Small sample precision evaluation data enhancement method and system based on deep learning

Abstract

The application discloses a small sample precision evaluation data enhancement method and system based on deep learning, relates to the technical field of detection data processing, and comprises the following steps: acquiring original evaluation data and review range data, performing review pairing processing on the original evaluation data to obtain review pairing data, performing boundary state migration identification on the review pairing data to obtain boundary state migration data, training a deep learning data enhancement model according to the review pairing data and the boundary state migration data, performing pairing enhancement on the review pairing data to obtain enhanced pairing data, merging the enhanced pairing data and the original evaluation data to obtain enhanced evaluation data, performing boundary state migration statistics on the enhanced evaluation data to obtain a precision evaluation result, and the application can expand the precision evaluation data volume of the small sample while maintaining the pairing relationship between the primary review record and the review record, so that the precision evaluation result can represent the discrete degree of the review pairing difference.

Description

Technical Field

[0001] This application relates to the field of detection data processing technology, and more specifically, to a method and system for small-sample precision assessment data augmentation based on deep learning. Background Technology

[0002] In the processes of precision parts dimensional inspection, laboratory testing, metrological calibration, online sensor monitoring, and product outgoing quality inspection, it is necessary to evaluate the precision of the test results based on the test data. Precision assessment is used to characterize the degree of consistency between test values ​​when the same object is repeatedly tested under specified test conditions. In test scenarios close to the judgment boundary, the initial test value of the object is likely to fall into the retest trigger range, which is determined based on the judgment boundary and the retest width. When the initial test value is within the retest trigger range, the same object needs to be retested, and the stability of the test process should be evaluated by combining the initial test value, the retest value, the initial test judgment result, and the retest judgment result.

[0003] In small-sample precision evaluation scenarios, the number of paired retest data is less than the minimum number of pairs required for precision evaluation. When precision evaluation is performed solely based on the original evaluation data, the results are easily affected by fluctuations in a small number of paired retest data. To expand the amount of evaluation data, existing processing methods typically employ deep learning data augmentation models to augment individual detection records or the overall data distribution, supplementing the original evaluation data with augmented samples. Data augmentation methods that use individual detection records or the overall data distribution as augmentation targets usually enhance samples by constraining them with mean, variance, or overall distribution similarity, making the augmented samples closer to the original evaluation data at the overall numerical level.

[0004] However, the detection records within the re-inspection trigger range are not independent single records. There is a pairing relationship between the initial and re-inspection records for the same detection object. The positional relationship between the initial and re-inspection values ​​relative to the judgment boundary, the migration relationship between the initial and re-inspection boundary states, and the consistency relationship between the initial and re-inspection judgment results all affect the precision assessment results. When the data augmentation model generates only a single detection record, or only constrains the overall distribution similarity between the augmented sample and the original assessment data, the augmented sample generated by the data augmentation model may meet the overall mean or overall variance requirements, but it cannot maintain the pairing relationship between the initial and re-inspection records for the same detection object.

[0005] When only overall distribution similarity is constrained without constraining the re-inspection pairing relationship, candidate enhanced pairing data may alter the state transition characteristics near the judgment boundary in the re-inspection pairing data formed from the original evaluation data. For example, in the re-inspection pairing data formed from the original evaluation data, when both the initial inspection boundary state and the re-inspection boundary state are in a qualified state, the candidate enhanced pairing data may form a qualified-to-unqualified migration type that is inconsistent with the boundary state migration type corresponding to the re-inspection pairing data. Similarly, in the re-inspection pairing data formed from the original evaluation data, when the initial inspection judgment result and the re-inspection judgment result are consistent, the candidate enhanced pairing data may form an enhanced pairing result that is inconsistent with the initial inspection judgment result and the re-inspection judgment result. For precision part dimensional inspection, enhanced pairing results inconsistent with the boundary state migration type corresponding to the re-inspection pairing data may lead to parts nearing the dimensional tolerance boundary being incorrectly released or incorrectly rejected. For laboratory testing, enhanced pairing results inconsistent with the initial inspection judgment result and the re-inspection judgment result may lead to samples nearing the limit boundary being incorrectly judged. For online sensor monitoring and metrological calibration, enhanced pairing results inconsistent with the initial inspection judgment result and the re-inspection judgment result may lead to inconsistencies between alarm triggering results, re-inspection judgment results, or calibration judgment results.

[0006] Therefore, in small-sample precision evaluation scenarios where the initial detection value is within the re-inspection trigger range, existing data augmentation methods using deep learning mainly focus on the overall distribution similarity between the augmented sample and the original evaluation data. They lack constraints on the pairing relationship between the initial detection record and the re-inspection record, lack constraints on the boundary state transition relationship near the judgment boundary, and lack verification of the boundary state transition relationship of the candidate augmented pairing data. This may cause the augmented sample to disrupt the pairing relationship between the initial detection record and the re-inspection record, as well as the consistency relationship between the initial detection judgment result and the re-inspection judgment result.

[0007] In view of this, this application proposes a data augmentation method and system for small-sample precision evaluation based on deep learning to solve the above problems. Summary of the Invention

[0008] To overcome the aforementioned shortcomings of the prior art and achieve the above objectives, this application provides the following technical solution: a deep learning-based method for small-sample precision evaluation data augmentation, comprising: Obtain the original assessment data and the re-inspection scope data, and perform re-inspection pairing processing on the original assessment data based on the re-inspection scope data to obtain the re-inspection pairing data; Based on the re-inspection range data, boundary state migration identification is performed on the re-inspection paired data to obtain boundary state migration data; The re-examination pairing data and boundary state transition data are processed to train the model, resulting in a deep learning data augmentation model. Based on boundary state transition data and deep learning data augmentation models, pairing enhancement processing is performed on the re-examination pairing data to obtain enhanced pairing data. The enhanced paired data and the original evaluation data are merged to obtain the enhanced evaluation data. Based on the re-inspection range data, boundary state migration statistical processing is performed on the enhanced assessment data to obtain the precision assessment results.

[0009] Furthermore, the method for obtaining the re-examination pairing data includes: Extract the detection object identifier, detection item identifier, detection batch identifier, record type, detection value, detection time, and judgment result from the original assessment data; Based on the record type, the original assessment data is divided into preliminary inspection records and re-inspection records. Preliminary inspection values ​​and preliminary inspection judgment results are extracted from the preliminary inspection records, and re-inspection values ​​and re-inspection judgment results are extracted from the re-inspection records. Extract the re-inspection trigger range corresponding to the test item identifier from the re-inspection range data; Initial inspection records whose initial inspection values ​​fall within the re-inspection trigger range are identified as initial inspection records to be paired. The initial inspection records and re-inspection records to be paired are matched for consistency of identifiers and for the order of detection time to obtain the paired re-inspection records; Based on the initial inspection record to be matched and the re-inspection record of the matched pairs, the re-inspection matching data is obtained.

[0010] Furthermore, the method for obtaining paired re-examination records includes: Extract the test object identifier, test item identifier, and test batch identifier from the initial inspection record and re-inspection record to be matched; Re-inspection records with consistent test object identification, consistent test item identification, and consistent test batch identification are identified as candidate re-inspection records; From the candidate re-examination records, retain the candidate re-examination records whose testing time is later than the initial test record to be matched, and obtain the valid candidate re-examination records; The detection time intervals between valid candidate re-examination records and the initial examination records to be paired are sorted to obtain the time interval sorting results; Based on the sorting results of the time intervals, the valid candidate re-examination record with the smallest detection time interval is determined as the paired re-examination record.

[0011] Furthermore, the method for obtaining boundary state transition data includes: Extract the initial test value, retest value, initial test judgment result, retest judgment result, and test item identifier from the retest paired data; Extract the judgment boundary, qualified side range, and unqualified side range corresponding to the test item identifier from the re-inspection range data; Based on the judgment boundary, the initial inspection value and the re-inspection value are subjected to boundary selection processing to obtain the target judgment boundary; Based on the target determination boundary, the boundary distance between the initial inspection value and the re-inspection value is calculated to obtain the initial inspection boundary distance and the re-inspection boundary distance. State transition identification is performed based on the initial inspection boundary distance, re-inspection boundary distance, qualified side range, unqualified side range, initial inspection judgment result, and re-inspection judgment result to obtain the state transition identification result. By associating the state transition identification results with the re-examination pairing data, boundary state transition data is obtained.

[0012] Furthermore, the method for obtaining the target determination boundary includes: When the decision boundary is a single decision boundary, the single decision boundary is determined as the target decision boundary; When the decision boundary includes both the lower and upper decision boundaries, the distances between the initial and re-inspection values ​​and the lower decision boundary are calculated and summed to obtain the lower limit distance sum. The distances between the initial inspection value and the re-inspection value and the upper limit judgment boundary are calculated and summed to obtain the upper limit distance sum. When the sum of the lower limit distances is less than or equal to the sum of the upper limit distances, the lower limit judgment boundary is determined as the target judgment boundary; When the sum of the lower limit distances is greater than the sum of the upper limit distances, the upper limit judgment boundary is determined as the target judgment boundary.

[0013] Furthermore, the method for obtaining the state transition recognition result includes: The accuracy of the test value record is obtained from the re-inspection and pairing data, and at least one of the minimum measurement division of the test item and the reading resolution of the test equipment is obtained. The maximum value of the obtained items is determined as the boundary width. Based on the equal width of the boundary, the qualified side range, and the unqualified side range, the initial inspection boundary distance and the re-inspection boundary distance are divided into states to obtain the initial inspection boundary state and the re-inspection boundary state. By combining the initial boundary state and the re-inspection boundary state, the boundary state transition type is obtained; By comparing the initial inspection results and the re-inspection results, a consistency marker is obtained. The boundary width, boundary state transition type, and decision consistency marker are used as the state transition recognition results.

[0014] Furthermore, the method for obtaining the deep learning data augmentation model includes: A training sample set is constructed based on the re-examination pairing data and boundary state transition data; Encode the boundary state transition type and decision consistency marker as conditional input features; The re-inspection width and judgment direction coefficient are obtained based on the re-inspection range data; Based on the target judgment boundary, re-inspection width, boundary equal width and judgment direction coefficient, the initial inspection value and re-inspection value are scaled to obtain the normalized boundary distance training label; Input the conditional input features and random input values ​​into the model to be trained to obtain the initial normalized boundary distance and the re-normalized boundary distance. The training error is calculated based on the generated initial normalized boundary distance, the generated re-normalized boundary distance, and the normalized boundary distance training labels, and the transfer state is verified to obtain the transfer state error. The deep learning data augmentation model is obtained by updating and stopping the training based on the training error and the transfer state error.

[0015] Furthermore, the method for calculating the training error includes: The error is calculated between the generated initial detection normalized boundary distance and the initial detection normalized boundary distance in the training sample set, and the error is calculated between the generated re-detection normalized boundary distance and the re-detection normalized boundary distance in the training sample set to obtain the normalized distance error. Based on the target judgment boundary, re-inspection width, boundary equal width and judgment direction coefficient, the detection values ​​of the generated initial inspection normalized boundary distance and the generated re-inspection normalized boundary distance are converted to obtain the generated initial inspection value and the generated re-inspection value. The normalized difference between the generated initial inspection value and the generated re-inspection value is calculated to obtain the generated normalized difference. The normalized difference is calculated by taking the initial and retest values ​​in the training sample set to obtain the training normalized difference. The error between the generated normalized difference and the training normalized difference is calculated to obtain the difference relationship error. The training error is obtained by weighted summation of the normalized distance error and the difference relationship error.

[0016] Furthermore, the method for obtaining enhanced pairing data includes: Extract the target boundary state migration type and target determination consistency flag from the boundary state migration data to obtain the target migration combination; Select the target re-examination pairing data corresponding to the target migration combination; Based on the boundary state migration data and re-inspection range data of the target re-inspection pairing data, the detection value transformation parameters and re-inspection trigger range are obtained; By combining the target transfer and random input values ​​and inputting them into the deep learning data augmentation model, we can obtain the candidate initial detection normalized boundary distance and the candidate re-detection normalized boundary distance. Based on the conversion parameters of the detection values, the candidate initial detection normalized boundary distance and the candidate re-detection normalized boundary distance are converted to obtain candidate enhancement pairing data; The candidate enhanced pairing data are re-identified by migration combination to obtain candidate migration combinations; When the initial detection value of the candidate enhancement is within the range of the re-detection trigger and the candidate migration combination is consistent with the target migration combination, the enhancement pairing data is retained.

[0017] Deep learning-based small-sample precision assessment data augmentation systems include: The data pairing processing module is used to acquire the original evaluation data and the re-inspection range data, and to perform re-inspection pairing processing on the original evaluation data based on the re-inspection range data to obtain the re-inspection paired data. The boundary state migration identification module is used to identify the boundary state migration of the re-inspection paired data based on the re-inspection range data, and obtain the boundary state migration data. The model training and processing module is used to train and process the re-examination pairing data and boundary state transition data to obtain a deep learning data augmentation model. The pairing enhancement processing module is used to enhance the model based on boundary state transfer data and deep learning data, and to perform pairing enhancement processing on the re-examination pairing data to obtain enhanced pairing data; The data merging and processing module is used to merge the enhanced paired data and the original evaluation data to obtain the enhanced evaluation data; The precision assessment processing module is used to perform boundary state migration statistical processing on the enhanced assessment data based on the re-inspection range data to obtain the precision assessment results.

[0018] Compared with existing technologies, the technical effects and advantages of the deep learning-based small-sample precision evaluation data augmentation method and system of this application are as follows: This application uses the re-examination range data as a constraint, forming re-examination paired data from the initial examination records and re-examination records in the original evaluation data, and completes model training, pairing enhancement, and statistical processing around the boundary state transfer relationship. Therefore, this application can expand the amount of small-sample precision evaluation data while retaining the pairing relationship between the initial examination records and the re-examination records in the enhanced evaluation data, reducing the number of candidate enhanced paired data that do not meet the re-examination trigger range requirements or target transfer combinations being retained as enhanced paired data.

[0019] Through re-inspection pairing processing, the initial inspection records and re-inspection records, which were scattered in the original evaluation data, are converted into re-inspection pairing data based on the same inspection object. The re-inspection pairing data retains the inspection object identifier, inspection item identifier, inspection batch identifier, initial inspection value, re-inspection value, initial inspection judgment result, and re-inspection judgment result. This allows model training and pairing enhancement processing to focus on paired inspection records, reducing the possibility of losing pairing relationships due to the initial inspection records and re-inspection records being split during data enhancement.

[0020] The boundary state migration identification process determines the positional changes of the initial and re-inspection values ​​relative to the judgment boundary based on the re-inspection range data, forming boundary state migration data that includes the initial inspection boundary state, the re-inspection boundary state, the boundary state migration type, and a judgment consistency marker. This boundary state migration data can distinguish between qualified-side retention, unqualified-side retention, qualified-side migration to unqualified-side, unqualified-side migration to qualified-side, and boundary overlap migration, reducing the likelihood of neglecting state changes near the judgment boundary when evaluating enhancement data solely based on overall mean, overall variance, or overall distribution similarity.

[0021] In model training and pairing augmentation, the deep learning data augmentation model uses boundary state transition type and decision consistency marker as conditional input features, and the positional relationship between the initial and re-inspection values ​​and the decision boundary as the training basis. After candidate augmentation pairing data is generated, it is retained based on the re-inspection trigger range, target transition combination, and candidate transition combination. Only when the candidate augmentation initial inspection value is within the re-inspection trigger range and the candidate transition combination is consistent with the target transition combination, is the candidate augmentation pairing data included in the merging process. This reduces the possibility that candidate augmentation pairing data may change the target boundary state transition type or alter the consistency between the initial and re-inspection decision results.

[0022] Through merging, the enhanced paired data is converted into pairs of enhanced initial inspection records and enhanced re-inspection records, which, together with the original evaluation data, form enhanced evaluation data. Through boundary state migration statistical processing, the re-inspection paired records in the enhanced evaluation data are re-extracted and statistically analyzed according to boundary state migration type. The precision evaluation results are formed by combining the dispersion of the re-inspection paired difference, the boundary state migration ratio, the number of consistent judgments, the number of inconsistent judgments, the consistency ratio, and the inconsistency ratio. This is used to characterize the dispersion between repeated detection results of the same detection object and the state migration near the judgment boundary. Attached Figure Description

[0023] Figure 1 This is a flowchart of a deep learning-based small-sample precision evaluation data augmentation method according to an embodiment of this application. Figure 2 This is a schematic diagram of a deep learning-based small-sample precision evaluation data augmentation system according to an embodiment of this application. Detailed Implementation

[0024] The technical solutions of this application will be described in detail, clearly, and completely below with reference to the accompanying drawings of the embodiments. It should be particularly noted that the specific embodiments described below are only used to better illustrate and explain the technical solutions of this application, and are intended to enable those skilled in the art to better understand and implement this application, and should not be construed as limiting the scope of protection of this application. Without departing from the spirit and substance of this application, those skilled in the art can modify, adjust, or make equivalent substitutions based on the content disclosed in this application, and these modifications, adjustments, or equivalent substitutions should all be considered within the scope of protection of this application. Example

[0025] This application provides a data augmentation method for small-sample precision evaluation based on deep learning. An electronic device acquires the original evaluation data and re-inspection range data, and performs re-inspection pairing processing, boundary state transition recognition, model training processing, pairing enhancement processing, merging processing, and boundary state transition statistical processing on the initial inspection records and re-inspection records of the same detection object.

[0026] The original assessment data comes from testing logs, re-inspection logs, metrological records, online monitoring records, or factory testing records. Each testing record includes at least the following: testing object identifier, testing item identifier, testing batch identifier, test value, testing time, record type, and judgment result. Record types include initial inspection records and re-inspection records. The testing object identifier is used to distinguish different testing objects. The testing item identifier is used to distinguish different testing items. The testing batch identifier is used to distinguish different testing batches. The test value is used for comparison of judgment boundaries. The testing time is used to determine the order of initial inspection records and re-inspection records. The judgment result indicates whether the test value meets the testing requirements.

[0027] The re-inspection range data includes the judgment boundary, re-inspection width, re-inspection trigger range, acceptable side range, and unacceptable side range. The judgment boundary distinguishes whether the test value meets the test requirements. The re-inspection width determines the range of values ​​near the judgment boundary that require re-inspection. The re-inspection trigger range is determined based on the judgment boundary and the re-inspection width. The acceptable side range indicates the side where the test value meets the test requirements. The unacceptable side range indicates the side where the test value does not meet the test requirements.

[0028] The re-inspection width should preferentially adopt the width specified in the testing procedures, metrological calibration requirements, or quality control documents. If the testing procedures, metrological calibration requirements, or quality control documents do not specify the re-inspection width, the electronic equipment shall determine the re-inspection width based on the accuracy of the recorded test value and the permissible error of the test item. When the accuracy of the recorded test value is less than or equal to the permissible error of the test item, the re-inspection width shall be greater than or equal to the accuracy of the recorded test value and less than or equal to the permissible error of the test item. When the accuracy of the recorded test value is greater than the permissible error of the test item, the re-inspection width shall be determined according to the permissible error of the test item. By determining the re-inspection width, the re-inspection trigger range can be adapted to the formal judgment requirements of the test item.

[0029] The judgment boundaries and acceptable / unacceptable ranges are obtained from the testing procedures, metrological calibration requirements, or quality control documents. The accuracy of the recorded test values ​​is obtained from the testing record format, metrological records, or the number of decimal places retained for the test values. The minimum metrological division for the tested items is obtained from the testing procedures, metrological calibration requirements, or testing equipment configuration records. The resolution of the testing equipment readings is obtained from the testing equipment manual, metrological calibration certificate, or testing equipment configuration records.

[0030] Please see Figure 1 This application provides a data augmentation method for small-sample precision evaluation based on deep learning, comprising: S1. Obtain the original assessment data and the re-inspection range data, and perform re-inspection pairing processing on the original assessment data based on the re-inspection range data to obtain the re-inspection pairing data.

[0031] S101. Electronic equipment obtains original evaluation data from testing logs, re-inspection logs, metrological records, online monitoring records, or factory inspection records. Testing logs provide the identification of the tested object, the tested item, the tested batch, the tested value, and the tested time. Re-inspection logs provide the tested value, tested time, and judgment result of the re-inspection record. Metrological records provide the accuracy of the tested value recording. Online monitoring records provide the record type and tested time for the same tested object during continuous testing. Factory inspection records provide the initial inspection and re-inspection records in the product's factory quality inspection. All data sources originate from records generated during the testing process, supporting the traceable acquisition of original evaluation data.

[0032] S102. The electronic equipment obtains re-inspection range data from the testing procedures, metrological calibration requirements, accuracy of test value recording, allowable error of test items, or quality control documents. Based on the test item identifier, the electronic equipment matches the judgment boundary, re-inspection width, re-inspection trigger range, pass / fail range, and fail / unqualified range with the test records in the original evaluation data. Different test items have different judgment boundaries and re-inspection widths; the electronic equipment matches them using the test item identifier, reducing the mixing of judgment boundaries between different test items.

[0033] S103, the electronic equipment organizes the original evaluation data according to the detection object identifier, detection item identifier, and detection batch identifier, and retains the detection value, detection time, record type, and judgment result. The electronic equipment distinguishes between initial inspection records and re-inspection records based on the record type. The detection value in the initial inspection record is determined as the initial inspection value, the detection value in the re-inspection record is determined as the re-inspection value, the judgment result in the initial inspection record is determined as the initial inspection judgment result, and the judgment result in the re-inspection record is determined as the re-inspection judgment result. The detection object identifier, detection item identifier, and detection batch identifier jointly define the repeated detection relationship of the same detection object under the same detection conditions, and the detection time is used to define the order of initial inspection records and re-inspection records.

[0034] S104, the electronic device filters the initial inspection records based on the re-inspection trigger range in the re-inspection range data. When the initial inspection value is within the re-inspection trigger range, the electronic device identifies the initial inspection record as a pending initial inspection record. When the initial inspection value is outside the re-inspection trigger range, the electronic device does not include the initial inspection record in the re-inspection pairing process. Initial inspection records within the re-inspection trigger range are selected because the re-inspection trigger range is determined by the judgment boundary and the re-inspection width. Initial inspection values ​​within the re-inspection trigger range are close to the judgment boundary, and the consistency between the initial inspection record and the re-inspection record will affect the precision evaluation result.

[0035] S105, the electronic equipment establishes a pairing relationship between the initial inspection record to be paired and the re-inspection record based on the consistency of the detection object identifier, the detection item identifier, the detection batch identifier, and the detection time of the re-inspection record being later than the detection time of the initial inspection record to be paired. When one initial inspection record to be paired corresponds to multiple candidate re-inspection records, the electronic equipment selects the candidate re-inspection record whose detection time is closest to the initial inspection record to be paired as the paired re-inspection record. When one re-inspection record corresponds to multiple candidate initial inspection records, the electronic equipment selects the initial inspection record to be paired that has the closest detection time to the re-inspection record and whose detection time is earlier than the re-inspection record as the paired initial inspection record. Detection records that cannot form a unique pairing relationship are not included in the re-inspection pairing data.

[0036] S106, the electronic device generates re-inspection pairing data based on the established pairing relationship. The re-inspection pairing data includes the detection object identifier, detection item identifier, detection batch identifier, detection value recording precision, initial detection value, re-inspection value, initial detection judgment result, re-inspection judgment result, initial detection time, re-inspection time, the direction of the difference between the initial detection value and the re-inspection value, and the magnitude of the difference between the initial detection value and the re-inspection value. The direction of the difference indicates whether the re-inspection value increases, decreases, or remains unchanged relative to the initial detection value. The magnitude of the difference indicates the degree of numerical difference between the initial detection value and the re-inspection value. When the initial detection record and the re-inspection record in the same re-inspection pairing data correspond to different detection value recording precisions, the electronic device determines the maximum value between the detection value recording precision corresponding to the initial detection record and the detection value recording precision corresponding to the re-inspection record as the detection value recording precision corresponding to the re-inspection pairing data.

[0037] Through the processing in S1, the scattered initial inspection records and re-inspection records in the original evaluation data are converted into re-inspection paired data. The re-inspection paired data uses the initial inspection records and re-inspection records of the same test object as the processing unit, which can reduce the situation where the re-inspection pairing relationship is lost due to treating the initial inspection records and re-inspection records as independent test records for enhancement.

[0038] S2, based on the re-inspection range data, perform boundary state migration identification on the re-inspection paired data to obtain boundary state migration data.

[0039] S201, the electronic device extracts the corresponding judgment boundary, re-inspection trigger range, qualified side range, and unqualified side range from the re-inspection range data based on the detection item identifier in the re-inspection pairing data. For a single judgment boundary, the electronic device determines the single judgment boundary as the target judgment boundary. For a two-sided judgment scenario, the electronic device calculates the distances between the initial inspection value and the lower limit judgment boundary, the re-inspection value and the lower limit judgment boundary, the initial inspection value and the upper limit judgment boundary, and the re-inspection value and the upper limit judgment boundary, respectively. If the sum of the distances between the initial inspection value and the lower limit judgment boundary and the re-inspection value and the lower limit judgment boundary is less than or equal to the sum of the distances between the initial inspection value and the upper limit judgment boundary and the re-inspection value and the upper limit judgment boundary, the electronic device determines the lower limit judgment boundary as the target judgment boundary; otherwise, the electronic device determines the upper limit judgment boundary as the target judgment boundary. By using the target judgment boundary determination method, the initial inspection value and the re-inspection value are used together for judgment boundary selection, which can reduce the inconsistency in the boundary state migration recognition caliber caused by selecting the target judgment boundary solely based on the initial inspection value or solely based on the re-inspection value.

[0040] S202, the electronic equipment determines the initial inspection boundary distance and the re-inspection boundary distance based on the target judgment boundary. The initial inspection boundary distance represents the position of the initial inspection value relative to the target judgment boundary. The re-inspection boundary distance represents the position of the re-inspection value relative to the target judgment boundary. Both the initial inspection boundary distance and the re-inspection boundary distance retain directional information. The directional information is used to indicate whether the detected value is located in the acceptable or unacceptable range. The reason for retaining the directional information is that precision assessment not only needs to determine the numerical difference between the initial inspection value and the re-inspection value, but also needs to determine whether the initial inspection value and the re-inspection value are distributed on different sides of the judgment boundary.

[0041] S203, the electronic equipment determines the boundary equivalence width based on at least one of the following: the accuracy of the recorded test value corresponding to the re-inspection and pairing data, the minimum metrological division of the test item, and the reading resolution of the test equipment. The boundary equivalence width is denoted as... The accuracy of the detection value record corresponding to the re-inspection paired data is denoted as . The minimum measurement division for the test item is denoted as The resolution of the detection equipment readings is denoted as Boundary equal to width Represented as: ; in, , and All values ​​are converted to the corresponding unit of measurement for the test items. When any one of the following is missing: the accuracy of the recorded test value, the minimum measurement division of the test item, or the resolution of the test equipment reading, the electronic device will determine the maximum value among the acquired items as the boundary width. The maximum value among the acquired items is chosen as the boundary width because the accuracy of the recorded test value, the minimum measurement division of the test item, and the resolution of the test equipment reading all limit the distinguishing ability of the test value near the judgment boundary. The boundary width corresponding to each pair of re-examination data is denoted as . .

[0042] S204, the electronic equipment determines the initial inspection boundary state and the re-inspection boundary state based on the initial inspection boundary distance, the acceptable side range, the unacceptable side range, and the boundary equivalent width. The initial inspection boundary state and the re-inspection boundary state include the acceptable side state, the unacceptable side state, and the boundary coincidence state. The electronic equipment performs boundary state judgment on the initial inspection value and the re-inspection value respectively. For the initial inspection value, the electronic equipment determines whether the distance between the initial inspection value and the target judgment boundary is less than or equal to the boundary equivalent width; if the distance between the initial inspection value and the target judgment boundary is less than or equal to the boundary equivalent width, the initial inspection boundary state is determined to be the boundary coincidence state; if the distance between the initial inspection value and the target judgment boundary is greater than the boundary equivalent width, the initial inspection boundary state is determined to be the acceptable side state or the unacceptable side state based on whether the initial inspection value is located within the acceptable side range or the unacceptable side range. For re-inspection values, the electronic equipment determines whether the distance between the re-inspection value and the target judgment boundary is less than or equal to the boundary's width. If the distance is less than or equal to the boundary's width, the re-inspection boundary state is determined to be a boundary coincidence state. If the distance is greater than the boundary's width, the re-inspection boundary state is determined to be either a qualified or unqualified state, depending on whether the re-inspection value is within the qualified or unqualified range. The re-inspection trigger range is used to determine the initial inspection record participating in the re-inspection pairing process. When the re-inspection value is outside the re-inspection trigger range, the electronic equipment still determines the re-inspection boundary state based on the qualified or unqualified range.

[0043] S205, the electronic device combines the initial inspection boundary state and the re-inspection boundary state from the same re-inspection pairing data to obtain a boundary state migration type. Boundary state migration types include qualified-side retention type, unqualified-side retention type, qualified-side to unqualified-side migration type, unqualified-side to qualified-side migration type, and boundary coincidence migration type. Qualified-side retention type indicates that both the initial inspection boundary state and the re-inspection boundary state are qualified-side states. Unqualified-side retention type indicates that both the initial inspection boundary state and the re-inspection boundary state are unqualified-side states. Qualified-side to unqualified-side migration type indicates that the initial inspection boundary state is qualified-side and the re-inspection boundary state is unqualified-side. Unqualified-side to qualified-side migration type indicates that the initial inspection boundary state is unqualified-side and the re-inspection boundary state is qualified-side. Boundary coincidence migration type indicates that either the initial inspection boundary state or the re-inspection boundary state is a boundary coincidence state.

[0044] S206, the electronic equipment determines a consistency flag based on whether the initial inspection result and the re-inspection result are consistent. If both the initial inspection result and the re-inspection result indicate that the detection requirements are met, or both indicate that the detection requirements are not met, the consistency flag is "consistent." If the initial inspection result and the re-inspection result differ, with one indicating that the detection requirements are met and the other indicating that the detection requirements are not met, the consistency flag is "inconsistent." The electronic equipment correlates the target judgment boundary, boundary width, initial inspection boundary state, re-inspection boundary state, boundary state transition type, consistency flag, initial inspection boundary distance, re-inspection boundary distance, direction of the difference between the initial inspection value and the re-inspection value, and magnitude of the difference between the initial inspection value and the re-inspection value to obtain boundary state transition data.

[0045] Through the processing in S2, the re-inspection paired data is converted into boundary state transition data that can represent changes in position near the decision boundary. The boundary state transition data simultaneously preserves the positional relationship between the initial and re-inspection values ​​relative to the decision boundary, the difference relationship between the initial and re-inspection values, and the consistency relationship between the initial and re-inspection decision results. This can reduce the possibility of changing the pairing relationship near the decision boundary simply by using the similarity of the overall numerical distribution as an enhancement criterion.

[0046] S3 is used to train the model on the re-examination pairing data and boundary state transition data to obtain a deep learning data augmentation model.

[0047] S301, the electronic device forms a training sample set based on the re-inspection pairing data and the boundary state transition data. Each training sample in the training sample set corresponds to a set of re-inspection pairing data and a set of boundary state transition data. The re-inspection pairing data provides the detection object identifier, detection item identifier, detection batch identifier, detection value recording accuracy, initial detection value, re-inspection value, initial detection judgment result, re-inspection judgment result, the direction of the difference between the initial detection value and the re-inspection value, and the magnitude of the difference between the initial detection value and the re-inspection value. The boundary state transition data provides the target judgment boundary, boundary equal width, initial detection boundary distance, re-inspection boundary distance, initial detection boundary state, re-inspection boundary state, boundary state transition type, and judgment consistency flag. Since both the re-inspection pairing data and the boundary state transition data are obtained from S1 and S2, the model training process no longer repeatedly collects the original evaluation data and re-inspection range data. The small sample scenario in this embodiment is a scenario where the number of re-inspection pairing data is greater than 0 and less than the minimum number of pairs required for precision evaluation.

[0048] S302, the electronic device obtains the re-inspection width from the re-inspection range data based on the detection item identifier corresponding to the training sample, and obtains the boundary equivalent width based on the boundary state transition data corresponding to the training sample. It then performs scale adjustment on the initial detection value, re-inspection value, initial detection boundary distance, and re-inspection boundary distance in the training sample set. Let the... The initial detection value in each training sample is The retest value was The target determination boundary is The width of the re-inspection is The boundary width is equal to The direction coefficient is determined as The electronic device obtains the qualified side range corresponding to the target judgment boundary from the re-inspection range data based on the detection item identifier corresponding to the training samples, and determines the judgment direction coefficient based on the position of the qualified side range relative to the target judgment boundary. When the acceptable range is located on the side of the target judgment boundary where the value increases, Set to 1; when the qualified side range is located on the side of the target judgment boundary where the value decreases, Set to -1. The determination direction coefficient is used to ensure that the initial inspection normalized boundary distance and the re-inspection normalized boundary distance, after scaling, have the same direction within the qualified side range. Initial normalized boundary distance of each training sample and the re-inspection normalized boundary distance Represented as: ; ; in, A value greater than 0, or a value greater than or equal to 0. and They have the same unit of measurement for detection values. The scaled initial detection normalized boundary distance and re-detection normalized boundary distance are used to represent the position of the initial detection value and the re-detection value relative to the target judgment boundary, so that the detection values ​​in different detection items can be entered into the model training process according to a unified scale.

[0049] S303, the electronic device uses the boundary state transition type, decision consistency label, and random input value as input features for a deep learning data augmentation model, and the initial detection normalized boundary distance and the re-detection normalized boundary distance as training labels to train the deep learning data augmentation model. The random input value is a one-dimensional random number obtained from a standard normal distribution with a mean of 0 and a variance of 1. The deep learning data augmentation model is implemented using a multi-layer feedforward neural network generation model. This model is used to generate the initial detection normalized boundary distance and the re-detection normalized boundary distance based on the boundary state transition type, decision consistency label, and random input value. Boundary state transition types are converted into boundary state transition type codes using one-hot encoding. The dimension of the boundary state transition type code is determined according to the number of boundary state transition types. The decision consistency label is converted into a decision consistency label code using binary encoding, where consistency corresponds to 1 and inconsistency corresponds to 0. The boundary state transition type code, decision consistency label code, and random input value are jointly input into the multilayer feedforward neural network generation model. The multilayer feedforward neural network generation model has two output nodes, which correspond to generating the initial normalized boundary distance and the re-normalized boundary distance, respectively.

[0050] S304, the electronic device obtains the generated initial inspection value and the generated re-inspection value based on the generated initial inspection normalization boundary distance and the generated re-inspection normalization boundary distance. Let the generated initial inspection normalization boundary distance be... The generated re-inspection normalized boundary distance is The initial detection value is generated. The generated re-inspection value is The generated initial test value and the generated retest value are represented as follows: ; ; By generating initial and re-inspection values, the data output by the deep learning data augmentation model can be returned to the original detection value unit of the corresponding detection item. The generated initial and re-inspection values ​​can then be used to continue boundary state migration recognition based on the re-inspection range data.

[0051] S305, the electronic device uses the normalized distance error and the difference relationship error as the target quantities for iteratively updating the model parameters, and employs gradient descent to iteratively update the model parameters. The normalized distance error measures the difference between the generated initial detection normalized boundary distance and the initial detection normalized boundary distance in the training samples, as well as the difference between the generated re-detection normalized boundary distance and the re-detection normalized boundary distance in the training samples. The difference relationship error measures the difference between the generated normalized difference between the generated initial detection value and the generated re-detection value, and the difference between the training normalized difference between the initial detection value and the re-detection value in the training samples. The generated normalized difference is determined based on the decision direction coefficient, the generated initial detection value, the generated re-detection value, the re-detection width, and the boundary equal width. The training normalized difference is determined based on the decision direction coefficient, the initial detection value in the training samples, the re-detection value in the training samples, the re-detection width, and the boundary equal width. For example, the first... The generated normalized difference corresponding to each training sample is denoted as . , No. The training normalization difference corresponding to each training sample is denoted as . Then a normalized difference is generated. and training normalized difference Represented as: ; ; After each round of model training, the electronic device, based on the generated initial detection value, generated re-detection value, the target judgment boundary corresponding to the current training sample, and the boundary equal width, and obtaining the qualified and unqualified side ranges from the re-detection range data according to the detection item identifier in the re-detection pairing data corresponding to the current training sample, redetermines the boundary state transition type. Then, it determines the transition state error based on whether the redetermined boundary state transition type is consistent with the boundary state transition type in the input features. Let the number of current training samples in each round of model training be... Electronic devices generate based on the current training samples The generated sample. The generated sample corresponds to the first... Let there be a current training sample. The number of generated samples with inconsistent boundary state transition types among the generated samples is Migration state error Represented as: ; in, Values ​​greater than 0 greater than or equal to 0 and less than or equal to The value. The number. The generation initial detection normalization boundary distance corresponding to each generated sample is: The generated re-inspection normalized boundary distance is , No. The initial normalized boundary distance corresponding to each current training sample is: The re-inspection normalized boundary distance is , No. The generation normalization difference corresponding to each generated sample is: , No. The training normalization difference corresponding to each current training sample is: Normalized distance error Sum and difference relationship error Represented as: ; ; The training error function is expressed as: ; in, For training error, For normalized distance error, Let be the difference relationship error, and α and β be the error weights, and In one implementation, and Take 1 / 2 for all; when the testing procedures, metrological calibration requirements, or quality control documents specify a weight for the normalized distance error or difference relationship error, determine the weight according to the specified weight. and Migration state error It does not participate in determining the direction of model parameter updates, and the transition state error... Used for determining stop conditions.

[0052] S306, the electronic device compares the training error with an error threshold and the transition state error with a transition error limit. If the training error is greater than the error threshold, the electronic device continues model training. If the training error is less than or equal to the error threshold and the transition state error is greater than the transition error limit, the electronic device continues model training. If the training error is less than or equal to the error threshold and the transition state error is less than or equal to the transition error limit, the electronic device stops model training and fixes the model parameters that meet the stopping conditions, thus obtaining a deep learning data augmentation model. The error threshold is determined based on the detection value recording accuracy, re-detection width, and boundary equivalence width. For example, the error threshold... Represented as: ; in, The number of training samples. For the first The accuracy of recording the detection values ​​corresponding to each training sample. For the first The boundaries corresponding to each training sample have the same width. For the first The re-examination width corresponding to each training sample This is the error scaling factor. Obtained from the allowable percentage of re-inspection difference given in the testing procedures, metrological calibration requirements, or quality control documents; if the testing procedures, metrological calibration requirements, or quality control documents do not give an allowable percentage of re-inspection difference, the error proportion coefficient. The transfer error limit is set to 1. It limits the proportion of inconsistent boundary state transfer types in the generated data after each round of model training. When the quality control document specifies a transfer error limit, it is determined according to the document. When the quality control document does not specify a transfer error limit, it is set to 0. A transfer error limit of 0 means that the boundary state transfer type redefined in the generated data after each round of model training must be consistent with the boundary state transfer type in the input features. When the number of training rounds reaches the upper limit, the electronic device stops model training and selects the model parameters with the smallest training error from the completed training rounds for solidification, thus obtaining the deep learning data augmentation model. The upper limit for the number of training rounds is obtained from the quality control document. When the quality control document does not specify an upper limit, it is determined based on the number of training samples, and is an integer multiple of the number of training samples. The integer multiple is obtained from the quality control document; when the document does not specify an integer multiple, it is set to 10.

[0053] Through the processing in S3, the training data source, input features, training labels, model type, training error function, model parameter iteration update method, and stopping condition of the deep learning data augmentation model are all limited. The deep learning data augmentation model uses boundary state transition type and decision consistency label as generation conditions, and initial detection normalized boundary distance and re-detection normalized boundary distance as training labels, which can reduce the situation where the augmented data only meets the overall numerical similarity but does not meet the re-detection pairing relationship.

[0054] S4. Based on the boundary state transition data and the deep learning data enhancement model, the re-examination pairing data is enhanced to obtain enhanced pairing data.

[0055] S401, the electronic device extracts the target boundary state migration type and the target judgment consistency flag from the boundary state migration data. The target boundary state migration type is either a qualified side retention type, a non-qualified side retention type, a qualified side to non-qualified side migration type, a non-qualified side to qualified side migration type, or a boundary overlap migration type. The target judgment consistency flag is either consistent or inconsistent. The target boundary state migration type and the target judgment consistency flag form a target migration combination. The target migration combination is used to define the re-examination pairing relationship that the candidate enhancement initial detection value and the candidate enhancement re-examination value need to maintain.

[0056] S402, in this embodiment, the small sample scenario refers to a scenario where the number of re-examination paired data is less than the minimum number of pairs required for precision evaluation. The electronic device determines the total number of target enhancements based on the number of re-examination paired data and the minimum number of pairs required for precision evaluation. Let the number of re-examination paired data be... The minimum number of pairs required for precision assessment is The total number of target enhancements is The total number of target enhancements is then expressed as: ; in, Values ​​greater than 0 greater than The numerical value. Minimum number of pairs required for precision assessment. Obtained from testing procedures, metrological calibration requirements, or quality control documents. The total number of target enhancements for electronic equipment is allocated according to the proportion of each target migration combination in the boundary state migration data. The number of retest pairing data corresponding to the target migration combination is , No. The number of enhancements corresponding to the target migration combination is , Represented as: ; in, This indicates rounding to the nearest integer. When the first... Number of re-examination pairing data corresponding to the target migration combination When equal to 0, the first The number of enhancements corresponding to the target migration combination If the sum of the enhancements corresponding to each target migration combination is less than the total number of target enhancements, the electronic device replenishes the enhancements in descending order of the number of re-inspection pairing data corresponding to each target migration combination. If the sum of the enhancements corresponding to each target migration combination is greater than the total number of target enhancements, the electronic device deducts enhancements from the target migration combinations with enhancements greater than 0 in ascending order of the number of re-inspection pairing data corresponding to each target migration combination, and the enhancements after each deduction are not less than 0, until the sum of the enhancements corresponding to each target migration combination equals the total number of target enhancements.

[0057] In step S403, the electronic device generates random input values ​​according to the method for obtaining random input values ​​in step S303, and inputs the target boundary state transition type, target judgment consistency marker, and random input values ​​into the deep learning data augmentation model. The deep learning data augmentation model outputs the candidate initial detection normalized boundary distance and the candidate re-detection normalized boundary distance based on the target boundary state transition type, target judgment consistency marker, and random input values. The electronic device selects target re-detection paired data from the re-detection paired data corresponding to the target transition combination, and obtains the target judgment boundary and boundary equal width from the boundary state transition data corresponding to the target re-detection paired data. Based on the detection item identifier corresponding to the target re-detection paired data, it obtains the re-detection width, re-detection trigger range, qualified side range, and unqualified side range from the re-detection range data, and determines the judgment direction coefficient based on the position of the qualified side range relative to the target judgment boundary. When a target migration combination corresponds to multiple re-inspection pairing data, the electronic device arranges the re-inspection pairing data corresponding to the target migration combination in ascending order of the initial inspection time. If the initial inspection times are the same, the re-inspection pairing data is arranged in ascending order of the re-inspection time. If both the initial inspection time and the re-inspection time are the same, the re-inspection pairing data is arranged according to the sorting result of the detection object identifier. The electronic device then cyclically selects the target re-inspection pairing data in the sorted order. Based on the target judgment boundary, re-inspection width, boundary equal width, and judgment direction coefficient, the electronic device converts the candidate initial inspection normalized boundary distance and the candidate re-inspection normalized boundary distance into candidate enhanced initial inspection values ​​and candidate enhanced re-inspection values. The candidate enhanced initial inspection values ​​and candidate enhanced re-inspection values ​​together form candidate enhanced pairing data. Candidate enhanced pairing data correspond to the same enhanced detection object identifier, the same enhanced detection item identifier, and the same enhanced detection batch identifier. The enhanced detection item identifier and enhanced detection batch identifier are determined based on the detection item identifier and detection batch identifier in the target re-inspection pairing data. The enhanced detection object identifier is determined based on the detection object identifier in the target re-examination pairing data and the generation order of the candidate enhanced pairing data.

[0058] S404, the electronic device re-identifies the boundary state migration of the candidate enhanced initial inspection value and the candidate enhanced re-inspection value based on the re-inspection range data to obtain the candidate boundary state migration type. The electronic device determines the candidate initial inspection boundary distance based on the candidate enhanced initial inspection value and the target judgment boundary, determines the candidate re-inspection boundary distance based on the candidate enhanced re-inspection value and the target judgment boundary, determines the candidate initial inspection boundary state and the candidate re-inspection boundary state based on the candidate initial inspection boundary distance, the candidate re-inspection boundary distance, the qualified side range, the unqualified side range, and the boundary equal width, and determines the candidate boundary state migration type based on the candidate initial inspection boundary state and the candidate re-inspection boundary state.

[0059] In S405, the electronic equipment determines the candidate initial inspection result based on the candidate enhancement initial inspection value, the target judgment boundary, the qualified side range, and the unqualified side range, and determines the candidate re-inspection result based on the candidate enhancement re-inspection value, the target judgment boundary, the qualified side range, and the unqualified side range. The electronic equipment obtains a candidate judgment consistency mark based on whether the candidate initial inspection result and the candidate re-inspection result are consistent. The determination rule for the candidate judgment consistency mark is the same as the determination rule for the judgment consistency mark in S2, so that the candidate enhancement pairing data and the re-inspection pairing data use the same judgment caliber.

[0060] S406, the electronic device retains candidate enhancement pairing data based on the candidate enhancement initial detection value, re-detection trigger range, candidate boundary state transition type, target boundary state transition type, candidate judgment consistency flag, and target judgment consistency flag. When the candidate enhancement initial detection value is within the re-detection trigger range, the candidate boundary state transition type is consistent with the target boundary state transition type, and the candidate judgment consistency flag is consistent with the target judgment consistency flag, the electronic device retains the candidate enhancement initial detection value and the candidate enhancement re-detection value, obtaining enhancement pairing data. When the candidate enhancement initial detection value is outside the re-detection trigger range, or the candidate boundary state transition type is inconsistent with the target boundary state transition type, or the candidate judgment consistency flag is inconsistent with the target judgment consistency flag, the electronic device does not retain the candidate enhancement initial detection value and the candidate enhancement re-detection value. For the first... Target migration combinations, when the quantity is increased When the value is 0, the electronic device does not target the first... Several target migration combinations generate candidate augmentation pairings. When the number of augmentations G_ When the value is greater than 0, the electronic device is determined based on the enhancement quantity. The maximum number of generation attempts is determined, and the maximum number of generation attempts is the number of enhancements. The integer multiples are obtained from the quality control document. If the quality control document does not specify an integer multiple, the integer multiple is 10. Electronic devices repeatedly generate candidate enhancement pairing data within the maximum number of generation attempts. Data is retained based on the following conditions: the initial detection value of the candidate enhancement is within the re-detection trigger range; the candidate boundary state transition type is consistent with the target boundary state transition type; and the candidate judgment consistency flag is consistent with the target judgment consistency flag. This process continues until the [number]th generation. The number of augmented pairings corresponding to the target migration combinations reaches the augmentation quantity. Alternatively, the number of times candidate enhancement pairings can be generated may reach the maximum number of generation attempts. If candidate enhancement pairings are not retained, the electronic device regenerates random input values ​​and continues to generate candidate enhancement pairings based on the regenerated random input values. For the first... For a given target migration combination, the number of times candidate augmentation pairing data is generated reaches the upper limit of the number of generation times, and the th... The number of augmented pairings corresponding to the target migration combinations is still less than the number of augmentations. At that time, the electronic device uses the number of enhanced pairing data already retained as the first... The actual number of enhancements corresponding to each target migration combination. The electronic device uses the actual retained enhancement pairing data as the enhancement pairing data to enter the merging process. Enhancement pairing data includes enhancement detection object identifier, enhancement detection item identifier, enhancement detection batch identifier, initial enhancement value, re-inspection value, initial enhancement judgment result, re-inspection judgment result, enhancement boundary state migration type, enhancement judgment consistency flag, and enhancement pairing identifier.

[0061] Through the processing in S4, the enhanced pairing data is still re-identified according to the re-inspection range data after generation, and the boundary state migration relationship is retained according to the re-inspection trigger range and target migration combination. This can reduce the situation where candidate enhanced pairing data deviates from the re-inspection pairing formation caliber or changes the consistency state between the initial inspection judgment result and the re-inspection judgment result.

[0062] S5 merges the enhanced paired data and the original evaluation data to obtain the enhanced evaluation data.

[0063] S501, the electronic device converts the enhanced initial inspection value in the enhanced pairing data into an enhanced initial inspection record, and converts the enhanced re-inspection value in the enhanced pairing data into an enhanced re-inspection record. The enhanced initial inspection record and the enhanced re-inspection record use the same enhanced detection object identifier, the same enhanced detection item identifier, the same enhanced detection batch identifier, and the same enhanced pairing identifier. The record type of the enhanced initial inspection record is an initial inspection record. The record type of the enhanced re-inspection record is a re-inspection record.

[0064] S502, the electronic equipment unifies the data items of the enhanced initial inspection record, enhanced re-inspection record, and original evaluation data. The objects of data item unification include the detection object identifier, detection item identifier, detection batch identifier, detection value, detection time, record type, and judgment result. The enhanced initial inspection record and enhanced re-inspection record use the same data item names and data item arrangement order as the original evaluation data. The reason for unifying the data items is that the enhanced evaluation data needs to include both the original evaluation data and the enhanced paired data; inconsistencies in data items will prevent the boundary state transition statistical processing from reading the inspection records according to the same rules.

[0065] S503, the electronic device organizes the initial enhancement inspection records and re-inspection records in pairs according to the enhancement pairing identifier. Each enhancement pairing identifier corresponds to one initial enhancement inspection record and one re-inspection record. When the same enhancement pairing identifier lacks an initial enhancement inspection record or a re-inspection record, the electronic device does not include the corresponding enhancement pairing data in the merging process. When the same enhancement pairing identifier has both an initial enhancement inspection record and a re-inspection record, the electronic device retains the corresponding enhancement pairing data. The reason for organizing in pairs is that the enhancement object in this embodiment is re-inspection pairing data, and the enhancement evaluation data needs to retain the pairing structure between the initial inspection record and the re-inspection record.

[0066] S504, the electronic equipment merges the paired enhanced initial inspection records, enhanced re-inspection records, and original evaluation data to obtain enhanced evaluation data. Enhanced evaluation data includes the inspection records, enhanced initial inspection records, and enhanced re-inspection records from the original evaluation data. The enhanced evaluation data retains the inspection object identifier, inspection item identifier, inspection batch identifier, inspection value, inspection time, record type, and judgment result from the original evaluation data, and also retains the enhanced pairing identifier from the enhanced pairing data.

[0067] Through processing in S5, the augmented paired data is incorporated into the augmentation evaluation data in the form of paired augmented initial inspection records and augmented re-inspection records. The augmentation evaluation data not only expands the amount of data used for precision assessment but also preserves the correspondence between the augmented initial inspection records and the augmented re-inspection records, thereby reducing the number of augmented data records that are processed into unrelated single inspection records during the merging process.

[0068] S6. Based on the re-inspection range data, perform boundary state migration statistical processing on the enhanced evaluation data to obtain the precision evaluation results.

[0069] S601, the electronic device extracts re-inspection pairing records from the enhancement evaluation data. These re-inspection pairing records include re-inspection pairing data formed from the original evaluation data and enhanced re-inspection pairing records formed from the enhanced pairing data. For detection records in the original evaluation data, the electronic device forms re-inspection pairing records according to the re-inspection pairing processing rules in S1. For enhanced initial inspection records and enhanced re-inspection records, the electronic device forms enhanced re-inspection pairing records based on the enhanced pairing identifier. If the electronic device fails to extract re-inspection pairing records from the enhancement evaluation data, it does not generate a precision evaluation result, outputs a result indicating insufficient re-inspection pairing records, and stops boundary state transition statistical processing. When the electronic device extracts re-inspection pairing records, the original evaluation data and enhanced pairing data jointly participate in the boundary state transition statistical processing.

[0070] S602, the electronic device, based on the re-inspection range data, redetermines the initial inspection boundary state, re-inspection boundary state, and boundary state transition type for each re-inspection pairing record. The electronic device employs the boundary state transition identification rules in S2, processing based on the judgment boundary, the acceptable side range, the unacceptable side range, and the boundary equal width. The electronic device combines the initial inspection boundary state and the re-inspection boundary state in the same re-inspection pairing record to obtain the boundary state transition type for the corresponding re-inspection pairing record.

[0071] S603, the electronic device groups and statistically analyzes the re-inspection pairing records according to the boundary state migration type, obtaining boundary state migration group data. Let the total number of re-inspection pairing records be... , No. The number of re-checked paired records corresponding to the boundary state transition type is , No. Boundary state transition ratio corresponding to each boundary state transition type Represented as: ; in, It is a value greater than 0. When the first... Number of re-checked paired records corresponding to each boundary state transition type When equal to 0, the first Boundary state transition ratio corresponding to each boundary state transition type Equal to 0, and the first Boundary state transition types are not included in the difference discrete value calculation. Boundary state transition grouping data includes the corresponding re-inspection pairing records, pairing quantity, and boundary state transition ratio for each boundary state transition type. The pairing quantity represents the number of re-inspection pairing records under the corresponding boundary state transition type. The boundary state transition ratio represents the proportion of the corresponding boundary state transition type in all re-inspection pairing records. The normalized difference mean and difference discrete value are calculated in S605 based on the re-inspection pairing records and normalized difference corresponding to each boundary state transition type. The number of consistent and inconsistent decisions are statistically obtained in S604 based on the re-inspection pairing records and consistency markers corresponding to each boundary state transition type.

[0072] S604, the electronic device, based on the consistency flag, counts the number of consistent and inconsistent decisions for each boundary state transition type, and adds the number of consistent and inconsistent decisions for each boundary state transition type to the boundary state transition grouping data; the electronic device also counts the number of consistent and inconsistent decisions for all re-inspection pairing records. Let the consistency flag in the i-th re-inspection pairing record be... . A value of 1 indicates that the initial inspection result and the re-inspection result are consistent. A value of 0 indicates that the initial inspection result and the re-inspection result are inconsistent. (Inconsistency rate) Represented as: ; Determine the proportion of inconsistencies Represented as: ; in, This represents the total number of paired records for all re-examinations. Used to indicate the proportion of initial inspection results that are consistent with re-inspection results. This indicates the proportion of discrepancies between the initial inspection result and the re-inspection result. The electronic device supplements the boundary state transition grouping data with the proportions of consistent and inconsistent results.

[0073] S605, the electronic device determines the dispersion of the re-inspection pairing difference based on the boundary state migration grouping data. Let the first... The number of re-checked paired records under the boundary state transition type is , No. The initial test value in each retested pairing record is The retest value was The width of the re-inspection is The boundary width is equal to The direction coefficient is determined as The electronic device first determines the first... Normalized difference of each paired record in the re-examination Normalized difference Represented as: ; When the Number of re-examination paired records under the boundary state migration type When greater than or equal to 1, the first Normalized difference mean under the boundary state transition type Represented as: ; When the Number of re-examination paired records under the boundary state migration type When greater than or equal to 2, the first Discrete value of difference under boundary state transition type Represented as: ; in, Indicates the first The re-examination matching record belongs to the first Types of boundary state transitions. When the first... Number of re-examination paired records under the boundary state migration type When less than 2, the first Some boundary state transition types do not participate in the difference discrete value calculation. Let the set of boundary state transition types that participate in the difference discrete value calculation be . , This includes boundary state transition types where the number of duplicated paired records is greater than or equal to 2. The dispersion of the duplicated paired difference. Represented as: ; in, For the first The boundary state transition ratio corresponding to each boundary state transition type. The set of boundary state transition types used in the differential discrete value calculation. When empty, the electronic device does not generate the re-inspection paired difference dispersion and records it as empty. The electronic device supplements the boundary state transition grouping data with the already generated normalized difference mean, the already generated difference dispersion, and the re-inspection paired difference dispersion to obtain the boundary state transition statistics. Through the processing in S605, the re-inspection paired difference dispersion is formed based on the normalized difference. The difference between the initial inspection value and the re-inspection value of different inspection items is sorted out by the judgment direction coefficient, the re-inspection width, and the boundary equal width, which can reduce the impact of different dimensions and judgment directions of different inspection items on the precision evaluation results.

[0074] S606, the electronic device determines the precision assessment result based on the generation status of the re-inspection pairing difference dispersion. When generating the re-inspection pairing difference dispersion, the electronic device obtains the precision assessment result based on the re-inspection pairing difference dispersion, the boundary state migration ratio corresponding to each boundary state migration type, the number of consistent decisions corresponding to each boundary state migration type, the number of inconsistent decisions corresponding to each boundary state migration type, the consistency ratio, and the inconsistency ratio from the boundary state migration statistics. The precision assessment result includes the re-inspection pairing difference dispersion, the boundary state migration ratio corresponding to each boundary state migration type, the number of consistent decisions corresponding to each boundary state migration type, the inconsistency ratio, and the inconsistency ratio. When the electronic device records the discrepancy of the re-inspection pairing difference as empty, it obtains a precision evaluation result excluding the discrepancy of the re-inspection pairing difference based on the boundary state migration ratio, the number of consistent judgments, the number of inconsistent judgments, the consistency ratio, and the inconsistency ratio corresponding to each boundary state migration type in the boundary state migration statistics. The precision evaluation result excluding the discrepancy of the re-inspection pairing difference includes the boundary state migration ratio, the number of consistent judgments, the number of inconsistent judgments, the consistency ratio, and the inconsistency ratio corresponding to each boundary state migration type. The discrepancy of the re-inspection pairing difference is used to characterize the degree of dispersion between repeated detection results of the same detection object. The boundary state migration ratio is determined based on the number of pairs corresponding to each boundary state migration type and the total number of re-inspection pairing records, and is used to characterize the distribution of the detection value from the initial detection boundary state to the re-inspection boundary state. The consistency ratio is used to characterize the degree of consistency between the initial detection judgment result and the re-inspection judgment result. The inconsistency ratio is used to characterize the degree of change between the initial detection judgment result and the re-inspection judgment result.

[0075] Through the processing in S6, the enhanced assessment data undergoes re-processing of boundary state migration statistics according to the re-inspection range data. The resulting precision assessment result is formed based on the pairing relationship between the initial inspection record and the re-inspection record of the same test object. The precision assessment result is not formed solely based on the overall mean or overall variance of the enhanced assessment data. Instead, it is formed when generating the re-inspection pairing difference dispersion based on the re-inspection pairing difference dispersion, the boundary state migration ratio, the number of consistent judgments corresponding to each boundary state migration type, the number of inconsistent judgments corresponding to each boundary state migration type, the consistency ratio, and the inconsistency ratio. When the re-inspection pairing difference dispersion is recorded as empty, it is formed based on the boundary state migration ratio, the number of consistent judgments corresponding to each boundary state migration type, the number of inconsistent judgments corresponding to each boundary state migration type, the consistency ratio, and the inconsistency ratio. This reduces the possibility of enhanced data being directly used for precision assessment after disrupting the pairing relationship near the judgment boundary. Example

[0076] This embodiment, based on Embodiment 1, describes the training of the deep learning data augmentation model constrained by boundary state transition data and the retention processing of candidate augmentation pairing data. In this embodiment, the original evaluation data, re-examination range data, re-examination pairing data, and boundary state transition data are all obtained according to the processing method in Embodiment 1. This embodiment does not change the technical chain of re-examination pairing processing, merging processing, and boundary state transition statistical processing in Embodiment 1.

[0077] In this embodiment, the electronic device uses re-inspection pairing data as the pairing enhancement object and boundary state transition data as the generation constraint condition. The boundary state transition data includes the target judgment boundary, boundary equal width, initial inspection boundary state, re-inspection boundary state, boundary state transition type, judgment consistency flag, initial inspection boundary distance, re-inspection boundary distance, the direction of the difference between the initial inspection value and the re-inspection value, and the magnitude of the difference between the initial inspection value and the re-inspection value. The boundary state transition data is used to constrain the positional relationship between the enhanced initial inspection value and the re-inspection value relative to the target judgment boundary, the combination relationship between the initial inspection boundary state and the re-inspection boundary state, the consistency relationship between the initial inspection judgment result and the re-inspection judgment result, and the pairing difference relationship between the initial inspection value and the re-inspection value.

[0078] The deep learning data augmentation model training in this embodiment includes the following processes.

[0079] The electronic device extracts the boundary state transition type and decision consistency marker from the boundary state transition data. It then converts the boundary state transition type into a boundary state transition type code using one-hot encoding, and converts the decision consistency marker into a decision consistency marker code using binary encoding. These boundary state transition type and decision consistency marker codes serve as conditional inputs to constrain the deep learning data augmentation model to generate the initial detection normalized boundary distance and the subsequent detection normalized boundary distance, based on the combination relationship between the initial detection boundary state and the subsequent detection boundary state represented by the boundary state transition type.

[0080] The electronic device extracts the initial inspection value and the re-inspection value from the re-inspection pairing data, and extracts the target judgment boundary and boundary equivalent width from the boundary state migration data. Based on the inspection item identifier corresponding to the re-inspection pairing data, the electronic device obtains the re-inspection width, the qualified side range, and the unqualified side range from the re-inspection range data, and determines the judgment direction coefficient based on the position of the qualified side range relative to the target judgment boundary. The judgment direction coefficient is used to unify the direction of the qualified side range relative to the target judgment boundary in different inspection items.

[0081] The electronic device performs scaling on the initial and re-inspection values ​​based on the target determination boundary, re-inspection width, boundary equal width, and determination direction coefficient, obtaining the initial and re-inspection normalized boundary distances. These distances are used as training labels to represent the paired positions of the initial and re-inspection values ​​relative to the target determination boundary. The electronic device uses boundary state transition type encoding, determination consistency label encoding, and random input values ​​as model inputs, and the initial and re-inspection normalized boundary distances as training labels to train the deep learning data augmentation model.

[0082] During model training, the deep learning data augmentation model outputs the initial detection normalized boundary distance and the re-detection normalized boundary distance based on the boundary state transition type encoding, the decision consistency marker encoding, and the random input value. The electronic device then obtains the initial detection value and the re-detection value based on these distances, the target decision boundary, the re-detection width, the boundary equal width, and the decision direction coefficient.

[0083] The electronic device determines the normalization distance error based on the difference between the generated initial detection normalized boundary distance and the initial detection normalized boundary distance in the training samples, and the difference between the generated re-detection normalized boundary distance and the re-detection normalized boundary distance in the training samples. The electronic device determines the difference relationship error based on the generation normalization difference between the generated initial detection value and the generated re-detection value, and the training normalization difference between the initial detection value and the re-detection value in the training samples. The normalization distance error is used to constrain the position of the generated initial detection value and the generated re-detection value relative to the target judgment boundary. The difference relationship error is used to constrain the pairing difference relationship between the generated initial detection value and the generated re-detection value. The electronic device determines the training error based on the normalization distance error and the difference relationship error, and iteratively updates the model parameters of the deep learning data augmentation model based on the training error. The normalization distance error, the difference relationship error, and the training error are all obtained according to the determination method in Example 1.

[0084] After each round of model training, the electronic device re-determines the initial detection boundary state and the re-detection boundary state based on the generated initial detection value, generated re-detection value, target judgment boundary, boundary equal width, qualified side range, and unqualified side range. It then re-determines the boundary state transition type based on these two boundary state transition types. The electronic device compares the generated boundary state transition type with the boundary state transition type corresponding to the model input to obtain the transition state error. The transition state error indicates whether the paired detection values ​​generated by the deep learning data augmentation model maintain the combination relationship between the initial detection boundary state and the re-detection boundary state represented by the boundary state transition type. The transition state error does not change the parameter update objects of the normalized distance error and the difference relationship error; it is used to determine whether the model training process meets the stopping condition.

[0085] The error threshold, transfer error limit, and upper limit of training epochs are all determined according to the method described in Example 1. When the training error is less than or equal to the error threshold and the transfer error is less than or equal to the transfer error limit, the electronic device stops model training and solidifies the model parameters that currently meet the stopping conditions, thus obtaining a deep learning data augmentation model. When the number of training epochs reaches the upper limit of training epochs, the electronic device stops model training and selects the model parameters with the smallest training error from the completed training epochs for solidification, thus obtaining a deep learning data augmentation model. Through the above training process, the deep learning data augmentation model can be constrained simultaneously by the decision boundary position, the difference relationship between the initial and re-inspection pairings, and the combination relationship between the initial and re-inspection boundary states when generating detection values.

[0086] The generation of candidate enhancement pairing data in this embodiment includes the following processes.

[0087] The electronic device extracts the target boundary state migration type and target determination consistency flag from the boundary state migration data, and forms a target migration combination by combining the target boundary state migration type and target determination consistency flag. Following the enhancement quantity determination method corresponding to the target migration combination in Embodiment 1, the electronic device allocates the enhancement quantity corresponding to the target migration combination from the total number of target enhancements based on the number of re-examination pairing data corresponding to the target migration combination. The enhancement quantity corresponding to the target migration combination is used to limit the number of enhancement pairing data that needs to be retained under the target migration combination.

[0088] The electronic device selects target re-inspection pairing data from the corresponding re-inspection pairing data according to the target migration combination. The electronic device obtains the target judgment boundary and boundary equal width from the boundary state migration data corresponding to the target re-inspection pairing data, and obtains the re-inspection width, re-inspection trigger range, qualified side range, and unqualified side range from the re-inspection range data based on the detection item identifier corresponding to the target re-inspection pairing data. The electronic device determines the judgment direction coefficient based on the position of the qualified side range relative to the target judgment boundary.

[0089] The electronic device generates random input values ​​and inputs the boundary state transition type code corresponding to the target boundary state transition type, the decision consistency mark code corresponding to the target decision consistency mark, and the random input values ​​into the deep learning data augmentation model. The deep learning data augmentation model outputs the candidate initial detection normalized boundary distance and the candidate re-detection normalized boundary distance. The electronic device converts the candidate initial detection normalized boundary distance and the candidate re-detection normalized boundary distance into candidate augmented initial detection values ​​and candidate augmented re-detection values ​​based on the target decision boundary, re-detection width, boundary equal width, and decision direction coefficient. The candidate augmented initial detection values ​​and candidate augmented re-detection values ​​together form candidate augmentation paired data.

[0090] The candidate enhancement pairing data retention process in this embodiment includes the following processes.

[0091] The electronic device re-identifies the boundary state transition of candidate enhanced initial inspection values ​​and candidate enhanced re-inspection values ​​based on the re-inspection range data. The electronic device determines the candidate initial inspection boundary distance based on the distance between the candidate enhanced initial inspection value and the target judgment boundary, and determines the candidate re-inspection boundary distance based on the distance between the candidate enhanced re-inspection value and the target judgment boundary. The electronic device determines the candidate initial inspection boundary state and the candidate re-inspection boundary state based on the candidate initial inspection boundary distance, the candidate re-inspection boundary distance, the boundary equal width, the qualified side range, and the unqualified side range. The electronic device combines the candidate initial inspection boundary state and the candidate re-inspection boundary state to obtain the candidate boundary state transition type.

[0092] The electronic equipment determines the candidate initial inspection result based on the candidate enhanced initial inspection value, the target judgment boundary, the acceptable range, and the unacceptable range. The electronic equipment determines the candidate re-inspection result based on the candidate enhanced re-inspection value, the target judgment boundary, the acceptable range, and the unacceptable range. The electronic equipment obtains a candidate judgment consistency mark based on whether the candidate initial inspection result and the candidate re-inspection result are consistent.

[0093] The electronic device compares the initial detection value of candidate enhancements with the re-detection trigger range, compares the candidate boundary state transition type with the target boundary state transition type, and compares the candidate judgment consistency flag with the target judgment consistency flag. If the initial detection value of a candidate enhancement is within the re-detection trigger range, the candidate boundary state transition type is consistent with the target boundary state transition type, and the candidate judgment consistency flag is consistent with the target judgment consistency flag, the electronic device retains the candidate enhancement pairing data, thus obtaining enhancement pairing data. If the initial detection value of a candidate enhancement is outside the re-detection trigger range, or the candidate boundary state transition type is inconsistent with the target boundary state transition type, or the candidate judgment consistency flag is inconsistent with the target judgment consistency flag, the electronic device does not retain the candidate enhancement pairing data.

[0094] The electronic device repeatedly performs candidate enhancement pair data generation and retention processes for the same target migration combination until the number of enhancement pair data corresponding to the target migration combination reaches the number of enhancements corresponding to the target migration combination, or the number of times candidate enhancement pair data is generated reaches the upper limit of the generation count. The upper limit of the generation count is obtained according to the method for determining the upper limit of the generation count in Embodiment 1. For target migration combinations that have not reached the number of enhancements after reaching the upper limit of the generation count, the electronic device uses the number of enhancement pair data that has been retained as the actual number of enhancements. The electronic device uses the actually retained enhancement pair data as the input data for the merging process in Embodiment 1, and obtains enhancement evaluation data through the merging process in Embodiment 1.

[0095] In this embodiment, the deep learning data augmentation model does not generate a single detection record based solely on the overall distribution of the original evaluation data. Instead, it generates pairs of candidate initial augmentation detection values ​​and candidate re-assertion detection values ​​under the constraints of boundary state transition type and decision consistency marker. The candidate augmentation pairing data then needs to undergo boundary state transition identification again according to the re-assertion range data. After preserving the consistency of the re-assertion trigger range, boundary state transition type, and decision consistency marker, it can be included in the augmentation evaluation data as augmentation pairing data. Therefore, while expanding the amount of small-sample precision evaluation data, the augmentation pairing data continues to conform to the re-assertion pairing processing rules, the combination relationship between the initial and re-assertion boundary states, and the decision consistency relationship, reducing the possibility of augmentation data disrupting the pairing relationship between the initial and re-assertion records. Example

[0096] Please see Figure 2 This embodiment provides a deep learning-based small-sample precision evaluation data augmentation system to implement a deep learning-based small-sample precision evaluation data augmentation method. The deep learning-based small-sample precision evaluation data augmentation system includes a data pairing processing module, a boundary state transition recognition module, a model training processing module, a pairing augmentation processing module, a data merging processing module, and a precision evaluation processing module.

[0097] The data pairing processing module is used to acquire the original evaluation data and re-inspection range data, and to perform re-inspection pairing processing on the original evaluation data based on the re-inspection range data to obtain re-inspection paired data. The original evaluation data includes initial inspection records and re-inspection records. Each inspection record includes at least the inspection object identifier, inspection item identifier, inspection batch identifier, inspection value, inspection time, record type, and judgment result. The re-inspection range data includes the judgment boundary, re-inspection width, re-inspection trigger range, qualified side range, and unqualified side range.

[0098] The data pairing processing module is also used to match the judgment boundary, re-inspection width, re-inspection trigger range, qualified side range, and unqualified side range with the test records in the original evaluation data according to the test item identifier; organize the original evaluation data according to the test object identifier, test item identifier, and test batch identifier, and distinguish between initial inspection records and re-inspection records according to the record type; determine the test value in the initial inspection record as the initial inspection value, the test value in the re-inspection record as the re-inspection value, the judgment result in the initial inspection record as the initial inspection judgment result, and the judgment result in the re-inspection record as the re-inspection judgment result; filter the initial inspection records according to the re-inspection trigger range, and determine the initial inspection records whose initial inspection values ​​are within the re-inspection trigger range as the initial inspection records to be paired; establish a pairing relationship between the initial inspection records to be paired and the re-inspection records according to the consistency of the test object identifier, the test item identifier, the test batch identifier, and the test time of the re-inspection record being later than the test time of the initial inspection records to be paired, to obtain the re-inspection pairing data.

[0099] The re-inspection pairing data includes the detection object identifier, detection item identifier, detection batch identifier, detection value recording precision, initial inspection value, re-inspection value, initial inspection judgment result, re-inspection judgment result, initial inspection time, re-inspection time, direction of the difference between the initial inspection value and the re-inspection value, and magnitude of the difference between the initial inspection value and the re-inspection value. When the initial inspection record and the re-inspection record in the same re-inspection pairing data correspond to different detection value recording precisions, the data pairing processing module determines the maximum value between the detection value recording precision corresponding to the initial inspection record and the detection value recording precision corresponding to the re-inspection pairing data as the detection value recording precision corresponding to the re-inspection pairing data.

[0100] The boundary state migration identification module is used to identify boundary state migrations in the re-inspection paired data based on the re-inspection range data, thereby obtaining boundary state migration data. The module is also used to extract the corresponding judgment boundary, re-inspection trigger range, qualified side range, and unqualified side range from the re-inspection range data based on the detection item identifiers in the re-inspection paired data. For a single judgment boundary, the single judgment boundary is determined as the target judgment boundary. For a two-sided judgment scenario, the target judgment boundary is determined based on the distances between the initial inspection value and the re-inspection value and the lower and upper limit judgment boundaries, respectively.

[0101] The boundary state migration identification module is also used to determine the initial inspection boundary distance and the re-inspection boundary distance based on the target judgment boundary; determine the boundary equivalent width based on the maximum value of the acquired items among the detection value recording accuracy, minimum measurement division of the detection item, and reading resolution of the detection equipment corresponding to the re-inspection paired data; determine the initial inspection boundary state and the re-inspection boundary state respectively based on the initial inspection boundary distance, the re-inspection boundary distance, the qualified side range, the unqualified side range, and the boundary equivalent width; combine the initial inspection boundary state and the re-inspection boundary state in the same re-inspection paired data to obtain the boundary state migration type; determine the judgment consistency mark based on whether the initial inspection judgment result and the re-inspection judgment result are consistent; and associate the target judgment boundary, the boundary equivalent width, the initial inspection boundary state, the re-inspection boundary state, the boundary state migration type, the judgment consistency mark, the initial inspection boundary distance, the re-inspection boundary distance, the direction of the difference between the initial inspection value and the re-inspection value, and the magnitude of the difference between the initial inspection value and the re-inspection value to obtain the boundary state migration data.

[0102] The model training processing module is used to train the model on the re-inspection paired data and boundary state transition data to obtain a deep learning data augmentation model. The module also forms a training sample set based on the re-inspection paired data and boundary state transition data; obtains the re-inspection width from the re-inspection range data based on the detection item identifier corresponding to the training sample, and obtains the boundary equivalent width based on the boundary state transition data corresponding to the training sample; determines the judgment direction coefficient based on the position of the qualified side range corresponding to the training sample relative to the target judgment boundary; and performs scaling on the initial inspection value and the re-inspection value in the training sample set based on the target judgment boundary, the re-inspection width, the boundary equivalent width, and the judgment direction coefficient to obtain the initial inspection normalized boundary distance and the re-inspection normalized boundary distance.

[0103] The model training module also uses boundary state transition types, decision consistency labels, and random input values ​​as input features for the deep learning data augmentation model, and the initial and re-inspection normalized boundary distances as training labels to train the model. The boundary state transition types are converted to boundary state transition type codes using one-hot encoding, and the decision consistency labels are converted to decision consistency label codes using binary encoding. The deep learning data augmentation model is implemented using a multi-layer feedforward neural network. Based on the boundary state transition type codes, decision consistency label codes, and random input values, the deep learning data augmentation model outputs the initial and re-inspection normalized boundary distances.

[0104] The model training processing module is also used to obtain generated initial detection values ​​and generated re-detection values ​​based on the generated initial detection normalized boundary distance, generated re-detection normalized boundary distance, target judgment boundary, re-detection width, boundary equal width, and judgment direction coefficient; to determine the normalization distance error based on the difference between the generated initial detection normalized boundary distance and the initial detection normalized boundary distance in the training samples, and the difference between the generated re-detection normalized boundary distance and the re-detection normalized boundary distance in the training samples; to determine the difference relationship error based on the generation normalization difference between the generated initial detection value and the generated re-detection value, and the training normalization difference between the initial detection value and the re-detection value in the training samples; to determine the training error based on the normalized distance error and the difference relationship error, and to iteratively update the model parameters of the deep learning data augmentation model based on the training error.

[0105] The model training module is also used to, after each round of model training, redetermine the initial detection boundary state and the re-detection boundary state based on the generated initial detection value, generated re-detection value, target judgment boundary, boundary equal width, qualified side range, and unqualified side range, and determine the generation boundary state transition type based on the generated initial detection boundary state and the generated re-detection boundary state; determine the transition state error based on whether the generation boundary state transition type is consistent with the boundary state transition type in the input features; compare the training error with the error threshold, and compare the transition state error with the transition error limit; when the training error is less than or equal to the error threshold, and the transition state error is less than or equal to the transition error limit, stop the model training process and solidify the model parameters that currently meet the stopping conditions to obtain the deep learning data augmentation model; when the number of model training rounds reaches the upper limit of the number of training rounds, stop the model training process, and select the model parameters with the smallest training error from the completed model training rounds for solidification to obtain the deep learning data augmentation model.

[0106] The pairing enhancement module is used to enhance the model based on boundary state transition data and deep learning data, performing pairing enhancement processing on the re-examination pairing data to obtain enhanced pairing data. The pairing enhancement module also extracts the target boundary state transition type and target decision consistency marker from the boundary state transition data, forming target transition combinations from these two types. When the number of re-examination pairing data is less than the minimum number of pairs required for precision evaluation, the pairing enhancement module determines the total target enhancement amount as the difference between the minimum number of pairs required for precision evaluation and the number of re-examination pairing data. When the number of re-examination pairing data is greater than or equal to the minimum number of pairs required for precision evaluation, the pairing enhancement module sets the total target enhancement amount to zero. The total target enhancement amount is allocated according to the proportion of each target transition combination in the boundary state transition data to obtain the enhancement amount corresponding to each target transition combination.

[0107] The pairing enhancement processing module is also used to select target re-inspection pairing data from the corresponding re-inspection pairing data according to the target migration combination; obtain the target judgment boundary and boundary equal width from the boundary state migration data corresponding to the target re-inspection pairing data; obtain the re-inspection width, re-inspection trigger range, qualified side range, and unqualified side range from the re-inspection range data according to the detection item identifier corresponding to the target re-inspection pairing data; determine the judgment direction coefficient according to the position of the qualified side range relative to the target judgment boundary; generate random input values, and input the boundary state migration type code corresponding to the target boundary state migration type, the judgment consistency mark code corresponding to the target judgment consistency mark, and the random input values ​​into the deep learning data augmentation model to obtain the candidate initial inspection normalized boundary distance and the candidate re-inspection normalized boundary distance; convert the candidate initial inspection normalized boundary distance and the candidate re-inspection normalized boundary distance into candidate enhanced initial inspection value and candidate enhanced re-inspection value according to the target judgment boundary, re-inspection width, boundary equal width, and judgment direction coefficient to obtain candidate enhanced pairing data.

[0108] The pairing enhancement processing module is also used to re-identify the boundary state migration of candidate enhancement initial inspection values ​​and candidate enhancement re-inspection values ​​based on the re-inspection range data to obtain candidate boundary state migration types; determine candidate initial inspection judgment results based on candidate enhancement initial inspection values, target judgment boundaries, qualified side range, and unqualified side range, and determine candidate re-inspection judgment results based on candidate enhancement re-inspection values, target judgment boundaries, qualified side range, and unqualified side range; obtain candidate judgment consistency markers based on whether candidate initial inspection judgment results and candidate re-inspection judgment results are consistent; when candidate enhancement initial inspection values ​​are within the re-inspection trigger range, candidate boundary state migration types are consistent with target boundary state migration types, and candidate judgment consistency markers are consistent with target judgment consistency markers, candidate enhancement initial inspection values ​​and candidate enhancement re-inspection values ​​are retained to obtain enhancement pairing data; when candidate enhancement initial inspection values ​​are outside the re-inspection trigger range, or candidate boundary state migration types are inconsistent with target boundary state migration types, or candidate judgment consistency markers are inconsistent with target judgment consistency markers, candidate enhancement initial inspection values ​​and candidate enhancement re-inspection values ​​are not retained.

[0109] The data merging module is used to merge the enhanced paired data and the original evaluation data to obtain enhanced evaluation data. This module also converts the enhanced initial inspection values ​​in the enhanced paired data into enhanced initial inspection records, and the enhanced re-inspection values ​​into enhanced re-inspection records; it configures the same enhanced paired identifier for the enhanced initial inspection records and enhanced re-inspection records corresponding to the same enhanced paired data; it unifies the data items of the enhanced initial inspection records, enhanced re-inspection records, and original evaluation data, including the detection object identifier, detection item identifier, detection batch identifier, detection value, detection time, record type, and judgment result; it pairs the enhanced initial inspection records and enhanced re-inspection records according to the enhanced paired identifier; and it merges the paired enhanced initial inspection records and enhanced re-inspection records with the original evaluation data to obtain the enhanced evaluation data.

[0110] The precision assessment processing module is used to perform boundary state migration statistical processing on the enhancement assessment data based on the re-inspection range data to obtain the precision assessment results. The precision assessment processing module is also used to extract re-inspection paired records from the enhancement assessment data; for the detection records in the original assessment data, re-inspection paired records are formed according to the re-inspection paired processing rules in the data paired processing module; for the enhancement initial inspection records and enhancement re-inspection records, enhancement re-inspection paired records are formed based on the enhancement paired identifier; based on the re-inspection range data, the initial inspection boundary state, re-inspection boundary state, and boundary state migration type are redefined for each re-inspection paired record; and the re-inspection paired records are grouped and statistically analyzed according to the boundary state migration type to obtain boundary state migration group data.

[0111] The precision evaluation processing module is also used to count the number of consistent and inconsistent decisions for each boundary state migration type based on the consistency marker, and to count the consistency and inconsistency ratios for all re-examination pairing records; it determines the discrepancy of the re-examination pairing difference based on the boundary state migration grouping data; when the precision evaluation processing module generates the discrepancy of the re-examination pairing difference, the precision evaluation results include the discrepancy of the re-examination pairing difference, the boundary state migration ratio corresponding to each boundary state migration type, the number of consistent decisions corresponding to each boundary state migration type, the number of inconsistencies corresponding to each boundary state migration type, the consistency ratio, and the inconsistency ratio; when the precision evaluation processing module records the discrepancy of the re-examination pairing difference as empty, the precision evaluation results include the boundary state migration ratio corresponding to each boundary state migration type, the number of consistent decisions corresponding to each boundary state migration type, the number of inconsistencies corresponding to each boundary state migration type, the consistency ratio, and the inconsistency ratio.

[0112] In some embodiments, the deep learning-based few-sample precision evaluation data augmentation system can be implemented by a processor and a memory. The memory stores a computer program, and when the processor executes the computer program, it implements the functions corresponding to the data pairing processing module, boundary state transition recognition module, model training processing module, pairing augmentation processing module, data merging processing module, and precision evaluation processing module. The data pairing processing module, boundary state transition recognition module, model training processing module, pairing augmentation processing module, data merging processing module, and precision evaluation processing module can be implemented as software programs, hardware circuits, or a combination of software programs and hardware circuits.

[0113] The deep learning-based small-sample precision assessment data augmentation system in this embodiment can generate re-examination paired data through a data pairing processing module, generate boundary state transition data through a boundary state transition recognition module, obtain a deep learning data augmentation model through a model training processing module, obtain augmented paired data through a pairing augmentation processing module, obtain augmented assessment data through a data merging processing module, and obtain precision assessment results through a precision assessment processing module. Therefore, it can expand the amount of small-sample precision assessment data while maintaining the pairing relationship between the initial and re-examination records, the combination relationship between the initial and re-examination boundary states, and the consistency between the initial and re-examination judgment results.

[0114] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A data augmentation method for small-sample precision evaluation based on deep learning, characterized in that, include: Obtain the original assessment data and the re-inspection scope data, and perform re-inspection pairing processing on the original assessment data based on the re-inspection scope data to obtain the re-inspection pairing data; Based on the re-inspection range data, boundary state migration identification is performed on the re-inspection paired data to obtain boundary state migration data; The re-examination pairing data and boundary state transition data are processed to train the model, resulting in a deep learning data augmentation model. Based on boundary state transition data and deep learning data augmentation models, pairing enhancement processing is performed on the re-examination pairing data to obtain enhanced pairing data. The enhanced paired data and the original evaluation data are merged to obtain the enhanced evaluation data. Based on the re-inspection range data, boundary state migration statistical processing is performed on the enhanced assessment data to obtain the precision assessment results.

2. The method for small-sample precision evaluation data augmentation based on deep learning according to claim 1, characterized in that, The method for obtaining the re-examination and pairing data includes: Extract the detection object identifier, detection item identifier, detection batch identifier, record type, detection value, detection time, and judgment result from the original assessment data; Based on the record type, the original assessment data is divided into preliminary inspection records and re-inspection records. Preliminary inspection values ​​and preliminary inspection judgment results are extracted from the preliminary inspection records, and re-inspection values ​​and re-inspection judgment results are extracted from the re-inspection records. Extract the re-inspection trigger range corresponding to the test item identifier from the re-inspection range data; Initial inspection records whose initial inspection values ​​fall within the re-inspection trigger range are identified as initial inspection records to be paired. The initial inspection records and re-inspection records to be paired are matched for consistency of identifiers and for the order of detection time to obtain the paired re-inspection records; Based on the initial inspection record to be matched and the re-inspection record of the matched pairs, the re-inspection matching data is obtained.

3. The method for small-sample precision evaluation data augmentation based on deep learning according to claim 2, characterized in that, The method for obtaining paired re-examination records includes: Extract the test object identifier, test item identifier, and test batch identifier from the initial inspection record and re-inspection record to be matched; Re-inspection records with consistent test object identification, consistent test item identification, and consistent test batch identification are identified as candidate re-inspection records; From the candidate re-examination records, retain the candidate re-examination records whose testing time is later than the initial test record to be matched, and obtain the valid candidate re-examination records; The detection time intervals between valid candidate re-examination records and the initial examination records to be paired are sorted to obtain the time interval sorting results; Based on the sorting results of the time intervals, the valid candidate re-examination record with the smallest detection time interval is determined as the paired re-examination record.

4. The method for small-sample precision evaluation data augmentation based on deep learning according to claim 1, characterized in that, The method for obtaining boundary state transition data includes: Extract the initial test value, retest value, initial test judgment result, retest judgment result, and test item identifier from the retest paired data; Extract the judgment boundary, qualified side range, and unqualified side range corresponding to the test item identifier from the re-inspection range data; Based on the judgment boundary, the initial inspection value and the re-inspection value are subjected to boundary selection processing to obtain the target judgment boundary; Based on the target determination boundary, the boundary distance between the initial inspection value and the re-inspection value is calculated to obtain the initial inspection boundary distance and the re-inspection boundary distance. State transition identification is performed based on the initial inspection boundary distance, re-inspection boundary distance, qualified side range, unqualified side range, initial inspection judgment result, and re-inspection judgment result to obtain the state transition identification result. By associating the state transition identification results with the re-examination pairing data, boundary state transition data is obtained.

5. The method for small-sample precision evaluation data augmentation based on deep learning according to claim 4, characterized in that, The method for obtaining the target determination boundary includes: When the decision boundary is a single decision boundary, the single decision boundary is determined as the target decision boundary; When the decision boundary includes both the lower and upper decision boundaries, the distances between the initial and re-inspection values ​​and the lower decision boundary are calculated and summed to obtain the lower limit distance sum. The distances between the initial inspection value and the re-inspection value and the upper limit judgment boundary are calculated and summed to obtain the upper limit distance sum. When the sum of the lower limit distances is less than or equal to the sum of the upper limit distances, the lower limit judgment boundary is determined as the target judgment boundary; When the sum of the lower limit distances is greater than the sum of the upper limit distances, the upper limit judgment boundary is determined as the target judgment boundary.

6. The method for small-sample precision evaluation data augmentation based on deep learning according to claim 4, characterized in that, The method for obtaining the state transition recognition result includes: The accuracy of the test value record is obtained from the re-inspection and pairing data, and at least one of the minimum measurement division of the test item and the reading resolution of the test equipment is obtained. The maximum value of the obtained items is determined as the boundary width. Based on the equal width of the boundary, the qualified side range, and the unqualified side range, the initial inspection boundary distance and the re-inspection boundary distance are divided into states to obtain the initial inspection boundary state and the re-inspection boundary state. By combining the initial boundary state and the re-inspection boundary state, the boundary state transition type is obtained; By comparing the initial inspection results and the re-inspection results, a consistency marker is obtained. The boundary width, boundary state transition type, and decision consistency marker are used as the state transition recognition results.

7. The method for small-sample precision evaluation data augmentation based on deep learning according to claim 6, characterized in that, The method for obtaining the deep learning data augmentation model includes: A training sample set is constructed based on the re-examination pairing data and boundary state transition data; Encode the boundary state transition type and decision consistency marker as conditional input features; The re-inspection width and judgment direction coefficient are obtained based on the re-inspection range data; Based on the target judgment boundary, re-inspection width, boundary equal width and judgment direction coefficient, the initial inspection value and re-inspection value are scaled to obtain the normalized boundary distance training label; Input the conditional input features and random input values ​​into the model to be trained to obtain the initial normalized boundary distance and the re-normalized boundary distance. The training error is calculated based on the generated initial normalized boundary distance, the generated re-normalized boundary distance, and the normalized boundary distance training labels, and the transfer state is verified to obtain the transfer state error. The deep learning data augmentation model is obtained by updating and stopping the training based on the training error and the transfer state error.

8. The method for small-sample precision evaluation data augmentation based on deep learning according to claim 7, characterized in that, The method for calculating training error includes: The error is calculated between the generated initial detection normalized boundary distance and the initial detection normalized boundary distance in the training sample set, and the error is calculated between the generated re-detection normalized boundary distance and the re-detection normalized boundary distance in the training sample set to obtain the normalized distance error. Based on the target judgment boundary, re-inspection width, boundary equal width and judgment direction coefficient, the detection values ​​of the generated initial inspection normalized boundary distance and the generated re-inspection normalized boundary distance are converted to obtain the generated initial inspection value and the generated re-inspection value. The normalized difference between the generated initial inspection value and the generated re-inspection value is calculated to obtain the generated normalized difference. The normalized difference is calculated by taking the initial and retest values ​​in the training sample set to obtain the training normalized difference. The error between the generated normalized difference and the training normalized difference is calculated to obtain the difference relationship error. The training error is obtained by weighted summation of the normalized distance error and the difference relationship error.

9. The method for small-sample precision evaluation data augmentation based on deep learning according to claim 7, characterized in that, The method for obtaining enhanced pairing data includes: Extract the target boundary state migration type and target determination consistency flag from the boundary state migration data to obtain the target migration combination; Select the target re-examination pairing data corresponding to the target migration combination; Based on the boundary state migration data and re-inspection range data of the target re-inspection pairing data, the detection value transformation parameters and re-inspection trigger range are obtained; By combining the target transfer and random input values ​​and inputting them into the deep learning data augmentation model, we can obtain the candidate initial detection normalized boundary distance and the candidate re-detection normalized boundary distance. Based on the conversion parameters of the detection values, the candidate initial detection normalized boundary distance and the candidate re-detection normalized boundary distance are converted to obtain candidate enhancement pairing data; The candidate enhanced pairing data are re-identified by migration combination to obtain candidate migration combinations; When the initial detection value of the candidate enhancement is within the range of the re-detection trigger and the candidate migration combination is consistent with the target migration combination, the enhancement pairing data is retained.

10. A deep learning-based small-sample precision evaluation data augmentation system, used to implement the deep learning-based small-sample precision evaluation data augmentation method according to any one of claims 1-9, characterized in that, include: The data pairing processing module is used to acquire the original evaluation data and the re-inspection range data, and to perform re-inspection pairing processing on the original evaluation data based on the re-inspection range data to obtain the re-inspection paired data. The boundary state migration identification module is used to identify the boundary state migration of the re-inspection paired data based on the re-inspection range data, and obtain the boundary state migration data. The model training and processing module is used to train and process the re-examination pairing data and boundary state transition data to obtain a deep learning data augmentation model. The pairing enhancement processing module is used to enhance the model based on boundary state transfer data and deep learning data, and to perform pairing enhancement processing on the re-examination pairing data to obtain enhanced pairing data; The data merging and processing module is used to merge the enhanced paired data and the original evaluation data to obtain the enhanced evaluation data; The precision assessment processing module is used to perform boundary state migration statistical processing on the enhanced assessment data based on the re-inspection range data to obtain the precision assessment results.

Images