An automatic substation image data non-inductive transmission control system

Abstract

The present application relates to the field of automation control, and more particularly to a kind of substation image data automation non-inductive transmission control system, including image acquisition module, intelligent video analysis module, automation transmission control module, security and privacy protection module and non-inductive interaction design module;The present application adopts multi-scale Flow model to deeply analyze image information, by distinguishing the needs of image anomaly detection and pixel-level anomaly positioning, flexibly uses multi-scale aggregation strategy, improves the accuracy and efficiency of anomaly identification;Optimized adaptive data packet scheduling method is adopted to ensure the continuity and stability of video and image data in the transmission process, and the improved intelligent optimizer algorithm is used to optimize the parameter vector of the optimized adaptive data packet scheduling method, to speed up the parameter convergence speed, enhance the reliability of data transmission, and bring smooth and unblocked user experience, maximize the use of existing network resources.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automation control, in particular to a substation image data automatic non-sensing transmission control system. BACKGROUND

[0002] In recent years, with the deepening reform of power system management and the development of automation technology, many substations have realized unattended operation, which promotes the image monitoring system as an important supplement to the monitoring of unattended substations. The huge amount of image data generated by substation monitoring increases the difficulty of quickly and accurately identifying abnormalities from massive images, which may delay the system response time and thus weaken the real-time monitoring efficiency. In remote and harsh environments, the instability of network connection further interferes with the continuity of data transmission, constituting another obstacle. In addition, the transmission of high-definition images greatly occupies the bandwidth, and the limited nature of network resources may become a bottleneck restricting the overall performance of the system. SUMMARY

[0003] In view of the above situation, the present application provides a substation image data automatic non-sensing transmission control system. In view of the problem that the huge amount of image data generated by substation monitoring increases the difficulty of quickly and accurately identifying abnormalities from massive images, the present application uses a multi-scale Flow model to deeply analyze image information, flexibly uses a multi-scale aggregation strategy by distinguishing the needs of image anomaly detection and pixel-level anomaly positioning, not only realizes comprehensive image anomaly scanning, but also accurately locates abnormal pixels, and improves the accuracy and efficiency of anomaly identification. In view of the problem that in remote and harsh environments, the instability of network connection further interferes with the continuity of data transmission, and the problem that the transmission of high-definition images greatly occupies the bandwidth, but the network resources are limited, the present application adopts an optimized adaptive data packet scheduling method to ensure the continuity and stability of video and image data during transmission, and uses an improved intelligent optimizer algorithm to optimize the parameter vector of the optimized adaptive data packet scheduling method, accelerates the parameter convergence speed, enhances the reliability of data transmission, and brings a smooth and unobstructed user experience, maximizing the use of existing network resources.

[0004] The substation image data automatic non-sensing transmission control system provided by the present application comprises an image acquisition module, an intelligent video analysis module, an automatic transmission control module, a security and privacy protection module, and a non-sensing interaction design module.

[0005] The image acquisition module deploys a high-definition camera in the key area of the substation for all-weather, multi-angle acquisition of high-definition video images, ensuring clear images under various lighting conditions.

[0006] The intelligent video analysis module is responsible for analyzing the collected high-definition video images using a multi-scale Flow model, identifying normal and abnormal data in the high-definition video images, obtaining abnormal data information and triggering an abnormal alarm;

[0007] The automatic transmission control module transmits high-definition video images as video data and uses an optimized adaptive data packet scheduling method to automatically adjust transmission parameters according to network conditions;

[0008] The security and privacy protection module has strict security measures built in, and for high-definition video images involving personal privacy, anonymization or specific area masking processing is performed;

[0009] The non-sensing interaction design module provides a visual interface, performs automatic scheduling and preset rule execution, reduces the need for manual operation, minimizes the impact of the entire monitoring process on daily operation and maintenance work, and realizes a truly non-sensing experience.

[0010] Further, the high-definition cameras in the image acquisition module include fixed cameras, panoramic cameras, and thermal imaging cameras, and unmanned aerial vehicles are used to conduct regular inspections within the scope of the substation, image acquisition is performed in places where ordinary equipment cannot be deployed, and infrared sensors and illumination sensors are used to assist image acquisition, optimizing the environmental conditions for image acquisition;

[0011] Further, the network devices required by the automatic transmission control module for transmission include switches and routers for building stable local area network or wide area network connections; wireless communication modules including Wi-Fi, 4G / 5G modules for remote transmission; and software-defined network controllers for intelligent management of network resources to ensure the security and reliability of data transmission;

[0012] Further, the visual interface includes a system status overview module, a real-time monitoring view module, a historical data playback module, an event management module, and a rule configuration module. The system status overview module displays the overall running status of the system, including whether data is being transmitted and whether there is a fault alarm.

[0013] The real-time monitoring view module provides real-time video stream display of the camera, allowing users to instantly view the on-site situation of the substation, supports simultaneous display of multiple video streams, and monitors key positions.

[0014] The historical data playback module allows users to retrieve and play video records within a certain time period in the past, provides search functions, and can quickly locate the required data segment according to date and time;

[0015] The event management record module displays abnormal events and alarm triggering events occurring in the substation, and supports classification and screening of the events;

[0016] The rule configuration module provides an interface for a user to define automatic scheduling and preset rules, and provides a test mode for verifying validity of the rules.

[0017] Further, the intelligent video analysis module is responsible for analyzing the collected high-definition video images using a multi-scale Flow model, and specifically includes the following steps:

[0018] Step S1: primary feature extraction, using a pre-trained ViT model as a feature extractor, forward propagating the high-definition video images, extracting feature maps of different levels, and obtaining primary feature maps;

[0019] Step S2: multi-scale feature extraction, using an average pooling layer to downsample the primary feature maps to multi-scale, and obtaining multi-scale feature maps;

[0020] Step S3: multi-scale Flow model, using asymmetric parallel streams to enhance feature expression of each scale, and using a fusion stream to fuse multi-scale information, and constructing a multi-scale Flow model;

[0021] Step S4: learning and optimization, setting minimization of negative log likelihood as a learning goal, operating Jacobian determinants of the parallel streams and the fusion stream, and optimizing the multi-scale Flow model to distinguish normal data and abnormal data distribution;

[0022] Step S5: anomaly detection and scoring, using pixel-level anomaly positioning and image-level anomaly detection methods to calculate an anomaly score;

[0023] Further, in step S3, the multi-scale Flow model specifically includes the following steps:

[0024] Step S31: parallel stream design, designing an asymmetric parallel stream, and constructing a parallel stream containing L stream blocks, each of which is constructed using an asymmetric architecture;

[0025] Step S32: in each parallel stream, using a 3*3 convolution ST network to automatically capture spatial context information, and performing position encoding on each feature map in the multi-scale feature maps to convert to obtain multi-scale latent feature maps;

[0026] Step S33: multi-scale information fusion, using a fusion stream to integrate the multi-scale latent feature maps, the fusion stream splicing after adjusting all input latent feature maps to the minimum size using average pooling, and using two convolution layers with similar structures to perform feature fusion to generate a fusion feature map;

[0027] Step S34: feature segmentation, segmenting the fused feature map and adjusting it to a multi-scale feature map to obtain a segmented feature map;

[0028] Step S35: scale and shift weight learning, learning the scale weight and shift weight of each scale using group convolution to adjust the features of the segmented feature map, the group convolution divides the input channels into N groups and performs convolution operation on each group independently, reducing the amount of calculation and the number of parameters;

[0029] Further, in step S5, anomaly detection and scoring, specifically including the following steps:

[0030] Step S51: log-likelihood estimation calculation, using log-likelihood to calculate the Jacobian determinant of any pixel on the segmented feature map to obtain a log-likelihood map, the formula used is as follows:

[0031] ;

[0032] In the formula, 、 and are natural numbers, represents the pixel value at the th scale of the multi-scale feature map, located at the th position, and the transformed latent feature; represents the Euclidean square of the latent feature ; represents the latent feature space, represents the log-likelihood estimation of the latent feature space ;

[0033] Step S52: log-likelihood map processing, using bilinear interpolation to enlarge the log-likelihood map of all scales to the original image size to obtain a rescaled log-likelihood map;

[0034] Step S53: pixel-level anomaly positioning, converting the rescaled log-likelihood map into a probability map and calculating a pixel-level anomaly map through additive aggregation, the formula used is as follows:

[0035] ;

[0036] ;

[0037] In the formula, represents the probability map of the th scale, is the additive aggregation probability map of the pixel-level anomaly, represents the maximum value, is the pixel-level anomaly map; ​

[0038] Step S54: image-level anomaly detection, the re-scaled log-likelihood map is first added and then converted into a probability map, and a global anomaly map is calculated, and the formula is as follows:

[0039] ;

[0040] ;

[0041] In the formula, all are multiplied together, is a multiplication aggregation probability map, is a global anomaly map at the image level;

[0042] Step S55: calculate the anomaly score, and the formula is as follows:

[0043] ;

[0044] In the formula, is a set parameter, the maximum K values are selected from , indicates the final anomaly score obtained.

[0045] The automatic transmission control module uses an optimized adaptive data packet scheduling method to automatically adjust the transmission parameters according to the network conditions, and the optimized adaptive data packet scheduling method specifically includes the following steps:

[0046] Step M1: network monitoring, continuously monitoring the round-trip time, congestion window, in-transit byte number and packet loss rate of each path during data transmission and updating in real time;

[0047] Step M2: video analysis, extracting the video characteristics of the video data in the data transmission process, calculating the real-time bit rate, and setting the frame priority;

[0048] Step M3: define the context and action space, construct the state vector, and define the action set, the state vector includes network state information and video characteristics, and the action set includes four actions of default scheduling, duplicate frame, discard frame and packet loss compensation;

[0049] Step M4: initialization, initializing a linear model parameter vector for each action, and setting an initial confidence boundary value;

[0050] Step M5: predict the action value, for each action, use the current state and the parameter vector corresponding to the action to predict the expected reward using linear regression, obtain the prediction result, and obtain the action corresponding to the expected reward, calculate and add the confidence boundary;

[0051] Step M6: Select the optimal action, according to the last prediction result, select the action with the highest expected reward as the output this time;

[0052] Step M7: Perform scheduling and calculate the actual reward, schedule the data packet according to the action with the highest expected reward, and calculate the actual reward value of this decision;

[0053] Step M8: Update optimization, record the actual reward and state vector when the scheduling is performed, and use the improved intelligent optimizer algorithm to optimize and update the parameter vector of the corresponding action to adjust the confidence boundary according to the obtained actual reward and state vector;

[0054] Further, in step M8, the parameter vector of the corresponding action is optimized and updated using the improved intelligent optimizer algorithm to adjust the confidence boundary, which specifically includes the following steps:

[0055] Step M81: Initialization stage, set the upper and lower bounds of the search space, the total number of iterations, the balance factor, randomly generate the initial candidate solution and the position of the candidate solution;

[0056] Step M82: Evaluate the initial candidate solution, calculate the fitness value of the initial candidate solution, set the initial candidate solution as the global optimal solution, and calculate and record the fitness value of the global optimal solution;

[0057] Step M83: Exploration stage, in the exploration stage iteration is performed, and the position is updated, wherein is less than the total number of iterations;

[0058] Further, in step M83, the exploration stage specifically includes the following steps:

[0059] Step M831: Position update, update the position of the candidate solution using the update rule, and the formula is as follows:

[0060] ;

[0061] ;

[0062] In the formula, is a natural number, is the global optimal solution, is the position of the candidate solution, is a random number in the range [0, 1], is the weight, is the balance factor, is the current iteration number, is the total number of iterations, is the natural exponential function, represents the weight factor at the iteration number ;

[0063] Step M832: Check the boundary. If the updated position exceeds the upper and lower bounds of the search space, set the updated position to the upper and lower bound values of the search space, otherwise, keep it unchanged.

[0064] Step M833: Update the global best solution. Calculate the fitness value of the updated candidate solution. If f(x) is better than f(gb), update the global best solution and the fitness value of the global best solution, otherwise, maintain it unchanged, where f(x) is the fitness value of the updated candidate solution, and f(gb) is the fitness value of the global best solution.

[0065] Step M834: End of iteration. Repeat steps M831 to M833 until the total number of iterations is reached.

[0066] Step M84: Development and dynamic adjustment. Perform the remaining iterations, dynamic adjustment and position update, where The sum of is equal to the total number of iterations.

[0067] Further, in step M84, development and dynamic adjustment, specifically includes the following steps:

[0068] Step M841: Update the position of the candidate solution. The formula used is as follows:

[0069] ;

[0070] In the formula, is a random number in the range [0, 1], and are the upper and lower bounds of the search space, respectively.

[0071] Step M842: Check the fitness stagnation. If the fitness value has not improved in m consecutive iterations, perform dynamic adjustment, where m is less than the total number of iterations.

[0072] Step M843: Dynamically adjust the position of the candidate solution. The formula used is as follows:

[0073] ;

[0074] In the formula, is a random number in the range [0, 1];

[0075] Step M844: Check the boundary and update. Confirm that the updated position of the candidate solution is still within the search space, calculate the fitness value of the updated candidate solution, and update the global best solution and the fitness value of the global best solution.​​

[0076] Step M845: Iterative optimization, repeat step M841 to step M844, when the total number of iterations is reached, the iteration is ended;

[0077] Step M85: Output the final global optimal solution and the fitness value of the optimal solution.

[0078] The beneficial results achieved by the above-mentioned scheme are as follows:

[0079] (1) For the problem that the image data generated by substation monitoring is huge, which leads to the increase of the difficulty of quickly and accurately identifying abnormalities from massive images, the scheme adopts a multi-scale Flow model to deeply analyze image information, and skillfully distinguishes and meets the dual requirements of image anomaly detection and pixel-level accurate positioning. By adopting a dynamic multi-scale aggregation strategy, not only the wide range of image abnormal screening is realized, but also the single abnormal pixel is accurately positioned, greatly enhancing the accuracy and response speed of abnormal identification. In addition, the pre-trained ViT model is integrated as a feature extractor, which effectively reduces the feature learning stage, ensures efficient and accurate feature extraction, and further optimizes the overall performance.

[0080] (2) For the problem that the instability of network connection in remote and harsh environments will further interfere with the continuity of data transmission, and the transmission of high-definition images greatly occupies the bandwidth, but the network resources are limited, the scheme adopts an optimized adaptive data packet scheduling method to ensure the continuity and stability of video and image data during transmission. This method carefully designs four sets of mechanisms, each of which solves different network problems, thereby ensuring the coherence and reliability of video and image data in complex transmission environments. Through continuous self-optimization to adapt to the dynamic changes of the network environment, the reliability of data transmission is effectively enhanced, and the optimization of network resource utilization efficiency is realized, bringing a smooth and uninterrupted user experience, and maximizing the use of existing network resources.

[0081] (3) In the optimized adaptive data packet scheduling method, the improved intelligent optimizer algorithm is used to adjust the parameter vector of the action, improve the network performance and data transmission efficiency, and adjust the confidence boundary. The SCO algorithm continuously adjusts the parameter vector through the iterative optimization process until it converges to the global optimal solution, thereby dynamically adjusting the confidence boundary to ensure that the scheduling strategy can maintain robustness and quickly respond to real-time changes in uncertain network environments, ultimately achieving more efficient and stable data transmission. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 A schematic diagram of a substation image data automatic non-inductive transmission control system is provided for the present application;

[0083] Figure 2 Flow model is shown in the figure;

[0084] Figure 3 Flow model is shown in the figure;

[0085] The accompanying drawings are included to provide a further understanding of the application, and are incorporated in and constitute a part of this specification, illustrate embodiments of the application, and together with the description serve to explain the application, and do not limit the application. DETAILED DESCRIPTION

[0086] The technical solutions in the embodiments of the application will be apparently and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only some of the embodiments of the application, but not all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0087] Embodiment one, refer to Figure 1 The application provides a substation image data automatic non-sense transmission control system, which comprises an image acquisition module, an intelligent video analysis module, an automatic transmission control module, a security and privacy protection module and a non-sense interaction design module.

[0088] The image acquisition module deploys a high-definition camera in a key area of the substation, and is used for acquiring high-definition video images in all-weather and multi-angle, so that clear images can be obtained under various illumination conditions.

[0089] The intelligent video analysis module is responsible for analyzing the collected high-definition video images using a multi-scale Flow model, identifying normal data and abnormal data in the high-definition video images, obtaining abnormal data information and triggering an abnormal alarm.

[0090] The automatic transmission control module transmits the high-definition video images as video data, and automatically adjusts transmission parameters according to network conditions using an optimized adaptive data packet scheduling method.

[0091] The security and privacy protection module is built-in strict security measures, and for high-definition video images involving personal privacy, anonymization or specific area shielding processing is performed.

[0092] The non-sense interaction design module provides a visual interface, performs automatic scheduling and preset rule execution, reduces the demand for manual operation, minimizes the influence of the entire monitoring process on daily operation and maintenance work, and realizes a real non-sense experience.

[0093] In the second embodiment, based on the above-mentioned embodiment, the high-definition camera in the image acquisition module includes a fixed camera, a panoramic camera, and a thermal imaging camera, and a drone is used to conduct regular inspection within the scope of the substation. In places where ordinary equipment cannot be deployed, image acquisition is conducted. At the same time, an infrared sensor and an illumination sensor are used to assist in image acquisition, and the environmental conditions for image acquisition are optimized.

[0094] The network devices required by the automatic transmission control module during transmission include switches and routers for building stable local area network or wide area network connections; wireless communication modules including Wi-Fi, 4G / 5G modules for remote transmission; and software-defined network controllers for intelligent management of network resources to ensure the security and reliability of data transmission.

[0095] The visualization interface includes a system status overview module, a real-time monitoring view module, a historical data playback module, an event management module, and a rule configuration module. The system status overview module displays the overall running status of the system, including whether data is being transmitted and whether there is a fault alarm.

[0096] The real-time monitoring view module provides real-time video stream display of the camera, allowing users to instantly view the on-site situation of the substation, supports simultaneous display of multiple video streams, and monitors key positions.

[0097] The historical data playback module allows users to retrieve and play video recordings within a certain time period in the past, provides search functions, and can quickly locate the required data segments according to date and time.

[0098] The event management record module displays abnormal events and alarm triggering events that occur in the substation and supports classification and filtering of events.

[0099] The rule configuration module provides an interface for users to define automatic scheduling and preset rules and provides a test mode for verifying the effectiveness of the rules.

[0100] In the third embodiment, referring to Figure 1 and Figure 2 , the intelligent video analysis module is responsible for analyzing the collected high-definition video images using a multi-scale Flow model. The specific steps include:

[0101] Step S1: Primary feature extraction, using a pre-trained ViT model as a feature extractor, forward propagating the high-definition video image, extracting different levels of feature maps, and obtaining primary feature maps;

[0102] Step S2: Multi-scale feature extraction, using an average pooling layer to downsample the primary feature maps to multiple scales to obtain multi-scale feature maps.

[0103] Step S3: Multi-scale Flow model, using asymmetric parallel flow to enhance feature expression of each scale, and using fusion flow to fuse multi-scale information, constructing multi-scale Flow model;

[0104] Step S4: Learning and optimization, setting minimizing negative log-likelihood as learning goal, operating Jacobian determinant of parallel flow and fusion flow, optimizing multi-scale Flow model to distinguish normal data and abnormal data distribution;

[0105] Step S5: Abnormal detection and scoring, using pixel-level abnormal positioning and image-level abnormal detection method to calculate abnormal score.

[0106] Embodiment four, based on the above embodiment, in step S3, the multi-scale Flow model specifically includes the following steps:

[0107] Step S31: Parallel flow design, designing asymmetric parallel flow, constructing parallel flow containing L flow blocks, and using asymmetric architecture to construct each flow block inside;

[0108] Step S32: In each parallel flow, using ST network of 3*3 convolution to automatically capture spatial context information, and performing position coding on each feature map in multi-scale feature map, to obtain multi-scale latent feature map;

[0109] Step S33: Multi-scale information fusion, using fusion flow to integrate multi-scale latent feature map, fusion flow using average pooling to splice all input latent feature maps after adjusting to minimum size, and using two structure similar convolution layers to perform feature fusion, to generate fusion feature map;

[0110] Step S34: Feature segmentation, segmenting and readjusting fusion feature map to multi-scale feature map, to obtain segmented feature map;

[0111] Step S35: Scale and shift weight learning, using group convolution to learn scale weight and shift weight of each scale, adjusting features of segmented feature map, group convolution dividing input channels into N groups and performing convolution operation on each group independently, to reduce calculation amount and parameter quantity.

[0112] Embodiment five, based on the above embodiment, in step S5, abnormal detection and scoring, specifically includes the following steps:

[0113] Step S51: Log-likelihood estimation calculation, using log-likelihood to calculate Jacobian determinant of any pixel on segmented feature map, to obtain log-likelihood map, and the formula used is as follows:

[0114] ;

[0115] wherein, , and are natural numbers, represents the pixel value at the th scale in the multi-scale feature map, and represents the pixel value at the th scale in the multi-scale feature map, and represents the Euclidean square of the latent feature represents the latent feature space, represents the log-likelihood estimate of the latent feature space

[0116] Step S52: Log-likelihood map processing, using bilinear interpolation to enlarge the log-likelihood map of all scales to the original image size to obtain a rescaled log-likelihood map;

[0117] Step S53: Pixel-level anomaly localization, converting the rescaled log-likelihood map into a probability map, and calculating a pixel-level anomaly map through additive aggregation, using the following formula:

[0118] ;

[0119] ;

[0120] wherein, represents the probability map of the th scale, is the additive aggregation probability map of the pixel-level anomaly, represents the max-removal, is the pixel-level anomaly map;

[0121] Step S54: Image-level anomaly detection, first adding the rescaled log-likelihood map and then converting it into a probability map, and calculating a global anomaly map, using the following formula:

[0122] ;

[0123] ;

[0124] wherein, represents the multiplication of all values from 1 to 3, is the multiplicative aggregation probability map, is the global anomaly map at the image level;

[0125] Step S55: Calculation of anomaly score, using the following formula:

[0126] ​​ ;

[0127] In the formula, is a set parameter, represents selecting the maximum K values from , and represents the final abnormal score obtained.

[0128] In a specific embodiment, the parallel stream cascades more stream blocks for more high-level feature maps of the channel in Embodiment Six based on the above-mentioned embodiment. In the process of feature extraction, 2, 5 and 8 full convolution stream blocks are stacked as asymmetric parallel streams for primary feature maps, each full convolution block includes two 3*3 convolutions, a RELU activation function and a layer normalization layer, and a 2D position encoding with a channel number of 64 is inserted to capture absolute position information.

[0129] In order to exchange information, a fusion stream is connected to fuse the outputs of the three parallel streams. The stream model is trained from scratch, and the training is performed on a 2080Ti GPU with a batch size of 16. In terms of optimization, we use the Adam optimizer with an initial learning rate of 1e-4, and reduce the learning rate by 3 times at 70% and 90% of the training period. Different training rounds are set according to the size of the data set. In this embodiment, the training round is 100 rounds.

[0130] In the scale and shift weight learning process, the group convolution divides the input channels into 10 groups, and the number of channels in each group depends on the total number of original input channels. For example, if the original input has 100 channels, each group will have 10 channels. Each group performs convolution operation independently, thereby performing fast convolution operation on multi-scale feature maps.

[0131] In view of the problem that the image data generated by the substation monitoring is huge, which increases the difficulty of quickly and accurately identifying abnormalities from the massive images, the scheme uses a multi-scale Flow model to deeply analyze image information, and skillfully distinguishes and meets the dual requirements of image anomaly detection and pixel-level accurate positioning. By using a dynamic multi-scale aggregation strategy, not only the wide range of image abnormal screening is realized, but also the single abnormal pixel is accurately positioned, greatly enhancing the accuracy and response speed of abnormal identification. In addition, the scheme integrates a pre-trained ViT model as a feature extractor, which effectively reduces the feature learning stage, ensures efficient and accurate feature extraction, and further optimizes the overall performance.

[0132] Embodiment Seven, based on the above-mentioned embodiment, the automatic transmission control module uses an optimized adaptive data packet scheduling method to automatically adjust the transmission parameters according to the network conditions. The optimized adaptive data packet scheduling method specifically includes the following steps:

[0133] Step M1: Network monitoring, continuously monitoring the round-trip time, congestion window, in-transit byte number and packet loss rate of each path during data transmission and updating in real time;

[0134] Step M2: Video analysis, extracting video characteristics of video data during data transmission and calculating real-time bit rate, setting frame priority;

[0135] Step M3: Define context and action space, construct state vector, and define action set, state vector includes network state information and video characteristics, action set includes default scheduling, duplicate frame, discard frame and packet loss compensation four actions;

[0136] Step M4: Initialization, initialize a linear model parameter vector for each action, and set the initial confidence boundary value;

[0137] Step M5: Predict action value, for each action, use the current state and the parameter vector corresponding to the action to predict the expected reward using linear regression, obtain the prediction result, and get the action corresponding to the expected reward, calculate and add the confidence boundary;

[0138] Step M6: Select the optimal action, according to the prediction result of the last time, select the action with the highest expected reward as the output of this time;

[0139] Step M7: Execute scheduling and calculate actual reward, schedule data packets according to the action with the highest expected reward, and calculate the actual reward value of this decision;

[0140] Step M8: Update optimization, record the actual reward and state vector when executing the scheduling, and according to the obtained actual reward and state vector, use the improved intelligent optimizer algorithm to optimize and update the parameter vector of the corresponding action, and adjust the confidence boundary.

[0141] Embodiment eight, based on the above embodiment, in step M3, the action set includes four actions of default scheduling, duplicate frame, discard frame and packet loss compensation, corresponding to four mechanisms including scheduling mechanism, repetition mechanism, discard mechanism and compensation mechanism, in step M7, according to the mechanism corresponding to the different actions to schedule data packets, the four mechanisms include the following contents:

[0142] Scheduling mechanism: First, collect the round-trip time of all established transmission paths, and sort these paths according to the round-trip time of the transmission path from small to large. For each pair of adjacent paths in the sorted order, calculate the time interval ratio, and according to the time interval ratio and the in-transit bytes and congestion window of the path, calculate the amount of data that needs to be pre-allocated by the path to ensure that the data can arrive in order; According to the calculated pre-allocated data amount, set the pre-allocated data size for each path, while considering the current to-be-sent data amount S, and update S to reflect the allocated data amount, and create a buffer for each path to store the pre-allocated data, ensuring that data can be continuously taken out from the buffer and sent, optimizing path utilization efficiency;

[0143] Repeat mechanism: Analyze the packet loss rate of each path, identify paths with poor network conditions, determine the path with the lowest packet loss rate as the main path for transmitting original data packets, and other paths with available congestion windows as standby paths for sending redundant copies of data, only high-priority data packets are replicated and distributed to standby paths to increase the probability of data packets reaching the receiving end, and redundant data packets are marked as semi-reliable to avoid unnecessary retransmissions when they are lost, saving bandwidth resources;

[0144] Discard mechanism: Determine which data is critical and which is non-critical based on the priority of video frames. In the case of limited bandwidth, only the header information of M frames is transmitted and the rest is discarded. The header is retained to ensure that the application layer can correctly process the data. When the bandwidth is insufficient to support the transmission of all data, high-priority data is given priority for complete transmission, while non-critical data is discarded;

[0145] Compensation mechanism: When the path packet loss rate is extremely high and the bandwidth is insufficient, consider taking compensation measures. When packaging video data, mark the reliability attribute of the QUIC packet according to the data priority. High-priority data and control information are marked as reliable, and the rest are unreliable. For packets marked as unreliable, do not trigger retransmission when they are lost, thereby saving bandwidth resources. Set a deadline based on the longest round-trip time of all paths. When the time waiting for data to fill the hole exceeds this deadline, use zero padding to fill the hole, reducing the impact on user experience.

[0146] Embodiment nine, based on the above embodiment, in step M8, the parameter vector corresponding to the action is optimized by using an improved intelligent optimizer algorithm, and the confidence boundary is adjusted, which specifically includes the following steps:

[0147] Step M81: initialization stage, set the upper and lower bounds of the search space, the total number of iterations, the balance factor, randomly generate an initial candidate solution and the position of the candidate solution;

[0148] Step M82: evaluate the initial candidate solution, calculate the fitness value of the initial candidate solution, set the initial candidate solution as the global best solution, calculate and record the fitness value of the global best solution;

[0149] Step M83: exploration phase, in the exploration phase, the following steps are performed: sub-iterations, position updating is performed, wherein is less than the total number of iterations;

[0150] Step M84: development and dynamic adjustment, the remaining sub-iterations are performed, dynamic adjustment and position updating are performed, wherein is greater than or equal to the total number of iterations;

[0151] Step M85: output the final global best solution and the fitness value of the best solution.

[0152] Example Ten, see Figure 3 , which is based on the above-mentioned examples, in step M83, the exploration phase, specifically includes the following steps:

[0153] Step M831: position updating, the position of the candidate solution is updated using the updating rule, and the formula used is as follows:

[0154] ;

[0155] ;

[0156] In the formula, is a natural number, is the global best solution, is the position of the th candidate solution, is a random number in the range [0, 1], is the weight, is the balance factor, is the current iteration number, is the total number of iterations, is the natural exponential function, is the weight factor at iteration number ;

[0157] Step M832: boundary check, if the updated position exceeds the upper and lower bounds of the search space, set the updated position to the upper and lower bound values of the search space, otherwise, keep it unchanged;

[0158] ​Step M833: update the global best solution, calculate the fitness value of the updated candidate solution, if f(x) is better than f(gb), update the global best solution and the fitness value of the global best solution, otherwise, keep unchanged, where f(x) is the fitness value of the updated candidate solution, f(gb) is the fitness value of the global best solution;

[0159] Step M834: iteration ends, repeat steps M831 to M833, when the total iteration number is reached, iteration ends. After m iterations, iteration ends.

[0160] In step M84, development and dynamic adjustment, specifically including the following steps:

[0161] Step M841: update the position of the candidate solution, the formula used is as follows:

[0162] ;

[0163] In the formula, is a random number in the range of [0, 1], and are the upper and lower bounds of the search space, respectively;

[0164] Step M842: check the fitness stagnation, if the fitness value has not improved in m consecutive iterations, dynamic adjustment is performed, where m is less than the total iteration number;

[0165] Step M843: dynamically adjust the position of the candidate solution, the formula used is as follows:

[0166] ;

[0167] In the formula, is a random number in the range of [0, 1];

[0168] Step M844: check the boundary and update, confirm that the updated position of the candidate solution is still within the search space, calculate the fitness value of the updated candidate solution, update the global best solution and the fitness value of the global best solution;

[0169] Step M845: iteration optimization, repeat steps M841 to M844, when the total iteration number is reached, iteration ends.

[0170] For remote and harsh environment, the instability of network connection can further interfere with the continuity of data transmission problem and the transmission of high-definition image greatly occupies the bandwidth, but the network resource is limited, the scheme adopts the optimized adaptive data packet scheduling method to ensure the continuity and stability of video and image data in the transmission process, the method carefully designs four sets of mechanism, each set solves different network problems, so as to guarantee the coherence and reliability of video and image data in complex transmission environment, through continuous self-optimization to adapt to the dynamic change of network environment, effectively enhance the reliability of data transmission, at the same time realize the optimization of network resource utilization efficiency, and bring smooth and unobstructed user experience, maximize the use of existing network resources.

[0171] The above describes the present application and its embodiments, which are not limited, and the drawings shown are only one of the embodiments of the present application, and the actual structure is not limited thereto. In summary, if a person skilled in the art is inspired thereby, without departing from the spirit of the present application, without creative design, similar structure and embodiments similar to the technical solution can be designed, which shall belong to the protection scope of the present application.

Claims

1. A substation image data automatic non-sensing transmission control system, characterized in that: The system comprises an image acquisition module, an intelligent video analysis module, an automatic transmission control module, a security and privacy protection module, and a non-sensing interaction design module. The image acquisition module deploys high-definition cameras in key areas of the substation for all-weather, multi-angle acquisition of high-definition video images. The intelligent video analysis module is responsible for analyzing the collected high-definition video images using a multi-scale Flow model, identifying normal and abnormal data in the high-definition video images, obtaining abnormal data information and triggering an abnormal alarm. The automatic transmission control module transmits high-definition video images as video data and uses an optimized adaptive data packet scheduling method to automatically adjust transmission parameters according to network conditions. The security and privacy protection module has strict security measures built in, and for high-definition video images involving personal privacy, anonymization or specific area masking is performed. The non-sensing interaction design module provides a visual interface for automatic scheduling and preset rule execution, reducing the need for manual operation. The automatic transmission control module uses an optimized adaptive data packet scheduling method to automatically adjust transmission parameters according to network conditions. Step M1: Network monitoring, continuously monitor the round-trip time, congestion window, in-transit byte count, and packet loss rate of each path during data transmission and update in real time. Step M2: Video analysis, extract video characteristics of video data during data transmission and calculate real-time bit rate, set frame priority. Step M3: Define context and action space, construct state vector, and define action set, state vector includes network state information and video characteristics. Step M4: Initialization, initialize a linear model parameter vector for each action and set the initial confidence boundary value. Step M5: Predict action value, for each action, use the current state and the parameter vector corresponding to the action to predict the expected reward using linear regression, get the prediction result, and get the action corresponding to the expected reward, calculate and add the confidence boundary. Step M6: Select the optimal action, select the action with the highest expected reward as the output according to the last prediction result. Step M7: Execute scheduling and calculate actual reward, schedule data packets according to the action with the highest expected reward selected, and calculate the actual reward value of this decision. Step M8: Update optimization, record the actual reward and state vector when scheduling, and use the improved intelligent optimizer algorithm to optimize and update the parameter vector of the corresponding action according to the obtained actual reward and state vector, adjust the confidence boundary.

2. The substation image data automatic non-sensing transmission control system according to claim 1, characterized in that: The intelligent video analysis module is responsible for analyzing the collected high-definition video images using a multi-scale Flow model, specifically including the following steps: Step S1: Primary feature extraction, use a pre-trained ViT model as a feature extractor to forward propagate the high-definition video image, extract feature maps at different levels, and obtain primary feature maps. Step S2: Multi-scale feature extraction, use an average pooling layer to downsample the primary feature maps to multiple scales to obtain multi-scale feature maps. Step S3: Multi-scale Flow model, using asymmetric parallel flow to enhance feature expression of each scale, and using fusion flow to fuse multi-scale information, constructing a multi-scale Flow model; Step S4: Learning and optimization, setting the minimization of negative log-likelihood as the learning goal, operating the Jacobian determinant of parallel flow and fusion flow, optimizing the multi-scale Flow model to distinguish normal data and abnormal data distribution; Step S5: Abnormal detection and scoring, using pixel-level abnormal positioning and image-level abnormal detection method to calculate abnormal score.

3. The substation image data automatic non-sensing transmission control system according to claim 2, characterized in that: In step S3, the multi-scale Flow model specifically includes the following steps: Step S31: Parallel flow design, designing an asymmetric parallel flow, constructing a parallel flow containing L flow blocks, and using an asymmetric architecture to build each flow block inside; Step S32: In each parallel flow, using the ST network of 3*3 convolution to automatically capture spatial context information, and performing position encoding on each feature map in the multi-scale feature map to obtain a multi-scale latent feature map; Step S33: Multi-scale information fusion, using fusion flow to integrate the multi-scale latent feature map, using average pooling to adjust all input latent feature maps to the minimum size after splicing, and using two convolution layers with similar structures to fuse features to generate a fusion feature map; Step S34: Feature segmentation, segmenting the fusion feature map and adjusting it to a multi-scale feature map to obtain a segmented feature map; Step S35: Scale and shift weight learning, using group convolution to learn the scale weight and shift weight of each scale to adjust the features of the segmented feature map.

4. The substation image data automatic non-sensing transmission control system of claim 2, wherein: In step S5, abnormal detection and scoring, specifically includes the following steps: Step S51: Log-likelihood estimation calculation, using log-likelihood to calculate the Jacobian determinant of any pixel on the segmented feature map to obtain a log-likelihood map; Step S52: Log-likelihood map processing, using bilinear interpolation to enlarge all scale log-likelihood maps to the original image size to obtain a rescaled log-likelihood map; Step S53: Pixel-level abnormal positioning, converting the rescaled log-likelihood map to a probability map and calculating a pixel-level abnormality map through additive aggregation; Step S54: Image-level anomaly detection, adding the rescaled log-likelihood map first and then converting it to a probability map, and calculating a global anomaly map; Step S55: Calculate the abnormal score.

5. The substation image data automatic and unconscious transmission control system according to claim 1, characterized in that: In step M8, the parameter vector corresponding to the action is optimized and updated using the improved intelligent optimizer algorithm to adjust the confidence boundary, specifically including the following steps: Step M81: Initialization stage, setting the upper and lower bounds of the search space, the total number of iterations, the balance factor, randomly generating the initial candidate solution and the position of the candidate solution; Step M82: Evaluate the initial candidate solution, calculate the fitness value of the initial candidate solution, set the initial candidate solution as the global optimal solution, calculate and record the fitness value of the global optimal solution; Step M83: Exploration phase, in which the exploration phase is performed sub-iterations, a position update is performed, in which is less than the total number of iterations; Step M84: Develop and dynamically adjust, for the remaining sub-iterations, the dynamic adjustments and position updates, where add is equal to the total number of iterations; Step M85: Output the final global optimal solution and the fitness value of the optimal solution.

6. The substation image data automatic and unconscious transmission control system according to claim 5, characterized in that: In step M83, the exploration stage, specifically includes the following steps: Step M831: Position update, updating the position of the candidate solution using the update rule; Step M832: Check the boundary, if the updated position exceeds the upper and lower bounds of the search space, set the updated position to the upper and lower bounds of the search space, otherwise, keep it unchanged; Step M833: Update the global optimal solution, calculate the fitness value of the updated candidate solution, if f(x) is better than f(gb), update the global optimal solution and the fitness value of the global optimal solution, otherwise, maintain it unchanged, wherein f(x) is the fitness value of the updated candidate solution, f(gb) is the fitness value of the global optimal solution; Step M834: The iteration ends, and steps M831 to M833 are repeated until After the sub-iteration, the iteration ends.

7. The substation image data automatic and unconscious transmission control system according to claim 5, characterized in that: In step M84, development and dynamic adjustment, specifically including the following steps: Step M841: Update the position of the candidate solution; Step M842: Check the fitness stagnation, if the fitness value has not improved in m consecutive iterations, perform dynamic adjustment, wherein m is less than the total number of iterations; Step M843: Dynamically adjust the position of the candidate solution; Step M844: Check the boundary and update, confirm that the updated position of the candidate solution is still within the search space, calculate the fitness value of the updated candidate solution, update the global optimal solution and the fitness value of the global optimal solution; Step M845: Iterative optimization, repeat steps M841 to M844, when the total number of iterations is reached, the iteration ends.

Images