Intelligent proofreading method and system for power transmission and transformation project drawings

Abstract

The invention relates to a power transmission and transformation project drawing intelligent proofreading method and system, and the method specifically comprises the steps: collecting various historical drawing data of a power transmission and transformation project, and automatically marking the information of electrical elements in a drawing through comparative learning; preprocessing the collected various historical drawing data of the power transmission and transformation project; constructing a composite drawing error judgment model based on deep learning, wherein the composite drawing error judgment model comprises a convolutional neural network and a recurrent neural network; extracting image features of the electrical elements in the target drawing by using a convolutional neural network, extracting text features of the electrical elements in the target drawing by using a recurrent neural network in combination with a self-attention mechanism, and comparing historical data to judge whether errors exist or not; and classifying the errors identified in the target drawing by using a classification algorithm, and performing quantitative evaluation on the severity of the errors identified in the target drawing based on a fuzzy comprehensive evaluation method in combination with a time sequence weight.

Description

Technical Field

[0001] The present invention relates to the technical field of power transmission and transformation engineering, and particularly to an intelligent proofreading method and system for power transmission and transformation engineering drawings. Background Art

[0002] In the field of power transmission and transformation engineering, traditional drawing proofreading relies on manual item-by-item verification, which has problems such as low efficiency, high labor costs, and inevitable human omissions. Design drawings contain a large number of electrical components (such as transformers, circuit breakers, etc.), complex topological connection relationships, and parameter markings. Proofreading a single drawing takes 3 - 5 person-hours, and the cumulative time for large projects can reach hundreds of person-days. Manual proofreading is easily affected by factors such as fatigue and limitations in professional knowledge, with a missed inspection rate as high as 12% - 18%, resulting in a 30% - 50% increase in rework costs. In addition, power transmission and transformation projects need to strictly follow more than a dozen national standards such as GB / T50062 - 2008, covering 236 mandatory verification rules in 7 categories (such as the consistency of neutral point grounding methods and the matching of protection device setting values). However, existing CAD tools only provide format templates and lack the ability to embed intelligent rules, making it difficult to achieve automated compliance checks. In terms of multi-modal data processing, the recognition accuracy of traditional OCR technology for engineering symbols is less than 85%, and it is unable to perform numerical verification by associating with parameter tables; existing AI solutions mostly focus on a single modality (such as pure images or pure text) and cannot achieve multi-dimensional association verification of "drawing - text - rules". With the rapid development of new power systems (such as new energy grid connection and flexible DC transmission), the traditional manual proofreading mode can no longer meet the rapidly evolving engineering requirements, and there is an urgent need to improve the efficiency and accuracy of drawing proofreading through intelligent technologies.

[0003] The Chinese invention patent with the patent publication number CN110334865A discloses a method and system for predicting the failure rate of power equipment based on a convolutional neural network, which includes a training step and a prediction step. Among them, the training step includes: (1) collecting the case PRPS maps of power equipment; (2) preprocessing the collected case PRPS map data; (3) constructing a first convolutional neural network module and training the first convolutional neural network module so that its output is the defect type corresponding to the case PRPS map data; (4) constructing a data set for each defect type based on the defect type; (5) respectively constructing their own fault binary classification sub-modules for each defect type, and each fault binary classification sub-module is constructed based on a second convolutional neural network module; training the second convolutional neural network so that each fault binary classification sub-module outputs a judgment on whether the power equipment has failed based on the probability value of the occurrence of a fault obtained from the case PRPS map data.

[0004] In the prior art, when using the traditional CNN structure for defect classification, only the image features in the drawing data are extracted as the classification criteria, and the text features in the drawing data are not utilized; for fault prediction, it is based on historical data within a fixed time period, and the dynamic changes in the occurrence probabilities of various fault types over a long time axis are not utilized, and the prediction results are not accurate enough. Summary of the Invention

[0005] To solve the problems existing in the above prior art, the present invention proposes an intelligent proofreading method and system for transmission and transformation engineering drawings.

[0006] The technical solution of the present invention is as follows:

[0007] On the one hand, the present invention proposes an intelligent proofreading method for transmission and transformation engineering drawings, and the specific steps include:

[0008] Collect various historical drawing data of transmission and transformation engineering, and use contrastive learning to automatically label the electrical component information in the drawings; and preprocess the collected various historical drawing data of transmission and transformation engineering.

[0009] Construct a composite drawing error judgment model based on deep learning, and the composite drawing error judgment model includes a convolutional neural network and a recurrent neural network; use the preprocessed various historical drawing data of transmission and transformation engineering as input, and the correct drawings of various historical drawing data of transmission and transformation engineering as output, and pre-train the composite drawing error judgment model in combination with the exponentially decaying adaptive learning rate adjustment strategy and Dropout to prevent overfitting technology, and at the same time learn the feature boundaries of different error types to complete error classification.

[0010] Use a convolutional neural network to extract the image features of electrical components in the target drawing, and use a recurrent neural network combined with a self-attention mechanism to extract the text features of electrical components in the target drawing; compare the image features and text features in the target drawing with historical data, and judge whether there are errors according to a preset error threshold.

[0011] Use a classification algorithm to classify the errors identified in the target drawing, and quantitatively evaluate the severity of the errors identified in the target drawing based on the fuzzy comprehensive evaluation method combined with time series weights.

[0012] Finally, output the error type, error location and severity evaluation result of the error of the target drawing as the modification object of the target drawing.

[0013] As a preferred embodiment, the preprocessing of the collected various historical drawing data of transmission and transformation engineering is specifically:

[0014] Use the histogram equalization method to enhance the image contrast, adopt Gaussian filtering to remove noise, convert all drawings to a unified format, use OCR technology to extract the text information in the drawings, and classify the text information.

[0015] As a preferred embodiment, the adaptive learning rate adjustment strategy of exponential decay is specifically as follows:

[0016] Preset the initial learning rate as α0. During the training process, as the number of training rounds increases, the learning rate α k Decays according to the following formula:

[0017] α k = α0 × λ k ;

[0018] In the formula, α0 is the initial learning rate, λ is the decay coefficient, and its value range is between (0, 1); k is the number of training rounds.

[0019] As a preferred embodiment, the Dropout technique for preventing overfitting is specifically as follows:

[0020] During the model training process, for the fully connected layer or other layers that are prone to overfitting, apply the Dropout technique. Preset the input of the input layer as x and the output as y. Use the Dropout operation to randomly set some elements in the input x to 0 with probability p. Preset the Dropout mask as m. The calculation formula for the output y after Dropout is specifically as follows:

[0021]

[0022] In the formula, ⊙ is element-wise multiplication.

[0023] As a preferred embodiment, the specific steps for quantitatively evaluating the severity of the errors identified in the target drawings based on the fuzzy comprehensive evaluation method combined with time series weights are as follows:

[0024] First, determine the evaluation factor set U = {u1, u2,..., u n}, where u n represents different types of errors, and n is the number of error types; then determine the evaluation grade set V = {v1, v2,..., v m}, where v m represents the error severity level, and m is the number of levels for classifying the error severity;

[0025] Based on the analytic hierarchy process, determine the initial weight vector A = (a1, a2,..., a n ), where a n is the weight coefficient;

[0026] Introduce a time decay factor f(t) to dynamically adjust the initial weight vector A of each evaluation factor. The function formula of the time decay factor f(t) is specifically as follows:

[0027] f(t) = e -λ(T-t) ;

[0028] In the formula, λ is the decay rate control factor, λ > 0; t is the occurrence time of the corresponding error type, and T is the current time;

[0029] Combine the time decay factor f(t) with the initial weight vector A to obtain a time-weighted weight vector A time , and the specific formula is as follows:

[0030] A time = (a1·f(t1), a2·f(t2), …, a n ·f(t n ));

[0031] Construct a fuzzy relation matrix R based on historical error statistical data. Each element r ij in the matrix R represents the membership degree of the i-th evaluation factor belonging to the j-th evaluation level;

[0032] Calculate the comprehensive evaluation result through fuzzy composition operation. The specific calculation formula is as follows:

[0033] B = A time ·R.

[0034] On the other hand, the present invention proposes an intelligent proofreading system for transmission and transformation engineering drawings, including:

[0035] A data acquisition and preprocessing module that collects various historical drawing data of transmission and transformation engineering, automatically labels the electrical component information in the drawings using contrastive learning; and preprocesses the collected various historical drawing data of transmission and transformation engineering;

[0036] A drawing error judgment model training module that constructs a composite drawing error judgment model based on deep learning. The composite drawing error judgment model includes a convolutional neural network and a recurrent neural network, and introduces a self-attention mechanism; uses the preprocessed various historical drawing data of transmission and transformation engineering as input and the correct drawings of various historical drawings of transmission and transformation engineering as output, and pre-trains the composite drawing error judgment model by combining an exponentially decaying adaptive learning rate adjustment strategy and Dropout to prevent overfitting technology, and simultaneously learns the feature boundaries of different error types to complete error classification;

[0037] The drawing error recognition module extracts the image features of electrical components in the target drawing using a convolutional neural network, and extracts the text features of electrical components in the target drawing using a recurrent neural network combined with a self-attention mechanism; compares the image features and text features in the target drawing with historical data, and determines whether there are errors according to a preset error threshold;

[0038] The drawing error classification and evaluation module classifies the errors identified in the target drawing using a classification algorithm, and quantitatively evaluates the severity of the errors identified in the target drawing based on the fuzzy comprehensive evaluation method combined with time series weights;

[0039] The error result output module finally outputs the error type, error location, and severity evaluation result of the target drawing as the object to be modified for the target drawing.

[0040] As a preferred implementation manner, the preprocessing of various historical drawing data of the power transmission and transformation project collected specifically includes:

[0041] Using the histogram equalization method to enhance the image contrast, using Gaussian filtering to remove noise, converting all drawings into a unified format, using OCR technology to extract the text information in the drawings, and classifying the text information.

[0042] As a preferred implementation manner, the exponentially decaying adaptive learning rate adjustment strategy specifically includes:

[0043] Preset the initial learning rate as α0. During the training process, as the number of training rounds increases, the learning rate α k Decays according to the following formula:

[0044] α k = α0 × λ k ;

[0045] In the formula, α0 is the initial learning rate, θ is the decay coefficient, and its value range is between (0, 1); k is the number of training rounds.

[0046] As a preferred implementation manner, the Dropout anti-overfitting technique specifically includes:

[0047] During the model training process, for the fully connected layer or other layers that are prone to overfitting, apply the Dropout technique. Preset the input of the input layer as x and the output as y. Use the Dropout operation to randomly set some elements in the input x to 0 with a probability p. Preset the Dropout mask as m. The calculation formula for the output y after Dropout is specifically:

[0048]

[0049] In the formula, ⊙ is element-wise multiplication.

[0050] As a preferred embodiment, the step of quantitatively evaluating the severity of the errors identified in the target drawing based on the fuzzy comprehensive evaluation method combined with the time series weight is specifically as follows:

[0051] First, determine the evaluation factor set U = {u1, u2,..., u n}, where u n represents different types of errors, and n is the number of error types; then determine the evaluation grade set V = {v1, v2,..., v m}, where v m represents the error severity level, and m is the number of levels of the error severity classification;

[0052] Based on the analytic hierarchy process, determine the initial weight vector A = (a1, a2,..., a n ) of each evaluation factor, where a n is the weight coefficient;

[0053] Introduce the time decay factor f(t) to dynamically adjust the initial weight vector A of each evaluation factor. The function formula of the time decay factor f(t) is specifically as follows:

[0054] f(t) = e -λ(T-t) ;

[0055] In the formula, λ is the decay speed control factor, λ > 0; t is the occurrence time of the corresponding error type, and T is the current time;

[0056] Combine the time decay factor f(t) with the initial weight vector A to obtain the time-weighted weight vector A time , and the specific formula is as follows:

[0057] A time = (a1·f(t1), a2·f(t2),..., a n ·f(t n ));

[0058] Based on the historical error statistical data, construct the fuzzy relation matrix R. Each element r ij in the matrix R represents the membership degree of the i-th evaluation factor belonging to the j-th evaluation grade;

[0059] Calculate the comprehensive evaluation result through the fuzzy synthesis operation. The specific calculation formula is as follows:

[0060] B = A time ·R.

[0061] The present invention has the following beneficial effects:

[0062] 1. The present invention realizes intelligent proofreading of power transmission and transformation engineering drawings by means of artificial intelligence technology, greatly improving the proofreading efficiency, shortening the original manual proofreading time that takes several days to several hours, significantly compressing the design cycle, and reducing the labor cost.

[0063] 2. In the present invention, the deep learning model can accurately identify various errors in the drawings, including component parameter errors, line connection errors, etc., by learning a large amount of historical drawing data, effectively reducing the engineering error rate.

[0064] 3. The present invention uses mathematical algorithms to quantitatively evaluate errors, providing users with accurate error information and severity judgments, greatly facilitating users to modify and improve the drawings, and improving the design quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 is a schematic diagram of the step flow of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0066] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a 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 of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0067] It should be understood that the step numbers used in the text are only for convenient description and do not limit the execution order of the steps.

[0068] It should be understood that the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0069] The terms "comprising" and "including" indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their combinations.

[0070] The term "and / or" refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0071] Embodiment 1:

[0072] See Figure 1 , an intelligent proofreading method for power transmission and transformation engineering drawings, and the specific steps include:

[0073] Collect various historical drawing data of power transmission and transformation projects, and automatically label the electrical component information in the drawings using contrastive learning; and preprocess the various historical drawing data of power transmission and transformation projects collected.

[0074] In this embodiment, the various historical drawing data include electrical schematic diagrams, wiring diagrams, and layout diagrams of each voltage level, different substation scales, and different types of transmission lines.

[0075] During actual collection, widely collect at least 1000 pieces of historical drawing data of power transmission and transformation projects, covering drawings of different voltage levels from 110kV to 500kV, different types of substations such as conventional substations and intelligent substations, and different forms of transmission lines such as overhead transmission lines and cable transmission lines.

[0076] Construct a composite drawing error judgment model based on deep learning. The composite drawing error judgment model includes a convolutional neural network and a recurrent neural network; use the preprocessed various historical drawing data of power transmission and transformation projects as input, and the correct drawings of various historical drawings of power transmission and transformation projects as output. Combine the exponentially decaying adaptive learning rate adjustment strategy and Dropout to prevent overfitting technology to pre-train the composite drawing error judgment model, and at the same time learn the feature boundaries of different error types to complete error classification.

[0077] In this embodiment, build a deep learning model with ResNet containing 10 residual blocks as the convolutional neural network part and LSTM as the recurrent neural network part.

[0078] Divide the preprocessed data into a training set, a validation set, and a test set according to 7:2:1. Set the initial learning rate to 0.001, adopt the adaptive learning rate algorithm, and set the number of training rounds to 200 rounds. During the training process, use a Dropout probability of 0.2 to prevent overfitting.

[0079] Use the test set to perform performance testing on the trained model. When the accuracy rate of the model on the test set reaches more than 95% and the recall rate reaches more than 90%, it is determined that the model training is successful.

[0080] Use the convolutional neural network to extract the image features of electrical components in the target drawing, and use the recurrent neural network combined with the self-attention mechanism to extract the text features of electrical components in the target drawing; compare the image features and text features in the target drawing with the historical data, and judge whether there are errors according to the preset error threshold.

[0081] In this embodiment, according to the convolutional neural network algorithm, in each convolutional layer, perform a convolutional operation on the input image through the convolutional kernel. Taking the l-th convolutional layer as an example, the pixel value at the coordinate (i,j) in the output feature map From the formula Calculated. Among them, is the weight of the convolutional kernel at the position (m, n), is the input at the corresponding position in the previous layer, and b l is the bias, and M and N are the sizes of the convolutional kernel. After multiple convolutions, the model can extract key image features such as the shape, size, position, and line direction of electrical components in the drawing.

[0082] The steps to extract the text features of electrical components in the target drawing using the recurrent neural network algorithm are as follows: Taking the time step t as an example, the output of the RNN hidden layer h t follows the formula h t =σ(W xh x t +W hh h t-1 +b h ). Among them, σ is the activation function, W xh is the weight matrix input to the hidden layer, W hh is the weight matrix from the hidden layer to the hidden layer, x t is the input at the current time step, h t-1 is the output of the hidden layer at the previous moment, and b h is the bias. In this way, the recurrent neural network can process the context relationship of the text and extract text features such as component parameters and technical descriptions, such as identifying the logical connection between the component name and the corresponding parameters.

[0083] Make a judgment on the identified errors: Compare the extracted image and text features with the standards and common error patterns of substation engineering drawings learned by the model during training. For image features, if the shape and position of the component deviate from the standard template by more than a set threshold (such as the component position deviation exceeds 1% of the image size), it is determined that there may be an error. For text features, if the parameter value is not within a reasonable range (such as the voltage level parameter does not match the drawing design voltage range), or the logical relationship between parameters is incorrect (such as the resistance value does not match the power calculation result), it is also identified as an error.

[0084] Use a classification algorithm to classify the errors identified in the target drawing, and quantitatively evaluate the severity of the errors identified in the target drawing based on the fuzzy comprehensive evaluation method combined with time series weights;

[0085] In this embodiment, a classification algorithm is used to classify the identified errors. For example, using the support vector machine (SVM) algorithm, for the input error feature vector x, the decision function is used to judge the error type. Among them, α i is the Lagrange multiplier, y i is the sample label, K(x i, x) is the kernel function and b is the bias. The model learns the feature boundaries of different error types during the training phase, enabling accurate classification of newly identified errors, such as component labeling errors, circuit connection errors, parameter errors, etc.

[0086] Finally, the error type, error location, and evaluation result of the severity of the errors in the target drawing are output as the modification objects of the target drawing.

[0087] As a preferred implementation manner of this embodiment, the preprocessing of various historical drawing data of the power transmission and transformation project collected is specifically as follows:

[0088] The histogram equalization method is used to enhance the image contrast, Gaussian filtering is used to remove noise, all drawings are converted into a unified format, OCR technology is used to extract the text information in the drawings, and the text information is classified.

[0089] As a preferred implementation manner of this embodiment, the exponentially decaying adaptive learning rate adjustment strategy is specifically as follows:

[0090] The initial learning rate is preset as α0. During the training process, as the number of training rounds increases, the learning rate α k decays according to the following formula:

[0091] α k = α0 × λ k ;

[0092] In the formula, α0 is the initial learning rate, θ is the decay coefficient, and its value range is between (0, 1); k is the number of training rounds.

[0093] In this embodiment, at the beginning of training, a larger learning rate α0 can quickly update the model parameters and accelerate convergence; as the training progresses, the learning rate gradually decreases, which helps the model to fine-tune near the optimal solution and improve the accuracy.

[0094] As a preferred implementation manner of this embodiment, the Dropout overfitting prevention technique is specifically as follows:

[0095] During the model training process, for the fully connected layer or other layers that are prone to overfitting, the Dropout technique is applied. The input of the input layer is preset as x and the output is y. The Dropout operation is used to randomly set some elements in the input x to 0 with a probability p. The preset Dropout mask is m. The calculation formula for the output y after Dropout is specifically as follows:

[0096]

[0097] In the formula, ⊙ is element-wise multiplication.

[0098] In this embodiment, during training, the Dropout masks m for different batches are different, enabling the model to learn different combinations of features during each training session, thereby preventing overfitting. During testing, the Dropout operation is not performed to ensure the stability of the model output.

[0099] As a preferred implementation manner of this embodiment, the step of quantitatively evaluating the severity of the errors identified in the target drawing based on the fuzzy comprehensive evaluation method combined with time series weights is specifically as follows:

[0100] First, determine the evaluation factor set U = {u1, u2,..., u n}, where u n represents different types of errors, and n is the number of error types; then determine the evaluation grade set V = {v1, v2,..., v m}, where v m represents the error severity level, and m is the number of levels for classifying the error severity.

[0101] In this embodiment, u n represents different types of errors. For example, u1 is an electrical connection error, u2 is a component parameter error, etc.; v m represents the error severity level. For example, v1 is a minor error, v2 is a general error, and v3 is a serious error.

[0102] Based on the analytic hierarchy process, determine the initial weight vector A = (a1, a2,..., a n ) of each evaluation factor, where a n is the weight coefficient;

[0103] In this embodiment, for example, for electrical connection errors and component parameter errors, if it is analyzed that electrical connection errors have a greater impact on the project, perhaps a1 = 0.6 and a2 = 0.4.

[0104] In addition, the weight vector A of each evaluation factor can also be determined by other methods such as expert scoring. Taking the analytic hierarchy process as an example,

[0105] introduce the time decay factor f(t) to dynamically adjust the initial weight vector A of each evaluation factor. The function formula of the time decay factor f(t) is specifically as follows:

[0106] f(t) = e -λ(T-t) ;

[0107] In the formula, λ is the decay rate control factor, λ > 0; t is the occurrence time of the corresponding error type, and T is the current time;

[0108] Combine the time decay factor f(t) with the initial weight vector A to obtain the time-weighted weight vector A time , and the specific formula is as follows:

[0109] A time =(a1·f(t1), a2·f(t2), …, a n ·f(t n ));

[0110] In this embodiment, time series weights are introduced in the fuzzy comprehensive evaluation, and the weights of error types are dynamically adjusted according to the occurrence times and time of historical error types (for example, the weights of recent high-frequency error types are increased).

[0111] Based on the historical error statistical data, a fuzzy relation matrix R is constructed. Each element r ij in the matrix R represents the membership degree of the i-th evaluation factor belonging to the j-th evaluation level;

[0112] In this embodiment, for example, for electrical connection errors, according to experience, it is judged that there is a 70% possibility of belonging to serious errors, 20% belonging to general errors, and 10% belonging to minor errors. Then in the fuzzy relation matrix, r 11 = 0.1, r 12 = 0.2, r 13 = 0.7.

[0113] The comprehensive evaluation result is calculated through fuzzy composition operation. The specific calculation formula is as follows:

[0114] B = A time ·R.

[0115] In this embodiment, B = A time ·R = (b1, b2, …, b m ) where the larger the value of b m , the higher the possibility that the error belongs to the j-th evaluation level, and the severity of the error is judged accordingly. For example, if B = (0.15, 0.25, 0.6) is calculated, then the error is more inclined to the serious error level.

[0116] Embodiment 2:

[0117] An intelligent proofreading system for transmission and transformation engineering drawings, comprising:

[0118] A data acquisition and preprocessing module that acquires various historical drawing data of transmission and transformation projects, automatically labels the electrical component information in the drawings using contrastive learning; and preprocesses the acquired various historical drawing data of transmission and transformation projects;

[0119] Drawing error judgment model training module, constructing a composite drawing error judgment model based on deep learning, the composite drawing error judgment model includes a convolutional neural network and a recurrent neural network, and introduces a self-attention mechanism; using the preprocessed various historical drawing data of the power transmission and transformation project as input, and the correct drawings of various historical drawings of the power transmission and transformation project as output, combining the exponentially decaying adaptive learning rate adjustment strategy and Dropout to prevent overfitting technology to pre-train the composite drawing error judgment model, and at the same time learning the feature boundaries of different error types to complete error classification;

[0120] Drawing error recognition module, using a convolutional neural network to extract the image features of electrical components in the target drawing, and using a recurrent neural network combined with a self-attention mechanism to extract the text features of electrical components in the target drawing; comparing the image features and text features in the target drawing with historical data, and judging whether there is an error according to a preset error threshold;

[0121] Drawing error classification and evaluation module, using a classification algorithm to classify the errors identified in the target drawing, and quantitatively evaluating the severity of the errors identified in the target drawing based on the fuzzy comprehensive evaluation method combined with time series weights;

[0122] Error result output module, finally outputting the error type, error location and severity evaluation result of the target drawing as the modification object of the target drawing.

[0123] As a preferred implementation manner of this embodiment, the preprocessing of the collected various historical drawing data of the power transmission and transformation project is specifically as follows:

[0124] Using the histogram equalization method to enhance the image contrast, using Gaussian filtering to remove noise, converting all drawings into a unified format, using OCR technology to extract the text information in the drawings, and classifying the text information.

[0125] As a preferred implementation manner of this embodiment, the exponentially decaying adaptive learning rate adjustment strategy is specifically as follows:

[0126] Presetting the initial learning rate as α0, during the training process, as the number of training rounds increases, the learning rate α k Decays according to the following formula:

[0127] α k =α0×λ k ;

[0128] In the formula, α0 is the initial learning rate, θ is the decay coefficient, and the value range is between (0,1); k is the number of training rounds.

[0129] As a preferred implementation manner of this embodiment, the Dropout to prevent overfitting technology is specifically as follows:

[0130] During the model training process, for the fully connected layer or other layers that are prone to overfitting, the Dropout technique is applied. Assuming the input of the input layer is x and the output is y, the Dropout operation is used to randomly set some elements in the input x to 0 with probability p. Assuming the Dropout mask is m, the specific calculation formula for the output y after Dropout is as follows:

[0131]

[0132] In the formula, ⊙ represents element-wise multiplication.

[0133] As a preferred implementation manner of this embodiment, the specific steps for quantitatively evaluating the severity of the errors identified in the target drawing based on the fuzzy comprehensive evaluation method combined with the time series weight are as follows:

[0134] First, determine the evaluation factor set U = {u1, u2,..., u n}, where u n represents different types of errors, and n is the number of error types; then determine the evaluation grade set V = {v1, v2,..., v m}, where v m represents the error severity level, and m is the number of levels for classifying the error severity;

[0135] Based on the analytic hierarchy process, determine the initial weight vector A = (a1, a2,..., a n ) of each evaluation factor, where a n is the weight coefficient;

[0136] Introduce the time decay factor f(t) to dynamically adjust the initial weight vector A of each evaluation factor. The function formula of the time decay factor f(t) is specifically as follows:

[0137] f(t) = e -λ(T-t) ;

[0138] In the formula, λ is the decay rate control factor, λ > 0; t is the occurrence time of the corresponding error type, and T is the current time;

[0139] Combine the time decay factor f(t) with the initial weight vector A to obtain the time-weighted weight vector A time , and the specific formula is as follows:

[0140] A time = (a1·f(t1), a2·f(t2),..., a n ·f(t n ));

[0141] Construct a fuzzy relation matrix R based on historical error statistical data. Each element r in matrix R ij represents the membership degree that the i-th evaluation factor belongs to the j-th evaluation level;

[0142] Calculate the comprehensive evaluation result through fuzzy composition operation. The specific calculation formula is as follows:

[0143] B = A time ·R.

[0144] The above are only the embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. An intelligent proofreading method for power transmission and transformation engineering drawings, characterized in that, The specific steps include: Collect various historical drawing data of the power transmission and transformation project, and automatically label the electrical component information in the drawings using contrastive learning; and preprocess the collected various historical drawing data of the power transmission and transformation project; Construct a composite drawing error judgment model based on deep learning. The composite drawing error judgment model includes a convolutional neural network and a recurrent neural network; use the preprocessed various historical drawing data of the power transmission and transformation project as input, and the correct drawings of various historical drawings of the power transmission and transformation project as output, and pre-train the composite drawing error judgment model by combining the exponentially decaying adaptive learning rate adjustment strategy and the Dropout technique to prevent overfitting, and at the same time learn the feature boundaries of different error types to complete error classification; Use a convolutional neural network to extract the image features of electrical components in the target drawing, and use a recurrent neural network combined with a self-attention mechanism to extract the text features of electrical components in the target drawing; compare the image features and text features in the target drawing with historical data, and judge whether there are errors according to the preset error threshold; Use a classification algorithm to classify the errors identified in the target drawing, and quantitatively evaluate the severity of the errors identified in the target drawing based on the fuzzy comprehensive evaluation method combined with time series weights; Finally, output the error type, error location and severity evaluation result of the target drawing as the modification object of the target drawing.

2. The intelligent proofreading method for power transmission and transformation engineering drawings according to claim 1, wherein The preprocessing of the collected various historical drawing data of the power transmission and transformation project is specifically: Use the histogram equalization method to enhance the image contrast, use Gaussian filtering to remove noise, convert all drawings to a unified format, use OCR technology to extract the text information in the drawings, and classify the text information.

3. The intelligent proofreading method for power transmission and transformation engineering drawings according to claim 1, wherein, The exponentially decaying adaptive learning rate adjustment strategy is specifically: The preset initial learning rate is α0. During the training process, as the number of training rounds increases, the learning rate α k decays according to the following formula: α k = α0 × λ k ; In the formula, α0 is the initial learning rate, λ is the decay coefficient, and its value range is between (0,1); k is the number of training rounds.

4. The intelligent proofreading method for transmission and transformation engineering drawings according to claim 1, wherein The Dropout technique to prevent overfitting is specifically: During the model training process, for the fully connected layer or other layers that are prone to overfitting, apply the Dropout technique. Assume that the input of the input layer is x and the output is y. Use the Dropout operation to randomly set some elements in the input x to 0 with probability p. Assume that the Dropout mask is m. The calculation formula for the output y after Dropout is specifically: In the formula, ⊙ is element-wise multiplication.

5. A method for intelligent proofreading of power transmission and transformation engineering drawings according to claim 1, characterized in that, The step of quantitatively evaluating the severity of the errors identified in the target drawing based on the fuzzy comprehensive evaluation method combined with time series weights is specifically: First, determine the evaluation factor set \(U = \{u_1, u_2, \ldots, u_n\}\), where \(u_i\) represents different types of errors and \(n\) is the number of error types; then determine the evaluation grade set \(V = \{v_1, v_2, \ldots, v_m\}\), where \(v_j\) represents the severity level of errors and \(m\) is the number of levels for classifying the severity of errors. n \(n\) n \(i\) m \(m\) m \(j\) Determine the initial weight vector A=(a1,a2,…,a n ) of each evaluation factor based on the analytic hierarchy process, where a n is the weight coefficient; Introduce a time decay factor f(t) to dynamically adjust the initial weight vector A of each evaluation factor. The function formula of the time decay factor f(t) is: f(t) = e -λ(T-t) ; In the formula, λ is the decay rate control factor, λ>0; t is the occurrence time of the corresponding error type, and T is the current time; Combine the time decay factor f(t) with the initial weight vector A to obtain a time-weighted weight vector A time , and the specific formula is as follows: A time =(a1·f(t1), a2·f(t2), …, a n ·f(t n )); Construct a fuzzy relation matrix R based on historical error statistical data. Each element r in the matrix R ij represents the membership degree that the i-th evaluation factor belongs to the j-th evaluation level; Calculate the comprehensive evaluation result through fuzzy composition operation. The specific calculation formula is as follows: B = A time ·R.

6. An intelligent proofreading system for power transmission and transformation engineering drawings, characterized in that, Include: A data collection and preprocessing module that collects various historical drawing data of the power transmission and transformation project and automatically labels the electrical component information in the drawings using contrastive learning; And preprocess the collected various historical drawing data of the power transmission and transformation project; The drawing error judgment model training module constructs a composite drawing error judgment model based on deep learning. The composite drawing error judgment model includes a convolutional neural network and a recurrent neural network, and introduces a self-attention mechanism. Using the preprocessed historical drawing data of various types of power transmission and transformation projects as input and the correct drawings of various historical drawings of power transmission and transformation projects as output, the composite drawing error judgment model is pre-trained by combining an exponentially decaying adaptive learning rate adjustment strategy and Dropout to prevent overfitting technology. At the same time, the feature boundaries of different error types are learned to complete error classification. The drawing error recognition module uses a convolutional neural network to extract the image features of electrical components in the target drawing, and uses a recurrent neural network combined with a self-attention mechanism to extract the text features of electrical components in the target drawing. The image features and text features in the target drawing are compared with historical data, and it is judged whether there is an error according to a preset error threshold. The drawing error classification and evaluation module uses a classification algorithm to classify the errors identified in the target drawing, and quantitatively evaluates the severity of the errors identified in the target drawing based on the fuzzy comprehensive evaluation method combined with time series weights. The error result output module finally outputs the error type, error location and severity evaluation result of the target drawing as the modification object of the target drawing.

7. An intelligent proofreading system for power transmission and transformation engineering drawings according to claim 6, characterized in that The specific preprocessing of the collected historical drawing data of various types of power transmission and transformation projects is as follows: The histogram equalization method is used to enhance the image contrast, Gaussian filtering is used to remove noise, all drawings are converted into a unified format, OCR technology is used to extract the text information in the drawings, and the text information is classified.

8. An intelligent proofreading system for power transmission and transformation engineering drawings according to claim 6, characterized in that The specific exponentially decaying adaptive learning rate adjustment strategy is as follows: The preset initial learning rate is α0. During the training process, as the number of training rounds increases, the learning rate α k decays according to the following formula: α k = α0 × λ k ; In the formula, α0 is the initial learning rate, λ is the decay coefficient, and the value range is between (0,1); k is the number of training rounds.

9. An intelligent proofreading system for transmission and transformation engineering drawings according to claim 6, characterized in that, The specific Dropout technology to prevent overfitting is as follows: During the model training process, for the fully connected layer or other layers prone to overfitting, the Dropout technology is applied. It is assumed that the input of the input layer is x and the output is y. The Dropout operation is used to randomly set some elements in the input x to 0 with a probability p. It is assumed that the Dropout mask is m. The calculation formula of the output y after Dropout is as follows: In the formula, ⊙ is element-wise multiplication.

10. The intelligent proofreading system for power transmission and transformation engineering drawings according to claim 6, characterized in that, The specific steps for quantitatively evaluating the severity of the errors identified in the target drawing based on the fuzzy comprehensive evaluation method combined with time series weights are as follows: First, determine the evaluation factor set U = {u1, u2, …, u n}, where u n represents different types of errors, and n is the number of error types; then determine the evaluation grade set V = {v1, v2, …, v m}, where v m represents the severity level of errors, and m is the number of levels for classifying the severity of errors. Determine the initial weight vector A=(a1,a2,…,a n ) of each evaluation factor based on the analytic hierarchy process, where a n is the weight coefficient; A time decay factor f(t) is introduced to dynamically adjust the initial weight vector A of each evaluation factor. The function formula of the time decay factor f(t) is as follows: f(t) = e -λ(T-t) ; In the formula, λ is the decay rate control factor, λ>0; t is the occurrence time of the corresponding error type, and T is the current time; Combine the time decay factor f(t) with the initial weight vector A to obtain the time-weighted weight vector A time , and the specific formula is as follows: A time =(a1·f(t1), a2·f(t2), …, a n ·f(t n )); Construct a fuzzy relation matrix R based on historical error statistical data. Each element r ij in the matrix R represents the membership degree that the i-th evaluation factor belongs to the j-th evaluation level; The comprehensive evaluation result is calculated through fuzzy composition operation. The specific calculation formula is as follows: B = A time ·R.

Images