Intelligent Recognition Method for Secondary Principle Drawings of Substations Based on Deep Learning

Abstract

The present invention relates to the field of intelligent recognition. More specifically, the present invention relates to an intelligent recognition method for secondary principle drawings of substations based on deep learning. The method includes: obtaining an electrical graphic element recognition network in the first stage obtained in advance; performing one round of training on the electrical graphic element recognition network in the first stage to obtain an output result and a feature image of each convolutional layer; calculating the discrimination difficulty between two electrical graphic elements; obtaining a difference normalization image of each two electrical graphic element regions in the secondary principle drawing image; using the discrimination difficulty as a weight to perform weighted calculation on the difference normalization image to obtain an attention image; using the attention image to extract key information of the secondary principle drawing image; setting the weight of the network parameters in the regularization term according to the correlation between the feature image and the key information; to achieve intelligent recognition. By reasonably setting the regularization weight, while realizing network lightweight, it can also have a small impact on the network recognition accuracy.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recognition. More specifically, the present invention relates to an intelligent recognition method for secondary principle drawings of substations based on deep learning. Background Art

[0002] Secondary principle drawings of substations play an important role in the construction, expansion, technical transformation, operation and maintenance, etc. of substations. However, the management of secondary principle drawings of substations is a time-consuming and laborious manual task. Therefore, how to intelligently manage secondary principle drawings of substations has become a new development requirement. To effectively manage secondary principle drawings of substations, it is necessary to accurately identify the effective information in the secondary principle drawings of substations first.

[0003] Since there are many electrical graphic elements and various types in the secondary principle drawings of substations, a relatively large recognition network is required to accurately identify the electrical graphic elements in the secondary principle drawings of substations. However, the recognition efficiency of a large recognition network is low and cannot guarantee real-time performance. To improve the recognition efficiency of secondary principle drawings of substations, it is necessary to perform lightweight processing on the recognition network. Among them, network regularization can reduce the number of non-zero network parameters by suppressing network parameters. Therefore, network pruning effects can be achieved through network regularization to realize network lightweight. During the process of network regularization, the regularization weight parameter will affect the suppression degree of network parameters. If the regularization weight parameter is set improperly, it is easy to suppress some important network parameters and some unimportant parameters are not suppressed, thus reducing the accuracy of the recognition network. Therefore, how to reasonably set the regularization weight parameter has become the research focus of the present invention.

[0004] The patent application document with the publication number of CN115223189A discloses a method and system for recognizing secondary drawings of substations, a retrieval method and system. The focus of the method in this patent application document is to improve the text recognition accuracy by enhancing the text data in the secondary drawings of substations. The method in this patent application document does not involve the content of network lightweight processing. Therefore, the method in this patent application cannot solve the problems in the present invention. Summary of the Invention

[0005] To solve the problem of how to reasonably set the regularization weight parameter, the present invention proposes an intelligent recognition method for secondary principle drawings of substations based on deep learning. The method includes the following steps:

[0006] Obtain a data set, where the data set is composed of labeled secondary principle drawing images;

[0007] Obtain the pre - obtained electrical graphic element recognition network in the first stage; use the secondary principle drawing image to train the electrical graphic element recognition network in the first stage for one round, obtain the output result and the feature image of each convolutional layer; calculate the discrimination difficulty between two electrical graphic elements , denote the recognition box containing any one of the two electrical graphic elements in the output result as the standby recognition box, respectively represent the probabilities that the area within the k - th standby recognition box in the s - th secondary principle drawing image is the j - th type of electrical graphic element and the i - th type of electrical graphic element, denote the category recognition error of the k - th standby recognition box in the s - th secondary principle drawing image, respectively represent the label values of the j - th category and the i - th category corresponding to the area of the k - th standby recognition box in the s - th secondary principle drawing image, denote the number of standby recognition boxes in the s - th secondary principle drawing image, denote the number of secondary principle drawing images, denote the preset anti - zero coefficient, denote the linear normalization function;

[0008] Obtain the difference normalization image of the areas of every two types of electrical graphic elements in the secondary principle drawing image; use the discrimination difficulty as the weight to perform weighted calculation on the difference normalization image to obtain the attention image;

[0009] Use the attention image to extract the key information of the secondary principle drawing image; set the weight of the network parameters in the regularization term according to the correlation between the feature image and the key information; to achieve intelligent recognition.

[0010] The present invention accurately evaluates the necessity of each network parameter based on the contribution of the information extracted from each network parameter to the primitive recognition, and then sets appropriate regularization weights for each network parameter according to the evaluation results. Further, considering that the correlation between the information extracted from each network parameter and the important information required for primitive recognition can reflect the contribution of the information extracted from each network parameter, the key information index of the secondary principle drawing image is introduced to accurately reflect the important information required for primitive recognition. Further, when considering the key information of the secondary principle drawing image, considering that the discrimination information of the primitives with greater recognition difficulty should be focused on, the discrimination difficulty and other indexes are introduced to accurately reflect the primitive recognition difficulty. Further, when obtaining the key information of the secondary principle drawing image, considering that the difference information of the primitives can better distinguish different types of primitives, the difference analysis of different types of electrical primitives is introduced to accurately screen out the key information for primitive recognition in the secondary principle drawing image. Further, when analyzing the discrimination difficulty of different types of primitives, considering the characteristics that greater discrimination difficulty leads to lower recognition accuracy and the contribution of the recognition error of the primitive with greater discrimination difficulty is larger, the recognition error of the primitive and the error difference analysis are introduced to accurately analyze the discrimination difficulty of the primitive.

[0011] Preferably, the obtaining of the electrical primitive recognition network in the first stage obtained in advance includes:

[0012] Obtain the pre-constructed electrical primitive recognition network; use the data set to train the electrical primitive recognition network for one round to obtain the electrical primitive recognition network in the first stage.

[0013] The present invention enables the electrical primitive recognition network to learn part of the information through one round of training of the electrical primitive recognition network, providing a basis for subsequent analysis of the information extraction ability of each network parameter.

[0014] Preferably, the obtaining of the difference normalization image of each two electrical primitive regions in the secondary principle drawing image includes:

[0015] Extract each electrical primitive region in the secondary principle drawing image according to the label;

[0016] Adjust all types of electrical primitive regions to the same size; subtract each two adjusted-size electrical primitive regions to obtain a difference image;

[0017] Perform normalization processing on the pixels in the difference image to obtain the difference normalization image of each two electrical primitive regions.

[0018] The present invention accurately extracts the difference information of different types of primitives through difference analysis, providing a basis for subsequent extraction of the key information for primitive recognition in the secondary principle drawing image.

[0019] Preferably, taking the discrimination difficulty as a weight, a weighted calculation is performed on the difference-normalized image to obtain an attention image, including:

[0020] Taking the discrimination difficulty as a weight, a weighted sum is performed on the difference-normalized images of any one electrical graphic element region and all other electrical graphic element regions to obtain an initial attention image;

[0021] Obtain the outer envelope pixels of the electrical graphic elements in the electrical graphic element region;

[0022] Adjust the value of the outer envelope pixels in the initial attention image to 1 to obtain the attention image.

[0023] When the present invention obtains the attention image, the discrimination difficulty and difference information are introduced, so that the information for distinguishing different types of graphic elements can be extracted from the attention image, and at the same time, the information with great discrimination difficulty can be extracted emphatically, providing a basis for extracting key information in the subsequent stage; further, considering that the outer envelope pixels will affect the positioning of the graphic element region, the value of the outer envelope pixels is adjusted, so that the information of the outer envelope pixels can be extracted when extracting key information.

[0024] Preferably, the extracting of the key information of the secondary principle drawing image by using the attention image includes:

[0025] Multiply each electrical graphic element region in the secondary principle drawing image by the attention image to obtain the key information of the secondary principle drawing image.

[0026] Preferably, the setting of the weights of the network parameters in the regularization term according to the correlation between the feature image and the key information includes:

[0027] Obtain all the feature images corresponding to each network parameter, and calculate the correlation between the feature image and the key information; take the reciprocal of the mean value of the correlations obtained from all the feature images of each network parameter to obtain the weights of each network parameter, denote the electrical graphic element recognition network in the first stage after one round of training as the electrical graphic element recognition network in the second stage, and construct the regularization term of the loss function of the electrical graphic element recognition network in the second stage based on the weights.

[0028] The present invention reduces the number of non-zero network parameters and improves the network recognition efficiency by adding a regularization term; further, by setting the weights of each network parameter in the regularization term, when the network is lightweight processed, the network recognition accuracy can be affected less.

[0029] Preferably, the regularization term adopts L2 regularization.

[0030] Preferably, the achieving of intelligent recognition includes:

[0031] Based on the loss function after adding the regularization term, use the data set to continue training the electrical graphic element recognition network in the second stage until the preset training completion cut-off condition is met, and obtain the finally trained electrical graphic element recognition network;

[0032] Input the newly collected secondary principle drawing image into the finally trained electrical graphic element recognition network to obtain the recognition result.

[0033] The present invention has the following beneficial effects:

[0034] The present invention accurately evaluates the necessity of each network parameter according to the contribution of the information extracted by each network parameter to the graphic element recognition, and then sets appropriate regularization weights for each network parameter according to the evaluation result;

[0035] Furthermore, considering that the correlation between the information extracted by each network parameter and the important information required for graphic element recognition can reflect the contribution of the information extracted by each network parameter, the key information index of the secondary principle drawing image is introduced to accurately reflect the important information required for graphic element recognition;

[0036] Furthermore, when considering the key information of the secondary principle drawing image, considering that the discrimination information of the graphic elements with greater recognition difficulty should be focused on, the discrimination difficulty and other indexes are introduced to accurately reflect the graphic element recognition difficulty;

[0037] Furthermore, when obtaining the key information of the secondary principle drawing image, considering that the difference information of the graphic elements can better distinguish different types of graphic elements, the difference analysis of different types of electrical graphic elements is introduced to accurately screen out the key information for graphic element recognition in the secondary principle drawing image;

[0038] Furthermore, when analyzing the discrimination difficulty of different types of graphic elements, considering the characteristics that greater discrimination difficulty leads to lower recognition accuracy and the contribution of the recognition error of the graphic element with greater discrimination difficulty is larger, the recognition error of the graphic element and the error difference analysis are introduced to accurately analyze the discrimination difficulty of the graphic element. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] By referring to the drawings and reading the following detailed description, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become easy to understand. In the drawings, several embodiments of the present invention are shown in an exemplary rather than restrictive manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0040] Figure 1 is the step flow chart of the intelligent recognition method for substation secondary principle drawings based on deep learning in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts belong to the protection scope of the present invention.

[0042] The following will describe in detail the specific implementation manners of the present invention in conjunction with the accompanying drawings.

[0043] Please refer to Figure 1 , which shows a flowchart of the steps of an intelligent recognition method for secondary principle drawings of a substation based on deep learning provided by an embodiment of the present invention. The method includes the following steps:

[0044] S1: Obtain a data set.

[0045] Specifically, obtain the secondary principle drawing images of several substations, and manually label and process the secondary principle drawing images. The labels of the secondary principle drawing images are the width and height of each electrical graphic element in the secondary principle drawing image, the central position of the electrical graphic element, and the one-hot encoding of the classification of the electrical graphic element. All the labeled secondary principle drawing images form a data set.

[0046] S2: Obtain the pre-obtained electrical graphic element recognition network in the first stage; use the secondary principle drawing images to train the electrical graphic element recognition network in the first stage, obtain the output result and the feature images of each convolutional layer; calculate the discrimination difficulty between two electrical graphic elements.

[0047] S20: Obtain the pre-obtained electrical graphic element recognition network in the first stage.

[0048] It should be noted that in order to implement the recognition of secondary principle drawings, it is necessary to first construct a recognition network.

[0049] Preferably, as an example, obtaining the pre-obtained electrical graphic element recognition network in the first stage includes:

[0050] Obtain the pre-constructed electrical graphic element recognition network; use the data set to train the electrical graphic element recognition network for one round to obtain the electrical graphic element recognition network in the first stage.

[0051] To construct the electrical graphic element recognition network, in this embodiment, the YoloV2 network is used as the electrical graphic element recognition network. Other embodiments can use other networks, and this embodiment does not make specific limitations.

[0052] The labeled secondary principle drawing images in the dataset are sequentially input into the electrical graphic element recognition network to train the electrical graphic element recognition network. The electrical graphic element recognition network after one round of training with all the labeled secondary principle drawing images is denoted as the electrical graphic element recognition network in the first stage.

[0053] It should be noted that training the electrical graphic element recognition network by sequentially inputting the labeled secondary principle drawing images in the dataset into the electrical graphic element recognition network is a prior art and will not be elaborated here.

[0054] It should be further noted that in order to set the regularization weights of the network parameters in the recognition network, it is necessary to first judge the importance of each network parameter. Since the network parameters in the initial recognition network are randomly initialized, it is impossible to judge the importance of the network parameters through the initial recognition network. Therefore, it is necessary to train the recognition network for a period of time to make it have a certain recognition ability, providing a basis for subsequent determination of the importance of the network parameters.

[0055] S21: Obtain the electrical graphic element recognition network in the first stage obtained in advance; use the secondary principle drawing image to perform one round of training on the electrical graphic element recognition network in the first stage, obtain the output result and the feature image of each convolutional layer; calculate the discrimination difficulty between two electrical graphic elements.

[0056] It should be noted that in order to set appropriate regularization weights for the network parameters, it is necessary to judge the contribution of each network parameter to the drawing recognition. If the information extracted by the network parameter contributes more to the recognition, the contribution of this network parameter to the drawing recognition is more. Therefore, this network parameter should not be suppressed, and the regularization weight of this network parameter should be set smaller. If the information extracted by the network parameter contributes less to the recognition, or even interferes with the recognition, the contribution of this network parameter to the drawing recognition is less. Therefore, this network parameter should be suppressed, and the regularization weight of this network parameter should be set larger.

[0057] It should be further noted that the differences between different types of electrical graphic elements are different. Some types of electrical graphic elements have larger differences, and some electrical graphic elements have smaller differences. Among them, the classification difficulty of electrical graphic elements with larger differences is greater, and the classification difficulty of electrical graphic elements with smaller differences is smaller. If the network parameter has a stronger ability to extract the discrimination information of electrical graphic elements with greater classification difficulty, this network parameter is more important. Therefore, it is necessary to judge the extraction of the discrimination information of different network parameters for electrical graphic elements with different classification difficulties. First, obtain the discrimination difficulty of different types of electrical graphic elements.

[0058] Preferably, as an example, input the secondary principle drawing image into the electrical graphic element recognition network in the first stage for one round of training, and obtain the output result and the feature image of each convolutional layer, including:

[0059] The tagged secondary principle drawing images are sequentially input into the electrical graphic element recognition network in the first stage, and the electrical graphic element recognition network in the first stage is trained to obtain the output results obtained when each secondary principle drawing image is input and the feature images of each convolutional layer.

[0060] It can be understood that the output results can reflect the recognition situation of each component graphic element in the secondary principle drawing image by the electrical graphic element recognition network in the first stage, and the feature images of each convolutional layer reflect the extraction situation of each piece of information in the secondary principle drawing image by the electrical graphic element recognition network in the first stage. These data are the data used to analyze the extraction situation of each network parameter for the discrimination information of electrical graphic elements with different classification difficulties, so they should be extracted first.

[0061] Preferably, as an example, calculate the discrimination difficulty between every two electrical graphic elements, including:

[0062]

[0063] Among them, the recognition frame containing any one of the two electrical graphic elements in the output result is denoted as the spare recognition frame. represents the probability that the area within the k-th spare recognition frame in the s-th secondary principle drawing image is the j-th type of electrical graphic element. represents the probability that the area within the k-th spare recognition frame in the s-th secondary principle drawing image is the i-th type of electrical graphic element. represents the label value of the j-th category corresponding to the area of the k-th spare recognition frame in the s-th secondary principle drawing image. respectively represent the label values of the i-th category corresponding to the area of the k-th spare recognition frame in the s-th secondary principle drawing image. represents the category recognition error of the k-th spare recognition frame in the s-th secondary principle drawing image. represents the number of spare recognition frames in the s-th secondary principle drawing image. represents the number of secondary principle drawing images. represents the linear normalization function. represents a preset anti-zero coefficient, which is used to prevent the denominator from being 0. In this embodiment, the preset anti-zero is taken as 0.001 as an example for description, and other values can be taken in other embodiments, and this embodiment does not make specific limitations. represents the discrimination difficulty between the i-th type of electrical graphic element and the j-th type of electrical graphic element.

[0064] It should be added that, for example, the process of obtaining the label value of each category in the area corresponding to the spare recognition frame is as follows: Suppose there are a total of three categories of electrical graphic elements, and the one-hot encodings of these three categories of electrical graphic elements are [1, 0, 0], [0, 1, 0], and [0, 0, 1] respectively. According to the label, obtain the area where each electrical graphic element is located in the secondary principle drawing. If there is an area where the second category of electrical graphic elements is located in the spare recognition frame, then the label value of the first category in the area corresponding to the spare recognition frame is 0, the label value of the second category in the area corresponding to the spare recognition frame is 1, and the label value of the third category in the area corresponding to the spare recognition frame is 0.

[0065] It should be further added that the process of obtaining the category recognition error of the spare recognition frame is as follows: Obtain the graphic element category in the area within the spare recognition frame according to the label, which is denoted as the reference graphic element category; obtain the probability that the area within the spare recognition frame is the reference graphic element category, and subtract the label value from the probability to obtain the category recognition error of the spare recognition frame. The method of obtaining the probability that the area within the spare recognition frame is various electrical graphic elements is prior art and will not be elaborated here.

[0066] It can be understood that reflects the category recognition error situation of the k-th spare recognition frame in the s-th secondary principle drawing image. The larger this value is, the greater the difficulty for the recognition network to recognize the type of graphic elements within the spare recognition frame; reflects the recognition error of the recognition network for the j-th type of electrical graphic element, reflects the recognition error of the recognition network for the i-th type of electrical graphic element, The closer it is to 1, the more similar the error rates of the two types of electrical graphic elements are, which further indicates that the misrecognition of the i-th or j-th type of electrical graphic element is caused by mutual misrecognition. For example, in the area within the spare recognition frame, the probability that the area is recognized as the i-th type of electrical graphic element is 0.8, and the probability of being recognized as the j-th type of electrical graphic element is 0.2. The recognition errors of these two types of electrical graphic elements are both 0.2, which indicates that the j-th type of electrical graphic element interferes with the recognition result of the i-th type of electrical graphic element, resulting in a large error. It is reflected by the output result corresponding to the k-th spare recognition frame in the s-th secondary principle drawing image, where the discrimination difficulty between the j-th type of electrical graphic element and the i-th type of electrical graphic element. The larger this value is, the greater the difficulty for the recognition network to recognize the type of graphic elements in the k-th spare recognition frame in the s-th secondary principle drawing image, and the large recognition difficulty is caused by the difficulty in distinguishing between the j-th type of electrical graphic element and the i-th type of electrical graphic element, thus further proving the discrimination difficulty between the j-th type of electrical graphic element and the i-th type of electrical graphic element.

[0067] S3: Obtain the difference normalized images of every two electrical graphic element regions in the secondary principle drawing image; calculate the weighted sum of the difference normalized images with the differentiation difficulty as the weight to obtain the attention image.

[0068] It should be noted that, in order to judge the extraction of the electrical graphic element differentiation information by each network parameter, it is necessary to first analyze which information should be focused on for differentiating different types of electrical graphic elements, and then judge the accuracy of the extraction of the electrical graphic element differentiation information by each network parameter according to the relevance between the information extracted by the network parameter and the information that should be focused on for electrical graphic element differentiation.

[0069] S30: Obtain the difference normalized images of every two electrical graphic element regions in the secondary principle drawing image.

[0070] Preferably, as an example, obtaining the difference normalized images of every two electrical graphic element regions in the secondary principle drawing image includes:

[0071] Extract each electrical graphic element region in the secondary principle drawing image according to the label;

[0072] Adjust all electrical graphic element regions to the same size; subtract every two adjusted-size electrical graphic element regions to obtain the difference image;

[0073] Normalize the pixels in the difference image to obtain the difference normalized images of every two electrical graphic element regions.

[0074] It can be understood that the difference normalized image can reflect the differentiation information of two electrical graphic element regions.

[0075] S31: Calculate the weighted sum of the difference normalized images with the differentiation difficulty as the weight to obtain the attention image.

[0076] Preferably, as an example, calculating the weighted sum of the difference normalized images with the differentiation difficulty as the weight to obtain the attention image includes:

[0077] Take the differentiation difficulty as the weight, and calculate the weighted sum of the difference normalized images of any one electrical graphic element region and all other electrical graphic element regions to obtain the initial attention image;

[0078] Obtain the outer envelope pixels of the electrical graphic elements in the electrical graphic element region;

[0079] Adjust the values of the outer envelope pixels in the initial attention image to 1 to obtain the attention image.

[0080] It should be noted that the outer envelope pixels of the electrical graphic element are the outermost circle of pixels of the electrical graphic element.

[0081] It can be understood that the greater the difficulty of differentiation, the greater the recognition difficulty of the recognition network. The greater the recognition difficulty, the more information in this aspect the recognition network should absorb to accurately recognize. Therefore, the weight of the information with a large differentiation difficulty is increased. By weighting the differentiation information with weights, the attention information for differentiating electrical graphic elements can be obtained. In addition, for graphic element recognition, not only the type needs to be recognized, but also the graphic element area needs to be accurately located. The outer contour of the graphic element is the main information basis for graphic element area location. Therefore, the outer contour of the graphic element also needs to be focused on, and thus the attention of the pixels of the outer contour of the graphic element should be increased.

[0082] S4: Extract the key information of the secondary principle drawing image using the attention image; set the weights of the network parameters in the regularization term according to the correlation between the feature image and the key information; to achieve intelligent recognition of the secondary principle drawings of the substation.

[0083] S40: Extract the key information of the secondary principle drawing image using the attention image.

[0084] Preferably, as an example, extracting the key information of the secondary principle drawing image using the attention image includes:

[0085] Performing a multiplication operation on the corresponding pixels of each electrical graphic element area in the secondary principle drawing image and the attention image to obtain the key information of the secondary principle drawing image.

[0086] It can be understood that the key information of the secondary principle drawing image is important information for differentiating the types of electrical graphic elements.

[0087] S41: Set the weights of the network parameters in the regularization term according to the correlation between the feature image and the key information.

[0088] It should be noted that in order to determine the accuracy of information extraction of each network parameter, it is necessary to judge the consistency between the information extracted by it and the information that should be focused on, and then judge the importance of each network parameter according to the consistency, and then set the regularization weight for the network parameter according to the importance.

[0089] Preferably, as an example, setting the weights of the network parameters in the regularization term according to the correlation between the feature image and the key information includes:

[0090] Obtain all the feature images corresponding to each network parameter, calculate the correlation between the feature image and the key information; take the reciprocal of the mean value of the correlations obtained from all the feature images of each network parameter to obtain the weight of each network parameter;

[0091] Record the electrical graphic element recognition network in the first stage after the completion of the new round of training as the electrical graphic element recognition network in the second stage;

[0092] Take the weight as the weight of the corresponding network parameter within the regularization term to construct the regularization term of the loss function for the electrical graphic element recognition network in the second stage.

[0093] It should be added that the calculation method for the correlation between the feature image and the key information includes: through upsampling and downsampling processing, adjust the feature image to the same size as the key information, splice the rows of the resized feature image together to obtain a feature sequence, splice the rows of the key information together to obtain a key information sequence, and take the absolute value of the Pearson correlation between the feature sequence and the key information sequence as the correlation between the feature image and the key information.

[0094] It should be further added that in this embodiment, the regularization term of the loss function is constructed based on L2 regularization. Other embodiments can construct the loss function based on other regularization methods, and this embodiment does not make specific limitations.

[0095] It should be noted that when the weight is known, the construction method of the regularization term is prior art and will not be elaborated here.

[0096] S42: To achieve intelligent recognition of the secondary principle drawings of the substation.

[0097] Preferably, as an example, to achieve intelligent recognition of the secondary principle drawings of the substation, it includes:

[0098] Based on the loss function with the added regularization term, use the data set to continue training the electrical graphic element recognition network in the second stage until the preset training completion cut-off condition is met, and obtain the finally trained electrical graphic element recognition network; in this embodiment, the convergence of the loss value is used as the preset training completion cut-off condition. Other embodiments can adopt other conditions, and this embodiment does not make specific limitations.

[0099] Input the newly collected secondary principle drawing image into the finally trained electrical graphic element recognition network to obtain the recognition result.

[0100] So far, this embodiment is completed.

[0101] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A deep learning-based intelligent recognition method for substation secondary schematic drawings, characterized in that: include: Obtain a data set, the data set consisting of labeled secondary schematic drawing images; Acquire a first-stage electrical element recognition network obtained in advance; Use the secondary schematic image to train the electrical element recognition network in the first stage, obtain the output results and the feature images of each convolution layer; calculate the difficulty of distinguishing the two electrical elements , the identification frame containing any one of the two electrical diagram elements in the output result is recorded as a spare identification frame, They represent the probability that the area within the kth spare identification box in the sth secondary schematic drawing image is the jth electrical element and the ith electrical element, respectively. represents the category recognition error of the kth alternative recognition box in the sth secondary schematic drawing image, They represent the label values ​​of the j-th category and the i-th category of the area corresponding to the k-th backup recognition box in the s-th secondary schematic drawing image, respectively. represents the number of spare recognition boxes in the s-th secondary schematic drawing image, Indicates the number of secondary schematic drawings images, Indicates the preset anti-zero coefficient. represents the linear normalization function; Obtain a difference normalized image of each two electrical element regions in the secondary schematic drawing image; perform weighted calculation on the difference normalized image using the difficulty of distinction as a weight to obtain an attention image; The focus image is used to extract the key information of the secondary schematic drawing image; The weights of network parameters in the regularization term are set according to the correlation between the feature image and the key information to achieve intelligent recognition.

2. The method for intelligent identification of secondary schematic drawings of substations based on deep learning according to claim 1 is characterized in that: The step of obtaining the first-stage electrical graphic element recognition network obtained in advance includes: A pre-built electrical graphic element recognition network is obtained; and a round of training is performed on the electrical graphic element recognition network using the data set to obtain a first-stage electrical graphic element recognition network.

3. The method for intelligent identification of secondary schematic drawings of substations based on deep learning according to claim 1 is characterized in that: The step of obtaining a difference normalized image of each two electrical graphic element regions in the secondary schematic drawing image includes: Extract each electrical element area in the secondary schematic drawing image according to the label; Adjust all electrical graphic element areas to the same size; and perform subtraction of every two electrical graphic element areas after the size adjustment to obtain a difference image; The pixels in the difference image are normalized to obtain a difference normalized image of each two electrical graphic element regions.

4. The method for intelligent identification of secondary schematic drawings of substations based on deep learning according to claim 1 is characterized in that: The method of performing weighted calculation on the difference normalized image to obtain the attention image by taking the difficulty of distinction as the weight includes: Taking the difficulty of distinction as a weight, the weighted sum of the difference normalized images between any electrical graphic element area and all other electrical graphic element areas is performed to obtain an initial attention image; Obtain outer envelope pixels of an electrical primitive in an electrical primitive region; The values ​​of the outer envelope pixels in the initial attention image are adjusted to 1 to obtain the attention image.

5. The method for intelligent identification of secondary schematic drawings of substations based on deep learning according to claim 1 is characterized in that: The method of extracting key information of the secondary schematic drawing image by using the attention image includes: The key information of the secondary schematic drawing image is obtained by multiplying each electrical element area in the secondary schematic drawing image by the attention image.

6. The method for intelligent identification of secondary schematic drawings of substations based on deep learning according to claim 1 is characterized in that: The step of setting the weight of the network parameter in the regularization term according to the correlation between the feature image and the key information includes: All feature images corresponding to each network parameter are obtained, and the correlation between the feature image and the key information is calculated; the weight of each network parameter is obtained by taking the inverse of the mean of the correlation obtained from all feature images of each network parameter, and the first-stage electrical element recognition network after one round of training is recorded as the second-stage electrical element recognition network, and the regularization term of the loss function of the second-stage electrical element recognition network is constructed based on the weight.

7. The method for intelligent identification of secondary schematic drawings of substations based on deep learning according to claim 6 is characterized in that: The regularization term adopts L2 regularization.

8. The method for intelligently identifying secondary schematic drawings of substations based on deep learning according to claim 6 is characterized in that: The intelligent identification is realized by: Based on the loss function after adding the regularization term, the electrical graph element recognition network of the second stage is continuously trained using the data set until the preset training completion cutoff condition is met, thereby obtaining the electrical graph element recognition network that has been finally trained. The newly acquired secondary schematic drawing image is input into the finally trained electrical graphic element recognition network to obtain the recognition result.

Images