Arc light detection method and device based on space-time fusion, storage medium and terminal

Abstract

The invention provides an arc light detection method and device based on space-time fusion, a storage medium and a terminal, and the method comprises the steps: obtaining a to-be-recognized arc light image, and obtaining the coordinate data of all preset key points in the to-be-recognized arc light image as a preset node data set; acquiring graph data corresponding to the arc light image to be identified based on the preset node data set, the graph data taking node coordinate data as nodes of a graph, and the graph data taking Euclidean distances between the nodes as edges of the graph; detecting whether a non-zero value exists in an edge in the graph data, if yes, extracting spatial features of the arc light image to be identified by using a CNN, and obtaining graph time sequence features based on the graph data; and fusing the graph time sequence features and the spatial features, and inputting the fused features into a preset classifier for classification to obtain an identification result corresponding to the to-be-identified arc light image. According to the invention, the accuracy of high-voltage arc light detection can be improved, and consumption of computing resources of arc light detection is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of arc monitoring, and relates to an arc detection method based on time-space fusion, and in particular to an arc detection method and device based on time-space fusion, a storage medium and a terminal. Background Art

[0002] Switchgear is the core equipment in the power system and is widely used in medium and low voltage distribution systems. Ensuring its stable operation is crucial to maintaining the health of the power system and improving the quality and reliability of power supply. Arc discharge is one of the most serious accidents in medium and low voltage switchgear. (What causes arc discharge?) It is usually accompanied by electric light, irritating sounds, and radiates in the form of electromagnetic waves, generating strong light and heat energy and over-limit pressure. The strong light and heat of arc faults can easily cause serious damage to the equipment in the distribution room, causing copper and aluminum connections to melt, switchgear and busbars to burn, cables to catch fire, toxic gas, explosions and other serious consequences, greatly threatening the safe and stable operation of the power grid and the safety of personnel. Therefore, extensive research has been carried out on arc monitoring and diagnosis at home and abroad.

[0003] In the early stage of arc monitoring research, countries such as Germany and the United States mainly focused on the monitoring of voltage and current changes. However, due to challenges in sensor sensitivity, electronic technology, communication transmission, etc., progress was slow. In the early 1990s, the British Electrical Test Center proposed a method of using photothermal sensors combined with sound for monitoring. Although it solved the problem of arc monitoring to a certain extent, the monitoring effect was not ideal due to the interference of external light and partial discharge. In recent years, the arc protection system has been gradually promoted in China, and the application field has expanded from distribution facilities to power generation facilities. For example, in a certain engineering expansion project in Henan in 1995, arc monitoring sensors were installed in the distribution cabinet for the first time, and overcurrent protection was performed through outgoing line protection, realizing the monitoring and alarm of arc; Nanjing Hongyi Electric Company developed the DPR360ARC arc protection system, and Baoding Unt Electric Company developed the UNT-EAP arc protection system. These systems have significantly improved the detection and response capabilities of arc faults.

[0004] In response to arc diagnosis research, many foreign manufacturers have launched arc protection and diagnostic products, such as Finland's UTU-Falcon and Australia's Rizner-EagleEye. After 2000, the VAMP220 series products launched by Finland's VAMP and ABB's arc protection products based on ultra-fast grounding switches applied advanced current and voltage monitoring technology to achieve faster and more accurate fault detection. China has also made certain progress in arc diagnosis technology. In 2008, Nanjing Hongyi Electric Company developed the DPR360ARC system, which used high-speed current limiting circuit breakers and fast vacuum switches to achieve rapid removal of arc faults within 50ms. The UNT-EAP system of Baoding Unt Electric Company and the EP200 system of Shaanxi Electric Power Research Institute have improved the detection accuracy and response speed of arc faults through advanced fault detection algorithms. However, at present, in the monitoring of arc single-phase grounding faults, there is still a need to further improve the detection accuracy and reliability of protection devices at home and abroad.

[0005] With the development of computer, Internet of Things and artificial intelligence technology, the collection of switchgear arc data has become more extensive, making arc discharge fault monitoring and detection based on big data and neural networks possible. Big data analysis technology can extract arc fault characteristics and patterns from historical data and real-time data, and identify early signs and potential risks through data mining and statistical analysis.

[0006] However, existing technologies require a large amount of high-quality data support. High-quality arc discharge data is scarce, and high-voltage arc discharge point data is difficult to obtain. In this case, arc detection requires continuous calls to neural networks, which consumes a lot of computing resources. At the same time, data reasons also lead to low accuracy in high-voltage arc detection. Summary of the invention

[0007] The purpose of the present invention is to provide an arc detection method and device based on time-space fusion, a storage medium and a terminal, which are used to solve the technical problems in the prior art that arc detection has large consumption of computing resources and low accuracy of high-voltage arc detection.

[0008] In a first aspect, the present invention provides an arc light detection method based on time-space fusion, comprising:

[0009] Acquire an arc light image to be identified, and acquire coordinate data of all preset key points in the arc light image to be identified as a preset node data set;

[0010] Based on the preset node data set, the graph data corresponding to the arc light image to be identified is acquired, wherein the graph data uses the node coordinate data as the nodes of the graph, and the graph data uses the Euclidean distance between the nodes as the edges of the graph;

[0011] Detect whether there is a non-zero value on the edge in the graph data, and if so, use CNN to extract the spatial features of the arc light image to be identified, and obtain the graph temporal features based on the graph data;

[0012] After the image temporal features are fused with the spatial features, the features are input into a preset classifier for classification to obtain a recognition result corresponding to the arc light image to be recognized.

[0013] In one embodiment of the present invention, obtaining graph time series features based on the graph data includes:

[0014] Acquire the image data corresponding to the arc image to be identified as current image data, and acquire the image data corresponding to the arc image at a previous moment of the arc image to be identified as previous moment image data;

[0015] Subtract the current graph data from the graph data at the previous moment to obtain a preliminary difference matrix;

[0016] Relu processing is performed on each element in the preliminary difference matrix to obtain a Euclidean distance difference matrix as the image temporal feature corresponding to the arc light image to be identified.

[0017] In one embodiment of the present invention, it further includes:

[0018] Each element in the Euclidean distance difference matrix is ​​subjected to exponential processing to obtain an exponentially processed matrix as a graph time series feature corresponding to the arc light image to be identified;

[0019] The base of the exponential processing is greater than 1, and the element to be exponentially processed in the Euclidean distance difference matrix is ​​used as the exponent, and the power result of the exponential processing is used as the corresponding new element in the Euclidean distance difference matrix.

[0020] In one embodiment of the present invention, fusing the graph temporal features with the spatial features includes:

[0021] Taking the feature map matrix size of the spatial feature as the target size, reshaping the size of the corresponding Euclidean distance difference matrix to obtain a reshaped Euclidean distance difference matrix;

[0022] Obtain the Hadamard product of the reshaped Euclidean distance difference matrix and the feature graph matrix of the spatial feature to obtain a fusion result of the graph time series feature and the spatial feature.

[0023] In one embodiment of the present invention, the preset classifier includes multiple fully connected layers, and a softmax layer connected to the last fully connected layer.

[0024] In one embodiment of the present invention, it further includes:

[0025] The model training method corresponding to the method is training based on the transfer migration strategy, and the transfer migration strategy includes:

[0026] Use the first voltage level data set to train the model until the model converges and obtain a converged model, and save the corresponding parameters as the first voltage level parameters;

[0027] Using the first voltage and other parameters as pre-training parameters, fine-tuning the convergence model using the second voltage level data set, and saving the corresponding parameters as the second voltage level parameters;

[0028] Wherein, the first voltage is smaller than the second voltage.

[0029] In one embodiment of the present invention, if the voltage in the recognition result includes multiple voltage levels, the multiple voltage levels are sorted to obtain level sorting data, and the model training method corresponding to the method is to perform level-by-level iterative training based on the transfer migration strategy and the level sorting data to obtain the result corresponding to the highest voltage level data in the level sorting data as the final result;

[0030] In each iteration, two adjacent data in the graded data are obtained as the first voltage level data set and the second voltage level data set for training based on the transfer migration strategy.

[0031] In a second aspect, the present invention further provides an arc detection device based on time-space fusion, characterized in that it comprises:

[0032] A data acquisition module, used for acquiring an arc image to be identified, and acquiring coordinate data of all preset key points in the arc image to be identified as a preset node data set;

[0033] A graph data construction module, used for acquiring graph data corresponding to the arc light image to be identified based on the preset node data set, wherein the graph data uses node coordinate data as graph nodes, and the graph data uses Euclidean distances between nodes as graph edges;

[0034] A trigger discrimination module is used to detect whether there is a non-zero value in the edge of the graph data. If so, CNN is used to extract the spatial features of the arc light image to be identified, and the graph temporal features are obtained based on the graph data;

[0035] The fusion classification module is used to fuse the image temporal features with the spatial features and then input them into a preset classifier for classification to obtain the recognition result corresponding to the arc light image to be recognized.

[0036] In a third aspect, the present invention further provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the arc detection method based on time-space fusion as described above.

[0037] In a fourth aspect, the present invention further provides a terminal, comprising a processor and a memory, wherein the memory is communicatively connected to the processor;

[0038] The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the terminal executes the arc detection method based on time-space fusion as described above.

[0039] As described above, the arc detection method and device based on time-space fusion, storage medium and terminal described in the present invention have the following beneficial effects:

[0040] 1. By judging the trigger situation, the CNN neural network is called to extract the spatial features of the arc image to be identified and the image timing features when the trigger conditions are met, which significantly reduces the consumption of computing resources by the empty signal. The Euclidean distance difference matrix is ​​used as the image timing features corresponding to the arc image to be identified, which further saves the computing resources required for detection.

[0041] 2. After fusing the image temporal features with the spatial features, they are input into the preset classifier for classification. This is to fuse the temporal features represented by the image temporal features with the spatial features extracted by CNN, and use the image temporal features as the spatiotemporal coupling enhancement coefficient to enhance the spatial features. This can significantly improve the accuracy of the recognition results of the electrode to which the arc belongs in the arc image to be identified, and also improve the accuracy of high-voltage arc detection.

[0042] 3. The model training through the step-by-step iterative transfer migration strategy can better solve the data scarcity problem caused by the difficulty in preparing and acquiring actual arc discharge data, thereby improving the accuracy of high-voltage arc detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A schematic flow chart of an arc detection method based on time-space fusion according to an embodiment of the present invention is shown.

[0044] Figure 2 A schematic diagram of data flow of an arc detection method based on time-space fusion described in an embodiment of the present invention is shown.

[0045] Figure 3 A schematic diagram of image digitization in the arc detection method based on time-space fusion described in an embodiment of the present invention is shown.

[0046] Figure 4 A schematic diagram of spatial feature extraction in an arc detection method based on time-space fusion described in an embodiment of the present invention is shown.

[0047] Figure 5 A schematic diagram of graph temporal feature extraction in the arc detection method based on spatiotemporal fusion described in an embodiment of the present invention is shown.

[0048] Figure 6 A schematic diagram of a step-by-step iterative transfer migration strategy in an arc light detection method based on time-space fusion according to an embodiment of the present invention is shown.

[0049] Figure 7 A schematic structural diagram of an arc detection device based on time-space fusion according to an embodiment of the present invention is shown.

[0050] Figure 8 A schematic diagram of the structure of a terminal according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0051] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0052] The principles and implementation methods of the arc detection method and device based on time-space fusion, storage medium and terminal of this embodiment will be explained in detail below, so that those skilled in the art can understand the arc detection method, device, storage medium and terminal based on time-space fusion of this embodiment without creative work.

[0053] In order to solve the above technical problems existing in the prior art, an embodiment of the present invention provides an arc light detection method based on time-space fusion.

[0054] Figure 1 FIG. 1 shows a flow chart of the arc light detection method based on time-space fusion according to an embodiment of the present invention, referring to FIG. Figure 1 As shown, the arc detection method based on time-space fusion in the embodiment of the present invention mainly includes the following steps:

[0055] Figure 2 FIG. 1 shows a schematic diagram of data flow of the arc detection method based on time-space fusion according to an embodiment of the present invention, referring to FIG. Figure 2As shown, the input image i+1 is used as the arc image to be identified, the input image i is the arc image at the previous moment, the image temporal features are obtained based on the arc image to be identified and the arc image at the previous moment, and the spatial features are extracted using CNN based on the arc image to be identified (input image i+1), and then the image temporal features and spatial features are fused to obtain the fusion result, and the fusion result is input into the preset classifier to obtain the final predicted recognition result. It should be noted that Figure 2 The data flow is explained by taking the case where there is a non-zero value on the edge in the image data corresponding to the arc light image to be identified as an example.

[0056] Step S100: Acquire an arc image to be identified (input image i+1), and acquire coordinate data of all preset key points in the arc image to be identified as a preset node data set.

[0057] Specifically, a high-speed camera is used to continuously capture images of the position to be detected, and all images are arranged in frames. For example, when an image captured at a certain moment is used as an arc image to be identified (input image i+1), the previous frame image of the arc image to be identified (input image i) is the corresponding arc image at the previous moment. According to actual needs, those skilled in the art can use other tools other than high-speed cameras to implement the above-mentioned image acquisition process, and can also independently set the time interval between different frame images. It should be noted that in this embodiment, only the situation when any one arc image to be identified is used for detection is described, but each of the above images can be used as an arc image to be identified and detected using the method of the present invention.

[0058] After obtaining the arc image to be identified, the coordinate data of all preset key points in the arc image to be identified are obtained as a preset node data set. All preset key points include arc starting point, arc turning point and arc termination point, wherein the number of arc turning points is at least one. When no key point is detected in the arc image to be identified, 0 or null value is used as the key point position coordinate, and the coordinate data of all preset key points in the arc image to be identified (that is, the position data of the preset key points in the image) are collected as the preset node data set corresponding to the arc image to be identified. It should be noted that the method used to obtain the coordinate data of all preset key points in the arc image to be identified in this embodiment is a traditional method. The traditional method is different from the deep learning algorithm, and its extraction speed is faster than the deep learning algorithm. There are many existing technologies that can be used to implement the traditional method of obtaining key point coordinate data from an image, such as Harris corner detection algorithm, ORB (Oriented FAST and Rotated BRIEF) algorithm, edge detection algorithm, etc. The present invention will not be described in detail here. It is only necessary to use a non-deep learning algorithm to obtain the coordinate data of all preset key points in the arc image to be identified. Optionally, the OpenCV library in Python integrates many traditional computer vision algorithms. This embodiment uses the Python programming language to call the open cv library to obtain the coordinate data of all preset key points in the arc image to be identified. These traditional methods detect key points in the image through mathematical analysis and image processing technology. Their advantages are fast speed and low hardware requirements, but compared with the neural network-based methods, their adaptability and accuracy to complex scenes are relatively low.

[0059] Step S200: based on a preset node data set, graph data corresponding to the arc image to be identified is obtained, wherein the graph data uses the node coordinate data as the nodes of the graph, and the graph data uses the Euclidean distance between the nodes as the edges of the graph.

[0060] Based on the preset node data set of the arc image to be identified obtained in step S100, the Euclidean distance between every two key points is obtained, and the preset key points in the preset node data set are assigned to the nodes of the graph, and the Euclidean distances between the preset key points are assigned to the edges of the graph to obtain the graph data corresponding to the arc image to be identified. The graph data is a matrix formed by the Euclidean distances between all the preset key points, which stores the distance information between every two preset key points. Figure 3 FIG. 1 shows a schematic diagram of the image data in the arc detection method based on time-space fusion according to an embodiment of the present invention, referring to FIG. Figure 3 As shown, the preset key point data in the arc light image to be identified can be converted into graph data. Specifically, any two preset key points a are obtained. i With b i The Euclidean distance between them is given by the following formula:

[0061]

[0062] Among them, x ai Indicates the preset key point a i The horizontal axis, X bi Indicates the preset key point b i The horizontal axis, y ai Indicates the preset key point a i The vertical coordinate, y bi Indicates the preset key point b i The vertical coordinate, d ai,bi Indicates the preset key point a i With b i The Euclidean distance between .

[0063] Step S300: Detect whether there is a non-zero value on the edge in the graph data. If so, use CNN to extract the spatial features of the arc image to be identified, and obtain the graph temporal features based on the graph data.

[0064] Specifically, if there is no arc in any arc image to be identified, then all edges in the corresponding graph data are 0. When the edges in the graph data at a certain moment (for example, the i+1th moment) are not all 0, that is, when the Euclidean distance has an increment, it means that an arc has appeared. Therefore, it is detected whether there is a non-zero value on the edge in the graph data. If there is a non-zero value, CNN is used to extract the spatial features of the arc image to be identified, and the graph temporal features are obtained based on the graph data. Figure 4 FIG. 1 shows a schematic diagram of spatial feature extraction in the arc light detection method based on time-space fusion according to an embodiment of the present invention, with reference to FIG. Figure 4 As shown, the arc light image to be identified is input into the CNN network after image preprocessing, and the feature map output by the preset convolution layer is extracted as the spatial feature. The specific structure of the CNN network is not further limited here. Those skilled in the art can construct a suitable CNN network according to actual needs, and select the corresponding preset convolution layer, for example, select the feature map output by the 5th convolution layer as the spatial feature.

[0065] The present invention detects whether there is a non-zero value on the edge of the graph data, which is equivalent to a trigger. By judging the trigger situation, the CNN neural network is called to extract the spatial features of the arc image to be identified and the temporal features of the graph when the trigger condition is met. When the trigger condition is not met, there is no need to call the CNN neural network. In the full process information of arc discharge collected by high-speed cameras, the proportion of valid signals, i.e. arc signals, is relatively low, and most of the signals are invalid empty signals, that is, the number of images taken during an arc discharge process may include hundreds of images, but only a small part of these images have obvious arc morphological changes. The technical solution of the present invention calls the CNN neural network to extract spatial features only when there is a valid signal, which significantly reduces the consumption of computing resources by empty signals, that is, reduces the computing resource consumption of the overall arc detection.

[0066] Optionally, Figure 5 FIG. 1 shows a schematic diagram of the time series feature extraction in the arc detection method based on time-space fusion according to an embodiment of the present invention, with reference to FIG. Figure 5 As shown, Figure 5 Where input graph i+1 represents the arc image to be identified, input graph i represents the arc image at the previous moment of the arc image to be identified, Matrix_i+1 represents the Euclidean distance difference matrix corresponding to the arc image to be identified, and obtaining graph time series features based on graph data includes the following steps:

[0067] Step S301: obtaining the image data corresponding to the arc image to be identified (input image i+1) as the current image data, and obtaining the image data corresponding to the arc image i at the previous moment of the arc image to be identified as the previous moment image data.

[0068] Since step S200 introduces the process of how to obtain graphic data, according to step S200, based on the arc image to be identified (input graphic i+1) (current moment), the corresponding graphic data is obtained as the current graphic data, and the graphic data corresponding to the arc image (input graphic i) at the previous moment (previous frame) of the arc image to be identified is obtained as the graphic data at the previous moment.

[0069] Step S302: Subtract the current graph data from the graph data at the previous moment to obtain a preliminary difference matrix.

[0070] The current graph data stores all the Euclidean distances at the current moment, and the previous moment graph data stores all the Euclidean distances at the previous moment. Subtracting the two is to subtract the elements of the edges at corresponding positions (that is, representing the Euclidean distances between the same two preset key points), and use the difference result as the preliminary difference matrix, which stores the change value of each Euclidean distance.

[0071] Step S303: Perform ReLU processing on each element in the preliminary difference matrix to obtain a Euclidean distance difference matrix as the image temporal features corresponding to the arc light image to be identified.

[0072] Specifically, the ReLU processing is performed on any element in the preliminary difference matrix according to the following formula:

[0073]

[0074] Among them, Matrix ij is the element in the i-th row and j-th column of the preliminary difference matrix, that is, when there is a positive increment in the element in the preliminary difference matrix, the increment value is retained, otherwise it is taken as zero. After Relu processing is performed on all elements in the preliminary difference matrix, the processing result is used as the Euclidean distance difference matrix, and the Euclidean distance difference matrix is ​​used as the image timing feature corresponding to the arc light image to be identified. Step S303 only needs to determine whether there is a positive increment in the timing feature, which further saves the computing resources required for detection.

[0075] Optionally, obtaining graph time series features based on graph data further includes step S304:

[0076] Step S304: perform index processing on each element in the Euclidean distance difference matrix to obtain the indexed matrix as the image temporal feature corresponding to the arc light image to be identified.

[0077] The base of the exponential processing is greater than 1, and the element to be exponentially processed in the Euclidean distance difference matrix is ​​used as the exponent, and the power result of the exponential processing is used as the corresponding new element in the Euclidean distance difference matrix. Specifically, the exponential processing of any element refers to the following formula:

[0078]

[0079] in, Represents the element in the i-th row and j-th column of the Euclidean distance difference matrix, Weight ij Represents the corresponding new elements in the Euclidean distance difference matrix, and e is the base of the exponential processing. The matrix composed of all the new elements is used as the image timing features corresponding to the arc image to be identified. The image timing features obtained by adding step S304 are compared with the image timing features obtained in step S303. The image timing features after the exponential processing can ensure that when there is a positive increment, the corresponding new elements in the Euclidean distance difference matrix are greater than 1, and the image timing features are amplified. The image timing features are strengthened during the subsequent fusion, which is beneficial to improving the recognition effect.

[0080] Step S400: After fusing the temporal features and spatial features of the image, the images are input into a preset classifier for classification to obtain a recognition result corresponding to the arc light image to be recognized.

[0081] Optionally, fusing the graph temporal features with the spatial features includes: taking the feature graph matrix size of the spatial features as the target size, reshaping the size of the Euclidean distance difference matrix to obtain a reshaped Euclidean distance difference matrix; obtaining the Hadamard product of the reshaped Euclidean distance difference matrix and the feature graph matrix of the spatial features to obtain the fusion result of the graph temporal features and the spatial features. The specific fusion process refers to the following formula:

[0082] CNN Final_map =Weight*CNN map

[0083] CNN Final_map Represents the fusion result, CNN map Represents the spatial features extracted by CNN, and Weight represents the reshaped Euclidean distance difference matrix. The characteristics of this fusion method are convenient reshaping, uniform size, and fast fusion. In addition, those skilled in the art can also set other feature fusion methods according to needs. This application fuses the graph time series features with the spatial features and inputs them into the preset classifier for classification. It fuses the time series features represented by the graph time series features with the spatial features extracted by CNN, and uses the graph time series features as the spatiotemporal coupling enhancement coefficient to enhance the spatial features, which can significantly improve the accuracy of the recognition results. The recognition result represents the source of the arc in the arc image to be identified, that is, the arc is emitted by what kind of electrode at what voltage. Optionally, the type of recognition result includes preset electrodes at different voltage levels, and the preset electrodes include copper electrode-copper electrode, copper electrode-stainless steel electrode, and copper electrode-galvanized steel plate electrode. The voltage level can be set according to actual needs. For example, when the voltage level includes 1kv, 2kv, 3kv, 4kv and 5kv, there are 3 preset electrodes at this time, then there are 15 types of recognition results, that is, 15 classification tasks. This application uses a trigger as a prerequisite for calling CNN and subsequent processing. On the one hand, it saves a lot of computing resources. On the other hand, when the trigger is triggered, that is, after using CNN to extract spatial features, extract graph temporal features, and subsequent feature fusion and classification tasks, it can significantly improve the accuracy of the recognition results of the electrode to which the arc belongs in the arc image to be identified, and also achieve an improvement in the accuracy of high-voltage arc detection.

[0084] Optionally, the preset classifier includes multiple fully connected layers and a softmax layer connected to the last fully connected layer. According to actual needs, those skilled in the art can select other preset classifiers.

[0085] Based on the above description of steps S100-S400, a person skilled in the art can understand the specific process of the arc detection method based on time-space fusion of the present invention, and can reproduce it according to the process, and the reproduction includes two processes of model training and model recognition. In practice, the two processes of model training and model recognition are corresponding, and the training parameters are obtained based on model training and saved (that is, the parameters are saved). During model recognition, the same model as the training process is used to call the saved parameters to realize recognition. However, the model training method can be further improved to improve the accuracy of high-voltage arc detection. Optionally, the model training method corresponding to the arc detection method based on time-space fusion in this embodiment is training based on the transfer migration strategy, and the transfer migration strategy includes: using the first voltage level data set to train the model until the model converges and obtain the converged model, that is, the converged model is trained to the converged model, and the corresponding parameters are saved as the first voltage level parameters; using the first voltage and other parameters as pre-training parameters, using the second voltage level data set to fine-tune the converged model, and saving the corresponding parameters as the second voltage level parameters; wherein the first voltage is less than the second voltage. In the actual data collection process, a large amount of arc discharge data can be collected under low voltage by designing low voltage arc discharge tests, but arc discharge data under high voltage is difficult to collect and is therefore relatively scarce. The above transfer migration strategy uses low voltage data for model training, and then transfers low voltage knowledge to high voltage arc detection scenarios through transfer learning technology, which can solve the problem of difficulty in obtaining high voltage arc data, resulting in a scarce data set, and further leading to inaccurate high voltage arc detection.

[0086] Optionally, Figure 6 FIG. 1 is a schematic diagram showing a step-by-step iterative transfer migration strategy in the arc light detection method based on time-space fusion according to an embodiment of the present invention, with reference to FIG. Figure 6As shown, if the voltage in the recognition result includes multiple voltage levels, the multiple voltage levels are sorted to obtain the ranking data. The model training method corresponding to the arc detection method based on time-space fusion in this embodiment is to perform step-by-step iterative training based on the transfer migration strategy and the ranking data to obtain the result corresponding to the highest voltage level data in the ranking data as the final result; wherein, in each iteration, two adjacent data in the ranking data are obtained as the first voltage level data set and the second voltage level data set for training based on the transfer migration strategy. Specifically, multiple voltage levels are sorted according to the voltage size to obtain ranking data T1, T2, T3...Tn from small to large voltages. Take two adjacent data in the ranking data in the order of voltage from small to large as the first voltage and the second voltage each time, and train based on the transfer migration strategy, iterate the training process until the second voltage level parameter corresponding to the highest voltage in the ranking data is obtained as the final saved parameter (that is, the last iteration), at which time the second highest voltage in the ranking data corresponds to the first voltage level parameter. Specifically, for the ranked data T1, T2, T3...Tn, in the first iteration, T1 is taken as the first voltage level data set and T2 is taken as the second voltage level data set, and the transfer migration strategy is used for training, and the storage parameter 2 is obtained as the training result corresponding to T2; in the second iteration, T2 is taken as the first voltage level data set and T3 is taken as the second voltage level data set, and the transfer migration strategy is used for training, and the storage parameter 3 is obtained as the training result corresponding to T3; it is iterated according to the above method until the storage parameter n is obtained as the training result corresponding to Tn. In the above process, the model structure has not changed, but by transferring the model step by step and fine-tuning the existing model with the new data set, the storage parameter n corresponding to the highest voltage level data set is finally obtained. Combined with the model and the final storage parameter n, it can be applied to the accurate recognition of the arc image to be identified corresponding to the highest level voltage. Therefore, the use of this method can achieve accurate recognition of arc images to be identified at higher voltages, such as accurate recognition of arc images to be identified at 40.5kv voltage.

[0087] This is because transfer learning also faces the problem of poor transfer effect when the gap between the source domain and the target domain is too large. Therefore, by designing experiments with different voltage levels and collecting arc discharge experimental data under different voltages, these data are trained with a step-by-step iterative transfer migration strategy from small to large voltages. In each iteration, the previous level (relatively low voltage) voltage arc discharge data is used to pre-train network parameters, and the next level (relatively high voltage) voltage arc discharge data is used to fine-tune the model parameters. The specific formula is as follows:

[0088] Para i+1 =FineTune(Para i )

[0089] Among them, Para i Indicates the saved parameters corresponding to the previous level voltage arc discharge data, FineTune indicates the training process of the transfer migration strategy, Para i+1 Indicates the storage parameters corresponding to the next level voltage arc discharge data.

[0090] The step-by-step iterative transfer migration strategy training model has low data volume requirements and can better solve the data scarcity problem caused by the difficulty in preparing and acquiring actual arc discharge data, thereby improving the accuracy of high-voltage arc detection.

[0091] The protection scope of the arc detection method based on time-space fusion in the embodiment of the present invention is not limited to the execution order of the steps listed in this embodiment. All solutions implemented by adding, reducing or replacing steps in the prior art based on the principles of the present invention are included in the protection scope of the present invention.

[0092] The arc detection method based on time-space fusion of the embodiment of the present invention judges the trigger situation, and only calls the CNN neural network to extract the spatial features of the arc image to be identified and extract the graph timing features when the trigger conditions are met, which significantly reduces the consumption of computing resources by the empty signal. At the same time, the timing features represented by the graph timing features are fused with the spatial features extracted by CNN, and the graph timing features are used as the time-space coupling enhancement coefficient to enhance the spatial features, thereby improving the accuracy of high-voltage arc detection. In addition, the model is trained through a step-by-step iterative transfer migration strategy to improve the accuracy of high-voltage arc detection.

[0093] In order to solve the above technical problems existing in the prior art, an embodiment of the present invention further provides an arc detection device based on time-space fusion.

[0094] Figure 7 FIG. 1 shows a schematic diagram of the structure of an arc detection device based on time-space fusion according to an embodiment of the present invention, referring to FIG. Figure 7 As shown, the arc detection device based on time-space fusion according to the embodiment of the present invention includes:

[0095] A data acquisition module is used to acquire the arc image to be identified, and acquire the coordinate data of all preset key points in the arc image to be identified as a preset node data set;

[0096] A graph data construction module is used to obtain graph data corresponding to the arc image to be identified based on a preset node data set, wherein the graph data uses node coordinate data as the nodes of the graph, and the graph data uses the Euclidean distance between the nodes as the edges of the graph;

[0097] The trigger discrimination module is used to detect whether there is a non-zero value on the edge of the graph data. If so, CNN is used to extract the spatial features of the arc image to be identified, and the graph temporal features are obtained based on the graph data.

[0098] The fusion classification module is used to fuse the temporal features and spatial features of the image and then input them into a preset classifier for classification to obtain the recognition result corresponding to the arc light image to be recognized.

[0099] The encoder of the arc detection device based on time-space fusion in the embodiment of the present invention judges the trigger situation, and only calls the CNN neural network to extract the spatial features of the arc image to be identified and extract the graph timing features when the trigger condition is met, which significantly reduces the consumption of computing resources by the empty signal. At the same time, the timing features represented by the graph timing features are fused with the spatial features extracted by CNN, and the graph timing features are used as the time-space coupling enhancement coefficient to enhance the spatial features, thereby improving the accuracy of high-voltage arc detection. In addition, the model is trained through a step-by-step iterative transfer migration strategy to improve the accuracy of high-voltage arc detection.

[0100] In order to solve the above-mentioned technical problems existing in the prior art, an embodiment of the present invention further provides a storage medium on which a computer program is stored, characterized in that when the program is executed by a processor, all steps of the arc detection method based on time-space fusion of the embodiment are implemented.

[0101] The specific steps of the arc detection method based on time-space fusion and the beneficial effects obtained by applying the readable storage medium provided by the embodiment of the present invention are the same as those in the above embodiment and will not be described in detail here.

[0102] A person of ordinary skill in the art can understand that all or part of the steps in the method for implementing the above-mentioned embodiment can be completed by instructing a processor through a program, and the program can be stored in a computer-readable storage medium, and the storage medium is a non-transitory medium, such as a random access memory, a read-only memory, a flash memory, a hard disk, a solid-state hard disk, a magnetic tape, a floppy disk, an optical disc, and any combination thereof. The above-mentioned storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that includes one or more available media. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a tape), an optical medium (for example, a digital video disc (DVD)), or a semiconductor medium (for example, a solid-state disk (SSD)), etc.

[0103] In order to solve the above technical problems existing in the prior art, an embodiment of the present invention further provides a terminal. Figure 8 A schematic diagram of the structure of the terminal according to an embodiment of the present invention is shown, referring to Figure 8 As shown, the terminal of the embodiment of the present invention includes a processor and a memory, and the memory and the processor are communicatively connected; the memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory, so that the terminal executes all steps of the arc detection method based on time-space fusion in the above embodiment.

[0104] The specific steps of the arc detection method based on time-space fusion and the beneficial effects obtained by applying the terminal provided by the embodiment of the present invention are the same as those in the above embodiment and will not be described in detail here.

[0105] It should be noted that the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage. Similarly, the processor may also be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.

[0106] Although the embodiments disclosed in the present invention are as above, the contents described are only embodiments adopted for facilitating the understanding of the present invention and are not intended to limit the present invention. Any technician in the technical field to which the present invention belongs can make any modifications and changes in the form and details of the implementation without departing from the spirit and scope disclosed in the present invention, but the protection scope of the present invention shall still be subject to the scope defined in the attached claims.

Claims

1. An arc detection method based on time-space fusion, comprising: Acquire an arc light image to be identified, and acquire coordinate data of all preset key points in the arc light image to be identified as a preset node data set; Based on the preset node data set, the graph data corresponding to the arc light image to be identified is acquired, wherein the graph data uses the node coordinate data as the nodes of the graph, and the graph data uses the Euclidean distance between the nodes as the edges of the graph; Detect whether there is a non-zero value on the edge in the graph data, and if so, use CNN to extract the spatial features of the arc light image to be identified, and obtain the graph temporal features based on the graph data; After the image temporal features are fused with the spatial features, the features are input into a preset classifier for classification to obtain a recognition result corresponding to the arc light image to be recognized.

2. The arc detection method according to claim 1, characterized in that: Acquiring graph time series features based on the graph data includes: Acquire the image data corresponding to the arc image to be identified as current image data, and acquire the image data corresponding to the arc image at a previous moment of the arc image to be identified as previous moment image data; Subtract the current graph data from the graph data at the previous moment to obtain a preliminary difference matrix; Relu processing is performed on each element in the preliminary difference matrix to obtain a Euclidean distance difference matrix as the image temporal feature corresponding to the arc light image to be identified.

3. The arc detection method according to claim 2, characterized in that: Also includes: Each element in the Euclidean distance difference matrix is ​​subjected to exponential processing to obtain an exponentially processed matrix as a graph time series feature corresponding to the arc light image to be identified; The base of the exponential processing is greater than 1, and the element to be exponentially processed in the Euclidean distance difference matrix is ​​used as the exponent, and the power result of the exponential processing is used as the corresponding new element in the Euclidean distance difference matrix.

4. The arc detection method according to claim 1, characterized in that: The fusing of the graph temporal features with the spatial features comprises: Taking the feature map matrix size of the spatial feature as the target size, reshaping the size of the corresponding Euclidean distance difference matrix to obtain a reshaped Euclidean distance difference matrix; Obtain the Hadamard product of the reshaped Euclidean distance difference matrix and the feature graph matrix of the spatial feature to obtain a fusion result of the graph time series feature and the spatial feature.

5. The arc detection method according to claim 1, characterized in that: The preset classifier includes multiple fully connected layers and a softmax layer connected to the last fully connected layer.

6. The arc detection method according to claim 1, characterized in that: Also includes: The model training method corresponding to the method is training based on the transfer migration strategy, and the transfer migration strategy includes: Use the first voltage level data set to train the model until the model converges and obtain a converged model, and save the corresponding parameters as the first voltage level parameters; Using the first voltage and other parameters as pre-training parameters, fine-tuning the convergence model using the second voltage level data set, and saving the corresponding parameters as the second voltage level parameters; Wherein, the first voltage is smaller than the second voltage.

7. The arc detection method according to claim 6, characterized in that: If the voltage in the recognition result includes multiple voltage levels, the multiple voltage levels are sorted to obtain level sorting data, and the model training method corresponding to the method is to perform level-by-level iterative training based on the transfer migration strategy and the level sorting data to obtain the result corresponding to the highest voltage level data in the level sorting data as the final result; In each iteration, two adjacent data in the graded data are obtained as the first voltage level data set and the second voltage level data set for training based on the transfer migration strategy.

8. An arc detection device based on time-space fusion, characterized in that: include: A data acquisition module, used for acquiring an arc image to be identified, and acquiring coordinate data of all preset key points in the arc image to be identified as a preset node data set; A graph data construction module, used for acquiring graph data corresponding to the arc light image to be identified based on the preset node data set, wherein the graph data uses node coordinate data as graph nodes, and the graph data uses Euclidean distances between nodes as graph edges; A trigger discrimination module is used to detect whether there is a non-zero value in the edge of the graph data. If so, CNN is used to extract the spatial features of the arc light image to be identified, and the graph temporal features are obtained based on the graph data; The fusion classification module is used to fuse the image temporal features with the spatial features and then input them into a preset classifier for classification to obtain the recognition result corresponding to the arc light image to be recognized.

9. A storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the arc light detection method based on time-space fusion described in any one of claims 1 to 7 is implemented.

10. A terminal, characterized in that: It includes a processor and a memory, and the memory is communicatively connected to the processor; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the terminal executes the arc detection method based on time-space fusion as described in any one of claims 1 to 7.

Images