Automatic dataset preprocessing

1. A data processing method for preparing a dataset implemented by a processor that receives a plurality of data input streams from different data providers, wherein the plurality of data input streams pass through said processor to prepare an output dataset, wherein the plurality of data input streams and the output dataset are different,

said data processing method comprising:

providing said processor, wherein via said processor,

receiving the plurality of data input streams from said different data providers that comprise industrial production sensors;

standardizing the plurality of data input streams to normalize the plurality of data input streams at an input of the processor;

encoding the plurality of data input streams that are normalized, to transform the plurality of data input streams that are normalized into a plurality of normalized and encrypted data input streams;

preprocessing the dataset from the plurality of normalized and encrypted data input streams, said preprocessing comprising

data augmentation to complete missing data taking into account a consistency of the plurality of normalized and encrypted data input streams with predefined correlation tables,

providing functions that comprise a plurality of different functions that allow a detection of outliers in the plurality of normalized and encrypted data input streams,

removing said outliers from the plurality of normalized and encrypted data input streams, by processing said plurality of normalized and encrypted data input streams with said functions,

verifying a fit of the plurality of normalized and encrypted data input streams processed with the functions, said verifying comprising providing feedback to take actions aiming at injecting the plurality of normalized and encrypted data input streams that are processed back into the processor when the fit of the plurality of normalized and encrypted data input streams that are processed with the functions is divergent, and

selecting a function of said functions for which the fit between said function and one of the plurality of data input streams comprises a smallest discrepancy;

generating a preprocessed dataset, said preprocessed dataset including

said plurality of normalized and encrypted data input streams,

new data generated during said data augmentation, and

new data replacing said outliers;

displaying to a user, on a display associated with said processor,

said functions,

said consistency of the plurality of normalized and encrypted data input streams with predefined correlation tables,

the fit of the plurality of normalized and encrypted data input streams processed with the functions, and

a table containing fit indices for each variable present in the plurality of normalized and encrypted data input streams; and,

transmitting said preprocessed dataset to a machine learning model that is trained to monitor an industrial process via supervised or unsupervised learning techniques,

such that said machine learning model is trained to one or more of predict maintenance, detect failure, detect fraud and detect cyber-attacks.

2. The data processing method according to claim 1 , wherein the industrial production sensors include connected objects, or machine sensors, or environmental sensors, or computing probes, or any combination thereof.

3. The data processing method according to claim 1 , wherein the industrial process is an agri-food production process, or a manufacturing production process, or a chemical synthesis process, or a packaging process, or a process for monitoring an IT infrastructure.

4. The data processing method according to claim 1 , further comprising automatically modifying one or more of said functions for said removing said outliers and said predefined correlation tables for said data augmentation.

5. The data processing method according to claim 4 , wherein the preprocessed dataset is used as input data to a learning model trained for monitoring said industrial process and wherein the automatically modifying is initiated when a variance is identified in the plurality of data input streams.

6. The data processing method according to claim 1 , further comprising, via said processor, determining a reliability indicator and a validity indicator, wherein when said reliability indicator and said validity indicator are respectively lower than a predetermined value, the plurality of data input streams of the dataset is injected back into the processor, via said processor.

7. The data processing method according to claim 1 , further comprising generating an evaluation of the plurality of data input streams to verify said fit, by said processor, wherein said generating said evaluation comprises comparing correlation indices for each data subsets of the plurality of data input streams and identifying, for each of the data subsets, predetermined distribution functions allowing a correlation higher than a predetermined threshold.

8. The data processing method according to claim 1 , wherein the encoding the plurality of data input streams that are normalized comprises

identifying data to be encoded,

modifying the data that is identified into encoded data, and

storing the encoded data.

9. The data processing method according to claim 1 , wherein, via said processor, the data augmentation to complete the missing data further comprises classifying the missing data by assigning a score or a ranking.

10. A data processing system for preparing a dataset implemented by a processor that receives a plurality of data input streams from different data providers, wherein the plurality of data input streams pass through said processor to prepare an output dataset, wherein the plurality of data input streams and the output dataset are different,

said data processing system comprising:

a processor configured to

receive the plurality of data input streams from said different data providers that comprise industrial production sensors;

standardize the plurality of data input streams to normalize the plurality of data input streams at an input of the processor;

encode the plurality of data input streams that are normalized, to transform the plurality of data input streams that are normalized into a plurality of normalized and encrypted data input streams;

preprocess the dataset from the plurality of normalized and encrypted data input streams, said preprocess comprising

completing missing data via data augmentation taking into account a consistency of the plurality of normalized and encrypted data input streams with predefined correlation tables,

providing functions that comprise a plurality of different functions that allow a detection of outliers in the plurality of normalized and encrypted data input streams,

removing said outliers from the plurality of normalized and encrypted data input streams, by processing said plurality of normalized and encrypted data input streams with said functions,

verifying a fit of the plurality of normalized and encrypted data input streams processed with the functions, said verifying comprising providing feedback to take actions aiming at injecting the plurality of normalized and encrypted data input streams that are processed back into the processor when the fit of the plurality of normalized and encrypted data input streams processed with the functions is divergent, and

selecting a function of said functions for which the fit between said function and one of the plurality of data input streams comprises a smallest discrepancy;

generate a preprocessed dataset, said preprocessed dataset including said plurality of normalized and encrypted data input streams, new data generated during said data augmentation, and new data replacing said outliers;

display to a user, on a display associated with said processor,

said functions,

said consistency of the plurality of normalized and encrypted data input streams with predefined correlation tables,

the fit of the plurality of normalized and encrypted data input streams processed with the functions, and

a table containing fit indices for each variable present in the plurality of normalized and encrypted data input streams; and,

transmit said preprocessed dataset to a machine learning model that is trained to monitor an industrial process via supervised or unsupervised learning techniques,

such that said machine learning model is trained to one or more of predict maintenance, detect failure, detect fraud and detect cyber-attacks.