PDF format value single automatic recognition and comparison method based on multi-scale feature extraction and parameter standardization

Abstract

The application discloses a PDF format fixed value single automatic recognition and comparison method based on multi-scale feature extraction and parameter standardization, multi-level visual features of a PDF fixed value single image are extracted by using a multi-scale convolutional neural network, a text and a table area are accurately positioned by using semantic segmentation, and character recognition and structured extraction are realized by using a recurrent network with an attention mechanism; unit uniform mapping and numerical standardization processing are performed on the recognized fixed value parameters, a standardized parameter vector is constructed, the standardized parameter vector and an operation fixed value vector collected by equipment are mapped to a same feature space, a difference degree is calculated by using Euclidean distance, and automatic comparison is realized by combining a dynamic tolerance threshold, and a visual comparison report is output. The application realizes high-robustness recognition and high-precision automatic proofreading of PDF fixed value singles in multiple formats and multiple versions, effectively solves the problems of low efficiency and easy errors in traditional manual checking, and significantly improves the automation level and reliability of protection fixed value management.

Description

Technical Field

[0001] This invention belongs to the field of power system relay protection technology, specifically involving an automatic identification and comparison method for PDF format setting sheets based on multi-scale feature extraction and parameter standardization. Background Technology

[0002] In the safe and stable operation of power systems, the correct operation of relay protection devices is crucial, and their behavior directly depends on the pre-set protection settings. To ensure that the protection settings operating in the field equipment are completely consistent with the setting sheets issued by the dispatching department after rigorous calculation and review, setting verification has become a fundamental, periodic, and critically important core task in the relay protection profession.

[0003] Currently, official value setting sheets issued by the dispatch center generally use tamper-proof PDF format files. However, this format is not designed for direct machine reading, which presents a primary obstacle to automated comparison. The traditional value setting verification process relies heavily on manual operation: maintenance personnel must first print out the current operating values ​​of the equipment into a paper list using a printer connected to the protection device, and then compare the paper list item by item with the PDF value setting sheet issued by the dispatch center, reading it aloud. This process is not only time-consuming and labor-intensive, but also prone to errors and omissions due to visual fatigue or human negligence when faced with a large number of value setting items, posing certain security risks.

[0004] While existing research has attempted to automate the processing of PDF-format value sheets using image recognition methods, significant limitations remain. General optical character recognition (OCR) technology is ill-suited to complex layouts, dense tables, and specialized electrical symbols. Recognition results are easily affected by inconsistent formats and varying fonts, leading to fragmented text extraction and a high error rate. Furthermore, current methods often separate the recognition and comparison stages, failing to construct a complete analytical link from image to structured semantics. The comparison process relies heavily on simple string matching, lacking a deep understanding of the electrical meaning of the value parameters, such as unit conversions and numerical tolerances. Consequently, extensive manual review is often required in practical applications, and the level of automation and reliability still needs improvement.

[0005] In summary, there is currently a lack of an integrated automated solution that can robustly identify PDF value sheets of various formats, accurately parse the structured value parameters within them, and perform intelligent tolerance comparison based on electrical semantics. Therefore, it is necessary to design an automatic identification and comparison method for PDF value sheets based on multi-scale feature extraction and parameter standardization to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization. This method aims to overcome the problems of poor robustness in identifying PDF format protected value sheets, unstructured parameter extraction, rigid comparison methods, and insufficient automation in the prior art. It realizes an end-to-end automated process from tamper-proof PDF files to intelligent comparison, significantly improving the accuracy and efficiency of verification.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: An automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization includes the following steps: S1. Obtain the PDF format protection setting file issued by the dispatch side and convert it into a high-resolution image sequence. At the same time, collect the original setting data of the current operation of the relay protection device from the device communication interface or virtual printing driver to form two types of input sources to be compared. S2, a multi-scale feature extraction model is constructed based on a parallel multi-branch convolutional neural network. Image features are extracted in parallel through convolutional kernels with different receptive fields to obtain the feature set F; S3, based on feature set F, performs semantic segmentation of text region and table line region, locates and extracts the coordinates of candidate fixed parameter region; S4, perform character recognition on the candidate fixed-value parameter area and convert it into a structured parameter set according to the logical relationship of "fixed-value name, value, unit"; S5, perform unit unification mapping and numerical standardization on the identified set of structured parameters, and construct an N-dimensional standard parameter vector P based on the processed parameters; S6, map the standard parameter vector P and the device setpoint vector Q to the same feature space, and calculate the Euclidean distance between the two as the difference degree D; S7, compare the difference degree D with the preset tolerance thresholds for each parameter dimension. Compare the values ​​to determine if they are consistent. S8 outputs a visual comparison report containing consistent, inconsistent, and outlier items.

[0008] Preferably, in step S2, the multi-scale feature extraction model includes at least three parallel convolutional branches with kernel sizes of 1×1, 3×3, and 5×5, respectively. The extracted features are then fused at multiple scales using a feature pyramid network to obtain a multi-scale fused feature set. F : ; in, I The input image is denoted by ⊕, which represents the feature concatenation operation. FPN is the feature pyramid fusion module.

[0009] Preferably, in step S3, semantic segmentation employs a lightweight segmentation network based on the U-Net++ architecture, where the network uses features... F As input, through an encoder-decoder structure and dense jump connections, the output is a probability map of each pixel belonging to text, table lines, or background.

[0010] Preferably, a hybrid loss function combining binary cross-entropy and Dice Loss is used to train the lightweight segmentation network, achieving pixel-level classification of text content, table lines, and cell regions, with a segmentation loss of [missing information]. A hybrid approach combining cross-entropy loss and the Dice coefficient is adopted: , in, y For real labels, To predict probabilities, C For the number of categories, To balance the weights, The smoothing constant is used; after segmentation, the precise bounding box coordinates of text blocks and table cells are obtained through morphological operations and connected component analysis. .

[0011] Preferably, in step S4, character recognition employs a sequence recognition model combining a convolutional neural network and a bidirectional long short-term memory network; for each bounding box... First, affine correction is performed using a spatial transformation network. Then, the sequence features are extracted by a CNN, followed by context encoding using a BiLSTM. Finally, the original character sequence is obtained by connecting it to a temporal classification decoder. .

[0012] Preferably, a predefined fixed-value professional dictionary is introduced during the decoding process. v A constrained search is performed to improve the accuracy of terminology recognition. The recognized text is then processed by a rule-based and statistical parser, which uses keyword matching, regular expressions, and context analysis to structure it into a set of triples. ; in M This represents the total number of identified constants.

[0013] Preferably, in step S5, performing unit-unified mapping on the identified set of structured parameters includes: According to the preset unit mapping rule table, units with different dimensions but the same physical meaning are uniformly converted. "kA", "A" and "mA" are uniformly mapped to the standard current unit "A", "kV", "V" and "mV" are uniformly mapped to the standard voltage unit "V", and "ms" and "s" are uniformly mapped to the second dimension.

[0014] Preferably, in step S5, the numerical standardization of the identified structured parameter set includes: Numerical standardization includes uniformly preserving floating-point numbers to a preset number of decimal places, performing decimal conversion on percentage parameters, and for fixed parameters with both linear and quadratic values, applying preset ratio coefficients. k The conversion process is performed, and the conversion relationship is as follows: ; in, These are the original identification values. These are the standardized values ​​after conversion. k For the variable ratio or proportionality coefficient; after unit unification and numerical standardization, each constant parameter is represented as a standard parameter vector: ; in, n The number of dimensions for a fixed parameter.

[0015] Preferably, in step S6, the device setpoint vector Q is obtained by parsing device communication messages or a plain text file generated by virtual printing, and is reconstructed with the unit standard in the same dimensional order as the standard parameter vector.

[0016] Preferably, in step S6, the Euclidean distance between the two is calculated as the degree of difference. D The calculation formula is: ; in, These are the standardized values ​​after conversion, representing the standardized components of the scheduling setpoint parameters. q i These are the corresponding operating setpoint parameter components on the equipment side. n The parameter dimension is denoted by 'x'. By using the above method, fixed-value data with different formats and representations are converted into a unified mathematical vector form, achieving precise quantitative comparison at the parameter level and avoiding misjudgment problems caused by relying solely on string matching.

[0017] Furthermore, before calculating the difference degree D, the system performs a parameter alignment operation: if the identified standard parameter vector P and the device fixed value vector Q have inconsistent dimensions, the system fills in the missing items using the preset default values ​​or removes redundant items that do not participate in the comparison to ensure that the semantics of dimension i correspond one-to-one.

[0018] Preferably, in step S7, a preset tolerance threshold is set. The tolerance threshold is dynamically adjusted according to the parameter type. For numerical parameters, the tolerance threshold is set based on the typical error range of the set value. For text parameters, the string edit distance is used as the difference measure.

[0019] Preferably, in step S7, the consistency determination of numerical parameters is performed using a three-level logic: If the difference If they are consistent, then it is considered consistent; if If so, it is determined that a review is required; if If they are inconsistent, then it is determined that they are not in agreement; among them and These are the tolerance thresholds for different levels.

[0020] Preferably, the text-type parameter uses string edit distance as the difference metric, and the calculation formula is as follows: ; in, and These represent the scheduling setting text and the device setting text, respectively.

[0021] Preferably, in step S8, the visual comparison report is output in tabular form, with each row corresponding to a setpoint parameter record. It includes at least the fields of setpoint name, scheduling setpoint, equipment setpoint, difference degree D, judgment result, and suggested processing information. The visual comparison report supports color coding: green marks consistent items, red marks inconsistent items, and yellow marks items that need to be reviewed. It also provides item-by-item difference values ​​and suggested correction information. At the same time, the report marks the source page number and table row number corresponding to each difference parameter so that maintenance personnel can quickly locate the corresponding position in the original PDF setpoint sheet and achieve traceability of the difference results.

[0022] Furthermore, the system will store each comparison result in a historical database for subsequent statistical analysis and model optimization, thereby achieving long-term traceability management of the protection setting verification process.

[0023] Preferably, an automatic identification and comparison system for PDF format value sheets based on multi-scale feature extraction and parameter standardization is used to execute the automatic identification and comparison method. The system includes: The data acquisition module is used to acquire the PDF format protection setting file issued by the dispatch side and convert it into a high-resolution image sequence. It is also used to collect the raw setting data of the current operation of the relay protection device from the device through the device communication interface or virtual printing driver. The multi-scale feature extraction module, whose input end is connected to the output end of the data acquisition module, is used to receive the high-resolution image sequence and construct a multi-scale feature extraction model based on a parallel multi-branch convolutional neural network. It extracts image features in parallel through convolutional kernels with different receptive fields to obtain a multi-scale fusion feature set F. The region segmentation and localization module has its input end connected to the output end of the multi-scale feature extraction module. It is used to receive the feature set F and perform pixel-level classification of text regions and table line regions in the image based on the semantic segmentation network, and locate and extract the coordinates of candidate fixed parameter regions. The character recognition and structuring module has its input end connected to the output end of the region segmentation and positioning module. It is used to receive the coordinates of the candidate fixed value parameter region, perform character recognition on the corresponding region, and convert the recognition result into a set of structured parameters according to the logical relationship of "fixed value name, value, and unit". The parameter standardization module, whose input end is connected to the output end of the character recognition and structuring module, is used to receive the structured parameter set, perform unit unification mapping and numerical standardization processing based on the transformation coefficient k on the recognized parameters, and construct and output an N-dimensional standard parameter vector P. The device setpoint parsing module has its input end connected to the output end of the data acquisition module. It is used to receive the raw setpoint data of the device currently in operation, and reconstruct the device setpoint vector Q by parsing the device communication message or the plain text file generated by virtual printing, according to the same dimension order as the standard parameter vector P and the unit standard. The difference comparison analysis module has its first input connected to the output of the parameter standardization module and its second input connected to the output of the device setpoint analysis module. It receives the standard parameter vector P and the device setpoint vector Q, maps them to the same feature space, calculates the Euclidean distance to obtain the difference degree D, and then compares the difference degree D with a preset tolerance threshold for each parameter dimension. Perform a comparison to determine whether the corresponding fixed values ​​are consistent, and output the determination result; The result generation and report output module has its input end connected to the output end of the difference comparison analysis module. It is used to receive the judgment results and generate and output a visual comparison report containing consistent, inconsistent and abnormal items.

[0024] Preferably, a computer device includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization.

[0025] Preferably, a computer-readable storage medium stores computer instructions that cause a computer to execute the described automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization.

[0026] The beneficial effects of the automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization provided by this invention are as follows: This invention uses PDF-format protected value sheets as the processing object. It establishes a complete technical process from image input to parameter-level automatic comparison output by constructing a multi-scale feature extraction, parameter structured recognition, unit unification and numerical standardization processing, and a value consistency judgment model based on vector distance. Compared with existing value verification methods that rely solely on character recognition or text matching, this invention transforms unstructured text into standard parameter vectors, thereby achieving semantic-level understanding and precise quantitative comparison of value information. By introducing a parameter standardization mapping mechanism and a multi-level tolerance judgment strategy, this invention effectively eliminates the impact of differences in value sheet formats, unit representations, and numerical precision on the comparison results. It solves the problem of insufficient comparison accuracy caused by format differences, inconsistent units, and inconsistent human interpretation in existing technologies, transforming the value verification process from manual experience-based judgment to an automated judgment process based on mathematical models. Therefore, this invention not only improves the accuracy and stability of protection setting identification and comparison, but also realizes intelligent and traceable management of the setting verification process. It has good engineering promotion value and practical application significance in the field of power system relay protection setting management, demonstrating significant technological innovation and practicality. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a flowchart illustrating the automatic recognition process for PDF format custom orders in this invention. Figure 3 This is a bar chart comparing the performance of the method in this invention with existing methods in this embodiment; Figure 4 This is a graph showing the accuracy of identification and comparison of each set of fixed values ​​in this embodiment of the invention. Figure 5 This is a schematic diagram of the structure of a computer device in an embodiment of the present invention. Detailed Implementation

[0028] Example 1: like Figure 1 As shown, the automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization includes the following steps: S1. Obtain the PDF format protection setting file issued by the dispatch side and convert it into a high-resolution image sequence. At the same time, collect the original setting data of the current operation of the relay protection device from the device communication interface or virtual printing driver to form two types of input sources to be compared. S2, a multi-scale feature extraction model is constructed based on a parallel multi-branch convolutional neural network. Image features are extracted in parallel through convolutional kernels with different receptive fields to obtain the feature set F; S3, based on feature set F, performs semantic segmentation of text region and table line region, locates and extracts the coordinates of candidate fixed parameter region; S4, perform character recognition on the candidate fixed-value parameter area and convert it into a structured parameter set according to the logical relationship of "fixed-value name, value, unit"; S5, perform unit unification mapping and numerical standardization on the identified set of structured parameters, and construct an N-dimensional standard parameter vector P based on the processed parameters; S6, map the standard parameter vector P and the device setpoint vector Q to the same feature space, and calculate the Euclidean distance between the two as the difference degree D; S7, compare the difference degree D with the preset tolerance thresholds for each parameter dimension. Compare the values ​​to determine if they are consistent. S8 outputs a visual comparison report containing consistent, inconsistent, and outlier items.

[0029] like Figure 2 As shown, in step S2, the multi-scale feature extraction model contains at least three parallel convolutional branches with kernel sizes of 1×1, 3×3, and 5×5, respectively. The extracted features are then fused at multiple scales using a feature pyramid network to obtain a multi-scale fused feature set. F : ; in, I The input image is denoted by ⊕, which represents the feature concatenation operation. FPN is the feature pyramid fusion module.

[0030] Preferably, in step S3, semantic segmentation employs a lightweight segmentation network based on the U-Net++ architecture, where the network uses features... F As input, through an encoder-decoder structure and dense jump connections, the output is a probability map of each pixel belonging to text, table lines, or background.

[0031] Preferably, a hybrid loss function combining binary cross-entropy and Dice Loss is used to train the lightweight segmentation network, achieving pixel-level classification of text content, table lines, and cell regions, with a segmentation loss of [missing information]. A hybrid approach combining cross-entropy loss and the Dice coefficient is adopted: , in, y For real labels, To predict probabilities, C For the number of categories, To balance the weights, The smoothing constant is used; after segmentation, the precise bounding box coordinates of text blocks and table cells are obtained through morphological operations and connected component analysis. .

[0032] Preferably, in step S4, character recognition employs a sequence recognition model combining a convolutional neural network and a bidirectional long short-term memory network; for each bounding box... First, affine correction is performed using a spatial transformation network. Then, the sequence features are extracted by a CNN, followed by context encoding using a BiLSTM. Finally, the original character sequence is obtained by connecting it to a temporal classification decoder. .

[0033] Preferably, a predefined fixed-value professional dictionary is introduced during the decoding process. v A constrained search is performed to improve the accuracy of terminology recognition. The recognized text is then processed by a rule-based and statistical parser, which uses keyword matching, regular expressions, and context analysis to structure it into a set of triples. ; in M This represents the total number of identified constants.

[0034] Preferably, in step S5, performing unit-unified mapping on the identified set of structured parameters includes: According to the preset unit mapping rule table, units with different dimensions but the same physical meaning are uniformly converted. "kA", "A" and "mA" are uniformly mapped to the standard current unit "A", "kV", "V" and "mV" are uniformly mapped to the standard voltage unit "V", and "ms" and "s" are uniformly mapped to the second dimension.

[0035] Preferably, in step S5, the numerical standardization of the identified structured parameter set includes: Numerical standardization includes uniformly preserving floating-point numbers to a preset number of decimal places, performing decimal conversion on percentage parameters, and for fixed parameters with both linear and quadratic values, applying preset ratio coefficients. k The conversion process is performed, and the conversion relationship is as follows: ; in, These are the original identification values. These are the standardized values ​​after conversion. k For the variable ratio or proportionality coefficient; after unit unification and numerical standardization, each constant parameter is represented as a standard parameter vector: ; in, nThe number of dimensions for a fixed parameter.

[0036] Preferably, in step S6, the device setpoint vector Q is obtained by parsing device communication messages or a plain text file generated by virtual printing, and is reconstructed with the unit standard in the same dimensional order as the standard parameter vector.

[0037] Preferably, in step S6, the Euclidean distance between the two is calculated as the degree of difference. D The calculation formula is: ; in, These are the standardized values ​​after conversion, representing the standardized components of the scheduling setpoint parameters. q i These are the corresponding operating setpoint parameter components on the equipment side. n The parameter dimension is denoted by 'x'. By using the above method, fixed-value data with different formats and representations are converted into a unified mathematical vector form, achieving precise quantitative comparison at the parameter level and avoiding misjudgment problems caused by relying solely on string matching.

[0038] Furthermore, before calculating the difference degree D, the system performs a parameter alignment operation: if the identified standard parameter vector P and the device fixed value vector Q have inconsistent dimensions, the system fills in the missing items using the preset default values ​​or removes redundant items that do not participate in the comparison to ensure that the semantics of dimension i correspond one-to-one.

[0039] Preferably, in step S7, a preset tolerance threshold is set. The tolerance threshold is dynamically adjusted according to the parameter type. For numerical parameters, the tolerance threshold is set based on the typical error range of the set value. For text parameters, the string edit distance is used as the difference measure.

[0040] Preferably, in step S7, the consistency determination of numerical parameters is performed using a three-level logic: If the difference If they are consistent, then it is considered consistent; if If so, it is determined that a review is required; if If they are inconsistent, then it is determined that they are not in agreement; among them and These are the tolerance thresholds for different levels.

[0041] Preferably, the text-type parameter uses string edit distance as the difference metric, and the calculation formula is as follows: ; in, and These represent the scheduling setting text and the device setting text, respectively.

[0042] Preferably, in step S8, the visual comparison report is output in tabular form, with each row corresponding to a setpoint parameter record. It includes at least the fields of setpoint name, scheduling setpoint, equipment setpoint, difference degree D, judgment result, and suggested processing information. The visual comparison report supports color coding: green marks consistent items, red marks inconsistent items, and yellow marks items that need to be reviewed. It also provides item-by-item difference values ​​and suggested correction information. At the same time, the report marks the source page number and table row number corresponding to each difference parameter so that maintenance personnel can quickly locate the corresponding position in the original PDF setpoint sheet and achieve traceability of the difference results.

[0043] Furthermore, the system will store each comparison result in a historical database for subsequent statistical analysis and model optimization, thereby achieving long-term traceability management of the protection setting verification process.

[0044] Preferably, an automatic identification and comparison system for PDF format value sheets based on multi-scale feature extraction and parameter standardization is used to execute the automatic identification and comparison method. The system includes: The data acquisition module is used to acquire the PDF format protection setting file issued by the dispatch side and convert it into a high-resolution image sequence. It is also used to collect the raw setting data of the current operation of the relay protection device from the device through the device communication interface or virtual printing driver. The multi-scale feature extraction module, whose input end is connected to the output end of the data acquisition module, is used to receive the high-resolution image sequence and construct a multi-scale feature extraction model based on a parallel multi-branch convolutional neural network. It extracts image features in parallel through convolutional kernels with different receptive fields to obtain a multi-scale fusion feature set F. The region segmentation and localization module has its input end connected to the output end of the multi-scale feature extraction module. It is used to receive the feature set F and perform pixel-level classification of text regions and table line regions in the image based on the semantic segmentation network, and locate and extract the coordinates of candidate fixed parameter regions. The character recognition and structuring module has its input end connected to the output end of the region segmentation and positioning module. It is used to receive the coordinates of the candidate fixed value parameter region, perform character recognition on the corresponding region, and convert the recognition result into a set of structured parameters according to the logical relationship of "fixed value name, value, and unit". The parameter standardization module, whose input end is connected to the output end of the character recognition and structuring module, is used to receive the structured parameter set, perform unit unification mapping and numerical standardization processing based on the transformation coefficient k on the recognized parameters, and construct and output an N-dimensional standard parameter vector P. The device setpoint parsing module has its input end connected to the output end of the data acquisition module. It is used to receive the raw setpoint data of the device currently in operation, and reconstruct the device setpoint vector Q by parsing the device communication message or the plain text file generated by virtual printing, according to the same dimension order as the standard parameter vector P and the unit standard. The difference comparison analysis module has its first input connected to the output of the parameter standardization module and its second input connected to the output of the device setpoint analysis module. It receives the standard parameter vector P and the device setpoint vector Q, maps them to the same feature space, calculates the Euclidean distance to obtain the difference degree D, and then compares the difference degree D with a preset tolerance threshold for each parameter dimension. Perform a comparison to determine whether the corresponding fixed values ​​are consistent, and output the determination result; The result generation and report output module has its input end connected to the output end of the difference comparison analysis module. It is used to receive the judgment results and generate and output a visual comparison report containing consistent, inconsistent and abnormal items.

[0045] Preferably, a computer device includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization.

[0046] Preferably, a computer-readable storage medium stores computer instructions that cause a computer to execute the described automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization.

[0047] Example 2: This embodiment aims to refine the character recognition and structured parsing model initially trained in Embodiment 1. By introducing prior knowledge feedback from the encoding library and visual-semantic consistency constraints, it improves the model's robustness and parsing accuracy when dealing with fuzzy characters, unconventional layouts, and low-quality images. Building upon the supervised training in Embodiment 1, this embodiment introduces a visual-semantic consistency loss function as an additional optimization objective, forcing the model's recognition results to not only match the image at the pixel level but also conform to predefined fixed-value professional semantic rules. The specific steps are as follows: Step 1: Construct the semantic encoding library and logical rule library: An enhanced character encoding library is established, containing standard terminology, unit symbols, and their visual morphological variations in the field of relay protection. This library stores the text encoding of characters and their corresponding visual feature vectors. Simultaneously, a parameter logic rule library is constructed, defining the logical constraints between the "name," "value," and "unit" of the setpoint parameters. That is, for any specific name, its corresponding parameter value must belong to the set of values, and its unit must belong to a preset set of physical units.

[0048] Step 2: Calculate the feedback loss of the encoding library: The character recognition module outputs a predicted character sequence and position confidence for the input image region; it then measures the visual feature vector of the predicted sequence against similar reference feature vectors retrieved from the encoding library, calculating the encoding library feedback loss Lcode. This loss term penalizes the deviation between the predicted sequence and known power industry visual-semantic patterns.

[0049] Step 3: Calculate the logical rule loss: The structured parsing module parses the predicted character sequence into triples (name, value, unit); and performs semantic validation on the input parameter logical rule base using these triples. If the parsing result violates constraints (such as misidentifying the "on / off" status as a value, or a mismatch between the unit and the parameter name), a logical rule loss is generated. .

[0050] Step 4: Construct the visual-semantic consistency loss function: By combining the above loss terms, a visual-semantic consistency loss is constructed. : ; in, and To balance hyperparameters, used to adjust the contribution ratio between visual pattern matching and business logic constraints.

[0051] Step 5: Hybrid Loss Optimization and Model Update Define the total loss function of the model. To monitor learning loss Visual-semantic consistency loss Weighted sum: ; in, and These are the weighting coefficients; by minimizing Backpropagation optimization of model parameters enables the model to spontaneously correct identification biases that do not conform to the logic of power business during training.

[0052] The advantage of this embodiment lies in that by deeply embedding domain expertise, coding libraries, and logical rules into the model training process in the form of a differentiable loss function, the model's recognition and parsing results are forced not only to fit the pixel-label pairs in the training data, but also to conform to the prior semantics and visual rules of the relay protection setting parameters; this effectively reduces the generation of errors in the model, such as misinterpreting the values ​​of pixels and labels. The probability of incorrect outputs such as misidentifying "0", misframing "kA" as "k" and "A", or parsing parameter combinations without physical meaning is significantly reduced. This significantly enhances the model's generalization ability and the interpretability of output results when faced with fonts, smudges, or non-standard formats that are not fully covered by the training data, thereby improving the reliability of the entire system in real-world complex environments.

[0053] Example 3: This embodiment is used to solve the problem of quickly applying a trained model to a new PDF value sheet with a completely different layout.

[0054] Suppose we already have a recognition model trained on a large number of source layout value sheets, which has visual-semantic consistency constraints as described in Example 2, and is called a pre-trained model. Now we need to apply it to a target layout value sheet with a completely new layout topology, such as a table nesting method and field distribution positions.

[0055] Step 1: Freeze general feature extraction capabilities: Analyzing the model architecture, the multi-scale feature extraction network primarily learns to extract basic character strokes, line geometry, and general electrical symbols from PDF images, such as... Visual features, these low-level features are highly universal across different document layouts. Therefore, to preserve the ability to extract these universal features, all parameters of the model's underlying convolutional network are frozen during the transfer adaptation process, keeping them constant during subsequent training.

[0056] Step 2: Reconstruct the layout topology model and fine-tune the spatial resolution units: Obtain sample values ​​from the target layout and identify its unique logical topology, such as the relative offset between value names and values, and the side-by-side distribution pattern of multiple columns. Using a small number of unlabeled samples of the target layout, fine-tune the units in the model responsible for spatial relationships and structured analysis, such as the coordinate projection-based association module. The goal of this fine-tuning is to enable the model to learn the unique "key-value" spatial alignment pattern of the target layout.

[0057] Step 3: Integrate verification and feedback closed loop: After fine-tuning, the frozen low-level feature units are re-integrated with the fine-tuned spatial parsing units to form a new model adapted to the target layout. In actual operation, the parameter logic rule base described in Example 2 is introduced for verification. If the recognition result under the new layout meets the constraints, the adaptation is deemed successful.

[0058] The effectiveness of this embodiment lies in achieving the reuse of the model's core visual feature extraction capabilities and rapid adaptation to new document layouts through a strategy of parameter freezing and local spatial unit fine-tuning. This significantly reduces the reliance on large amounts of manually annotated data for new format value sheets, greatly improving the engineering adaptation efficiency and practical value of the comparison system across different power stations and documents from different manufacturers.

[0059] Example 4: This embodiment takes a large hydropower plant with a total installed capacity of 2400MW, including six 400MW hydro-turbine generator units, as the application object. It selects 12 PDF format setting sheets issued by the dispatch center in the third quarter of 2024 for 220kV bus protection, main transformer protection, generator protection, etc., as well as the corresponding operation setting data of the protection devices. The protection devices are NARI PCS-985 series and NARI DGT801 series. The method of this invention is used for automatic identification and comparison. At the same time, it is compared with the traditional manual verification method and the general OCR-string matching method to statistically analyze various key performance indicators.

[0060] The actual operation data of the method of the present invention includes 12 setpoint sheets, each containing an average of 25 setpoint items, for a total of 300 setpoint items. The operation data is shown in Table 1 below.

[0061] Table 1: Actual operating data of the method of the present invention;

[0062] The same set of 12 fixed value sheets were tested using both the traditional manual verification method and the general OCR-string matching method. The results are shown in Table 2 below, which compares the data with the existing technical methods.

[0063] Table 2: Comparison data between existing technologies and this method;

[0064] As shown in Tables 1 and 2 above, the method of this invention demonstrates significant technological advancements in practical applications at hydropower plants. Firstly, in terms of recognition accuracy, the method of this invention achieves an average of 98.1%, nearly 20 percentage points higher than the 78.4% of the general OCR + string matching method. This is attributed to the character recognition mechanism based on multi-scale feature extraction and a professional dictionary constraint, effectively solving the difficulties in recognizing power industry symbols, dense tables, and fuzzy fonts. Secondly, in terms of comparison accuracy, the method of this invention reaches 99.0%, far exceeding the 72.9% of the comparison method. Its core lies in parameter standardization and vector distance calculation strategies, eliminating misjudgments caused by inconsistent units, differences in numerical precision, and format variations. Thirdly, in terms of efficiency, the method of this invention takes an average of only 17.75 seconds to process a single setpoint, while traditional manual verification takes 23.5 minutes, representing an efficiency improvement of nearly 80 times; compared to the general OCR method, processing time is reduced by approximately 60%. Furthermore, the method of this invention requires an average of only 0.58 items for manual review per document, greatly reducing the burden on maintenance personnel. In summary, this invention achieves breakthrough improvements in three dimensions: identification accuracy, comparison reliability, and processing efficiency, providing a highly automated and robust solution for the protection setting management of hydropower enterprises.

[0065] Figure 3 The time consumption of the three methods is clearly displayed: traditional manual verification takes an average of 1410 seconds, the general OCR + string matching method takes 44.8 seconds, while the method of this invention only takes 17.75 seconds, which is about 79 times more efficient than manual verification and about 60% more efficient than the general OCR method. The line graph shows the recognition accuracy and comparison accuracy: the general OCR method has a recognition accuracy of only 78.4% and a comparison accuracy of 72.9%; the method of this invention has a recognition accuracy of 98.1% and a comparison accuracy of 99.0%. This graph fully demonstrates that this invention achieves extremely high efficiency while maintaining high accuracy, which is unattainable by manual methods.

[0066] Figure 4 The performance of the method of this invention on 12 actual hydropower plant setting sheets was plotted. The red dotted line represents the recognition accuracy rate, and the blue square line represents the comparison accuracy rate. It can be seen that the recognition accuracy rate fluctuated between 96.4% and 100%, with an average of 98.1%; the comparison accuracy rate ranged between 95.8% and 100%, with an average of 99.0%. Except for a few setting sheets where the comparison accuracy rate was slightly lower due to minor damage to the original PDF image or unconventional table nesting, but still above 95%, the remaining setting sheets achieved 100% comparison accuracy. This figure demonstrates the strong robustness of the method of this invention to setting sheets of different formats and manufacturers, maintaining high reliability even under non-ideal conditions, and completely solving the problem of poor adaptability of traditional OCR methods to power industry symbols and complex formats.

[0067] Example 5: like Figure 5 As shown, embodiments of the present invention also provide a computer device. Figure 5 Taking a single processor 10 as an example, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other using different buses and can be mounted on a common motherboard or otherwise installed as needed. The processor can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices, such as display devices coupled to the interface. In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations, for example, as a server array, a group of blade servers, or a multiprocessor system.

[0068] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0069] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0070] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0071] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0072] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0073] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

Claims

1. An automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization, characterized in that: Includes the following steps: S1. Obtain the PDF format protection setting file issued by the dispatch side and convert it into a high-resolution image sequence. At the same time, collect the original setting data of the current operation of the relay protection device from the device communication interface or virtual printing driver to form two types of input sources to be compared. S2, a multi-scale feature extraction model is constructed based on a parallel multi-branch convolutional neural network. Image features are extracted in parallel through convolutional kernels with different receptive fields to obtain the feature set F; S3, based on feature set F, performs semantic segmentation of text region and table line region, locates and extracts the coordinates of candidate fixed parameter region; S4, perform character recognition on the candidate fixed-value parameter area and convert it into a structured parameter set according to the logical relationship of "fixed-value name, value, unit"; S5, perform unit unification mapping and numerical standardization on the identified set of structured parameters, and construct an N-dimensional standard parameter vector P based on the processed parameters; S6, map the standard parameter vector P and the device setpoint vector Q to the same feature space, and calculate the Euclidean distance between the two as the difference degree D; S7, compare the difference degree D with the preset tolerance thresholds for each parameter dimension. Compare the values ​​to determine if they are consistent. S8 outputs a visual comparison report containing consistent, inconsistent, and outlier items.

2. The automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization according to claim 1, characterized in that, In step S2, the multi-scale feature extraction model contains at least three parallel convolutional branches with kernel sizes of 1×1, 3×3, and 5×5, respectively. The extracted features are then fused at multiple scales using a feature pyramid network to obtain a multi-scale fused feature set. F : ； in, I The input image is denoted by ⊕, which represents the feature concatenation operation. FPN is the feature pyramid fusion module.

3. The automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization according to claim 2, characterized in that, In step S3, semantic segmentation employs a lightweight segmentation network based on the U-Net++ architecture. The network uses features... F As input, through an encoder-decoder structure and dense jump connections, the output is a probability map of each pixel belonging to text, table lines, or background.

4. The automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization according to claim 3, characterized in that, A lightweight segmentation network is trained using a hybrid loss function combining binary cross-entropy and Dice Loss to achieve pixel-level classification of text content, table lines, and cell regions. The segmentation loss... A hybrid approach combining cross-entropy loss and the Dice coefficient is adopted: ， in, y For real labels, To predict probabilities, C For the number of categories, To balance the weights, The smoothing constant is used; after segmentation, the precise bounding box coordinates of text blocks and table cells are obtained through morphological operations and connected component analysis. .

5. The automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization according to claim 1, characterized in that, In step S4, character recognition employs a sequence recognition model combining convolutional neural networks and bidirectional long short-term memory networks; for each bounding box... First, affine correction is performed using a spatial transformation network. Then, the sequence features are extracted by a CNN, followed by context encoding using a BiLSTM. Finally, the original character sequence is obtained by connecting it to a temporal classification decoder. .

6. The automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization according to claim 5, characterized in that, Introducing a predefined fixed-value professional dictionary during the decoding process v A constrained search is performed to improve the accuracy of terminology recognition. The recognized text is then processed by a rule-based and statistical parser, which uses keyword matching, regular expressions, and context analysis to structure it into a set of triples. ； in M This represents the total number of identified constants.

7. The automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization according to claim 1, characterized in that, In step S5, the unit-unified mapping of the identified structured parameter set includes: According to the preset unit mapping rule table, units with different dimensions but the same physical meaning are uniformly converted. "kA", "A" and "mA" are uniformly mapped to the standard current unit "A", "kV", "V" and "mV" are uniformly mapped to the standard voltage unit "V", and "ms" and "s" are uniformly mapped to the second dimension.

8. The automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization according to claim 7, characterized in that, In step S5, the numerical standardization of the identified structured parameter set includes: Numerical standardization includes uniformly preserving floating-point numbers to a preset number of decimal places, performing decimal conversion on percentage parameters, and for fixed parameters with both linear and quadratic values, applying preset ratio coefficients. k The conversion process is performed, and the conversion relationship is as follows: ； in, These are the original identification values. These are the standardized values ​​after conversion. k For the variable ratio or proportionality coefficient; after unit unification and numerical standardization, each constant parameter is represented as a standard parameter vector: ； in, n The number of dimensions for a fixed parameter.

9. The automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization according to claim 1, characterized in that, In step S6, the device setpoint vector Q is obtained by parsing device communication messages or generating plain text files through virtual printing, and is reconstructed with the unit standard according to the same dimensional order as the standard parameter vector.

10. The automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization according to claim 9, characterized in that, In step S6, the Euclidean distance between the two is calculated as the degree of difference. D The calculation formula is: ； in, These are the standardized values ​​after conversion, representing the standardized components of the scheduling setpoint parameters. q i These are the corresponding operating setpoint parameter components on the equipment side. n The number of parameter dimensions.

11. The automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization according to claim 1, characterized in that, In step S7, a preset tolerance threshold is set. The tolerance threshold is dynamically adjusted according to the parameter type. For numerical parameters, the tolerance threshold is set based on the typical error range of the set value. For text parameters, the string edit distance is used as the difference measure.

12. The automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization according to claim 11, characterized in that, In step S7, the consistency determination of numerical parameters is performed using three levels of logic: If the difference If they are consistent, then it is determined to be consistent; if If so, it is determined that a review is required; if If they are inconsistent, then it is determined that they are not consistent; among them and These are the tolerance thresholds for different levels.

13. The automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization according to claim 11, characterized in that, For text-type parameters, string edit distance is used as the difference metric, and the calculation formula is as follows: ； in, and These represent the scheduling setting text and the device setting text, respectively.

14. The automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization according to claim 1, characterized in that, In step S8, the visualization comparison report is output in tabular form, with each row corresponding to a setpoint parameter record. It includes at least the fields of setpoint name, scheduling setpoint, equipment setpoint, difference degree D, judgment result and suggested processing information. The visualization comparison report supports color coding: green marks consistent items, red marks inconsistent items, and yellow marks items that need to be reviewed. It also provides item-by-item difference values ​​and suggested correction information.

15. An automatic identification and comparison system for PDF format value sheets based on multi-scale feature extraction and parameter standardization, used to execute the automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization as described in any one of claims 1-14, characterized in that the system... include: The data acquisition module is used to acquire the PDF format protection setting file issued by the dispatch side and convert it into a high-resolution image sequence. It is also used to collect the raw setting data of the current operation of the relay protection device from the device through the device communication interface or virtual printing driver. The multi-scale feature extraction module has its input end connected to the output end of the data acquisition module. It is used to receive the high-resolution image sequence and construct a multi-scale feature extraction model based on a parallel multi-branch convolutional neural network. It extracts image features in parallel through convolutional kernels with different receptive fields to obtain a multi-scale fusion feature set F. The region segmentation and localization module has its input end connected to the output end of the multi-scale feature extraction module. It is used to receive the feature set F and perform pixel-level classification of text regions and table line regions in the image based on the semantic segmentation network, and locate and extract the coordinates of candidate fixed parameter regions. The character recognition and structuring module has its input end connected to the output end of the region segmentation and positioning module. It is used to receive the coordinates of the candidate fixed value parameter region, perform character recognition on the corresponding region, and convert the recognition result into a set of structured parameters according to the logical relationship of "fixed value name, value, and unit". The parameter standardization module, whose input end is connected to the output end of the character recognition and structuring module, is used to receive the structured parameter set, perform unit unification mapping and numerical standardization processing based on the transformation coefficient k on the recognized parameters, and construct and output an N-dimensional standard parameter vector P. The device setpoint parsing module has its input end connected to the output end of the data acquisition module. It is used to receive the raw setpoint data of the device currently in operation, and reconstruct the device setpoint vector Q by parsing the device communication message or the plain text file generated by virtual printing, according to the same dimension order as the standard parameter vector P and the unit standard. The difference comparison analysis module has its first input connected to the output of the parameter standardization module and its second input connected to the output of the device setpoint analysis module. It receives the standard parameter vector P and the device setpoint vector Q, maps them to the same feature space, calculates the Euclidean distance to obtain the difference degree D, and then compares the difference degree D with a preset tolerance threshold for each parameter dimension. Perform a comparison to determine whether the corresponding fixed values ​​are consistent, and output the determination result; The result generation and report output module has its input end connected to the output end of the difference comparison analysis module. It is used to receive the judgment results and generate and output a visual comparison report containing consistent, inconsistent and abnormal items.

16. A computer device, characterized in that, It includes a memory and a processor, which are interconnected and communicate with each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization as described in claims 1-14.

17. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the automatic identification and comparison method for PDF format value sheets based on multi-scale feature extraction and parameter standardization as described in claims 1-14.

Images