Pressing plate information identification method based on Ai

Abstract

The invention relates to an Ai-based pressing plate information identification method. The method comprises the following steps: acquiring an image; pressing plate image recognition: constructing an image recognition model special for the transformer substation pressing plate by adopting a deep learning algorithm; pressing plate label character recognition: recognizing characters on the signboard by using an optical character recognition technology; calculating the position of the pressing plate: calculating the accurate position of the pressing plate by combining the spatial layout information of the transformer substation based on the image recognition result of the pressing plate; pressing plate label association: associating the identified pressing plate label characters with corresponding pressing plates, and establishing a one-to-one correspondence relationship; identification error elimination: setting a plurality of verification mechanisms, and carrying out verification and error correction on an identification result; and model self-training: the system has self-learning ability, and automatically optimizes and updates the model along with continuous accumulation of new data. According to the invention, the efficiency and accuracy of pressing plate information identification are greatly improved, manual intervention is reduced, and the error rate is reduced.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a pressure plate information recognition method based on AI. Background Art

[0002] In the daily operation and maintenance of substations, accurate identification and management of pressure plates are crucial. At present, the identification of pressure plates mainly relies on manual work, which is inefficient and prone to errors. Manual identification of pressure plate positions and label text is not only time-consuming and labor-intensive, but also difficult to ensure accuracy and real-time performance in complex substation environments.

[0003] The main disadvantages are: slow manual recognition speed, which cannot meet the needs of rapid survey. It is easily affected by human factors, such as fatigue and lack of concentration, which leads to recognition errors. It is difficult to recognize labels with different fonts, sizes and colors, and the manual recognition accuracy is low. Therefore, how to accurately measure the working height of personnel in confined spaces is of great research significance. Summary of the invention

[0004] The purpose of this invention is to provide an efficient and accurate AI system that can automatically identify the location information and related text content of the medium voltage plate of the substation, improve the efficiency of preliminary surveys, and reduce the difficulty of construction. It can also improve the recognition accuracy of texts of different fonts, sizes, and colors, and ensure the reliability of label text recognition.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for identifying platen information based on Ai, the method comprising:

[0007] Image acquisition;

[0008] Platen image recognition: Use deep learning algorithms to build an image recognition model specifically for substation platens. Collect a large number of platen images at different angles and lighting conditions for training, so that the image recognition model can accurately identify the features of various platens and thus determine the position coordinates of the platens.

[0009] Platen label text recognition: Use optical character recognition (OCR) technology to recognize the text on the sign. OCR technology can recognize text information in images. For the platen label on the screen cabinet, the text content is recognized, such as the platen number, function description, etc. The model is trained to adapt to texts of different fonts, sizes and colors to improve recognition accuracy. At the same time, some image preprocessing operations may be performed, such as contrast enhancement and noise removal, to optimize the text recognition effect.

[0010] Platen position calculation: Based on the platen image recognition results and combined with the substation spatial layout information, a specific algorithm is used to calculate the exact position of the platen. For example, the three-dimensional coordinates of the platen are determined by fusing image information from multiple cameras and using triangulation and other methods.

[0011] Pressure plate label association: associate the identified pressure plate label text with the corresponding pressure plate, establish a one-to-one correspondence, and establish a complete information file of the pressure plate, including location coordinates, label text content, and information about the corresponding cabinet. This information can be stored in the database to facilitate subsequent query, management, and analysis, and can be combined with image feature matching and text content matching to ensure accuracy;

[0012] Eliminate recognition errors: Set up a variety of verification mechanisms (for example, check the rationality of location information, verify the accuracy of label text, etc. If abnormal data is found, it can be corrected by re-collecting images or manual intervention) to verify and correct recognition results; for example, by comparing information from multiple data sources, using historical data for trend analysis, etc., possible erroneous recognition results can be eliminated.

[0013] Model self-training: The system has self-learning capabilities and automatically optimizes and updates the model as new data continues to accumulate. You can set regular model update tasks, or make timely model adjustments when erroneous recognition results are found.

[0014] As a further technical solution of the present invention, the method also includes image acquisition: taking pictures of the screen cabinets in the substation based on an image acquisition device (digital camera, mobile phone or professional industrial camera, etc.) to obtain clear and complete screen cabinet images, performing standardized processing and saving them to an untrained data set, and the standardized processing includes grayscale processing.

[0015] As a further technical solution of the present invention, when a deep learning algorithm is used to construct an image recognition model specifically for substation pressure plates, CNN and RNN are fused in parallel to obtain MobileNet-Attention-GRU. MobileNet-Attention is an improved network model based on CNN, which is mainly constructed according to MobileNet and the attention mechanism Attention. MobileNet is a lightweight model in CNN and a lightweight deep convolutional neural network developed for embedded devices. Its core idea is to use lightweight deep separable convolution instead of ordinary convolution. The human brain will pay more attention to matters of interest. The attention mechanism is a method for obtaining features of areas of interest in neural networks, thereby increasing the weights of key features.

[0016] By embedding a lightweight channel attention mechanism (Coordinate Attention) containing position information in the depth-wise separable convolution, more sufficient features of the region of interest can be extracted. First, the input feature is dimensionally increased using a 1×1 convolution, thereby expanding the dimension; secondly, a 3×3 depth convolution is used to extract features; thirdly, a lightweight attention mechanism is used to adjust the feature weights; then, a 1×1 convolution is used to reduce the feature dimension; finally, the input and output features are fused through a residual structure; in the present invention, BN+h-swish is used after the convolution for normalization and to reduce the computational cost caused by deepening the network; for the attention mechanism, it is mainly divided into two parts, coordinate information embedding (Coordinate Information Embedding) and coordinate attention generation (Coordinate Information Embedding). In the Attention Generation part, the coordinate information embedding module uses pooling in two directions for the input sensor features, divides the features into two independent feature maps, and uses X-average pooling to learn the sequence relationship of all sensors in a single time, and uses Y-average pooling to learn the sequence relationship between single sensors in multiple times, thereby achieving effective feature extraction; then, 1×1 convolution is used to adjust the channels. In the coordinate attention generation module, the features extracted in the previous step are fused and split again, and 1×1 convolution is used to adjust different channels. The weight information of two different sequences is directly obtained through the Sigmoid activation function. Finally, the sequence weight information is multiplied by the input feature to obtain the weight-adjusted features.

[0017] As a further technical solution of the present invention, two structures, a "reset gate" and an "update gate", are used inside the GRU. The reset gate determines the historical information that needs to be forgotten, while the update gate determines the information that needs to be continuously transmitted backwards.

[0018] Finally, the model obtained by MobileNet-Attention and GRU model is fused through element summation operation to prevent overfitting. The model is sparsely processed using L1 regularization to obtain the estimated height. The cost function formula of L1 regularization is as follows:

[0019]

[0020] Where w is the parameter of the MobileNet-Attention model and α is the regularization coefficient. L(w, x, y) is the original loss function without regularization. Loss function after L1 regularization.

[0021] Compared with the prior art, the beneficial effects of the present invention are: the present invention provides a platen information recognition method based on Ai, the present invention greatly improves the efficiency and accuracy of platen information recognition, reduces manual intervention, and reduces the error rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flowchart of a pressure plate information recognition method based on Ai.

[0023] Figure 2 This is a structural diagram of the MobileNet-Attention-GRU of an Ai-based pressure plate information recognition method.

[0024] Figure 3 This is a structural diagram of MobileNet-Attention in an Ai-based pressure plate information recognition method.

[0025] Figure 4 This is a structural diagram of GRU in an Ai-based pressure plate information recognition method. DETAILED DESCRIPTION

[0026] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0027] See also Figure 1 , an embodiment of the present invention provides a platen information recognition method based on Ai, the method comprising:

[0028] Image acquisition: Use image acquisition equipment (digital cameras, mobile phones, or professional industrial cameras, etc.) to take photos of the cabinets in the substation to obtain clear and complete images of the cabinets, perform standardization processing, and save them in an untrained data set. Standardization processing includes grayscale processing. In actual use, users use appropriate image acquisition equipment to take photos of the cabinets in the substation. During the acquisition process, attention should be paid to factors such as lighting conditions, shooting angles, and distances to ensure that clear and complete images of the cabinets are obtained. These images will serve as input data for subsequent processing.

[0029] Platen image recognition: Use deep learning algorithms to build an image recognition model specifically for substation platens. Collect a large number of platen images at different angles and lighting conditions for training, so that the image recognition model can accurately identify the features of various platens and thus determine the position coordinates of the platens.

[0030] Platen label text recognition: Use optical character recognition (OCR) technology to recognize the text on the sign. OCR technology can recognize text information in images. For the platen label on the screen cabinet, the text content is recognized, such as the platen number, function description, etc. The model is trained to adapt to texts of different fonts, sizes and colors to improve recognition accuracy. At the same time, some image preprocessing operations may be performed, such as contrast enhancement and noise removal, to optimize the text recognition effect.

[0031] Platen position calculation: Based on the platen image recognition results, combined with the substation spatial layout information and possible other reference information (such as known equipment size, installation standards, etc.), a specific algorithm is used to calculate the exact position of the platen. For example, by fusing image information from multiple cameras, the three-dimensional coordinates of the platen are determined using triangulation and other methods. The calculated platen position information is expressed in the form of precise coordinates, providing an accurate position basis for subsequent operations.

[0032] Pressure plate label association: associate the identified pressure plate label text with the corresponding pressure plate, establish a one-to-one correspondence, and establish a complete information file of the pressure plate, including location coordinates, label text content, and information about the corresponding cabinet. This information can be stored in the database to facilitate subsequent query, management, and analysis, and can be combined with image feature matching and text content matching to ensure accuracy;

[0033] Eliminate recognition errors: Set up a variety of verification mechanisms (for example, check the rationality of location information, verify the accuracy of label text, etc. If abnormal data is found, it can be corrected by re-collecting images or manual intervention) to verify and correct recognition results; for example, by comparing information from multiple data sources, using historical data for trend analysis, etc., possible erroneous recognition results can be eliminated.

[0034] Model self-training: The system has self-learning capabilities and automatically optimizes and updates the model as new data continues to accumulate. You can set regular model update tasks, or make timely model adjustments when erroneous recognition results are found.

[0035] See also Figures 2 to 4 , Figure 2The original data is the processed data X, and X′ is the model data that is sent to the network training after the sample is divided by the time window. In this embodiment, when a deep learning algorithm is used to construct an image recognition model specifically for the substation pressure plate, CNN and RNN are fused in parallel to obtain MobileNet-Attention-GRU. MobileNet-Attention is an improved network model based on CNN, which is mainly constructed according to MobileNet and the attention mechanism Attention. MobileNet is a lightweight model in CNN and a lightweight deep convolutional neural network developed for embedded devices. Its core idea is to use lightweight deep separable convolution instead of ordinary convolution. The human brain will pay more attention to matters of interest. The attention mechanism is a method for obtaining features of regions of interest in neural networks, thereby increasing the weights of key features.

[0036] By embedding a lightweight channel attention mechanism (Coordinate Attention) containing position information in the depth-wise separable convolution, more sufficient features of the region of interest can be extracted. First, the input feature is dimensionally increased using a 1×1 convolution, thereby expanding the dimension; secondly, a 3×3 depth convolution is used to extract features; thirdly, a lightweight attention mechanism is used to adjust the feature weights; then, a 1×1 convolution is used to reduce the feature dimension; finally, the input and output features are fused through a residual structure; in the present invention, BN+h-swish is used after the convolution for normalization and to reduce the computational cost caused by deepening the network; for the attention mechanism, it is mainly divided into two parts, coordinate information embedding (Coordinate Information Embedding) and coordinate attention generation (Coordinate Information Embedding). In the Attention Generation part, the coordinate information embedding module uses pooling in two directions for the input sensor features, divides the features into two independent feature maps, and uses X-average pooling to learn the sequence relationship of all sensors in a single time, and uses Y-average pooling to learn the sequence relationship between single sensors in multiple times, thereby achieving effective feature extraction; then, 1×1 convolution is used to adjust the channels. In the coordinate attention generation module, the features extracted in the previous step are fused and split again, and 1×1 convolution is used to adjust different channels. The weight information of two different sequences is directly obtained through the Sigmoid activation function. Finally, the sequence weight information is multiplied by the input feature to obtain the weight-adjusted features.

[0037] In this embodiment, two structures, "reset gate" and "update gate", are used inside GRU. The reset gate determines the historical information that needs to be forgotten, while the update gate determines the information that needs to be passed on.

[0038] Finally, the model obtained by MobileNet-Attention and GRU model is fused through element summation operation to prevent overfitting. The model is sparsely processed using L1 regularization to obtain the estimated height. The cost function formula of L1 regularization is as follows:

[0039]

[0040] Where w is the parameter of the MobileNet-Attention model, and α is the regularization coefficient, which is used to control the strength of the regularization term. L(w, x, y) is the original loss function without regularization. Loss function after L1 regularization.

[0041] The training phase in the present invention includes:

[0042] (1) Training settings

[0043] The data source is the sensor data collected in the field, with a total of 30 hours of data. After data cleaning and screening, there are 106,540 valid data in total, and 5-fold cross validation is used for training.

[0044] (2) Training environment

[0045] The training environment uses the Pytorch deep learning framework in the Pycharm integrated development environment. The hardware configuration is Xeon 6252N (CPU), Nvidia A5000 24GB (GPU), and 128GB (memory).

[0046] (3) Evaluation indicators

[0047] The reliability of the detection results is determined by setting the confidence threshold. A higher confidence threshold will reduce false positives and the non-maximum suppression threshold (NMS threshold) is used to remove redundant detection boxes.

[0048] (4) Training methods

[0049] Use epochs=1000, batch=1000, lr0=0.01.

[0050] It should be noted that, in this article, the term "comprises" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0051] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A platen information recognition method based on Ai, characterized in that: The method comprises: Image acquisition; Platen image recognition: Use deep learning algorithms to build an image recognition model specifically for substation platens. Collect a large number of platen images at different angles and lighting conditions for training, so that the image recognition model can accurately identify the features of various platens and thus determine the position coordinates of the platens. Platen label text recognition: Use optical character recognition technology to recognize the text on the sign, and improve recognition accuracy by training the model to adapt to texts of different fonts, sizes and colors; Pressure plate position calculation: Based on the results of pressure plate image recognition and combined with the spatial layout information of the substation, the accurate position of the pressure plate is calculated; Platen label association: associate the recognized platen label text with the corresponding platen to establish a one-to-one correspondence, and ensure accuracy by combining image feature matching and text content matching; Identification error elimination: set up a variety of verification mechanisms to verify and correct the recognition results; Model self-training: The system has self-learning capabilities and automatically optimizes and updates the model as new data continues to accumulate.

2. The method for identifying platen information based on Ai according to claim 1, characterized in that: The method also includes image acquisition: taking pictures of the panel cabinets in the substation based on an image acquisition device to obtain clear and complete panel cabinet images, performing standardization processing and saving them in an untrained data set, wherein the standardization processing includes grayscale processing.

3. The method for identifying platen information based on Ai according to claim 1, characterized in that: When using deep learning algorithms to build an image recognition model specifically for substation pressure plates, CNN and RNN are fused in parallel to obtain MobileNet-Attention-GRU. MobileNet-Attention is an improved network model based on CNN, which is mainly constructed based on MobileNet and the attention mechanism Attention. MobileNet is a lightweight model in CNN and a lightweight deep convolutional neural network developed for embedded devices. Its core idea is to use lightweight deep separable convolution instead of ordinary convolution. By embedding a lightweight channel attention mechanism containing position information in the depthwise separable convolution, first, the input feature is dimensionally increased using a 1×1 convolution, thereby expanding the dimension; second, a 3×3 depthwise convolution is used to extract features; and third, a lightweight attention mechanism is used to adjust feature weights; Then, the feature dimension is reduced by 1×1 convolution; finally, the input and output features are fused through the residual structure; after the convolution, BN+h-swish is used for normalization and to reduce the computational cost caused by the deepening of the network; for the attention mechanism, it is mainly divided into two parts, the coordinate information embedding and the coordinate attention generation part. The coordinate information embedding module uses pooling in two directions for the input sensor features, divides the features into two feature maps in independent directions, and uses X average pooling to learn the sequence relationship of all sensors between a single time, and uses Y average pooling to learn the sequence relationship between single sensors at multiple times, so as to achieve effective feature extraction; then, 1×1 convolution is used to adjust the channel, and the coordinate attention generation module fuses the features extracted in the previous step and splits them again, and uses 1×1 convolution to adjust different channels. The weight information of two different sequences is directly obtained through the Sigmoid activation function, and finally the sequence weight information is multiplied by the input feature to obtain the weight-adjusted feature.

4. The method for identifying platen information based on Ai according to claim 3, characterized in that: GRU uses two structures, "reset gate" and "update gate". The reset gate determines the historical information that needs to be forgotten, while the update gate determines the information that needs to be passed on. Finally, the model obtained by MobileNet-Attention and GRU model is fused through element summation operation to prevent overfitting. The model is sparsely processed using L1 regularization to obtain the estimated height. The cost function formula of L1 regularization is as follows: Among them, w is the parameter of the MobileNet-Attention model, α is the regularization coefficient, and L(w, x, y) is the original loss function without regularization. Loss function after L1 regularization.

Images