Generating training data for machine learning a model to predict the risk of a rare event occurring

Abstract

Generation of training data for machine learning of a model predicting the risk of occurrence of a rare event. The invention relates to a method and a device for generating training data for machine learning of a model predicting the risk of occurrence of a rare event, said training data being formed of vectors of variables representing a state of at least one system monitored before the occurrence of said rare event, from input training data (4) comprising a first initial set (6) of data associated with an absence of occurrence of said rare event, and a second initial set (8) of data associated with an occurrence of said rare event.The device implements modules for determining (20), using a machine learning method, at least a subset of variables associated with a risk of occurrence of said rare event exceeding a risk threshold, and for generating (22) a third dataset by pseudo-random variations on said subset of data, thus providing an augmented training dataset. Figure for the abstract: Figure 1.

Description

Title of the invention: Generation of training data for machine learning of a model for predicting a risk of occurrence of a rare event

[0001] The invention relates to a method and a device for generating learning data for machine learning of a model for predicting a risk of occurrence of a dreaded rare event.

[0002] The invention also relates to a method and a device for automatic learning of a model for predicting a risk of occurrence of an associated rare event.

[0003] Finally, the invention relates to associated computer programs.

[0004] The invention lies in the general field of predicting the risk of occurrence of rare events, in particular the early detection of precursor signals to rare events.

[0005] Recently, the use of artificial intelligence, in particular classification or prediction algorithms trained by machine learning, has enabled great advances in the fields of diagnosis and failure prediction, whether in industrial or railway systems or early diagnosis enabling the detection of risks of developing various diseases by patients.

[0006] However, such artificial intelligence algorithms are efficient when they have been trained on so-called annotated or labeled training data, that is to say, input data, containing variables having values ​​characterizing the monitored data, possibly taken by temporal successions, which are associated with known classifications. When sufficient training data representative of each possible output class of a classification algorithm are provided in such a training phase, the parameters of the classification model implemented by the algorithm are calculated, to ensure the performance of the algorithm.

[0007] Intrinsically, the training data are present in mass for common events, and are rare and unbalanced for rare events. In other words, the distribution of the training data between data representative of current events and data representative of rare events is unbalanced.

[0008] Thus, in a case of application to an industrial system, numerous operating data, for example collected by sensors, are available, which makes it possible to obtain numerous data prior to the occurrence of frequent breakdowns. (wear etc) and very little data prior to the occurrence of exceptional breakdowns.

[0009] Similarly, in the context of a diagnosis of a rare disease, the databases contain little data relating to patients carrying the disease compared to patients not carrying the disease.

[0010] In these cases, we speak of unbalanced learning data, comprising a first set of data associated with an absence of occurrence of a feared rare event, and a second set of data associated with a presence of occurrence of said rare event, the cardinal of the first set being very largely greater than the cardinal of the second set, for example, the cardinal of the first set being greater than 95% of the total cardinal of the two sets, and the cardinal of the second set being less than 5% of the total cardinal.

[0011] Thus, the data of the second set are “drowned” in the training data set, and the classification or prediction algorithm trained by machine learning on such training data is less efficient for predicting the rare event.

[0012] One of the objectives of the invention is to remedy these drawbacks, in order to improve the performance of predicting the occurrence of rare events from initially unbalanced learning databases.

[0013] To this end, the invention proposes, according to one aspect, a method for generating learning data for machine learning of a model for predicting a risk of occurrence of a feared rare event, said learning data being formed from vectors of variables representative of a state of at least one system monitored before the occurrence of said feared rare event, from input learning data comprising a first initial set of data associated with an absence of occurrence of said rare event, and a second initial set of data associated with a presence of occurrence of said rare event.This method comprising steps, implemented by at least one calculation processor, of: - A) determining, from said first and second initial sets of data, by a machine learning method, at least one subset of variables associated with a risk of occurrence of said rare event greater than a risk threshold, said risk being calculated from the values ​​of said variables, . - B) generation of a third set of data by pseudo-random variations on said subset of data, and calculation of an associated risk, - C) grouping said first, second and third data sets into an augmented training data base, - D) application of a first classification model by unsupervised learning on the augmented basis of training data, allowing to obtain at least three groups of training data, comprising a first group of training data associated with a risk of occurrence of said rare event lower than a first threshold; a second group of training data associated with a risk of occurrence of said rare event higher than a second threshold; a third group of training data associated with a third risk, lower than the second threshold and higher than the first threshold.

[0014] Advantageously, this method makes it possible to obtain an augmented base of learning data, and thus to obtain more balanced groups of learning data. Advantageously, this makes it possible to obtain a better prediction of the occurrence of rare events by applying the automatic learning of a model or algorithm for predicting the dreaded rare event, when the balanced learning data are used.

[0015] The method for generating learning data for machine learning of a model for predicting a risk of occurrence of a dreaded rare event according to the invention may also have one or more of the characteristics below, taken independently or according to all technically conceivable combinations:

[0016] The method further comprises:

[0017] E) an application of at least one second machine learning classification model to reclassify at least the training data of the third group into the first group or into the second group, and estimation of an associated reclassification error probability, the first group of training data thus obtained forming a first output set of training data and the second group of training data thus obtained forming a second output set of training data.

[0018] Step E) comprises the application of a plurality of second classification models called classifiers: application of a classifier by semi-supervised learning on the training data of the first group and the training data of the third group, application of a classifier by supervised learning on the training data of the first group and the training data of the second group and application of a classifier by unsupervised learning on the training data of the second group and the training data of the third group.

[0019] If the probability of reclassification error is greater than an error threshold, steps A) to D) are iterated, said first group of training data forming the first initial set of data and said second group of training data forming said second initial set of data for one iteration next.

[0020] In step A) of determination, a random forest automatic learning method is applied.

[0021] According to another aspect, the invention relates to a device for generating learning data for machine learning of a model for predicting a risk of occurrence of a dreaded rare event, said learning data being formed from vectors of variables representative of a state of at least one system monitored before the occurrence of said dreaded rare event, from input learning data comprising a first initial set of data associated with an absence of occurrence of said rare event, and a second initial set of data associated with a presence of occurrence of said rare event, the device comprising at least one calculation processor configured to implement: - a determination module, from said first and second initial sets of data, by a machine learning method,of at least one subset of variables associated with a risk of occurrence of said rare event greater than a risk threshold, said risk being calculated from the values ​​of said variables, , - a module for generating a third set of data by pseudo-random variations on said subset of data, and calculating an associated risk, - a module for grouping said first, second and third sets of data into an augmented learning database, - a module for applying a first classification model by unsupervised learning on the augmented basis of training data, making it possible to obtain at least three groups of training data, comprising a first group of training data associated with a risk of occurrence of said rare event lower than a first threshold; a second group of training data associated with a risk of occurrence of said rare event higher than a second threshold; a third group of training data associated with a third risk, lower than the second threshold and higher than the first threshold.

[0022] The device for generating learning data for machine learning of a model for predicting a risk of occurrence of a dreaded rare event further comprises a module for applying at least a second classification model by machine learning to reclassify at least the learning data of the third group into the first group or into the second group, and estimation of an associated reclassification error probability, the first group of learning data thus obtained forming a first output set of training data and the second group of training data thus obtained forming a second output set of training data.

[0023] The device for generating learning data for machine learning of a model for predicting a risk of occurrence of a dreaded rare event is configured to implement the method for generating learning data for machine learning of a model for predicting a risk of occurrence of a dreaded rare event, according to all its embodiments.

[0024] According to another aspect, the invention relates to a computer program comprising software instructions which, when implemented by a programmable electronic device, implement a method of generating training data for machine learning a model for predicting a risk of occurrence of a feared rare event as briefly described above.

[0025] According to another aspect, the invention relates to a method for automatic learning of a model for predicting a risk of occurrence of a dreaded rare event, implementing a method for generating learning data as briefly described above and a step of learning the values ​​of the prediction model implementing the learning data obtained by the method.

[0026] According to another aspect, the invention relates to a device for automatic learning of an associated prediction model, this device for automatic learning of an automatic prediction model of a model for predicting a risk of occurrence of a dreaded rare event comprising a device for generating learning data as described above and a module for learning the values ​​of the prediction model implementing the learning data obtained by said device for generating learning data.

[0027] Another relates to a computer program comprising software instructions which, when executed by a programmable device, implement a machine learning method as briefly described above.

[0028] Other characteristics and advantages of the invention will emerge from the description given below, for information purposes only and in no way limiting, with reference to the appended figures, among which:

[0029] [Fig-1] [Fig.l] schematically represents a system comprising a device for generation of training data for machine learning of a model for predicting the risk of occurrence of a dreaded rare event.;

[0030] [Fig.2] [Fig.2] is a synopsis of the main steps of a method for generating learning data according to a first embodiment;

[0031] [Fig.3] [Fig.3] is a block diagram of the main steps of a method for generating learning data according to a second embodiment;

[0032] [Fig.4] [Fig.4] is a synopsis of the main stages of a process machine learning of a prediction model of a risk of occurrence of a dreaded rare event and implementation of said prediction model.

[0033] [Fig.l] schematically illustrates a system 2 for generating learning data for machine learning of a model for predicting a risk of occurrence of a dreaded rare event.

[0034] System 2 is presented in a generic manner. The applications of such a system are numerous, for example for predicting a risk of exceptional malfunction of a monitored industrial system or for diagnosing a patient's risk of suffering from a rare disease.

[0035] Generally speaking, a monitored system is a system in which the feared rare event is likely to occur.

[0036] For example, in industrial application, a monitored system is a type of plant or set of equipment to be monitored.

[0037] In the diagnostic application, the monitored system is a patient's organism, based on recorded data.

[0038] For the prediction of the risk of occurrence of a rare event, it is planned to use a prediction model, which is trained by machine learning on training data (or input data) representative of characteristics of monitored systems of the same type.

[0039] In other words, for the industrial application, for example if the monitored system is a coal-fired power plant, the training data are data representative of one, or preferably several, coal-fired power plants, taken over time. In addition, characteristics such as the type and age of the equipment used are also provided.

[0040] For example, for the diagnostic application, data relating to patients, for example results of in vitro analysis tests, and patient characteristics (eg age, sex) are stored.

[0041] It is clear that the training data are data adapted for the intended application.

[0042] Preferably, these data are organized in the form of vectors of variables having values, representative of a state of monitoring data of the monitored system. In other words, each vector is associated with a state of a monitored system.

[0043] Of course, a plurality of vectors can be associated with the same monitored system.

[0044] The values ​​of the variables are for example provided by sensors and recorded over time (e.g. alarm feedback, operating measurement reading, deviation or anomaly, technical fact).

[0045] The training data is recorded before the occurrence of a rare event feared, but for at least part of the training data, the occurrence of the rare event was observed (i.e. feared dysfunction or declared rare disease).

[0046] The training data are collected in an input training database 4, stored in a data storage system 5.

[0047] The database 4 comprises a first initial set 6 of data associated with an absence of occurrence of said rare event. A probability of occurrence of the rare event is also stored, this probability of occurrence coming from a priori knowledge, for example either from the opinion of an expert in the field in the case of supervised learning, or from the results resulting from the prior implementation of unsupervised Machine Learning algorithms, such as Association or Clustering, which makes it possible to associate, or to bring together in the same class, events with the same probability of occurrence and / or the same level of risk.

[0048] The database 4 comprises a second initial set 8 of data associated with the presence of occurrence of said rare event, i.e. these are data taken from monitored systems for which the feared rare event has occurred.

[0049] The first initial set 6 has a cardinal N1 (i.e. comprises N1 vectors of variables), the second initial set 8 has a cardinal N2 (i.e. comprises N2 vectors of variables).

[0050] In database 4, the number NI is substantially higher than N2, i.e. N2< X / 100 x (N1+N2), where X is for example less than 10, or even less than 5, and even less than 1.

[0051] One of the objectives of the training data generation system 2 is to provide better balanced training data sets, so as to better characterize the risk of occurrence of the feared rare event.

[0052] The system 2 for generating learning data comprises a device 10 for generating learning data, which is for example a programmable electronic device or a plurality of interconnected programmable electronic devices.

[0053] The device 10 comprises in particular a calculation unit 12, comprising one or more calculation processors, and an electronic memory 14, adapted to communicate via a data communication bus (not shown).

[0054] The device 10 is also adapted to communicate, in reading and writing, with the data storage system 5, which comprises the input training database 4.

[0055] Similarly, the device 10 is also adapted to communicate, in reading and writing, with a data storage system 15, configured to store an output training database 16, described in more detail below.

[0056] Each of the data storage systems 5, 15 is a computer-readable medium, for example, a medium capable of storing electronic instructions and of being coupled to a bus of a computer system. For example, the readable medium is an optical disk, a magneto-optical disk, a ROM memory, a RAM memory, any type of non-volatile memory (for example EPROM, EEPROM, FLASH, NVRAM), a magnetic card or an optical card.

[0057] The device 10 comprises the following modules, implemented by the computing unit 10:

[0058] - a module 20 for determining, from the first and second initial sets 6, 8 of data, by an automatic learning method, at least one subset of variables associated with a risk of occurrence of the rare event greater than a risk threshold, said risk being calculated from values ​​of said variables;

[0059] - a module 22 generation of a third set of data by variations pseudo-random on said subset of data, and calculation of an associated risk;

[0060] - a module 24 for grouping the first, second and third sets of data into an augmented training database 25;

[0061] - a module 26 for applying a first classification model by learning unsupervised on the database augmented with training data, allowing to obtain:

[0062] a first group of learning data 30 associated with a risk of occurrence of the rare event lower than a first threshold SI;

[0063] a second group of learning data 32 associated with a risk of occurrence of the rare event greater than a second threshold S2;

[0064] a third group of learning data 34 associated with a third risk, lower than the second threshold S2 and higher than the first threshold S1;

[0065] - a module 28 for applying at least a second classification model by machine learning to reclassify at least the training data of the third group 34 into the first group or the second group, and estimating an associated reclassification error probability.

[0066] In one embodiment, the module 28 implements three distinct second classification models (also called classifiers), comprising an unsupervised classifier, a semi-supervised classifier and an unsupervised classifier.

[0067] It is planned to use the modules 20, 22, 24, 26 and 28 iteratively, until the probability of reclassification error is considered satisfactory, i.e. lower than a predetermined error threshold.

[0068] At output is obtained a learning database 16 comprising at least a first set of learning data output 36 and a second set of learning data output 38, obtained after application of the module 28, each output set 36, 38 having an associated risk of occurrence of the dreaded rare event, 36', 38'. For example, the risk 36' of occurrence of the dreaded rare event, associated with the first output set 36, is low, for example lower than a first risk threshold, for example lower than 15% or lower than 5% probability of occurrence, and the risk 38' of occurrence of the dreaded rare event, associated with the second output set 38, is high, for example higher than 55% or 65% probability of occurrence.

[0069] Advantageously, the cardinal N' 1 of the first output set of training data 36 is less than the cardinal NI of the first initial set 6, and the cardinal N'2 of the second output set of training data 38 is greater than the cardinal N2 of the first initial set 8. In other words, a rebalancing of the training data is carried out, and an associated risk is also calculated. To the extent that the training data are associated with a risk of occurrence of the feared rare event, the rebalanced training database also includes information relating to the risk. It is also said that the training data obtained are "informed" of the risk of occurrence of the feared rare event.

[0070] In one variant, more than two output sets are obtained, each of the output sets having an associated risk of occurrence of the feared rare event.

[0071] In one embodiment, the modules 20, 22, 24, 26 and 28 are implemented in the form of software code, and form a computer program, comprising software instructions which, when implemented by a programmable electronic device, implement a method for generating learning data for machine learning of a model for predicting a risk of occurrence of a dreaded rare event.

[0072] In a variant not shown, the modules 20, 22, 24, 26 and 28 are each produced in the form of a programmable logic component, such as an FPGA (Field Programmable Gate Array), or a GPGPU (General-purpose processing on graphics processing), or in the form of a dedicated integrated circuit, such as an ASIC (Application Specific Integrated Circuit).

[0073] The computer program for generating learning data is further capable of being recorded on a medium, not shown, that is readable by a computer. The computer-readable medium is, for example, a medium capable of storing electronic instructions and of being coupled to a bus of a computer system. By way of example, the readable medium is an optical disk, a magneto-optical disk, a ROM memory, a RAM memory, any type of non-volatile memory (for example EPROM, EEPROM, FLASH, NVRAM), a magnetic card or an optical card.

[0074] [Fig.2] is a synopsis of the main stages of a process for generating training data in a first embodiment, for a monitored system and a feared rare event.

[0075] As input, the input training database 4 is provided, comprising the first initial set of training data corresponding to the absence of occurrence of the rare event, and the second initial set of data corresponding to the occurrence of the rare event.

[0076] Each set comprises the same types of data, including vectors of variables representing states of one or more monitored system(s) of the same type.

[0077] The method comprises a first step 40, implemented by the module 20, of determining from said first and second initial sets of data, by a machine learning method, at least one subset of variables associated with a risk of occurrence of said rare event greater than a risk threshold, said risk being calculated from values ​​of said variables.

[0078] In other words, step 40 performs artificial intelligence-assisted modeling of risk exposure.

[0079] For example, the machine learning method implemented in step 40 is a random forest learning method with bootstrap, a well-known random sampling technique with replacement. Advantageously, this makes it possible to generate a distribution of estimates instead of a single point estimate, and to obtain uncertainty values ​​of the associated estimates. Alternatively, the XGBoost (for “eXtreme Gradient Boosting”) algorithm is implemented.

[0080] According to other variants, in step 40 a logistic regression or a support vector machine or SVM is implemented.

[0081] Step 40 makes it possible to determine states of the monitored system that are precursors to the feared rare event.

[0082] These states are characterized by variables having given values.

[0083] A risk of occurrence of the associated dreaded rare event is calculated by a chosen formula.

[0084] For example, the risk is calculated as follows:

[0085] Risk In f arm£C iV ariable = Occ^Variable 2 ^A

[0086] With Variable designating the variable of interest associated with a given risk (an expression of oncogene or a biomarker in Health for example) or an impact variable (severity of a consequence induced by an anomaly in an industrial environment for example);

[0087] Occ designating the occurrence or frequency of realization of this risk, associated with this variable;

[0088] A designating the function of attenuation or amplification of the risk depending on the overall state of the system in industrial application or the overall state of health of a patient (comorbidities, environmental and / or occupational exposures at risk, etc.).

[0089] In one embodiment, A is a multiplying factor.

[0090] In an example carried out on clinical and genetic data for the prediction of patients at high risk of developing an aggressive form of clear cell kidney cancer, we first carried out a logarithmic normalization of gene expressions to rebalance the weight of each gene with the other available clinical data to ensure balanced consideration and a standardized predictive approach to the Machine Learning classification algorithms.

[0091] In this example, we have introduced and generated for each initial normalized data (data), a new associated data (data_risk) which is defined as a function modeling the risk carried by the initial data, according to the following formula:

[0092] data-i = Occurrence*data 2 *AJ

[0093] A predetermined risk threshold is used, depending on the field of application, for example a risk threshold between 5% and 15% of a predetermined risk level, for example maximum, depending on the recommendations and feedback from the field of application.

[0094] The method then comprises a step 42 of generating a third set of data by pseudo-random variations on the subset of data obtained in step 40, and calculating an associated risk.

[0095] This makes it possible to artificially generate data corresponding to states of the monitored system that are precursors to the dreaded rare event.

[0096] In one embodiment, step 42 implements a grouping algorithm (or “clustering”), for example the K-class grouping algorithm known as “K-means” (or K-means in French) to adapt the introduction of pseudo-random variations. Step 42 makes it possible to find hidden correlations between the different unannotated (or unlabeled) events by grouping them into equivalent risk classes.

[0097] The first, second initial sets and the third set of data generated in step 42 are grouped in the grouping step 44 into an augmented base.

[0098] The method then comprises a step 46 of applying a first classification model by unsupervised machine learning on the database augmented with training data, making it possible to obtain three groups of training data 30, 32, 34 defined previously, which are respectively:

[0099] a first group of learning data 30 associated with a risk of occurrence of the rare event lower than a first threshold SI;

[0100] a second group of learning data 32 associated with a risk of occurrence of the rare event greater than a second threshold S2, the threshold S2 being greater than S1;

[0101] a third group of learning data 34 associated with a third risk, lower than the second threshold S2 and higher than the first threshold SI.

[0102] For example, the threshold SI is between 5% and 10% probability of occurrence, the threshold S2 is between 15% and 25% probability of occurrence.

[0103] Unlike supervised learning, unsupervised learning allows the algorithm to learn and predict with unannotated (or unlabeled) examples, independently. Unsupervised learning algorithms are notably used to classify data, calculate the distribution density and its associated uncertainties, and reduce dimensions in the case of a large number of variables much greater than the number of individuals (example in Health, 20,000 genes for a few hundred cancer patients).

[0104] For example, step 46 implements a “clustering” or “association” method, this type of method being reference methods in the field of unsupervised machine learning. Other methods can also be applied, for example principal component analysis (PCA).

[0105] Step 46 is followed by a step 48 of applying at least one second classification model (or classifier) ​​by machine learning to reclassify the training data, and calculate an associated reclassification error probability.

[0106] Classification errors are associated with uncertainties associated with data, data collection and processing systems, and also models and algorithms. In the state of the art, several methods can address this point by deploying the so-called resampling strategy.

[0107] We thus distinguish between oversampling and undersampling.

[0108] Oversampling methods work by increasing the number of observations of the minority class(es) in order to achieve a satisfactory minority class / majority class ratio.

[0109] Undersampling methods work by decreasing the number of observations of the majority classes until a balanced ratio is reached.

[0110] There are several machine learning algorithms for generating synthetic samples automatically. The most popular of these algorithms is SMOTE (for Synthetic Minority Over-sampling Technique). SMOTE is an over-sampling method that works by creating syn- aesthetics from the minority class instead of creating simple copies.

[0111] The ClusterCentroids algorithm is an Undersampling algorithm based on Clustering methods to generate a number of centroids from the original data, without losing information about the majority class when reducing it.

[0112] In this embodiment, step 48 comprises a substantially parallel implementation of three distinct classifiers in sub-steps 50, 52, 54.

[0113] For example, sub-step 50 implements a semi-supervised machine learning classifier, on the training data of the first group and the training data of the third group. Semi-supervised learning makes it possible to train a model on a partially annotated dataset which includes some annotated data and a majority of unannotated data. The idea is to assign the annotations using the similarity between a small number of annotated data and a larger volume of unannotated data. For example, sub-step 50 implements a Generative Adversarial Networks (or GAN) model.

[0114] For example, sub-step 52 implements a classifier by supervised machine learning, for example a support vector machine (or SVM for “Support Vector Machine”) or a logistic regression, on the training data of the first group and the training data of the second group.

[0115] For example, sub-step 54 implements a classifier by unsupervised machine learning, for example a Clustering or Association algorithm, on the training data of the second group and the training data of the third group.

[0116] Advantageously, the respective classifiers are implemented independently, which makes it possible to evaluate and quantify the impact of uncertainties in order to reduce them.

[0117] The method then comprises a step 56 of comparing the minimum of the reclassification error probability, among the three sets, with a predetermined error threshold SP (or reclassification error probability threshold). If the minimum of the reclassification error probability is greater than the error threshold SP, steps 40 to 48 are iterated.

[0118] Otherwise, the method comprises a step 58 of obtaining at least two sets 36, 38 of learning data output, and of calculating the risk of occurrence of the associated rare event.

[0119] In a second embodiment of the method for generating learning data for machine learning of a model for predicting a risk of occurrence of a dreaded rare event, the method comprises steps 40 to 44

[0120]

[0121]

[0122]

[0123]

[0124]

[0125] analogous to the steps described above. Unlike the first embodiment, the method comprises a step 46' of applying a first classification model by unsupervised machine learning on the database augmented with training data, making it possible to obtain a number G greater than 3 of training data groups, for example a number G=4. Each data group is associated with a probability of risk of occurrence of the rare event included in a range of values, for example: - A first group of learning data 31 associated with a risk of occurrence of the rare event between 0% and 25% probability of occurrence of the rare event; - A second group of learning data 33 associated with a risk of occurrence of the rare event between 25% and 50% probability of occurrence of the rare event; - A third group of learning data 35 associated with a risk of occurrence of the rare event between 50% and 75% probability of occurrence of the rare event; - A fourth group of training data 37 associated with a risk of occurrence of the rare event between 75% and 100% probability of occurrence of the rare event. We understand that this is generalizable to any number G of learning groups. The method in the second embodiment then comprises a step 48', analogous to step 48 previously described, of applying at least one second classification model (or classifier) ​​by machine learning to reclassify the learning data, and calculate an associated reclassification error probability. In this embodiment, several groups of training data are provided as inputs to the first, second and third classifiers 50, 52, 54. Various groupings of the training data groups provided as input to the first, second and third classifiers are possible, these input groupings being carried out for example according to the number of unbalanced data. For example, in the illustrated embodiment, the training data of the first group 31 and the third group 35 are provided as input to the first classifier, the training data of the first group 31 and the fourth group 37 are provided as input to the second classifier, the training data of the second group 33 and the fourth group 37 are provided as input to the third classifier.

[0126] Alternatively, other groupings are envisaged, depending on the application and / or the respective cardinalities of the training data groups.

[0127] The following steps 56, 58 are analogous to the steps previously described with reference to [Fig.2].

[0128] [Fig.4] illustrates a method of using the generated training data.

[0129] This method comprises a machine learning phase of the prediction model and a subsequent phase of use for the effective prediction of the risk of occurrence of a rare event in a monitored system, for example an industrial monitored system.

[0130] The method comprises, in the learning phase, a reception of the learning database 16 comprising the at least two sets 36, 38 of learning data output generated by the learning data generation method described above. The learning data are of course data representative of characteristics of monitored systems of a given type, which have been balanced by the method described above.

[0131] The method comprises a step 80 of machine learning a model for predicting the risk of occurrence of a given dreaded rare event on the balanced learning data.

[0132] The method also comprises a storage 82 of the prediction model, i.e. of its architecture and of the parameter values ​​characterizing it.

[0133] The method comprises, in a second phase of effective use, the provision as input of input data 84 characteristics of a monitored system, for which the risk of occurrence of the rare event is not known a priori.

[0134] The method then comprises a step 86 of implementing the prediction model stored in step 82, on the input data 84, and obtaining an effective risk R of occurrence of the associated rare event.

[0135] If the actual risk is greater than a predetermined critical risk threshold (step 88), the method comprises the implementation 90 of a prevention action: for example, in the case of a monitored industrial system, sending an alert, with associated information where appropriate, making it possible to determine preventive actions: predictive maintenance, shutdown of targeted equipment, etc.

[0136] The critical risk threshold is defined according to the application.

[0137] The method of using the training data is implemented by one or more programmable electronic devices.

[0138] For example, the method is implemented by a programmable electronic device materially analogous to the device 10 for generating learning data, which further comprises a module for learning the values ​​of the prediction model implementing the learning data obtained, and / or a module for implementing the stored prediction model, and a module for applying a prevention action following the determination of an effective risk greater than a critical risk threshold.

[0139] Advantageously, the invention thus proposes an automated device and method for balancing unbalanced data by generating new data, informed of the risk (in English “Risk-informed”), i.e. stored in association with an associated risk value, for the automatic learning of a model or algorithm for predicting exposure to a dreaded rare event.

[0140] Advantageously, the inventors have shown that the performance score characterizing the performance of prediction algorithms (or models) trained by machine learning, for example the Fl-score, which evaluates the capacity of a classification model to effectively predict positive individuals by making a compromise between precision and recall, is systematically improved when the balanced learning database is used.

Claims

1. Claims Method for generating learning data for machine learning of a model for predicting a risk of occurrence of a dreaded rare event, said learning data being formed from vectors of variables representative of a state of at least one system monitored before the occurrence of said dreaded rare event, from input learning data (4) comprising a first initial set (6) of data associated with an absence of occurrence of said rare event, and a second initial set (8) of data associated with a presence of occurrence of said rare event, the method comprising steps, implemented by at least one calculation processor, of: - A) determination (40), from said first and second initial sets of data, by a machine learning method, of at least one subset of variables associated with a risk of occurrence of said rare event greater than a risk threshold, said risk being calculated from the values ​​of said variables, - B) generation (42) of a third set of data by pseudo-random variations on said subset of data, and calculation of an associated risk, - C) grouping (44) said first, second and third sets of data into an augmented training database, - D) application (46) of a first classification model by unsupervised learning on the augmented basis of training data, making it possible to obtain at least three groups (30, 32, 34; 31, 33, 35, 37) of training data, comprising a first group (30) of training data associated with a risk of occurrence of said rare event lower than a first threshold; a second group (32) of training data associated with a risk of occurrence of said rare event higher than a second threshold; a third group (34) of training data associated with a third risk, lower than the second threshold and higher than the first threshold.

2. The method of claim 1, further comprising A. applying (48) at least one second machine learning classification model to reclassify at least the training data of the third group into the first group or the second group, and estimating an associated reclassification error probability, the first group of training data thus obtained forming a first output set of training data (36) and the second group of training data thus obtained forming a second output set of training data (38).

3. Method according to claim 2, in which step E) comprises the application (48) of a plurality of second classification models called classifiers: application (50) of a classifier by semi-supervised learning on the training data of the first group and the training data of the third group, application (52) of a classifier by supervised learning on the training data of the first group and the training data of the second group and application (54) of a classifier by unsupervised learning on the training data of the second group and the training data of the third group.

4. A method according to claim 2 or 3, wherein if said reclassification error probability is greater than an error threshold, steps A) to D) are iterated, said first group of training data forming the first initial set of data and said second group of training data forming said second initial set of data for a subsequent iteration.

5. Method according to one of claims 1 to 4, in which in step A) of determination a method of automatic learning by random forests is applied.

6. Method for automatic learning of a model for predicting a risk of occurrence of a dreaded rare event, implementing a method for generating learning data according to claims 1 to 5 and a step of learning (80) the values ​​of the prediction model implementing the learning data obtained by the method for generating learning data.

7. A computer program comprising software instructions which, when executed by a programmable device, implement a method of generating training data according to claims 1 to 5.

8. A computer program comprising software instructions which, when executed by a programmable device, implement a machine learning method according to claim 6.

9. Device for generating learning data for machine learning of a model for predicting a risk of occurrence of a dreaded rare event, said learning data being formed from vectors of variables representative of a state of at least one system monitored before the occurrence of said dreaded rare event, from input learning data (4) comprising a first initial set (6) of data associated with an absence of occurrence of said rare event, and a second initial set (8) of data associated with a presence of occurrence of said rare event, the device comprising at least one calculation processor configured to implement: - a module (20) for determining, from said first and second initial sets of data, by a machine learning method, at least one subset of variables associated with a risk of occurrence of said rare event greater than a risk threshold,said risk being calculated from the values ​​of said variables, - a module (22) for generating a third set of data by pseudo-random variations on said subset of data, and calculating an associated risk, - a module (24) for grouping said first, second and third sets of data into an augmented base of learning data, - a module (26) for applying a first classification model by unsupervised learning on the augmented base of learning data, making it possible to obtain at least three groups of learning data, comprising a first group of learning data associated with a risk of occurrence of said rare event lower than a first threshold; a second group of learning data, associated with a risk of occurrence of said rare event greater than a second threshold; a third group of training data associated with a third risk, lower than the second threshold and higher than the first threshold.

10. Device for generating learning data for machine learning of a model for predicting a risk of occurrence of a dreaded rare event according to claim 9, further comprising a module (28) for applying at least a second classification model by machine learning to reclassify at least the learning data of the third group in the first group or in the second group, and estimation of an associated reclassification error probability, the first group of learning data thus obtained forming a first output set of learning data (36) and the second group of learning data thus obtained forming a second output set of learning data (38).

11. Device for automatic learning of a model for predicting a risk of occurrence of a dreaded rare event, comprising a device for generating learning data according to claim 9 or 10 and a module for learning the values ​​of the prediction model implementing the learning data obtained by said device for generating learning data.

Images