspection stream data, calculates the difference between the monitoring target and the historical normal model, and sends the calculation result to the inspection edge node in real time;

d), broadcasting dynamic information to neighborhood nodes in the edge block by the patrol edge node, periodically sending a part of collected data to the patrol host, identifying the transformer substation target by the patrol edge node by adopting a target intelligent identification method based on a lightweight deep network, calculating an abnormal value, and sending a calculation result to the patrol host in real time if the patrol edge node is calculated to be abnormal;

and e), the inspection host performs cooperative processing on the data of the cloud and the inspection edge nodes, compares the data with an abnormal threshold value predefined by the system, calculates a comprehensive abnormal index A, obtains an abnormal state of the inspection target, further judges the possibly occurred abnormality and updates the system scheduling in time.

Further, the step b) comprises the following specific steps:

step b1), the substation patrol cloud side system adopts a DAG task model, which comprises a strong real-time hot spot code analysis mechanism, an application program analysis process and mapping, and a cloud + side oriented patrol task scheduling model, wherein the strong real-time hot spot code analysis mechanism comprises two modules, namely a branch identification module and a path detection module;

step b2), adopting DAG task scheduling under the transformer substation distributed heterogeneous computing system, wherein the DAG task scheduling model under the transformer substation distributed heterogeneous computing system represents how to schedule task nodes in the DAG task graph to computing nodes in the hardware resource topological graph to execute so as to optimize the total operation completion time of the DAG tasks in the heterogeneous system.

Further, the step c) comprises the following specific steps:

step c1), establishing a transformer substation data model through a large amount of transformer substation historical inspection data;

step c2), processing the currently received substation patrol data: each convolution module is densely connected with all convolution modules on the upper layer in a mode of densely connecting the convolution modules, and the characteristics of each layer in the depth network model are multiplexed;

step c3), dense convolutional neural network based area sampling: on the basis of an RPN model structure, acquiring a higher-quality sampling area of the substation equipment based on a DRPN method of a dense convolutional neural network, inputting substation equipment images with any size into a DRPN model, and outputting the substation equipment images into a plurality of sampling areas corresponding to each class;

step c4), acquiring a target sampling area of the transformer substation: after transformer substation equipment feature pictures of an input channel are subjected to dense convolutional network layer transformation, generating a plurality of feature maps with W x H size, wherein each pixel of the feature maps has a wide receptive field, generating W x H x k anchor point frames for one W x H feature map, a sampling area comprises a large number of foreground areas containing transformer substation targets and background areas not containing the transformer substation targets, and selecting the anchor point frame which can represent the most sample features from W x H x k samples;

step c5), training DRPN by adopting a joint cost function, and realizing end-to-end training and testing: after cost functions of classification and regional sampling algorithms are respectively set, a combined cost function shown in a formula (1), L, is designed by jointly calculating class LOSS (LOSS) and position LOSS of a sampling region_clsAnd L_regRespectively representing the cost function of classification and anchor box regression, where N_clsNumber of anchor boxes selected in one training, N_regFeature map size representing the selection anchor box:

step c6), rapidly classifying the transformer substation targets based on the dense convolutional neural network;

step c7), feature conversion: performing convolution, pooling, normalization, linear or nonlinear conversion operation on each layer of feature map;

step c8), gradient diffusion and gradient expansion;

step c9), ROIP normalization transformation, which normalizes the feature matrixes of different dimensions to the same dimension;

step c10), joint cost function: DFCN has two sibling outputs: the category information of K +1 targets and the coordinate position of each category, and the joint cost function for training the DFCN are shown as the following formula (2):

L(p,u,t^u,v)＝L_cls(p,u)+λ[u≥1]L_loc(t^u,v)

classified ROIs are obtained by DRPN, each ROI has a class and position coordinate label, denoted u and v, t^uRepresenting the position coordinates to be predicted, in the form of tuplesSetting the background class as class 0, setting lambda as 1, and calculating the position coordinate without regression for the background ROI;

and c11), calculating the difference between the substation patrol target and the substation historical model, and if the calculation result is abnormal, transmitting the calculation result to the patrol edge node in real time.

Further, the step d) comprises the following specific steps:

step d1), the patrol edge node collects the video and sound multi-dimensional dynamic information in the edge block and periodically sends a part of collected data to the patrol host through the network;

step d2), using a Se-dressnet lightweight classification network: the convolution operation is utilized to compress the input characteristic channel, so that the calculation amount required in the subsequent module operation is reduced;

step d3), inputting the signals into 1 × 1 and 3 × 3 convolution kernel channels respectively, and realizing diversity of the sizes of the receptive fields through convolution kernels with different sizes, so that the characterization capability of extracting the patrol target characteristics of the transformer substation is more effective;

step d4), performing weighting operation on the characteristic diagram by using a channel weighting method provided in the SeNet network for reference, so that the characteristic diagram with excellent performance can be utilized to the maximum extent;

step d5), lightweight detection network: extracting the target characteristics of the transformer substation equipment by using a Se-DResNet lightweight classification network module;

step d6), selecting a lightweight SSD deep network as a basic framework, and improving the problem of poor SSD accuracy;

step d7), adjusting the positive and negative proportion of the input classifier sample by using a mode of reshaping cross entropy loss, and simultaneously, reasonably distinguishing the difficult samples from the easy-to-separate samples in the sample;

and d8), calculating the difference between the substation patrol target and a preset model of the patrol edge node, and if the patrol edge node calculates abnormal data, sending the calculation result to the patrol host in real time.

Further, the step e) comprises the following specific steps:

step e1), the inspection host machine adopts an evaluation mode combining cloud end and edge end detection, and two abnormal detection results are combined in a weighted mode to obtain a comprehensive abnormal index A:

A＝W_N*A_N+W_C*A_C

A_Nrepresenting the anomaly index, W, calculated at the edge_NRepresenting the proportion of the anomaly index calculated at the edge, A_CRepresenting an anomaly index, W, calculated by the cloud_cThe proportion of the abnormal index obtained by cloud computing is represented, and the W can be adjusted in real time according to the data volume collected by the edge end and the communication condition_cAnd W_NA value of (d);

and e2), further judging the possible transformer substation equipment abnormity according to the calculated comprehensive abnormity index A, and updating the system scheduling in time.

Under the environment of substation patrol cloud-edge cooperation, firstly, a substation target intelligent identification method based on a DRFCN (distributed resource control network) network at a patrol host end is designed according to the characteristics of a substation patrol target; in order to detect possible abnormity in the patrol process in real time, the invention establishes a real-time abnormity detection model through a lightweight deep network target intelligent identification method of patrol edge nodes and cooperation between cloud edges of a transformer substation patrol system, thereby reducing the problem of system real-time performance caused by large data transmission delay and heavy algorithm processing as far as possible. Experiments show that after the transformer substation inspection method based on the cloud-side system and video intelligent analysis is adopted, the real-time performance of the system is improved by 22.75%.

Drawings

FIG. 1 is a DAG task scheduling model diagram of the unmanned patrol cloud-side system of the substation of the present invention;

FIG. 2 is a cloud/edge resource coordination diagram of the unmanned inspection of the transformer substation according to the present invention;

FIG. 3 is a diagram of a DRPN algorithm model structure for intelligent recognition of a patrol host according to the present invention;

fig. 4 is a schematic diagram of the intelligent recognition DFCN convolution of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, are within the scope of the present invention.

The invention needs to solve the following technical problems: in a substation patrol scene with high real-time requirements, data is better processed in a cloud-edge resource cooperation mode; how to dig frequent patterns through the cloud end and identify the abnormal phenomenon of equipment in historical patrol data; when the characteristics of the tour stream data time-varying evolution are ignored, the global characteristic model constructed by enough amount of historical data has higher abnormality detection precision; how to detect a novel abnormal mode possibly existing in newly acquired patrol flow data and an abnormal target of a transformer substation in a specific scene.

The embodiment of the invention provides a transformer substation inspection method based on a cloud side system and video intelligent analysis, which comprises the following steps of:

step a), establishing a cloud edge cooperative system framework for substation inspection, wherein the cloud edge cooperative system framework is composed of an inspection host, an inspection edge block and an inspection edge node, and the inspection host is used for processing and analyzing large data of a substation image; the patrol edge block is used for gathering, storing, processing and intelligently applying the front patrol node sensing data and is independent and autonomous; the patrol edge node is used for collecting and intelligently processing multi-dimensional sensing data such as videos and audios, and sends patrol data to adjacent nodes and a patrol host in a certain period;

b), the transformer substation patrol cloud side system configures intelligent task scheduling between the patrol host and the patrol edge node according to the real-time occupation condition of system resources;

wherein, the step b) comprises the following specific steps:

step b1), the substation patrol cloud side system adopts a DAG task model, which comprises a strong real-time hot spot code analysis mechanism, an application program analysis flow and mapping, and a cloud + side oriented patrol task scheduling model; the strong real-time hot spot code analysis mechanism comprises two modules, namely a branch identification module and a path detection module. The cloud/edge migration technology supporting strong real-time and large samples mainly relates to strategy gradient reinforcement learning method design based on MCTS, migration strategy network structure design based on seq2seq models and migration algorithm input and output design. The computing-intensive cloud/edge scheduling algorithm optimization technology mainly comprises a neural network commonality acceleration algorithm optimization technology, an SVM commonality acceleration algorithm optimization technology and an integrated learning commonality acceleration algorithm scheduling optimization technology.

Step b2), a DAG task scheduling model under the substation distributed heterogeneous computing system is shown in FIG. 1. The DAG task graph shown in fig. 1 represents the parallel and data dependency relationship among tasks in the substation patrol application, and the hardware resource topology graph shown in fig. 2 represents the topology relationship among the computing nodes and the hardware resource state of the substation distributed heterogeneous computing system. The DAG task scheduling model under the transformer substation heterogeneous system can be expressed as how to schedule task nodes in the DAG task graph to be executed on computing nodes in the hardware resource topological graph so as to optimize the overall operation completion time of the DAG tasks in the heterogeneous system.

And c), establishing a historical data model by the inspection host through a historical inspection database of the transformer substation, identifying the monitoring target of the transformer substation by adopting a target intelligent identification method based on the DRFCN network for the current received inspection flow data, and calculating the difference between the monitoring target and the historical normal model. The patrol host sends the calculation result to the patrol edge node in real time;

wherein, the step c) comprises the following specific steps:

step c1), establishing a transformer substation data model through a large amount of transformer substation historical inspection data;

step c2), processing the currently received substation patrol data. And each convolution module is densely connected with all convolution modules on the upper layer in a mode of densely connecting the convolution modules, and the characteristics of each layer in the deep network model are multiplexed. The convolution module at the bottom learns the characteristics of all the convolution modules, so that the characteristic expression capability of the deep network model is necessarily improved, the average identification accuracy of the substation equipment can be improved, the deep network model can be reduced, and the problems of gradient dispersion and gradient expansion can be effectively solved.

Step c3), area sampling based on dense convolutional neural network. On the basis of an RPN model structure, a DRPN method based on a dense convolutional neural network is used for obtaining a higher-quality sampling area of the substation equipment. The DRPN model structure is as shown in fig. 3, and the DRPN inputs substation equipment images of any size and outputs a plurality of sampling regions corresponding to each class;

step c4), obtaining a transformer substation target sampling area. After the transformer substation equipment feature pictures of the input channels are subjected to dense convolution network layer transformation, a plurality of feature maps with W multiplied by H sizes are generated, and each pixel (neuron) of the feature maps has a wide receptive field. For a W × H feature map, W × H × k anchor blocks are generated, and these sampling areas include a large number of foreground areas (including the substation target) and background areas (not including the substation target). Selecting an anchor point frame which can represent the characteristics of the sample most from the W multiplied by H multiplied by k samples;

step c5), joint cost function: in order to share the calculated amount and the storage space of the dense convolutional network, a DRPN is trained by adopting a joint cost function, and end-to-end training and testing are realized. After cost functions of classification and regional sampling algorithms are respectively set, a combined cost function shown in a formula (1), L, is designed by jointly calculating class LOSS (LOSS) and position LOSS of a sampling region_clsAnd L_regRespectively representing the cost function of classification and anchor box regression, where N_clsAnchor representing one-training selectionNumber of dot frames, N_regThe size of the feature map representing the selected anchor box.

Step c6), transformer substation target fast classification based on the dense convolutional neural network. The model structure of the DFCN algorithm is shown in figure 4. After a high-quality transformer substation target sampling area is quickly and efficiently obtained by using a DRPN algorithm, characteristics are extracted from a deep characteristic diagram for classification, and the category and position information of a transformer substation inspection target is simultaneously output in a result.

Step c7), feature transformation. And performing operations such as convolution, pooling, normalization, linear or nonlinear conversion and the like on the feature map of each layer.

Step c8), gradient diffusion and gradient expansion. In the mode of densely connecting convolution layers, the gradient is also reversely transferred in the model in a mode of sum, namely the DFCN strengthens the feature expression capability of the deep learning model, but the calculation amount is not required to be increased when the gradient is reversely transferred.

Step c9), ROIP normalization transformation. The feature matrices of different dimensions are normalized to the same dimension.

Step c10), the joint cost function. DFCN has two sibling outputs: category information of K +1 substation targets and coordinate positions of each category. The joint cost function for training the DFCN is shown in formula (2):

L(p,u t,^u v,＝)L_cls p(u,+λ)u≥[L_loc 1t^u]v (2)

classified ROIs are obtained by DRPN, each ROI has a class and position coordinate label, denoted u and v, t^uRepresenting the position coordinates to be predicted, in the form of tuplesThe background class is set to class 0, λ is set to 1, and the position coordinates are not calculated back for the background ROI.

As mentioned above, the DRPN is used to generate a high-quality sampling region of the substation patrol target, and the DFCN is used to calculate the discrete probability distribution and the position coordinates of the substation patrol target sampling region category information. If the DRPN and DFCN are to share a dense convolutional layer, the DRFCN needs a training mechanism so that the DRPN and DFCN share weights. The invention adopts DRPN and DFCN united distribution parameter training method.

And c11), calculating the difference between the substation patrol target and the substation historical model, and if the calculation result is abnormal, transmitting the calculation result to the patrol edge node in real time.

And d), broadcasting dynamic information to the neighborhood nodes in the edge block by the patrol edge node, and periodically sending a part of collected data to the patrol host. And the patrol edge node adopts a target intelligent identification method based on a lightweight deep network to identify the transformer substation target and calculates an abnormal value. If the calculation of the patrol edge node is abnormal, the calculation result is sent to the patrol host in real time;

wherein, the step d) comprises the following specific steps:

step d1), collecting multi-dimensional dynamic information such as video, sound and the like in the edge block by the patrol edge node, and periodically sending a part of collected data to the patrol host through the network;

step d2), using a Se-DResNet lightweight classification network. The convolution operation is utilized to compress the input characteristic channel, so that the calculation amount required in the subsequent module operation is reduced;

step d3), inputting the signals into 1 × 1 and 3 × 3 convolution kernel channels respectively, and realizing diversity of the sizes of the receptive fields through convolution kernels with different sizes, so that the characterization capability of extracting the patrol target characteristics of the transformer substation is more effective;

step d4), performing weighting operation on the characteristic diagram by using a channel weighting method provided in the SeNet network for reference, so that the characteristic diagram with excellent performance can be utilized to the maximum extent;

step d5), the detection network is lightened. Extracting the target characteristics of the transformer substation equipment by using a Se-DResNet lightweight classification network module;

step d6), selecting a lightweight SSD deep network as a basic framework, and improving the problem of poor SSD accuracy. In a multi-scale target detection part of the transformer substation, the idea of feature fusion is used for reference, and the quality of input features is improved, so that the input features can express target information of the transformer substation more accurately;

step d7), aiming at the problem of unbalanced proportion of positive and negative samples in the classifier part, the positive and negative proportion of the input classifier sample is adjusted by using a mode of reshaping cross entropy loss, and meanwhile, the difficult samples and the easy samples in the sample can be reasonably distinguished;

and d8), calculating the difference between the substation patrol target and a preset model of the patrol edge node, and if the patrol edge node calculates abnormal data, sending the calculation result to the patrol host in real time.

And e), the inspection host performs cooperative processing on the data of the cloud and the inspection edge nodes, compares the data with an abnormal threshold value predefined by the system, calculates a comprehensive abnormal index A, obtains an abnormal state of the inspection target, further judges the possibly occurred abnormality and updates the system scheduling in time.

Wherein, the step e) comprises the following specific steps:

step e1), the inspection host machine adopts an evaluation mode combining cloud end and edge end detection, and two abnormal detection results are combined in a weighted mode to obtain a comprehensive abnormal index A:

A＝W_N*A_N+W_C*A_C

A_Nrepresenting the anomaly index, W, calculated at the edge_NRepresenting the proportion of the anomaly index calculated at the edge, A_CRepresenting an anomaly index, W, calculated by the cloud_cAnd the proportion of the abnormal index calculated by the cloud is represented. Can adjust W in real time according to the data volume and communication condition collected by the edge terminal_cAnd W_NThe value of (c).

And e2), further judging the possible transformer substation equipment abnormity according to the calculated comprehensive abnormity index A, and updating the system scheduling in time.

The invention designs a transformer substation inspection method based on a cloud side system and video intelligent analysis. A system framework based on cloud edge cooperation is designed according to target characteristics of a transformer substation, and through cooperation between cloud edge nodes and a target detection algorithm of a DRFCN (data communications network) network at a patrol host end and a lightweight deep network at patrol edge nodes, the problem that a traditional patrol host cannot timely process a large amount of original data acquired at edges is solved, so that the problem of unmanned intelligent patrol of the transformer substation under the requirement of high real-time performance is effectively solved. The method has high response speed and detection precision when the abnormal target of the substation of the inspection data flow is continuously detected.

The above description is only an embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.