TWI755995B - A method and a system for screening engineering data to obtain features, a method for screening engineering data repeatedly to obtain features, a method for generating predictive models, and a system for characterizing engineering data online - Google Patents

Abstract

The present invention provides a method and system for screening engineering data to obtain features, a method for screening engineering data multiple times to obtain features, a method for generating prediction models, and a system for characterizing engineering data online. The system for screening engineering data to obtain features includes: a judging unit, a statistical unit and a processing unit. The judging unit is used for judging whether a project data is numerical data or category data. If the judging unit judges that the engineering data is numerical data, the statistical unit performs Kolmogorov-Smirnov test for each category in the engineering data and its corresponding fields, so as to confirm the Whether the engineering data is normally distributed. The processing unit is used to separately check whether each of the categories reaches the threshold value corresponding to the first, second, third, and fourth threshold value groups, and defines a category that meets the threshold value as a valid feature .

Description

A method and system for screening engineering data to obtain features, a method for screening engineering data multiple times to obtain features, a method for generating predictive models, and a system for online characterization of engineering data

The present invention relates to a method for screening engineering data to obtain features, a system for screening engineering data to obtain features, a method for generating prediction models, and a system for characterizing engineering data online. Especially in various engineering fields, methods and systems for efficient and rapid screening and processing of engineering data.

Generally speaking, for various engineering data such as medical engineering, electrical engineering, and mechanical engineering, all kinds of data generated in the process need to go through the steps of obtaining data, processing data, selecting model, training, evaluating, and adjusting hyperparameters. Having access to the generation of predictive models, conventional data scientists need to spend a considerable amount of time to correctly construct predictive models. In other words, in the conventional technology, it takes too long to clean and organize the data, so that the technology of machine learning cannot be used effectively. On the other hand, the process of feature extraction is a tedious process, and the conventional technology usually relies on manual domain knowledge, experience and complicated data manipulation, and the final features will be subject to the subjective constraints of manual workers. Although machine learning has many proven benefits, Successfully harnessing machine learning requires enormous human effort, as no single algorithm or model can address every possible situation. For example, while researchers in medical engineering are familiar with clinical data, they still lack the machine learning expertise needed to apply these concatenated data to big data sources.

When confronted with supervised machine learning problems, data scientists are often tasked with creating explanatory variables (also known as features) that predict outcomes of interest. Ideal feature engineering entails constructing features that not only provide useful insights into the data itself, but also takes into account any limitations of the learning algorithm used. This is not a trivial process because the performance of a given machine learning algorithm depends heavily on the quality of the input data. That is, constructing features from raw data typically requires extensive domain knowledge and is therefore usually performed manually by human experts in a trial-and-error manner. This makes feature engineering a crucial and time-consuming step in the machine learning pipeline. Feature engineering, also known as feature construction, is the process of constructing new features from existing data to train machine learning models. Feature construction is more important than the actual model used, because a machine learning algorithm can only learn from given data, so how to construct a feature relevant to the desired goal is critical.

Furthermore, in the field of data analysis, researchers often use the relational database in the feature tool to automatically find potential features through the relationship between data tables and achieve automatic results that are close to the manual results of data scientists. However, the analysis method based on the relationship between the data tables in the relational database can only be applied to the relevant data of the relational database, that is, the numerical data, and cannot be used on the data with categorical characteristics.

Furthermore, in the prior art, for numerical features, various calculations are exhaustively listed, and then the model is used to verify whether there is an improvement result, and if so, the initial features of the next generation are included until Results no longer improve. However, this method is only applicable to numerical data and cannot be used for categorical data.

Therefore, in order to overcome the aforementioned problems, the present invention has been developed.

In order to overcome the aforementioned technical problems, the present invention adopts multiple directions to generate new features: time field data group, related field data group, field field data group: compared with the conventional technology, it provides specific feature generation for different technical fields. Processing; using statistical distribution to compare the similarity of each feature in the two data sets, removing dissimilar features, and supporting numerical and categorical features at the same time; so as not to use complex algorithms, correlation layers, or model testing methods in conventional techniques. Find, evaluate features, and do not need to use trained algorithms to check the found features, which not only greatly and comprehensively improves the efficiency of selecting effective features, but also automates processing to achieve better results than data scientists. Accuracy.

In order to achieve the foregoing purpose, the present invention provides a method for screening engineering data to obtain features, comprising:

A: It is judged that the engineering data is numerical data or categorical data. If the engineering data is numerical data, go to step A1, and if it is judged that the engineering data is categorical data, go to step A2;

Wherein the steps A1 and A2 are as follows:

A1: Perform the Kolmogorov-Smirnov Test on each category of the numerical data and its corresponding fields to confirm whether the engineering data is normally distributed; if If the test result is a normal distribution, go to step B1; if the test result is a non-normal distribution, go to step B2;

A2: Calculate the Kremo V coefficient for each category in the category data and its corresponding fields Check to obtain a first threshold value group composed of multiple threshold values, and then proceed to step C1;

The steps B1 and B2 are as follows:

B1: T-test is performed on each category of the numerical data and its corresponding multiple fields to obtain a second threshold value group consisting of multiple threshold values; then step C2 is performed;

B2: Check the dispersion of multiple fields corresponding to each category in the numerical data respectively; if it is judged that the dispersion of multiple fields corresponding to each category exceeds the preset value, perform KL divergence (Kullback-Leibler divergence) test to obtain a third threshold group consisting of multiple thresholds, and then go to step C3; if it is determined that the dispersion of multiple fields corresponding to each category in the numerical data is not If the preset value is exceeded, then the Mann-Whitney U test is performed to obtain a fourth threshold value group composed of multiple threshold values, and then step C4 is performed;

The steps C1, C2, C3, and C4 are as follows:

C1: Check whether each of the categories reaches the corresponding threshold value in the first threshold value group, and define the category that has reached the threshold value as a feature;

C2: Check whether each of the categories reaches the corresponding threshold value in the second threshold value group, and define the category that has reached the threshold value as a feature;

C3: Checking whether each of the categories has reached the corresponding threshold value in the third threshold value group, and defining the category that has reached the threshold value as a feature;

C4: Check whether each of the categories reaches the corresponding threshold value in the fourth threshold value group, and define the category that meets the threshold value as a feature.

During implementation, the step A further includes: judging whether the engineering data is numerical data or type data, and if the engineering data is numerical data, dividing the numerical data into two groups and then dividing the two groups into two groups. Carry out a similarity test, leaving the numerical data with significant similarities Step A1 is performed after the category and its corresponding field; if it is determined that the engineering data is category-type data, the category-type data is divided into two groups, and the two groups are checked for similarity, leaving Step A2 is performed after the significant categories in the category data and their corresponding fields.

The present invention also provides a method for performing multiple screening on engineering data to obtain features, which comprises the following steps:

a. Data cleaning for engineering data, including:

b. Compensate the missing value of the field in the engineering data; generate a plurality of field data groups from the engineering data, wherein the field data groups include: time field data group, field field data group or related field data group; wherein the step of generating the field field data group includes: disassembling and combining the value field data group and the category field data group to generate the field field data group; wherein the step of generating the associated field data group Including: from the numerical field data group and the category field data group with significant positive or negative correlation by statistical test of correlation; and calculating and combining the significant positive or negative correlation to generate correlation field data group;

c. Screen these time field data groups, field field data groups or related field data groups by the method described above to obtain the screened features.

During implementation, after the step c, it further includes:

d1: perform adversarial learning on the screened features with at least one machine learning algorithm to verify the similarity of the screened features; and

d2: Adjust the threshold value of the similarity verification by grid search (Grid Search) for those screened features with insignificant similarity, and perform step A1 again; keep the screened feature similarity Remarkable similarities.

When implementing, in this step a, the step of compensating for the missing value of the column in the engineering data comprises: using the median value, using the minimum value, using the maximum value, using the mode value, using Compensation of 4 quantiles or use of other similar columns; in the step a, it further includes removing the different fields corresponding to these categories but there is no change.

During implementation, the step b further includes a step of generating a category field data group, which includes: merging the first stroke in the category field data group, merging the last stroke in the category field data group, and merging the category The middle pen in the field data group, the one with the most occurrences in the combined field data group of this category, or the one with changes in the combined field data group of this category.

During implementation, the step of generating the time field data group in step b includes: taking the time corresponding to the fields including: year, month, day, week, hour, minute, second, or every 15 minutes .

When implemented, the step of generating and generating the field field data group further includes: the step of disassembling and combining the value field data group to generate the field field data group includes: generating binary digits from the value field data group The single-digit field data group or generating the single-digit field field data in decimal; wherein the step of dismantling and combining the type field data group includes: processing the character string in the type field data group Splitting to generate a plurality of split strings: counting the number of occurrences of each of the split strings, leaving split strings above a threshold of times; and the remaining splits encoding each of the strings of , each of the encodings is distinct from each other.

During implementation, the steps for generating the associated field data group further include:

A correlation test is performed on this numerical column data group, and numerical columns with significant positive or negative correlation will be found Bit data, perform the following operations to generate the associated field data group:

Add, subtract, multiply, divide, take the LOG value, take the angle of a trigonometric function; or

The correlation test is performed on the categorical column data group, and the categorical column data group with significant positive or negative correlation is subjected to the following operations to generate the associated column data group: rearrange the strings; Each of the permuted strings is encoded to generate a group of associated field data, wherein each of the encodings are distinct from each other.

The present invention also provides a method for generating a prediction model, comprising:

X: mark the screened features generated by the aforementioned method as a training group or a test group;

Y: mix the filtered features;

Z: Distinguish the training group or the test group through at least one machine learning algorithm, so as to establish a prediction model.

The present invention further provides a system for screening engineering data to obtain features, which includes: a processor, the processor includes a judging unit, a statistical unit and a processing unit. The judging unit is used for judging whether the engineering data is numerical data or type data; the statistical unit is used for judging each type of the engineering data and its corresponding multiple fields if the judging unit judges that the engineering data is numerical data. Kolmogorov-Smirnov Test is used to confirm whether the engineering data is normally distributed. If the judging unit judges that the engineering data is not numerical data, then the statistical unit performs the Cremo V coefficient test on each category of the engineering data and its corresponding fields to obtain a plurality of threshold values. The first threshold value group; wherein if the test result of the statistical unit 12 is a normal distribution, the statistical unit performs a T-test test on each category in the engineering data and its corresponding fields to obtain a plurality of thresholds A second threshold group of values. if the The test result of the statistical unit is a non-normal distribution, then the statistical unit tests the dispersion degree of the multiple fields corresponding to each category respectively; if the statistical unit determines that the dispersion degree of the multiple fields corresponding to each category exceeds the predetermined If the value is set, the Kullback-Leibler divergence test is performed to obtain a third threshold group consisting of multiple thresholds. If the statistic unit determines that the dispersion of the plurality of fields corresponding to each category does not exceed the preset value, the Mann-Whitney U test is performed to obtain a fourth threshold value group consisting of a plurality of threshold values. The processing unit is used to separately check whether each of the categories reaches the threshold value corresponding to the first, second, third, and fourth threshold value groups, and defines a category that meets the threshold value as a feature.

The present invention further provides a system for characterizing engineering data online, which includes: a server, the server includes a processor, and the processor includes a storage unit and a processing unit. The storage unit is used to receive the original engineering data input from the client and store the original engineering data; wherein the processing unit is used to read the original engineering data from the storage unit and process the original engineering data with the aforementioned method A number of filtered features are obtained.

In order to further understand the present invention, the following preferred embodiments are given, and the specific components of the present invention and the achieved effects are described in detail as follows in conjunction with the drawings and drawing numbers.

A, A1, A2, B1, B2, C1, C2, C3, C4, a, b, c, d, d1, d2: Steps

1: Processor

11: Judgment unit

12: Statistics Unit

13: Processing unit

2: A system to characterize engineering data online

FIG. 1 is a flow chart of an embodiment of a method for screening engineering data to obtain features of the present invention.

FIG. 2 is a flow chart of an embodiment of a method for screening engineering data for multiple times to obtain features of the present invention.

FIG. 3 is a schematic diagram of an embodiment of a system for screening features in engineering data according to the present invention.

FIG. 4 is a schematic diagram of an embodiment of the system for online characterization of engineering data of the present invention.

FIG. 5 is an original engineering data table of an embodiment of the method for screening engineering data for features of the present invention.

FIG. 6 is an engineering data table after time series sampling processing according to an embodiment of the method for screening engineering data to obtain features of the present invention.

FIG. 7 is the generation of engineering data through time features according to an embodiment of the method for screening engineering data to obtain features of the present invention.

Figures 8 and 9 are embodiments of the method for screening engineering data to obtain features of the present invention

FIG. 10 is a feature correlation statistics of engineering data according to an embodiment of the method for screening engineering data to obtain features of the present invention.

FIG. 11 is an embodiment of the method for screening engineering data to obtain features of the present invention, and generating engineering data through associated features.

Figures 12-15 illustrate filtering engineering data by features according to an embodiment of the method for screening engineering data to obtain features of the present invention.

FIG. 16 shows the features retained after screening the engineering data according to an embodiment of the method for screening engineering data to obtain features of the present invention.

Please refer to FIG. 1, the present invention provides a method for screening engineering data to obtain features, which includes:

A: Judging that the engineering data is numerical data or category data, if the engineering data is numerical data If it is judged that the engineering data is a category type data, then go to Step A2;

Wherein the steps A1 and A2 are as follows:

A1: Perform the Kolmogorov-Smirnov Test on each category of the numerical data and its corresponding fields to confirm whether the engineering data is normally distributed; if If the test result is a normal distribution, a step B1 is performed, and if the test result is a non-normal distribution, a step B2 is performed;

A2: Perform the Cremo V coefficient test on each category and its corresponding fields in the category data to obtain a first threshold value group composed of multiple threshold values, and then perform a step C1;

The steps B1 and B2 are as follows:

B1: T-test is performed on each category of the numerical data and its corresponding multiple fields to obtain a second threshold value group consisting of multiple threshold values; then step C2 is performed;

B2: Check the dispersion degree of the multiple fields corresponding to each category in the numerical data respectively; if it is determined that the dispersion degree of the multiple fields corresponding to each category exceeds a preset value, perform KL dispersion The Kullback-Leibler divergence test is performed to obtain a third threshold group consisting of multiple thresholds, and then step C3 is performed; if the dispersion of multiple fields corresponding to each category in the numerical data is determined If it does not exceed a preset value, then perform the Mann-Whitney U test to obtain a fourth threshold value group composed of multiple threshold values, and then perform step C4;

Wherein the steps C1, C2, C3, and C4 are as follows:

C1: Check whether each of the categories reaches the corresponding threshold value in the first threshold value group, and define a category that has reached the threshold value as a feature;

C2: Check whether each of the categories reaches the corresponding threshold value in the second threshold value group, and define a category that has reached the threshold value as a feature;

C3: Check whether each of the categories reaches the corresponding threshold value in the third threshold value group, and define a category that has reached the threshold value as a feature;

C4: Check whether each of the categories reaches the corresponding threshold value in the fourth threshold value group, and define a category that meets the threshold value as a feature.

Referring to FIG. 3 , the present invention further provides a system for screening engineering data to obtain features, which includes a processor 1 , and the processor 1 includes a judgment unit 11 , a statistics unit 12 and a processing unit 13 . The judging unit 11 is used for judging that a project data is numerical data or type data; the statistics unit 12 is used for determining each type of the project data and its corresponding multiple if the judging unit 11 judges that the project data is numerical data The Kolmogorov-Smirnov Test is performed on the field to confirm whether the engineering data is normally distributed. If the judging unit 11 judges that the engineering data is not numerical data, the statistics unit 12 performs the Kremo V coefficient test on each type of the engineering data and its corresponding fields to obtain a plurality of threshold values. A first threshold value group is formed; wherein if the test result of the statistical unit 12 is a normal distribution, the statistical unit 12 performs a T-test test on each category in the engineering data and its corresponding multiple fields to obtain A second threshold value group formed by a plurality of threshold values. If the test result of the statistical unit 12 is an abnormal distribution, the statistical unit 12 checks the dispersion of the plurality of fields corresponding to each category respectively; if the statistical unit 12 determines the plurality of fields corresponding to each category If the dispersion exceeds a predetermined value, a Kullback-Leibler divergence test is performed to obtain a third threshold group consisting of a plurality of thresholds. If the statistical unit 12 determines that the dispersion of the plurality of fields corresponding to each category does not exceed a predetermined value, the Mann-Whitney U test is performed to obtain a fourth threshold value group formed by a plurality of threshold values. The processing unit 13 is for checking, respectively, whether each of the categories reaches all of the first, second, third, fourth threshold groups For the corresponding threshold value, a category that reaches the threshold value is defined as a feature.

The method and system of the present invention are described in detail below. First of all, the method and system for screening engineering data to obtain features of the present invention are methods and systems for processing various engineering data. The types of engineering data include but are not limited to: financial engineering, chemical engineering, mechanical engineering, biomedical engineering, etc. Engineering data in various fields. First, in step A, the judging unit 11 judges the engineering data as numerical data or category data. If the engineering data is numerical data, proceed to step A1, and if the determining unit 11 determines that the engineering data is categorical data, proceed to step A2. The steps A1 and A2 are described as follows. In the step A1, the statistical unit 12 is used to perform the Kolmogorov-Smirnov Test for each category in the numerical data and its corresponding fields to confirm Whether the engineering data is normally distributed; if the test result is a normal distribution, go to step B1; if the test result is a non-normal distribution, go to step B2. The Kolmogorov-Smirnov test (hereinafter referred to as KS-test) is a statistical data analysis method, which is specially designed for testing distributed numerical data (distributed data set) rather than discrete data. KS-test does not need to make any assumptions about the distribution of numerical data, and has high sensitivity to the shape and position of the CDF (cumulative distribution function curve) of numerical data, and can accurately evaluate the relative distribution of numerical data. If the test result is a normal distribution (p value>0.05), then execute step B1; if the test result is a non-normal distribution (p value<0.05), execute step B2.

Furthermore, in the step A2, the statistical unit 12 performs the Cremo V coefficient test on each category and its corresponding fields in the category data to obtain a first threshold value group composed of a plurality of threshold values group, and then proceed to step C1. For statistical tests related to the Chi-Square test, the strength test used in such statistical tests is the Cremo V coefficient test. Each category produces the first door Thresholds, the first thresholds are used to measure the degree of correlation between multiple fields in at least two categories.

The steps B1 and B2 are described below. In the step B1, the statistical unit 12 performs T-test test on each type of the numerical data and its corresponding multiple fields to obtain a second threshold value group consisting of multiple threshold values; Step C2. The default conditions of the T-test test are that the field (dependent variable) corresponding to the category is a continuous variable, the field corresponding to the category is randomly sampled from the parent group, and the parent group is normally distributed. Since the numerical data in the present invention and this step have already passed the KS-test, they meet the preset conditions of the T-test. Furthermore, in step B2, the statistics unit 12 respectively checks the dispersion of the plurality of fields corresponding to each category in the numerical data. If it is determined that the dispersion of the plurality of fields corresponding to each category exceeds a predetermined value, the statistical unit 12 performs the KL divergence test to obtain a third threshold group consisting of a plurality of thresholds, and then performs Step C3. The K-L divergence test is used to assess the amount of information lost when one hypothetical distribution is used to approximate another. If the statistical unit 12 determines that the dispersion of the plurality of fields corresponding to each category in the numerical data does not exceed a predetermined value, then the Mann-Whitney U test is performed to obtain a fourth value consisting of a plurality of threshold values. Threshold value group, and then go to step C4. The purpose of the Mann-Whitney U test performed by the statistical unit 12 is to compare the differences of at least two random samples to infer the differences between the two parent groups. The basis for making inferences is based on the sampling distribution composed of engineering data, and the verification statistic value U is calculated according to the grade of the variable score in the sample.

The steps C1, C2, C3, and C4 are respectively as follows. In step C1, the processing unit 13 respectively checks whether each of the categories reaches the corresponding threshold value in the first threshold value group, and defines a category that has reached the threshold value as a feature. In step C2, the processing order Element 13 separately checks whether each of the categories reaches the corresponding threshold value in the second threshold value group, and defines a category that has reached the threshold value as a feature. In step C3, the processing unit 13 respectively checks whether each of the categories reaches the corresponding threshold value in the third threshold value group, and defines a category that has reached the threshold value as a feature. In step C4, the processing unit 13 respectively checks whether each of the categories reaches the corresponding threshold value in the fourth threshold value group, and defines a category that has reached the threshold value as a feature.

The aforementioned technical solution of the present invention uses statistical analysis methods to compare the similarity of features in multiple data sets, removes dissimilar engineering data, and is applicable to both numerical and categorical categories in engineering data, so that conventional techniques are not used. Various algorithms such as: algorithms trained with a large amount of data to find and evaluate features in engineering data can also achieve efficient technical results using algorithms to find. In other embodiments of the present invention, U-Test is commonly used for engineering data in the field of mechanical manufacturing (under the condition that the data values are similar to those of the same stage or of the same type), and U-Test is commonly used for engineering data in the medical field (similar to the value of instrument judgment values). The KL test is often used for engineering data in the financial field, because the numerical changes of financial engineering vary on different dates and in different regions.

In another embodiment, the step A further includes: judging whether the engineering data is numerical data or category data, and if the engineering data is numerical data, dividing the numerical data into two groups and dividing the numerical data into two groups. The two groups are tested for similarity, and the step A1 is performed after the significant categories in the numerical data and their corresponding fields are left. The similarity analysis of the present invention is used for the data processing flow of classification and clustering. Similarity needs to be analyzed according to the attribute value of engineering data itself, including: attribute value processing, similarity measurement standard, etc. If it is judged that the engineering data is category-type data, the category-type data is divided into two groups and the similarity between the two groups is checked, and the significant category in the category-type data and its corresponding field backward Do this step A2. The similarity metrics of the present invention include: Euclidean Distance, Manhattan Distance, Chebyshev Distance, Minkowski Distance, Standardized Euclidean distance , Mahalanobis Distance, Cosine, Hamming distance, Jaccard distance, Jaccard similarity coefficient, Correlation coefficient and Correlation distance, information entropy, etc. will not be repeated here.

Please refer to FIG. 2, the present invention also provides a method for multiple screening of engineering data to obtain features, which includes:

a. Data cleaning for a project data, including:

b. Compensate for missing values of fields in the engineering data; generate a plurality of field data groups from the engineering data, wherein the field data groups include: a time field data group, a field field data group or an association A field data group; wherein the step of generating a field field data group includes: disassembling and combining a value field data group or a category field data group to generate the field field data group;

Wherein, the step of generating the related field data group includes: using the statistical test of correlation to have a significant positive or negative correlation from the numerical field data group and the category field data group; and performing the significant positive or negative correlation on the data group. Operations and combinations to generate associated field data groups;

c. Screen the time field data group, the field field data group or the related field data group according to the above-mentioned embodiment of the method of screening engineering data to obtain features, so as to obtain features that have been screened multiple times.

The following describes in detail the method of the present invention to screen engineering data for many times to obtain features. Law. First, in step a, data cleaning is performed on engineering data, including engineering data in various fields such as financial engineering, chemical engineering, mechanical engineering, and biomedical engineering. Data cleaning is data science (DS) or machine learning (ML). ), so that the truly critical features can be found incrementally in subsequent steps. In another embodiment, the data cleaning step of the present invention includes: compensating for missing values of fields in the engineering data, removing categories of unchanged fields in the engineering data, checking outlier data groups, and the like. Generally speaking, it is difficult for engineering data in various fields to be complete, so it is necessary to compensate for the missing values of the fields in the engineering data or remove the categories with unchanged fields in the engineering data, so as not to affect the construction of the subsequent training model and find the features that help lead to overfitting.

Furthermore, in some cases, the engineering data will hide other data groups that are very different from the main data, so in other embodiments, in the step a, it is also necessary to check out the outlier data groups individually or separately deal with. In the step a, the step of compensating for the missing value of the column in the engineering data comprises: using the median to supplement the value, using the minimum value to supplement the value, using the maximum value to supplement the value, using the mode to supplement the value, and using the quartile. Numerical complement or use other similar column complements. In another embodiment, the step a further includes a step of removing the categories whose fields in the engineering data are unchanged, for example, removing the fields corresponding to different fields in the categories but not changing, so as to prevent the These categories will not affect the analysis and prediction results of subsequent data processing.

Furthermore, in step b, a plurality of field data groups are found from the project data, wherein the field data groups include: time field data group, field field data group or related field data group. The step of generating the field field data group includes: disassembling and combining a value field data group or a category field data group to generate the field field data group. In one embodiment, the step of generating a related field data group includes: using a statistical test of correlation to determine which of the numerical field data group and the category field data group has a significant positive or negative correlation; and the significantly positive correlation or negative correlation Operators perform operations and combinations to generate the associated field data group. Furthermore, in another embodiment, the step of generating the aforementioned numerical field data group includes: merging the maximum value, minimum value, average value, median value in the fields corresponding to the fields in the categories in the engineering data number or inner mode. In the step b, it further includes a step of generating a category field data group, which includes: merging the first stroke in the category field data group, merging the last stroke in the category field data group, and merging the category field The middle pen in the data group, the one with the most occurrences in the data group of the combined category field, or the one with changes in the combined data group of the category field. The step of generating the time field data group includes: taking the time corresponding to the fields including: year, month, day, week, hour, minute, second, or every 15 minutes.

Please refer to FIG. 4 , the present invention further provides a system for automatically characterizing engineering data. The system 2 includes: a server. The server includes a processor, the processor includes a storage unit and a processing unit, wherein the storage unit is used for receiving an original project data input from the client and storing the original project data; wherein the processing unit is used for reading Taking the raw engineering data from the storage unit and processing the raw engineering data in the method described in the previous embodiment to obtain a plurality of screened features. In other embodiments, the method claimed in the present invention can be set in a cloud system, allowing the client to input engineering data from a remote end by means of the Internet or a local area network, and the original project input by the client The data were screened for important features by the methods described in the preceding examples of the present invention.

The method and system of the present invention are described below based on engineering data in the field of mechanical engineering. Please refer to Figure 5. The table in Figure 5 lists the original engineering data of a tool, including time stamp, command speed, actual speed, spindle current, tool compensation ratio, idle reason, etc. Again, please refer to Figure 6. Figure 6 lists the original engineering data processed by the aforementioned tool according to the sampling rate of the data, automatically calculates the appropriate time frequency of the data, and merges the data time. For example: sampling According to the sampling theory, the sampling rate must be at least 10 times higher than the phenomenon to be observed before it can be observed.

Again, please refer to Fig. 7. Fig. 7 lists the steps of generating the time feature with the time stamp in the original engineering data processed by the aforementioned tool, including: taking the time corresponding to these fields including: year, month, day, Day, hour, minute, second, or every 15 minutes.

In another embodiment, after the step c, it further includes: in step d1, performing adversarial learning on the screened features with at least one machine learning algorithm to verify the similarity of the screened features. In step d2, the screened ones with insignificant feature similarity adjust the threshold for verifying similarity by grid search, and perform step A1 again; keep the screened ones with significant feature similarity . There are many classification models in machine learning, such as LR, SVM or XGBoost; and deep learning models CNN, LSTM, etc. However, different models have different parameter settings and need to be adjusted and selected by themselves. When the data volume of engineering data is too large, grid search can easily become a consumption of system resources, so the present invention uses grid search after the above-mentioned steps, which greatly improves the efficiency of grid search.

Furthermore, in other embodiments, the aforementioned step of generating the field field data group further includes: the step of disassembling and combining the value field data group to generate the field field data group includes using the value field The data group generates binary single-digit field field data group or generates decimal single-digit field field data. The step of disassembling and combining the category field data group includes: dividing the character strings in the category field data group to generate a plurality of divided character strings, and counting the number of divided character strings among the divided character strings. the number of occurrences of each, leaving a segmented string above the threshold of one count; and encoding each of the remaining segmented strings, each of the encodings being are different from each other, so that each code is independent, and the present invention uses Generate new, meaningful class field data from the class field data group.

Then, please refer to FIG. 8 , which lists the aforementioned steps of generating the field field data group: generating the binary-unit value field data group from the value field data group. Then, please refer to FIG. 9, which lists an embodiment of disassembling and combining the value field data group to generate the field field data group. Please refer to Fig. 10. Fig. 10 shows the steps of generating the related field data group of the present invention: after the correlation test is performed on the numerical data, the numerical data with significant correlation (p value < 0.05) will be tested for Steps such as adding/subtracting/multiplying/dividing, taking LOG, taking the angle of trigonometric functions, etc., produce the result shown in Figure 11. For example, in Fig. 10, the field column data group is generated by multiplying the maximum actual speed of the relevant category by the maximum value of the spindle current.

In another embodiment, the step of generating the associated field data group further includes: performing a correlation test on the numerical data, and performing the following operations on the numerical data with significant positive or negative correlation to generate the numerical field data group. The aforementioned operations include: adding/subtracting/multiplying/dividing, taking LOG values, taking various angles of various trigonometric functions; or performing correlation tests on this type of data group, including chi-square test related series tests, etc., will have significant positive or negative results. The negatively correlated categorical field data group is subjected to the following operations to generate the correlated field data group. The aforementioned arithmetic help includes: rearranging the strings, for example, recombining the letters A, B, C into ABC, ACB, BCA, etc.; or encoding each of the rearranged strings A group of associated field data is generated wherein each of the codes are distinct from each other. For example, the character strings AB, AC, correspond to 110, 101, and 111, respectively.

The present invention further provides a method for generating a prediction model, comprising the following steps. In step X, the screened features generated by the method in the foregoing embodiment of the present invention are marked as training group or test group, the training group is used as test data, and the test group is used as test data; Whether the features found by the invented method and system are accurate. The screened features are then combined in step Y. Furthermore, in step Z, at least one machine learning algorithm is used to distinguish the training group or the test group, thereby establishing a prediction model.

FIGS. 12 to 16 illustrate the process from step A to steps C1 , C2 , C3 and C4 in the aforementioned embodiment of the present invention. The processing unit 13 respectively checks whether each of the categories has reached the threshold value corresponding to the first, second, third, and fourth threshold value groups, and defines a category that has reached the threshold value as a feature , categories that do not meet the threshold are removed. For example, as shown in Figure 12, the maximum actual speed of the category that exceeds the set similarity threshold (0.5) is retained through KS-Test and U-Test; the actual speed that does not reach the set threshold is removed. value. Furthermore, FIG. 12 to FIG. 16 further show that the screened features are subjected to adversarial learning with at least one machine learning algorithm as described in steps d1 and d2 in the embodiment of the present invention to verify the selected features. Results of the similarity of the filtered features. For example, as shown in Figure 12, the AUC value of the adversarial verification result is 0.98, and then the aforementioned similarity threshold is adjusted by grid search. Please refer to FIG. 13 again to leave or remove the class with the adjusted similarity threshold, eg, the command speed in the class is removed in this step. Finally, please refer to Figure 16, until the value of AUC is stable, the 5 features finally screened out in this example are obtained.

Therefore, the present invention has the following advantages:

1. Compared with the existing feature screening method, the present invention adds time merging and domain feature decomposition processing from the original engineering data.

2. The present invention uses a combination of multiple statistical tests to quickly screen and evaluate features, and can quickly screen out effective features without using time-consuming multiple algorithms and conventional correlation layers or model tests.

3. Using similarity to select features can effectively remove features with high correlation but low similarity, reduce the phenomenon of overfitting of subsequent modules, and improve the accuracy of the model.

4. The present invention automates processing, achieves an accuracy similar to that of a data scientist, supports numerical and categorical features, and improves the efficiency of finding valid features

The above are the specific embodiments of the present invention and the technical means used. According to the disclosure or teaching herein, many changes and modifications can be derived and deduced, which can still be regarded as equivalent changes made by the concept of the present invention. If the function does not exceed the substantial spirit covered by the description and drawings, it should be regarded as being within the technical scope of the present invention, and should be stated first.

To sum up, according to the content disclosed above, the present invention can clearly achieve the intended purpose of the invention, and provides a method and system for screening engineering data to obtain features, and a method for screening engineering data multiple times to obtain features , The method for generating prediction models and the system for characterizing engineering data online can quickly evaluate effective features without the use of complex algorithms, which is very valuable for industrial use, and can file invention patent applications in accordance with the law.

A, A1, A2, B1, B2, C1, C2, C3, C4: Steps

Claims (12)

A method for screening engineering data to obtain features, comprising: A: It is judged that a project data is numerical data or categorical data, if the project data is numerical data, then go to step A1, if it is judged that the project data is categorical data, go to step one A2; Wherein the steps A1 and A2 are as follows: A1: Perform the Kolmogorov-Smirnov Test on each category of the numerical data and its corresponding fields to confirm whether the engineering data is normally distributed; if If the test result is a normal distribution, a step B1 is performed, and if the test result is a non-normal distribution, a step B2 is performed; A2: Perform the Cremo V coefficient test on each category and its corresponding fields in the category data to obtain a first threshold value group composed of multiple threshold values, and then perform a step C1; The steps B1 and B2 are as follows: B1: perform T-test test on each category in the numerical data and its corresponding fields to obtain a second threshold value group composed of a plurality of threshold values; then perform a step C2; B2: Check the dispersion degree of the multiple fields corresponding to each category in the numerical data respectively; if it is determined that the dispersion degree of the multiple fields corresponding to each category exceeds a preset value, perform KL dispersion The Kullback-Leibler divergence test is performed to obtain a third threshold group consisting of multiple thresholds, and then step C3 is performed; if the dispersion of multiple fields corresponding to each category in the numerical data is judged If the degree does not exceed a preset value, then perform the Mann-Whitney U test to obtain a fourth threshold value group composed of multiple threshold values, and then perform a step C4; The steps C1, C2, C3, and C4 are as follows: C1: Check whether each of the categories reaches the corresponding threshold value in the first threshold value group, and define a category that has reached the threshold value as a feature; C2: Check whether each of the categories reaches the corresponding threshold value in the second threshold value group, and define a category that has reached the threshold value as a feature; C3: Check whether each of the categories reaches the corresponding threshold value in the third threshold value group, and define a category that has reached the threshold value as a feature; C4: Check whether each of the categories reaches the corresponding threshold value in the fourth threshold value group, and define a category that meets the threshold value as a feature. The method of claim 1, wherein the step A further comprises: Judging that a project data is numerical data or category data, if the project data is numerical data, divide the numerical data into two groups, and perform similarity test on the two groups, and leave the After the significant categories in the numerical data and their corresponding fields, go to step A1; If it is judged that the engineering data is category-type data, the category-type data is divided into two groups and the similarity between the two groups is checked, leaving the category-type data with significantly the same category and its corresponding After entering the field, proceed to step A2. A method for performing multiple screening on engineering data to obtain features, comprising: a. Data cleaning for a project data, including: Compensate for missing values in the fields in the engineering data; b. Generate a plurality of field data groups from the engineering data, wherein the field data groups include: a time field data group, a field field data group or a related field data group; Wherein the step of generating the field field data group includes: generating a value field data group or a The category field data group is disassembled and combined to generate the field field data group; The step of generating the associated field data group includes: There is a significant positive or negative correlation between the numerical field data group and the categorical field data group by a statistical test of association; and Operations and combinations are performed on those with significant positive or negative correlations to generate groups of associated field data; c. Use such time field data groups, domain field data groups or associated field data groups as The method described in claim 1 performs screening to obtain features that have been screened multiple times. The method according to claim 3, further comprising after step c: d1: perform adversarial learning on the screened features with at least one machine learning algorithm to verify the similarity of the screened features; and d2: Adjust the threshold value of the similarity verification by grid search for the screened ones with insignificant feature similarity, and perform step A1 again; keep the screened ones with significant feature similarity. The method according to claim 3, wherein in the step a, the step of compensating for the missing value of the field in the engineering data comprises: using the median to make up the value, using the minimum value to make up the value, using the maximum value to make up the value, Use mode complement value, use quartile complement value or use other similar column complement value; wherein in the step a, it further includes removing the different fields corresponding to these categories but there is no change. The method of claim 3, wherein the step b further comprises the step of generating a category field data group, which includes: merging the first stroke in the category field data group, merging the category field data group The last in the group, the middle in the group of merged fields of this category Pen, merge the most frequent occurrences in the data group of this type of field, or merge the change in the data group of this type of field. The method of claim 3, wherein the step b further comprises the step of generating the time field data group comprising: taking the time corresponding to the fields including: year, month, day, week, hour , minutes, seconds or every 15 minutes. The method of claim 3, wherein the step of generating and generating a field field data group further comprises: The steps of disassembling and combining the numerical field data group to generate the field field data group include: Generate a binary single-digit field data group or generate a decimal single-digit field field data from the numerical field data group; The steps of dismantling and combining the category field data group include: splitting the strings in the class field data group to generate a plurality of split strings; counting the number of occurrences of each of the split strings, leaving split strings above the one count threshold; and Each of the remaining split strings is encoded, each of the encodings being distinct from each other. The method of claim 3, wherein the step of generating the associated field data group further comprises: The correlation test is performed on the numerical field data group, and the numerical field data with significant positive or negative correlation is subjected to the following operations to generate a related field data group: Add, subtract, multiply, divide, take the LOG value, take the angle of a trigonometric function; or Perform a correlation test on this category-type field data group, and there will be categories with significant positive or negative correlations The type field data group performs the following operations to generate the associated field data group: Rearranging the strings; or encoding each of the rearranged strings to generate a group of associated field data, wherein each of the encodings are distinct from each other. A method of producing a predictive model comprising: X: mark the feature generated by the method described in any one of request item 1 to request item 9 as a training group or a test group; Y: mix the filtered features; Z: Distinguish the training group or the test group by at least one machine learning algorithm, thereby establishing a prediction model. A system for screening engineering data to obtain features, comprising: a processor, the processor includes a judgment unit, a statistics unit and a processing unit; The judging unit is used for judging whether a project data is numerical data or category data; Wherein, the statistical unit is used to perform Kolmogorov-Smirnov test (Kolmogorov-Smirnov test) for each category of the engineering data and its corresponding fields if the judgment unit judges that the engineering data is numerical data. -Smirnov Test), to confirm whether the engineering data is normally distributed; If the judging unit judges that the engineering data is not numerical data, the statistical unit will perform the Cremo V coefficient test on each category of the engineering data and its corresponding fields to obtain a plurality of threshold values. The first threshold group of ; If the test result of the statistical unit is a normal distribution, the statistical unit performs T-test test on each category in the engineering data and its corresponding fields to obtain a second threshold value composed of a plurality of threshold values group; Wherein, if the test result of the statistical unit is an abnormal distribution, the statistical unit tests the dispersion degree of the fields corresponding to each category respectively; if the statistical unit determines the dispersion of the fields corresponding to each category If the degree exceeds a preset value, the KL divergence (Kullback-Leibler divergence) test is performed to obtain a third threshold value group composed of multiple threshold values; if the statistical unit determines the multiple fields corresponding to each category The dispersion does not exceed a preset value, then perform the Mann-Whitney U test to obtain a fourth threshold group consisting of multiple thresholds; Wherein the processing unit is used to separately check whether each of the categories has reached the threshold value corresponding to the first, second, third and fourth threshold value groups, and define a category that has reached the threshold value as a feature . A system for characterizing engineering data online, comprising: A server includes a processor, the processor includes a storage unit and a processing unit, wherein the storage unit is used to receive an original project data input from a client and store the original project data; wherein the processing unit Features are obtained by reading the original engineering data from the storage unit and processing the original engineering data with the method described in any one of claims 1 to 9.

Images