Metering detection data automatic identification method based on deep learning

Abstract

The invention relates to the technical field of measurement data, in particular to a measurement detection data automatic identification method based on deep learning, and the method comprises the steps: marking a distortion section through a signal intensity score, timely finding signal distortion in a data collection stage, and carrying out the correction and reconstruction of a distortion signal through wavelet transform. The accuracy of original data is improved, abnormal modes in the data are accurately mined through the generative adversarial network, meanwhile, abnormal reasons are associated, a causal report is formed, multivariate regression analysis starts from an error source, the strength of each influence factor is quantified, a clear direction and basis are provided for subsequent data optimization, and the accuracy of the abnormal modes in the data is improved. Through the classification analysis capability of the support vector machine, the corrected data features are gradually compared and accurately classified, and the classification marking result is further layered and sorted through the deep learning model, so that the reliability, accuracy and efficiency of the measurement data in the transmission, storage and analysis process are improved.

Description

Technical Field

[0001] The present invention relates to the field of metrology data technology, and in particular to a method for automatic recognition of metrology detection data based on deep learning. Background Art

[0002] The field of metrology data technology involves the precise measurement of various physical quantities, and the processing, analysis and management of the measured data. The goal is to obtain accurate data through measuring equipment or systems and ensure consistency and reliability in transmission, storage and processing.

[0003] The purpose of the automatic identification method of metrology and testing data based on deep learning is to realize the intelligent identification, classification and processing of complex data in the metrology and testing process. By introducing deep learning technology, it can realize the automatic analysis of large-scale metrology data, improve the accuracy and reliability of detection, and reduce the need for manual intervention, so as to achieve the effect of improving detection efficiency and optimizing quality control processes. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings existing in the prior art and to propose a method for automatic recognition of metrological detection data based on deep learning.

[0005] In order to achieve the above object, the present invention adopts the following technical solution: a method for automatic identification of metrological detection data based on deep learning, comprising the following steps:

[0006] Step 1: Receive raw data through the sensor, score the signal strength, mark and record the distorted sections, and generate a signal quality assessment report;

[0007] Step 2: According to the signal quality assessment report, wavelet transform is used to correct the marked segment and the distorted signal to obtain reconstructed signal data;

[0008] Step 3: Structural decomposition is performed on the reconstructed signal data, abnormal patterns and abnormal causes are identified through a generative adversarial network, and the generative adversarial network is continuously trained to obtain a deep learning model, and the relationship between the abnormal pattern and the abnormal cause is recorded to generate a report on the abnormal pattern and causal association;

[0009] Step 4: Using the abnormal pattern and causal association report, calculate the influence intensity of each factor through multiple regression analysis, and evaluate the error source at the same time to obtain the quantitative analysis result of the error source;

[0010] Step 5: According to the quantitative analysis results of the error sources, the original data is corrected by accurately adjusting the parameters, the accuracy and reliability of the original data are optimized, and the inverse corrected data is generated;

[0011] Step six: input the reverse corrected data into the deep learning model, and use a support vector machine to classify the model according to data features, output the classification results, and obtain the measurement data category recognition results.

[0012] As a further solution of the present invention, the specific steps of generating the signal quality assessment report are:

[0013] Based on the raw data received by the sensor, the intensity value of the signal data is extracted one by one, and the intensity value of each signal segment is recorded through segmented scanning detection. The signal strength is quantified and scored according to the preset scoring rules and archived. The start and end positions of all intensity abnormal segments are marked to generate signal scoring and distortion segment marking data;

[0014] Based on the signal score and distorted segment marking data, by traversing all score and marking records, each signal segment is integrated with its corresponding score and marking information, the number and length of the distorted segments are statistically classified, the overall coverage ratio of the distorted signal is calculated by aggregation, and the score statistics and segment marking information are generated;

[0015] Based on the scoring statistical data and the segment marking information, by screening the scoring information of each signal segment, a comprehensive analysis is performed on the coverage and quality distribution of the signal, and by normalizing the distribution ratio and quality score of the distorted signal, a signal quality assessment report is generated;

[0016] The signal quality assessment report includes the specific location of the distorted section, the signal strength level and the noise level.

[0017] As a further solution of the present invention, the specific steps of generating the reconstructed signal data are:

[0018] Based on the signal quality assessment report, filter the distorted segment signal data, segment the distorted signal into fixed lengths by segmenting, analyze and calculate the value variation range of each segment signal, generate a characteristic parameter table of each segment by counting its fluctuation rate and abnormal amplitude, and generate distorted signal segment characteristic data;

[0019] Based on the distorted signal segment characteristic data, the segment signal is processed by traversing the segment by segment, the abnormal jump value appearing in the signal is interpolated and supplemented, the signal change characteristic is corrected by adjusting the continuity of the signal boundary value, and the distribution within the signal is numerically equalized to generate segment signal correction data;

[0020] Based on the segment signal correction data, each segment of the corrected signal is spliced ​​with the normal signal data in the original order, and the data characteristics of the splicing boundary are corrected by point-by-point analysis, the connection values ​​and fluctuation rates between all segments are processed, and the overall consistency between the signal segments is adjusted to generate reconstructed signal data;

[0021] The reconstructed signal data includes a reconstructed waveform, an adjusted frequency response and a signal stability score.

[0022] As a further solution of the present invention, the specific steps of generating the abnormal pattern and causal association report are:

[0023] Based on the reconstructed signal data, the signal data is divided into multiple independent hierarchical data areas according to the time range and the signal amplitude change trend, and the change rate and structural characteristics of the signal are extracted and processed by scanning layer by layer, and the key parameters of the signal are extracted, and the abnormal data part existing in the signal is separated, and the dynamic characteristics separation of the abnormal data is completed to generate signal decomposition feature data;

[0024] Based on the signal decomposition feature data, a generator of a generative adversarial network is used to learn the potential distribution characteristics of the signal decomposition feature data, and the generated data and actual data are compared to capture the abnormal pattern characteristics in the signal distribution, and the abnormal pattern characteristics are extracted as key dynamic parameters, and the abnormal pattern characteristics are matched with the known characteristics of the potential abnormal causes, and the association mapping between the pattern and the cause is completed, and the analysis results are organized into a multi-dimensional parameter set to generate abnormal pattern and cause relationship data;

[0025] Based on the abnormal pattern and cause relationship data, the causal characteristics of the abnormal pattern are modeled using a generative adversarial network, and the causal characteristic distribution is reconstructed, and the consistency between the generated distribution and the actual distribution is evaluated, and the dynamic association path of the causal variable is determined. The variables in the causal path are analyzed item by item using a feature relationship mapping method, and the causal strength between the pattern characteristics and the abnormal causes is quantified. The modeling results are recorded as a causal association data set to obtain a causal relationship model, and an abnormal pattern and causal association report is generated at the same time;

[0026] The abnormal pattern and causal association report includes a list of identified abnormal patterns, related cause analysis and impact assessment.

[0027] As a further solution of the present invention, the data structure is processed by multi-level feature decomposition, feature parameters in each layer of data are extracted, the signal data is preliminarily layered according to time period and frequency range, the data is divided into multiple different feature intervals, the amplitude, change rate and structural parameter extraction operations are performed on the layered data by layer-by-layer scanning, the dynamic characteristics of each layer of data are grouped and aggregated, and the change range and change trend of the fluctuation amplitude are extracted at the same time, and the decomposed dynamic characteristics are compared with the abnormal parameters layer by layer, the change range of the abnormal characteristics is recorded, and the abnormal signal content with dynamic characteristics is separated;

[0028] The method identifies pattern features in abnormal signals, matches pattern features with potential abnormal causes, extracts dynamic characteristics related to time series changes in data, screens the characteristic distribution of data item by item by comparing amplitude differences, periodic changes and signal mutation points at different time points and frequency ranges, finds the abnormal signal characteristic range, further analyzes the fluctuation pattern, extracts repeated change cycles, peak spacing and change rules, and matches them to an existing set of characteristic parameters, compares the change patterns of potential abnormal causes, and records the specific association information between the pattern and the cause.

[0029] As a further solution of the present invention, the specific steps of generating the quantitative analysis result of the error source are:

[0030] Based on the abnormal pattern and causal association report, extract the causal factor information in the report, perform weight analysis of multivariate relationships, decompose the association strength and weight parameters between multiple variables, and generate causal influence weight distribution results;

[0031] Based on the causal influence weight distribution results, the error data in the variables are extracted, the statistical analysis of the error deviation is performed, the error source is evaluated by parameter difference decomposition, and the statistical results of the error source are generated;

[0032] Based on the statistical results of the error sources, the error sources are quantitatively decomposed, the error distribution characteristics of the causal variables are extracted, and the quantitative evaluation is completed in combination with the error weight calculation to generate the quantitative analysis results of the error sources;

[0033] The quantitative analysis results of the error sources include a list of major error sources, the contribution rate of each source, and environmental factors that affect the error.

[0034] As a further solution of the present invention, the specific steps of generating the reverse corrected data are:

[0035] Based on the quantitative analysis results of the error sources, specific parameters in the error distribution are extracted, each parameter is grouped and analyzed according to the deviation range, the deviation value of each group of parameters is calculated, the data range within the group is adjusted, and a correction deviation parameter set is generated;

[0036] Based on the correction deviation parameter set, each parameter in the original data is corrected one by one, the data is replaced by the correction value in the grouping range, and the deviation value is corrected by accumulating item by item to generate the corrected original data;

[0037] Based on the corrected original data, extracting the uncorrected residual outliers in the data, performing out-of-range deviation assessment on the outliers, gradually adjusting the deviation parameters of the outliers, and generating inversely corrected data;

[0038] The reverse corrected data includes adjusted data parameters, data comparison before and after optimization, and correction effect evaluation.

[0039] As a further solution of the present invention, the specific steps of generating the measurement data category identification result are:

[0040] Based on the reverse corrected data, the classification feature values ​​in the data are scanned item by item, the feature range of each data item is extracted, and a tag value is added to the data according to the classification standard to generate a data classification feature set;

[0041] Based on the data classification feature set, a support vector machine is used to compare and analyze the reverse corrected data with the classification feature value, and the classification mark of each data item is gradually confirmed. The data is classified and sorted based on the marking result to generate a data classification marking result;

[0042] Based on the data classification and labeling results, each classification label is input into the deep learning model, and the classified output is hierarchically sorted in combination with the input data to complete the extraction of the classified measurement data and generate the measurement data category recognition result;

[0043] The measurement data category recognition result includes the classified data type, the classification accuracy and the decision basis of the model judgment.

[0044] As a further solution of the present invention, the support vector machine is according to the formula: , in: To measure the classification and labeling results of the test data, is the feature vector of the metrological detection data to be identified, is the feature vector of the trained support vector in the support vector machine, is the Lagrange multiplier weight of the support vector, is the category label corresponding to the support vector, is the Gaussian kernel function, is the scale parameter of the Gaussian kernel, To dynamically adjust parameters, is the feature weight vector, is the feature weighted result of the linear part of the kernel function, is the bias term of the model, is the regularization term, is the regularization coefficient.

[0045] As a further solution of the present invention, the classification feature values ​​in the data are scanned item by item to extract the feature range of each data. The classification feature values ​​are attributes and indicators in the data used to distinguish categories. The extraction step is specifically to extract numerical and categorical attributes from the inversely corrected data according to predefined classification rules. For numerical attributes, the numerical values ​​are directly extracted. For categorical attributes, dictionary mapping is used to convert them into numerical codes. For missing values, interpolation and default values ​​are used to fill them. Finally, the feature values ​​are sorted and classified.

[0046] Compared with the prior art, the advantages and positive effects of the present invention are:

[0047] 1. In the present invention, the distortion section is marked by signal strength scoring, signal distortion is discovered in time during the data collection stage, and the distorted signal is corrected and reconstructed by wavelet transform, so that the accuracy of the original data is improved;

[0048] 2. In this invention, the generative adversarial network is used to accurately mine abnormal patterns in the data, and the abnormal causes are associated to form a causal report. The multivariate regression analysis starts from the source of error, quantifies the strength of each influencing factor, and provides a clear direction and basis for subsequent data optimization;

[0049] 3. In the present invention, the corrected data features are gradually compared and accurately classified through the classification analysis capability of the support vector machine, so that the classification results are more consistent with the regular characteristics of the data itself, and the reliability and accuracy of the classification are improved. The classification labeling results are further layered and sorted through the deep learning model, so that the classified output data is more hierarchical and organized, and the reliability, accuracy and efficiency of the measurement data in the transmission, storage and analysis process are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the main steps of the present invention;

[0051] Figure 2 It is a schematic diagram of the refinement of S1 of the present invention;

[0052] Figure 3 It is a schematic diagram of the refinement of S2 of the present invention;

[0053] Figure 4 It is a schematic diagram of the refinement of S3 of the present invention;

[0054] Figure 5 It is a schematic diagram of the refinement of S4 of the present invention;

[0055] Figure 6 It is a detailed schematic diagram of S5 of the present invention;

[0056] Figure 7 It is a detailed schematic diagram of S6 of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0058] See also Figure 1 The present invention provides a technical solution: a method for automatic identification of metrological detection data based on deep learning, comprising the following steps:

[0059] S1: Receives raw data from the sensor, scores the signal strength, marks and records the distorted sections, and generates a signal quality assessment report;

[0060] S2: According to the signal quality assessment report, wavelet transform is used to correct the marked segment and the distorted signal to obtain the reconstructed signal data;

[0061] S3: Perform structural decomposition on the reconstructed signal data, identify abnormal patterns and abnormal causes through generative adversarial networks, continuously train the generative adversarial networks to obtain a deep learning model, record the relationship between abnormal patterns and abnormal causes, and generate abnormal pattern and causal association reports;

[0062] S4: Using abnormal patterns and causal association reports, through multiple regression analysis, calculate the impact intensity of each factor, and evaluate the error sources at the same time to obtain quantitative analysis results of error sources;

[0063] S5: Based on the quantitative analysis results of the error sources, the original data is corrected by precisely adjusting the parameters to optimize the accuracy and reliability of the original data and generate the reverse corrected data;

[0064] S6: Input the reverse corrected data into the deep learning model, and use the support vector machine to classify the model according to the data characteristics, output the classification results, and obtain the measurement data category recognition results.

[0065] See also Figure 2 , the specific steps to generate a signal quality assessment report are:

[0066] S101: Based on the raw data received by the sensor, the strength values ​​of the signal data are extracted one by one, the strength value of each signal segment is recorded by segmented scanning detection, the signal strength is quantitatively scored and archived according to the preset scoring rules, the start and end positions of all strength abnormal segments are marked, and the signal score and distortion segment marking data are generated;

[0067] S102: Based on the signal score and the distorted segment marking data, by traversing all the score and marking records, each signal segment is integrated with its corresponding score and marking information, the number and length of the distorted segments are statistically classified, the overall coverage ratio of the distorted signal is calculated by aggregation, and the score statistics and segment marking information are generated;

[0068] S103: Based on the scoring statistical data and the segment marking information, by screening the scoring information of each signal segment, comprehensively analyzing the coverage and quality distribution of the signal, and generating a signal quality assessment report by normalizing the distribution ratio and quality score of the distorted signal;

[0069] Based on the raw data received by the sensor, a segmented scanning detection method is used to divide the signal data into fixed time windows with a window size of 5 milliseconds. The signal strength value is calculated in each signal segment. The sliding window technology is used with a window size of 5 milliseconds and a step size of 1 millisecond to detect the continuity of the signal and save the detection value in array form. By setting a threshold value of 0.2, the strength value of each signal segment is determined and quantified and scored. According to the scoring rules, the score value is archived as an integer value ranging from 0 to 100. The signal segment with a strength value less than the threshold is marked and its start and end positions are recorded to generate signal score and distortion segment marking data.

[0070] Based on the signal score and distortion segment mark data, by traversing all score records and segment marks, the signal score is classified and processed using a statistical classification algorithm, and the score is divided into 5 levels, 0-20, 21-40, 41-60, 61-80, 81-100. The corresponding relationship between the signal score and the distortion segment is mapped, and the segment merging algorithm is used to merge the segments according to the start and end positions. The merging condition is that the interval between the two segments is less than 2 milliseconds. The length and number of the merged segments are counted, and the distortion coverage is defined as the ratio of the total length of the distortion segment to the total length of the signal using an aggregate calculation method. All statistical results are stored in a dictionary form to generate score statistics and segment mark information;

[0071] Based on the scoring statistical data and segment marking information, by screening the numerical distribution range of the signal score, the score value is interval-statisticed using the histogram calculation algorithm, the score value is divided into 10 equally spaced intervals, and the statistical quantity of each interval is stored in a list. The normalized score is defined as the original score value divided by 100 using the normalization method, and the coverage is defined as the ratio of the signal length in a specific scoring range to the total signal length using the distribution ratio calculation method. The comprehensive analysis algorithm is used to define the comprehensive signal quality score as the weighted sum of the coverage of the scoring interval, and the weight is set to the upper limit of the scoring interval divided by 100. The distribution histogram of the signal score and the signal coverage range diagram are generated using the statistical function, and the signal quality assessment report is generated;

[0072] Among them, the signal quality assessment report includes the specific location of the distortion section, signal strength level and noise level.

[0073] See also Figure 3 , the specific steps to generate the reconstructed signal data are:

[0074] S201: based on the signal quality assessment report, filter the distorted segment signal data, segment the distorted signal into fixed lengths by segment-by-segment segmentation, analyze and calculate the value variation range of each segment signal, generate a characteristic parameter table of each segment by counting its fluctuation rate and abnormal amplitude, and generate distorted signal segment characteristic data;

[0075] S202: Based on the distorted signal segment characteristic data, the segment signal is processed by traversing the segment by segment, the abnormal jump value in the signal is interpolated and supplemented, the signal change characteristic is corrected by adjusting the continuity of the signal boundary value, and the distribution in the signal is numerically equalized to generate segment signal correction data;

[0076] S203: Based on the segment signal correction data, each segment of the corrected signal is spliced ​​with the normal signal data in the original order, and the data characteristics of the splicing boundary are corrected by point-by-point analysis, the connection values ​​and fluctuation rates between all segments are processed, and the overall consistency between the signal segments is adjusted to generate reconstructed signal data;

[0077] Based on the signal quality assessment report, filter the distorted segment signal data, use a fixed-length segmentation algorithm to segment the distorted signal segment by segment, set the segmentation length to 500 sampling points, use the segmentation function to divide the signal into segments according to the fixed length and store it in an array, calculate the value range of each segment signal by an analytical algorithm, use the maximum difference calculation formula to define the value range as the difference between the maximum and minimum values ​​in the segment, count the volatility of each segment signal and calculate it according to the formula, the volatility is defined as the ratio of the standard deviation of the value in the segment to the average value, use the abnormal amplitude statistical method to record the deviation value range in the segment, store the statistical value in a two-dimensional characteristic table, and generate the distorted signal segment characteristic data;

[0078] Based on the distorted signal segment characteristic data, the segment signal is traversed segment by segment, and the abnormal jump value in the signal is processed by the interpolation algorithm. The identification of the abnormal jump value is completed by setting the threshold. The threshold is defined as the mean value plus or minus two times the standard deviation. The linear interpolation method is used to supplement the value out of the range. The interpolation calculation formula is defined as the linear weighted sum of adjacent valid points. The signal boundary value continuity adjustment algorithm is used to smooth the start and end positions of the segment. The numerical distribution of the boundary points is adjusted by the convolution algorithm. The convolution window size is 3 sampling points. The internal distribution of the signal is processed by the equalization algorithm. The equalization processing is completed by the segmented standardization formula. The mean value in the segment is set to 0 and the standard deviation is set to 1 to generate the segment signal correction data.

[0079] Based on the segment signal correction data, each segment of the corrected signal is spliced ​​with the normal signal data in the original order, and the point-by-point analysis algorithm is used to detect the data characteristics of the correction splicing boundary. The boundary data characteristic detection is completed through volatility and mean difference analysis. The volatility detection is completed by calculating the variance point by point and comparing it with the neighborhood data variance. The mean difference is completed by calculating the mean difference between the boundary point and its neighborhood point. The connection value is adjusted using the data consistency adjustment method. The adjustment method is the moving average method. The window size is set to 5 sampling points. The overall volatility of the spliced ​​signal is globally detected, the deviation of the cross-segment value is corrected, and the reconstructed signal data is generated.

[0080] The reconstructed signal data includes the reconstructed waveform, the adjusted frequency response and the signal stability score.

[0081] See also Figure 4 ,The specific steps to generate abnormal pattern and causal association report are:

[0082] S301: Based on the reconstructed signal data, the signal data is divided into multiple independent hierarchical data areas according to the time range and the change trend of the signal amplitude, and the change rate and structural characteristics of the signal are extracted and processed by scanning layer by layer, and the key parameters of the signal are extracted. At the same time, the abnormal data part existing in the signal is separated, and the dynamic characteristics of the abnormal data are separated to generate signal decomposition feature data;

[0083] S302: Based on the signal decomposition feature data, the generator of the generative adversarial network is used to learn the potential distribution characteristics of the signal decomposition feature data, compare the generated data with the actual data, capture the abnormal pattern characteristics in the signal distribution, and extract the abnormal pattern characteristics as key dynamic parameters, match the abnormal pattern characteristics with the known characteristics of the potential abnormal causes, complete the association mapping between the pattern and the cause, and organize the analysis results into a multi-dimensional parameter set to generate abnormal pattern and cause relationship data;

[0084] S303: Based on the abnormal pattern and cause relationship data, the causal characteristics of the abnormal pattern are modeled using a generative adversarial network, and the causal characteristic distribution is reconstructed. The consistency between the generated distribution and the actual distribution is evaluated, and the dynamic association path of the causal variable is determined. The variable in the causal path is analyzed item by item using a feature relationship mapping method, and the causal strength between the pattern characteristics and the abnormal cause is quantified. The modeling results are recorded as a causal association data set to obtain a causal relationship model, and a report on the abnormal pattern and causal association is generated at the same time;

[0085] Based on the reconstructed signal data, the signal data is partitioned according to the time range and the signal amplitude change trend using a hierarchical partitioning algorithm. The signal change trend is extracted by setting a sliding window with a time window size of 100 milliseconds. Each partition is aggregated according to the amplitude change direction. The change rate of each signal segment is quantified using the change rate calculation formula. The quantization method is the amplitude difference in each time window divided by the time difference. The frequency characteristics of the signal are extracted using the fast Fourier transform method. The amplitude sequence is converted into a frequency domain signal and the main frequency value and high-frequency component are extracted. The absolute amplitude of the signal change is extracted using the amplitude difference calculation method. The amplitude calculation formula is the sum of the absolute values ​​of the amplitude difference of adjacent sampling points. The abnormal data part is screened by the signal strength comparison method, and the abnormal data is separated from the normal signal. The dynamic change characteristics of the abnormal data are recorded to generate signal decomposition feature data.

[0086] Based on the signal decomposition feature data, the generator of the generative adversarial network is used to learn the potential distribution characteristics of the signal decomposition feature data. The generator adopts a deep convolutional neural network structure, the convolution kernel size is set to 3×3, and the number of channels is set to 64. After generating comparison samples of the feature data, the discriminator is used to discriminate the distribution difference between the generated samples and the actual samples. The discriminator adopts a binary classification structure, and the loss function adopts the cross entropy loss. The feature comparison algorithm is used to capture the abnormal pattern characteristics existing in the distribution of the generated data and the actual data. The feature extraction function is used to extract the abnormal pattern features as key dynamic parameters. The parameters include the signal amplitude range, frequency offset and jump point position. The pattern matching method is used to match the abnormal pattern features with the known characteristics of the potential abnormal causes. The association mapping relationship between the pattern and the cause is constructed through the association matrix. The association results between the pattern and the cause are stored as a multidimensional parameter set to generate abnormal pattern and cause relationship data;

[0087] Based on the abnormal pattern and cause relationship data, the causal characteristics of the abnormal pattern are modeled using a generative adversarial network. The generator uses a bidirectional long short-term memory network structure to capture the signal causal distribution. The number of input layer nodes is set to 256, the number of hidden layer nodes is set to 128, and the time step is set to 10 steps. The mean square error is used as the loss function to optimize the generated distribution. After reconstructing the causal characteristic distribution, the consistency between the generated distribution and the actual distribution is calculated, and the dynamic association path of the causal variable is identified. The variables in the causal path are analyzed item by item using the feature relationship mapping method. The weight normalization algorithm is used to quantify the causal strength between the pattern characteristics and the abnormal causes. The weight normalization method is to divide the weight values ​​of all causal paths by the sum of the total weight values ​​to generate a causal association data set of the modeling results, record the distribution characteristics of the causal variables, obtain the causal relationship model, and generate an abnormal pattern and causal association report at the same time;

[0088] Among them, the abnormal pattern and causal association report includes a list of identified abnormal patterns, relevant cause analysis and impact assessment.

[0089] The data structure is processed by multi-level feature decomposition, and the feature parameters in each layer of data are extracted. The signal data is preliminarily stratified according to time period and frequency range, and the data is divided into multiple different feature intervals. The amplitude, change rate and structural parameters of the stratified data are extracted by layer-by-layer scanning. The dynamic characteristics of each layer of data are grouped and aggregated, and the range and trend of fluctuation amplitude are extracted. The decomposed dynamic characteristics are compared with the abnormal parameters layer by layer, and the range of abnormal characteristics is recorded to separate the abnormal signal content with dynamic characteristics.

[0090] Identify pattern features in abnormal signals, match pattern features with potential abnormal causes, extract dynamic characteristics related to time series changes in the data, screen the characteristic distribution of the data item by item by comparing the amplitude differences, periodic changes and signal mutation points at different time points and frequency ranges, find the abnormal signal feature range, and further analyze the fluctuation pattern, extract repeated change cycles, peak spacing and change rules, and match them to the existing characteristic parameter set, compare the change pattern of potential abnormal causes, and record the specific correlation information between the pattern and the cause.

[0091] See also Figure 5 , the specific steps to generate the quantitative analysis results of error sources are:

[0092] S401: Based on the abnormal pattern and causal association report, extract the causal factor information in the report, perform weight analysis on the multivariate relationship, decompose the association strength and weight parameters between the multivariate, and generate the causal influence weight distribution result;

[0093] S402: Based on the causal influence weight distribution results, the error data in the variables are extracted, and statistical analysis of the error deviation is performed. The error sources are evaluated through parameter difference decomposition, and statistical results of the error sources are generated;

[0094] S403: Based on the statistical results of error sources, quantitative decomposition of error sources is performed, error distribution characteristics of causal variables are extracted, quantitative evaluation is completed in combination with error weight calculation, and quantitative analysis results of error sources are generated;

[0095] Based on the abnormal pattern and causal association report, the causal factor information in the report is extracted, the relationship between the causal variables is analyzed using the multivariate weight analysis method, and the association strength between multiple variables is calculated using the weight decomposition algorithm. The weight decomposition process includes constructing a variable relationship matrix with a matrix dimension equal to the number of variables, calculating the mutual information value between each pair of variables and normalizing it into an association weight value, decomposing the weight parameter using the matrix decomposition method, and decomposing the weight parameter in a singular value decomposition manner. The eigenvector in the decomposition result is extracted and stored as a weight parameter table to generate the causal influence weight distribution result;

[0096] Based on the causal influence weight distribution results, the error data in the variables are extracted, and the deviation statistical analysis method is used to analyze and process the error data. The error source is analyzed through the difference decomposition method. The difference decomposition method is to calculate the absolute error between the observed value and the theoretical value of the variable. The absolute error is defined as the absolute value of the observed value minus the theoretical value. The distribution analysis algorithm is used to perform distribution statistics on the error data. The distribution statistics method is to construct an error histogram. The number of intervals of the histogram is set to 10, and the interval width is the total error range divided by the number of intervals. The statistical results are saved as an error distribution table to generate error source statistical results.

[0097] Based on the statistical results of the error sources, the error distribution characteristics of the causal variables are extracted, and the error sources are quantitatively analyzed using the quantitative decomposition method. The quantitative decomposition method is to perform weighted decomposition on the error data in combination with the weight values ​​in the error distribution table. The weighted method is to multiply each error value by its corresponding weight value. The error results are converted into percentage form using the normalization method, and the key items in the error distribution characteristics are marked. The marking method is to select items with an error percentage greater than 10. Combined with the error weight calculation, the quantitative evaluation of the error sources is completed, and all evaluation data are stored as an error quantification table to generate the quantitative analysis results of the error sources.

[0098] Among them, the quantitative analysis results of error sources include a list of major error sources, the contribution rate of each source, and environmental factors affecting the errors.

[0099] See also Figure 6 , the specific steps to generate the reverse corrected data are:

[0100] S501: Based on the quantitative analysis results of the error sources, extract specific parameters in the error distribution, perform group analysis on each parameter according to the deviation range, calculate the deviation value of each group of parameters, adjust the data range within the group, and generate a correction deviation parameter set;

[0101] S502: Based on the correction deviation parameter set, correct each parameter in the original data one by one, replace the data with the correction value in the grouping range, correct the deviation value by accumulating item by item, and generate corrected original data;

[0102] S503: extracting uncorrected residual outliers in the data based on the corrected original data, performing out-of-range deviation evaluation on the outliers, gradually adjusting deviation parameters of the outliers, and generating inversely corrected data;

[0103] Based on the quantitative analysis results of the error sources, the specific parameters in the error distribution are extracted, and the deviation grouping algorithm is used to group and analyze each parameter according to the deviation range. The grouping method is to set the grouping interval, and the interval range is divided into 10 equal parts from the minimum to the maximum value of the deviation value. The deviation value of each group of parameters is calculated using the intra-group deviation mean calculation formula. The formula is the weighted average of all deviation values ​​in the group, and the weight is the influencing factor of the parameter. The upper and lower limits of the data in the group are adjusted using the range adjustment algorithm. The adjustment rule is to expand the upper and lower limits to the deviation between the maximum and minimum values ​​in the group by 10%. All adjusted data are stored as a parameter set to generate a correction deviation parameter set.

[0104] Based on the correction deviation parameter set, each parameter in the original data is corrected one by one, and the correction value in the group range is replaced by the range correction method. The correction value is defined as the mean deviation within the group. The corresponding parameter position in the original data is located by the matching algorithm, and each matching item is replaced by the correction value. The deviation value is corrected by the item-by-item accumulation method. The accumulation method is to accumulate the current deviation value and the previous correction result and update the current parameter value. The accumulation result is stored in the correction parameter array to generate the corrected original data;

[0105] Based on the corrected original data, the uncorrected residual outliers in the data are extracted, and the outlier detection algorithm is used to evaluate the deviation of the residual outliers outside the range. The range of the outliers is calculated by setting upper and lower limits. The upper limit is the 95% quantile of the normal data, and the lower limit is the 5% quantile. The values ​​out of the range are marked as outliers, and the deviation parameters of the outliers are gradually adjusted. The adjustment method is that the difference between the outlier and its adjacent normal value decreases, and the decreasing step is set to 10% of the difference until the outlier returns to the range. The adjusted outliers are stored as a reverse correction data set to generate reverse corrected data;

[0106] Among them, the reverse corrected data includes adjusted data parameters, data comparison before and after optimization, and correction effect evaluation.

[0107] See also Figure 7 ,The specific steps to generate the measurement data category recognition results are:

[0108] S601: Based on the reverse corrected data, scan the classification feature values ​​in the data item by item, extract the feature range of each data item, add a tag value to the data according to the classification standard, and generate a data classification feature set;

[0109] S602: Based on the data classification feature set, support vector machine is used to compare and analyze the reverse corrected data with the classification feature value, and the classification mark of each data item is gradually confirmed. The data is classified and sorted based on the marking result to generate the data classification marking result;

[0110] S603: Based on the data classification and labeling results, each classification label is input into the deep learning model, and the classified output is hierarchically sorted in combination with the input data, and the classified measurement data is extracted to generate the measurement data category recognition result;

[0111] Based on the reverse corrected data, the classification feature values ​​in the data are scanned item by item, and the feature range of each data is calculated using the interval extraction algorithm. The feature range is defined as the difference between the minimum and maximum values ​​of the feature value. The feature value sequence is read one by one through the traversal algorithm, and the maximum and minimum values ​​of each feature value are counted. The feature value is matched and analyzed with the standard value using the set classification standard. The matching method is to compare the range of the feature value with the upper and lower limits of the classification standard, and add the corresponding mark value to the feature data table. The format of the mark value is integer data, which is stored in the classification feature matrix to generate a data classification feature set.

[0112] Based on the data classification feature set, support vector machine is used to compare and analyze the inverse corrected data with the classification feature value. The support vector machine model adopts radial basis function kernel, the kernel parameter γ is set to 0.1, and the penalty factor C is set to 1. The support vector machine model is trained with the training set data, and the eigenvalue vector in the classification feature set is input into the model to classify and predict each data item. The predicted category is output as integer data, and the classification mark of each data item is gradually confirmed. The data is classified and sorted through the grouping algorithm in combination with the classification mark. The grouping method is to partition and store according to the classification mark value, and the grouped data after classification and sorting is saved in the classification result table to generate the data classification marking result;

[0113] Based on the data classification labeling results, each classification label is input into the deep learning model. The deep learning model adopts a fully connected neural network. The network structure includes an input layer, two hidden layers and an output layer. The number of input layer nodes is the dimension of the classification label, the number of hidden layer nodes is 128 and 64 respectively, and the number of output layer nodes is the number of categories of the classification output. The loss function of the model is the cross entropy loss, and the learning rate is set to 0.001. The classification label and the original data are combined into an input feature vector and input into the deep learning model. The feature vector is classified and predicted by the model, and the classification results are stored in a hierarchical data table by category. The classification result data is hierarchically grouped according to the hierarchical sorting algorithm, and the hierarchical results are saved in the classification result data set to generate the measurement data category recognition results;

[0114] Among them, the results of measurement data category recognition include the classified data type, classification accuracy and decision basis of model judgment.

[0115] Support vector machine, according to the formula: , in: To measure the classification and labeling results of the test data, is the feature vector of the metrological detection data to be identified, is the feature vector of the trained support vector in the support vector machine, is the Lagrange multiplier weight of the support vector, is the category label corresponding to the support vector, is the Gaussian kernel function, is the scale parameter of the Gaussian kernel, To dynamically adjust parameters, is the feature weight vector, is the feature weighted result of the linear part of the kernel function, is the bias term of the model, is the regularization term, is the regularization coefficient;

[0116] Execution process: First, the support vector machine is trained through the training set to obtain the feature vector of the support vector , the corresponding category label , Lagrange multiplier weight and model bias , then for the metrological detection data to be identified , the kernel function is used to calculate the similarity between the data to be identified and the support vector in the feature space, where the nonlinear part is realized by the Gaussian kernel and the kernel scale parameter Optimized by cross-validation, the linear part By feature weight vector and dynamically adjust parameters Emphasize the importance of specific features, It is calculated based on the information gain of each feature in the measurement detection task. Dynamically optimize through gradient descent method to adjust the weight ratio of nonlinear and linear parts, and add regularization terms Suppress the influence of noise on the results, regularization coefficient Determined by grid search method, it is used to balance the complexity and generalization ability of the model. The kernel function results of all support vectors are accumulated and the bias term is added. and the regularization term , map the calculation results into classification labels, and output the classification labeling results of the measurement and detection data.

[0117] Scan the classification feature values ​​in the data item by item and extract the feature range of each data. The classification feature values ​​are the attributes and indicators in the data used to distinguish categories. The specific extraction steps are as follows: extract numerical and categorical attributes from the reverse-corrected data according to predefined classification rules. For numerical attributes, directly extract the values. For categorical attributes, use dictionary mapping to convert them into numerical codes. For missing values, use interpolation and default values ​​to fill them. Finally, organize and classify the feature values.

[0118] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A method for automatic identification of metrological detection data based on deep learning, characterized in that: The following steps are involved: Step 1: Receive raw data through the sensor, score the signal strength, mark and record the distorted sections, and generate a signal quality assessment report; Step 2: According to the signal quality assessment report, wavelet transform is used to correct the marked segment and the distorted signal to obtain reconstructed signal data; Step 3: Structural decomposition is performed on the reconstructed signal data, abnormal patterns and abnormal causes are identified through a generative adversarial network, and the generative adversarial network is continuously trained to obtain a deep learning model, and the relationship between the abnormal pattern and the abnormal cause is recorded to generate a report on the abnormal pattern and causal association; Step 4: Using the abnormal pattern and causal association report, calculate the influence intensity of each factor through multiple regression analysis, and evaluate the error source at the same time to obtain the quantitative analysis result of the error source; Step 5: According to the quantitative analysis results of the error sources, the original data is corrected by accurately adjusting the parameters, the accuracy and reliability of the original data are optimized, and the inverse corrected data is generated; Step six: input the reverse corrected data into the deep learning model, and use a support vector machine to classify the model according to data features, output the classification results, and obtain the measurement data category recognition results.

2. The method for automatic identification of metrological detection data based on deep learning according to claim 1 is characterized in that: The signal quality assessment report includes the specific location of the distorted section, the signal strength level and the noise level.

3. The method for automatic identification of metrological detection data based on deep learning according to claim 1 is characterized in that: The reconstructed signal data includes a reconstructed waveform, an adjusted frequency response, and a signal stability score.

4. The method for automatic identification of metrological detection data based on deep learning according to claim 1 is characterized in that: The specific steps of generating the abnormal pattern and causal association report are: Based on the reconstructed signal data, the signal data is divided into multiple independent hierarchical data areas according to the time range and the signal amplitude change trend, and the change rate and structural characteristics of the signal are extracted and processed by scanning layer by layer, and the key parameters of the signal are extracted, and the abnormal data part existing in the signal is separated, and the dynamic characteristics separation of the abnormal data is completed to generate signal decomposition feature data; Based on the signal decomposition feature data, a generator of a generative adversarial network is used to learn the potential distribution characteristics of the signal decomposition feature data, and the generated data and actual data are compared to capture the abnormal pattern characteristics in the signal distribution, and the abnormal pattern characteristics are extracted as key dynamic parameters, and the abnormal pattern characteristics are matched with the known characteristics of the potential abnormal causes, and the association mapping between the pattern and the cause is completed, and the analysis results are organized into a multi-dimensional parameter set to generate abnormal pattern and cause relationship data; Based on the abnormal pattern and cause relationship data, the causal characteristics of the abnormal pattern are modeled using a generative adversarial network, and the causal characteristic distribution is reconstructed, and the consistency between the generated distribution and the actual distribution is evaluated, and the dynamic association path of the causal variable is determined. The variables in the causal path are analyzed item by item using a feature relationship mapping method, and the causal strength between the pattern characteristics and the abnormal causes is quantified. The modeling results are recorded as a causal association data set to obtain a causal relationship model, and an abnormal pattern and causal association report is generated at the same time; The abnormal pattern and causal association report includes a list of identified abnormal patterns, related cause analysis and impact assessment.

5. The method for automatic identification of metrological detection data based on deep learning according to claim 4 is characterized in that: The method uses a multi-level feature decomposition method to process the data structure, extracts feature parameters from each layer of data, preliminarily stratifies the signal data according to time periods and frequency ranges, divides the data into multiple different feature intervals, extracts amplitude, change rate and structural parameters for the stratified data by scanning layer by layer, groups and aggregates the dynamic characteristics of each layer of data, extracts the range and trend of fluctuation amplitude, compares the decomposed dynamic characteristics with abnormal parameters layer by layer, records the range of abnormal characteristics, and separates abnormal signal content with dynamic characteristics; The method identifies pattern features in abnormal signals, matches pattern features with potential abnormal causes, extracts dynamic characteristics related to time series changes in data, screens the characteristic distribution of data item by item by comparing amplitude differences, periodic changes and signal mutation points at different time points and frequency ranges, finds the abnormal signal characteristic range, further analyzes the fluctuation pattern, extracts repeated change cycles, peak spacing and change rules, and matches them to an existing set of characteristic parameters, compares the change patterns of potential abnormal causes, and records the specific association information between the pattern and the cause.

6. The method for automatic identification of metrological detection data based on deep learning according to claim 1 is characterized in that: The quantitative analysis results of the error sources include a list of major error sources, the contribution rate of each source, and environmental factors that affect the errors.

7. The method for automatic identification of metrological detection data based on deep learning according to claim 1 is characterized in that: The reverse corrected data includes adjusted data parameters, data comparison before and after optimization, and correction effect evaluation.

8. The method for automatic identification of metrological detection data based on deep learning according to claim 1, characterized in that: The specific steps of generating the measurement data category identification result are: Based on the reverse corrected data, the classification feature values ​​in the data are scanned item by item, the feature range of each data item is extracted, and a tag value is added to the data according to the classification standard to generate a data classification feature set; Based on the data classification feature set, a support vector machine is used to compare and analyze the reverse corrected data with the classification feature value, and the classification mark of each data item is gradually confirmed. The data is classified and sorted based on the marking result to generate a data classification marking result; Based on the data classification and labeling results, each classification label is input into the deep learning model, and the classified output is hierarchically sorted in combination with the input data to complete the extraction of the classified measurement data and generate the measurement data category recognition result; The measurement data category recognition result includes the classified data type, the classification accuracy and the decision basis of the model judgment.

9. The method for automatic identification of metrological detection data based on deep learning according to claim 1, characterized in that: The support vector machine is based on the formula: ， in: To measure the classification and labeling results of the test data, is the feature vector of the metrological detection data to be identified, is the feature vector of the trained support vector in the support vector machine, is the Lagrange multiplier weight of the support vector, is the category label corresponding to the support vector, is the Gaussian kernel function, is the scale parameter of the Gaussian kernel, To dynamically adjust parameters, is the feature weight vector, is the feature weighted result of the linear part of the kernel function, is the bias term of the model, is the regularization term, is the regularization coefficient.

10. The method for automatic identification of metrological detection data based on deep learning according to claim 8, characterized in that: The classification feature values ​​in the data are scanned item by item to extract the feature range of each data. The classification feature values ​​are attributes and indicators in the data used to distinguish categories. The extraction step is specifically to extract numerical and categorical attributes from the reverse-corrected data according to predefined classification rules. For numerical attributes, the numerical values ​​are directly extracted. For categorical attributes, dictionary mapping is used to convert them into numerical codes. For missing values, interpolation and default value filling are used. Finally, the feature values ​​are sorted and classified.

Images