FR3145426A1 - Method for detecting operating anomalies of industrial equipment based on anomaly scores and corresponding installation - Google Patents

Abstract

Method for detecting operating anomalies of industrial equipment from anomaly scores and corresponding installation Method for controlling industrial equipment to detect operating anomalies of the industrial equipment, comprising the following steps: a) obtaining an initial model adapted to calculate anomaly scores from a plurality of operating parameters, the initial model resulting from a machine learning process, b) obtaining a plurality of values defining at least one operating point of the industrial equipment, c) calculating the anomaly score associated with the operating point, d) displaying an image showing anomaly scores calculated by the initial model, as a function of at least two operating parameters, e) determining, by a user, at least one desired correction in the displayed image, f) from the initial model and the desired correction, obtaining a corrected model, g) displaying a corrected image showing the corrected anomaly scores as a function of the two coordinates, steps b) to g) being optionally iterated, h) obtaining of a plurality of values of the operating parameters defining an operating point, and i) calculating the anomaly score associated with said operating point using the corrected model obtained.

Description

Method for detecting operating anomalies of industrial equipment from anomaly scores and corresponding installation

The present invention relates to a method for controlling industrial equipment to detect operating anomalies of the industrial equipment, comprising obtaining a model adapted to calculate anomaly scores from a plurality of operating parameters representative of operating states of the industrial equipment, the anomaly scores being respectively representative of probabilities that the operating states are abnormal, the initial model resulting from a machine learning process.

The invention also relates to a corresponding installation, comprising the industrial equipment, sensors adapted to provide the plurality of operating parameters, and a control system adapted to implement certain of the steps of such a method.

Industrial equipment is for example a production line, a fluid compression unit, or any equipment with an industrial purpose and whose health we wish to monitor, that is to say to know if it is functioning normally or if its functioning is abnormal and requires, for example, a shutdown, maintenance, or any act of administration of the industrial equipment. The operating parameters are for example pressures, temperatures, vibration levels, etc.

In this area, two main approaches are known.

According to a first main approach, operating parameters are identified by experts and monitored to ensure that these parameters evolve within ranges considered normal. The advantage of this first strategy is that it is based on a physical knowledge of industrial equipment and more generally on physical laws and standards. For example, maximum acceptable vibration levels are defined for a rotating machine, according to a standard, depending on the power of the machine.

However, a first drawback of this strategy is that the usual context of use of the rotating machine is not integrated, while the machine generally operates in a reduced range of the normal operating range. Thus, this strategy generally does not allow early detection of a change in behavior linked to a degradation of the industrial equipment, as long as the parameters remain in the normal range.

Furthermore, monitoring the joint evolution of several operating parameters can also be considered, but its implementation sometimes proves to be complex, or even prohibitive, particularly in the case of a large number of operating parameters to be monitored. This leads machine operators to monitor only a small number of operating parameters, and separately, which makes monitoring very partial.

In a second main approach, the normal behavior(s) of the machine are learned from one or more operating parameters using learning algorithms. This creates a model providing, for each possible operation of the machine, an anomaly score representing the probability that the operation is abnormal. The anomaly score is for example a probability between 0 and 1, the value “1” meaning that the operation is abnormal with a maximum probability, the value “0” meaning that the operation is normal with a maximum probability.

The advantage of this second approach is to know, after a learning phase on a data history, the normal operations defined by several operating parameters, whether by their individual values, but also in relation to each other. This allows to detect behavioral anomalies very early.

However, a major drawback of this second approach is that obtaining a high-performance model, combining both good sensitivity and good specificity, requires a large data history, representative of all normal behaviors of the machine. In particular, since the relationships between sensors are monitored, the size of this history is closely linked to the number of sensors monitored. Thus, it is unthinkable to have this representative history and obtain an optimal model for calculating anomaly scores. In practice, industrial equipment monitoring is carried out while only an imperfect model is available.

In order to improve the model, one or more re-parameterizations of the model are carried out over time, in order to take into account the good and bad anomaly detections carried out. By re-parameterization, we mean here a modification of the parameters defining the model.

However, these re-parameterizations are complex and tedious, due to their frequency and the quantity of information to be exchanged. As a result, they are not always well accepted by a user of the control system, typically a technician, who must qualify all the anomalies, and for whom the re-parameterizations can cause a significant work overload.

Another disadvantage of this second approach is that it does not integrate the physical or technical knowledge that one may have of industrial equipment, which can eventually lead to false alarms in trivial operating cases.

An aim of the invention is to remedy all or part of the above drawbacks, by providing a control method for detecting operating anomalies of industrial equipment, which is more reliable and/or easier to implement.

To this end, the invention relates to a method for controlling industrial equipment to detect operating anomalies of the industrial equipment, comprising the following steps:

a) obtaining an initial model suitable for calculating anomaly scores from a plurality of operating parameters representative of operating states of the industrial equipment, the anomaly scores being respectively representative of probabilities that the operating states are abnormal, the initial model resulting from a machine learning process,

(b) obtaining a plurality of values of the operating parameters defining at least one operating point of the industrial equipment,

c) calculation, using the initial model, of the anomaly score associated with the operating point,

(d) displaying at least one image showing anomaly scores calculated by the initial model, as a function of at least two of the operating parameters forming two coordinates of the image, the image showing the anomaly score of the operating point,

(e) determination, by a user, of at least one desired correction in the displayed image, the desired correction relating to at least one correction point, or at least one correction zone,

(f) from the initial model and the desired correction, obtaining a corrected model adapted to calculate, from the plurality of operating parameters, corrected anomaly scores,

g) display of a corrected image showing the corrected anomaly scores based on the two coordinates,

steps b) to g) being optionally iterated over time, the corrected model of one iteration becoming the initial model of a following iteration,

(h) obtaining a plurality of values of the operating parameters defining an operating point of the industrial equipment, and

i) calculation of the anomaly score associated with said operating point using the corrected model obtained during the last of the iterations.

According to particular embodiments, the method comprises one or more of the following characteristics, taken in isolation or in all technically possible combinations:

- step a) includes a machine learning sub-step to obtain the initial model;

- in step h), and optionally in step b), measurements are carried out by a plurality of sensors of the industrial equipment to obtain the values of the operating parameters;

- the method comprises a step j) of development by the user or a machine, from said anomaly score obtained in step i), of an administration decision of the industrial equipment intended to modify the operating state of the industrial equipment;

- in step e), the desired correction relates to an area of the displayed image; and in step f), the calculation of the corrected model comprises a summation of the initial model and a penalty function having as variables at least said coordinates of the image, the penalty function being adapted to carry out in the corrected image a correction of predefined shape, for example spherical, elliptical, rectangular, or by band;

- in step e), the desired correction relates to at least one correction point belonging to the displayed image; and in step f), the calculation of the corrected model comprises a summation of the initial model and a penalty function having as variables at least said coordinates of the image, the penalty function comprising a term associated with the correction point, the term being, in absolute value, maximum for said correction point and decreasing as a function of a distance from said correction point;

- said term is proportional to a Gaussian function having said distance as a variable;

- said term comprises a preselected parameter to define a propagation of the desired correction in the corrected image, the term being of the form , Or :

X is a current point having as coordinates the operating parameters, and where the penalty function is calculated,

Y is the correction point,

a is a coefficient,

K is a decreasing function, with K(0) = 1,

d(X,Y) is the distance between the correction point Y and the current point X, and

γ is the preselected parameter; and

- the method comprises a step f1) of determining the preselected parameter from a data set comprising a plurality of operating points having respectively as coordinates values of the operating parameters, the determination comprising the following sub-steps:

- for increasing values of a parameter of rank k , and for each of said operating points, calculation of a distance between said operating point and its kth ^closest neighbor in the plurality of operating points,

- for each of the values of the parameter of rank k , calculation of an average distance equal to an average of the distances obtained for the operating points,

- calculation of a first derivative of the mean distance with respect to the parameter of rank k and, if the first derivative has at least a first peak, choice of a value of the parameter of rank k located just before the first peak, the preselected parameter then being defined as the mean distance associated with said value of the parameter of rank k ,

- if the first derivative does not have a first peak, calculation of a second derivative of the mean distance with respect to the parameter of rank k and choice of a value of the parameter of rank k for which the second derivative becomes stable, the preselected parameter then being defined as the mean distance associated with said value of the parameter of rank k .

The invention also relates to an installation comprising:

- industrial equipment,

- sensors adapted to provide a plurality of operating parameters representative of operating states of the industrial equipment, and

- a control system adapted to detect operating anomalies of the industrial equipment, the control system comprising at least a central unit, a memory, a screen and a keyboard, the memory comprising instructions which when read by the central unit cause the installation to carry out steps a) to d) and f) to i) of a method as described above, the screen being adapted to display the image showing anomaly scores or the corrected image.

The invention will be better understood from reading the following description, given solely as an example and with reference to the attached drawings, in which:

there is a schematic view of an installation according to the invention,

there is a schematic view illustrating the steps of a method according to the invention, some of which are implemented by the installation shown in the ,

there is a graph giving anomaly scores, in value levels, of historical operating points of industrial equipment of the installation represented on the , depending on two of the operating parameters,

there is an image analogous to the graph shown in the , showing anomaly scores, in value levels, provided by the initial model of the process represented on the , the anomaly scores having been obtained by machine learning using historical operating points,

there is a graph similar to that shown in the , showing corrections desired by a user on correction points taken from the historical operating points represented on the ,

there is a graph similar to that shown in the , showing a penalty function obtained from the desired corrections shown in the , the preselected parameter to define the propagation of the correction being 0.25,

there is a graph similar to that shown in the , showing corrected anomaly scores provided by a corrected model of the process shown in the , the corrected anomaly scores being obtained from the anomaly scores represented on the and the penalty function represented on the ,

there is a graph similar to that shown in the , giving corrected anomaly scores for the historical operating points represented on the ,

there is a graph similar to that shown in the , showing a user-desired correction on elliptical and rectangular areas of the graph shown in the ,

there is a graph similar to that shown in the , showing corrected anomaly scores provided by the corrected model, the corrected anomaly scores being obtained from the anomaly scores shown in the and the correction shown on the ,

there is a variation of the graph shown in the , showing a correction desired by the user on a strip of the image represented on the ,

there is a graph similar to that shown in the , showing a penalty function obtained from a combination of the penalty functions shown in Figures 6 and 9,

there is a variation of the graph shown in the , showing a penalty function obtained from the desired corrections shown in the , the preselected parameter to define the correction propagation being 0.05 instead of 0.25,

there is a graph similar to that shown in the , showing corrected anomaly scores obtained from the anomaly scores represented on the and the penalty function represented on the ,

there is a variation of the graph shown in the , showing a penalty function obtained from the desired corrections shown in the , the preselected parameter to define the correction propagation being 1.0 instead of 0.25,

there is a graph similar to that shown in the , showing corrected anomaly scores obtained from the anomaly scores shown in the and the penalty function represented on the ,

there is a variation of the graph shown in the , showing a penalty function obtained from the desired corrections shown in the , the preselected parameter to define the correction propagation being 5.0 instead of 0.25,

there is a graph similar to that shown in the , showing corrected anomaly scores obtained from the anomaly scores shown in the and the penalty function represented on the ,

there is a three-part plot showing, as a function of the rank parameter k , at the top, the average distance obtained for the historical operating points represented on the , in the middle, the first derivative of the mean distance with respect to the parameter of rank k , and at the bottom, the second derivative of the mean distance with respect to the parameter of rank k ,

there is a variation of the graph shown in the , showing historical operating points not forming several distinct groups, but only one, and

there is a graph similar to that shown in the , but obtained for the historical operating points represented on the , and not those shown on the .

Facility

In reference to the , an installation 10 according to the invention is described.

Installation 10 is, for example, located on a site for the production of energy, raw materials, goods, or intermediate products intended for industry.

The installation 10 comprises industrial equipment 12 and sensors 14 adapted to provide a plurality of operating parameters C1, C2 … CN, representative, according to their possible values, of operating states of the industrial equipment. The installation 10 also comprises a control system 16 adapted to detect operating anomalies of the industrial equipment 12.

The industrial equipment 12 is for example a compressor, a turbine, a boiler, or any industrial equipment known per se.

There are N sensors 14, with N equal to one, two, three or more.

According to a particular embodiment, the or at least one of the sensors 14 can provide several operating parameters. For example, a multi-axis accelerometer provides accelerations along three axes, i.e. three operating parameters.

Operating parameters are, for example, temperatures, pressures, vibration levels, or other quantities typical of industrial equipment. Operating parameters are, for example, instantaneous values, or averaged over time. Operating parameters are advantageously standardized, i.e., for example, divided by a reference value.

For example, a pressure P is replaced by a pressure P*=P/P0, where:

P* is the normalized pressure,

P is the pressure in bar, varying for example between 1 and 2 bar,

P0 is a reference pressure, for example 1.5 bar.

More generally, P* = aP+b, where a and b are parameters, which are advantageously calculated automatically.

The control system 16 is located near the equipment 12, or alternatively remotely. The control system 16 is adapted to be used by at least one user 18 (hereinafter called “the user”, but it may be a team, or successive users). The control system 16 is advantageously adapted to act directly (arrow 19) on the industrial equipment 12, for example to stop it, start it, or modify an operating instruction. The control system 16 is advantageously adapted to receive and store historical operating data, or to receive values of the operating parameters C1, C2 … CN.

The control system 16 comprises at least a central unit 20, a memory 22, a screen 24 and a keyboard 26.

The user 18 is typically a technician responsible for controlling the industrial equipment 12, for example on site, or alternatively remotely.

The memory 22 comprises instructions which, when read by the central unit 20, cause the installation 10 to carry out steps a) to d) and f) to i) of a method according to the invention which will be described below with reference to the .

Step e) of this method is carried out by the user 18, who enters one or more desired corrections 28 into the control system 16.

The screen 24 is for example adapted to display images such as those shown in FIGS. 3 to 18, showing anomaly scores obtained from the operating parameters C1, C2 … CN, or showing desired corrections, depending on two of the operating parameters, in the example the parameters C1 and C2. Advantageously, the user 18 can select the pair of parameters serving as coordinates for the images that he wishes to see displayed.

According to a variant not shown, the images are produced as a function of three of the operating parameters, but such three-dimensional images are more difficult for the user 18 to read and interpret.

Process

In reference to the , we will now describe a method according to the invention, which is implemented in the example by the installation 10 and the user 18 (or users).

The method is suitable for controlling industrial equipment 12, in particular by detecting anomalies in its operation in the form of anomaly scores displayed on the screen 24.

The method comprises a step a) of obtaining an initial model F suitable for calculating anomaly scores from the operating parameters C1, C2 … CN, a step b) of obtaining a plurality of values 30 of the operating parameters defining at least one operating point M ( ) of the industrial equipment 12, and a step c) of calculating, using the initial model, the anomaly score F(M) associated with the operating point M ( ).

The method comprises a step d) of displaying at least one image 32 ( ) showing anomaly scores calculated by the initial model F, as a function of at least two of the operating parameters, here the parameters C1 and C2, forming two coordinates of the image 32, the image 32 showing the anomaly score of the operating point M and possibly being accompanied by an alert message.

The method comprises a step e) of determining, by the user 18, at least the desired correction 28 (figures 5, 9, 11 and 12) to be made to the displayed image 32, the desired correction relating to at least one correction point ( , showing several point corrections), or at least one correction zone (Figures 9 and 11 showing correction zones). The correction can also combine point and zone corrections ( ).

The method comprises a step f) of obtaining, from the initial model F and the desired correction 28, a corrected model Fp adapted to calculate, from the plurality of operating parameters C1, C2 … CN, corrected anomaly scores.

The method comprises a step g) of displaying a corrected image 36 (figures 7 and 10) showing the corrected anomaly scores, as a function of the two parameters C1 and C2.

Steps b) to g) are optionally iterated over time, the corrected model Fp of one iteration becoming the initial model F of a following iteration, so as to make successive corrections which accumulate.

The method comprises a step h) of obtaining a plurality of values 38 of the operating parameters 14 defining an operating point of the industrial equipment 12, for example the point M, and a step i) of calculating the anomaly score Fp(M) associated with said operating point using the corrected model, or that obtained during the last of the iterations. An alert message is optionally emitted by the control system 16 as a function of the value of Fp(M).

Advantageously, the method comprises a step j) of development by the user 18 or a machine, for example the control system 16, from the anomaly score obtained in step h), of an administration decision 40 of the industrial equipment 12, intended to modify the operating state of the industrial equipment 12, for example to stop it or to change an operating instruction.

Step a) corresponds to an initial start-up of the process. Steps b) to g), possibly iterated, correspond to a period, more or less long, of taking into account one or more feedbacks from the user 18 on anomaly scores, in particular one-off, and/or of taking into account his knowledge of the industrial equipment 12, in particular by one or more feedbacks on areas of the displayed images. Steps h) to j) are rather operating steps once the model F has been made reliable by the feedback(s) from the user 18.

It is understood that steps b) to g) allow for the reliability of anomaly scores in a flexible manner, without going through a tedious re-parameterization of the model F.

Step a) – Initial model

Anomaly scores are representative of probabilities that operating states (defined by operating parameters C1, C2 … CN) are abnormal. For example, anomaly scores have a value in the interval [0,1], with a value of 1 representing definitely abnormal operation, and a value of 0 meaning definitely normal operation.

At each point X defined by the operating parameters, the initial model F obtained in step a) (or the corrected model obtained during the previous iteration) provides an anomaly score F(X).

On the , image 32 represents the anomaly scores F(X) as a function of the coordinates C1 and C2 of each point.

In the example, parameter C1 varies between -10 and +10, and parameter C2 varies between -6 and +8. In a manner known per se, these are advantageously standardized values, in particular so as not to favor one operating parameter over another in the distance calculations which will be explained below.

Anomaly scores are represented by contour lines (as for altitude on geographic maps). In the example, the levels 0.0, 0.2… 1.0 have been chosen on the .

In practice, of course, we will rather use a color gradient, which is easier to read.

The initial model F results from a machine learning process.

The initial model F has for example been obtained previously and is located in the memory 22 of the control system 16.

According to a particular embodiment, step a) comprises a sub-step k) of automatic learning to obtain the initial model F. For example, historical data 42 represented on the , comprising a set of operating points. Using a machine learning known in itself (for example, one-class SVM, PCC, KNN methods, etc.), areas encompassing these points are determined. For each point X, a probability of anomaly is determined, generally as a function of the distance of point X from these areas. The learning process results in the initial model F whose image 32 on the provides a partial representation. It is a functional, probabilistic translation of the discrete historical data 42 represented on the .

Step b) and h) – Measurements

In step h), and optionally in step b), measurements are advantageously carried out by the sensors 14 of the industrial equipment 12 to obtain the values of the operating parameters C1 to CN.

Alternatively, the values of the operating parameters C1 to CN come for example from memory 22 and correspond for example to test cases, or to old measured values.

Steps c) and d) – Displaying anomaly scores

In step c), an anomaly score F(M) associated with the operating point M is calculated using the operating parameters C1 to CN. On the , F(M) has a value slightly greater than 0.4. For example, if the anomaly score F(M) is greater than a predefined threshold, an audible and/or visual alert is issued.

In step d), image 32 visible on the is displayed. Figure 32 shows the anomaly scores F(X) for any point X located in a plane, with point X defined by its coordinates C1 and C2 in that plane.

Steps e) and f) – specific corrections desired by the user

According to a first embodiment, in step e), the correction 28 desired by the user 18 takes the form of one or more one-off corrections. These corrections relate, for example, to the operating point M, or to other correction points Y.

For example, the shows, at the bottom left of the graph, a certain number of unlabeled points, to which are for example associated anomaly scores of 0.5, meaning that the initial model F cannot determine whether these operating points are normal or abnormal.

The user decides for example that these points correspond to abnormal operating modes, and labels these points as abnormal, as visible on the , showing a positive correction for these points, which will increase their anomaly scores.

In step f), the calculation of the corrected model Fp includes a summation of the initial model F and a penalty function Pen having as variables at least the coordinates C1, C2 of the image, advantageously all the operating parameters C1 to CN.

For each point X, the corrected anomaly score is given by the formula:

Or :

F is the initial model,

Pen is the desired correction,

Fp is the corrected model.

If Fp(X) is greater than 1.0 or less than 0.0, a normalization is advantageously carried out, so that the corrected anomaly score Fp(X) remains in the interval [0,1].

If the user decides to perform a single point correction at a point Y, then, for any point X, the penalty function Pen is advantageously, in absolute value, maximum for said correction point Y, and decreasing as a function of a distance d(X,Y) between the correction point Y and the point X. For example:

Or :

a is a coefficient such that the penalty is advantageously worth a correction b desired by the user at point Y, the correction b being between -1 and +1 in the example, with a=b in the case of a single point correction.

K is a decreasing function, with for example K(0) = 1,

d(X,Y) is for example the Euclidean distance between points X and Y,

γ is a preselected parameter whose meaning will be explained below.

Alternatively, other distances than the Euclidean distance could be used.

The correction is maximum in absolute value at point Y, and its absolute value worth . The further away from point Y, the higher the value decreases.

On the , the user has actually decided to perform several point corrections at a plurality of points Y _i . The corrections are positive for a point cloud located at the bottom left, and negative for another point cloud located on the right.

For any point X, the penalty function Pen is then given by the formula:

Or :

i is an integer used to index the points Y _i ,

a _i is a coefficient advantageously calculated so that the penalty at point Y _i is worth a correction b _i for example between -1 and +1 requested by the user, that is to say that, for any point labeled Y _i :

There shows the values of Pen(X) in the example, on a scale from -1.0 to +1.0.

The method allows to label any operating point. For example, the user labels a point as abnormal by choosing a correction b worth +0.5 or as normal by choosing a correction b worth -0.5.

Advantageously, if the user is not sure of the correction he wants, he can choose corrections lower than 0.5 in absolute value, so that the correction has only a relative effect on the corrected anomaly score. Conversely, if the correction point has a low initial anomaly score (therefore considered rather normal), while the user considers that this operating point is abnormal, the user can choose a correction greater than +0.5, or even a correction of +1.0. Normalization advantageously brings the Fp(X) value back into the interval [0,1] in all cases.

There shows the values of Fp(X), i.e. the corrected image 36.

Advantageously, each term of the penalty function Pen associated with a correction point Y is proportional to a Gaussian function having the distance d(X,Y) as a variable. In this case, the function K takes for example the following form:

Or :

x is a variable,

exp is the exponential function,

ln is the natural logarithm,

ord is a parameter greater than 1.

This leads, for each term associated with a correction point Y, to a penalty worth:

Alternatively, other K functions are possible. The inventors have observed that a Gaussian function is particularly advantageous.

As we have seen, in the case of a single correction point Y, Pen(Y) = a. Furthermore, if d(X,Y) = γ, Pen(X) = a/2 in the example. The preselected parameter γ therefore represents, in the example, the distance for which the correction at point X is only half of what it is at correction point Y in the case of a single correction point Y.

Alternatively, other definitions of the preselected parameter γ are possible by modifying the function K. For example, the preselected parameter γ can represent the distance for which the correction at point X is only a given fraction of what it is at correction point Y.

The method therefore allows, by preselecting the value of the parameter γ, to control the “propagation” or “diffusion” of a point correction from the correction point Y to points X located in the vicinity of the correction point Y. The lower the value of the parameter γ, the more the correction applied decreases rapidly as one moves away from the correction point Y.

In Figures 6 and 7, the value of the parameter γ is for example 0.25. We will return to the effect of the parameter γ and describe a method for choosing the value of this parameter.

There shows the corrected anomaly scores Fp(Y _i ) of the points Y _i represented on the .

Steps e) and f) – correction on an area desired by the user

According to a second embodiment (alternatively or in addition to the first embodiment), in step e), the correction 28 desired by the user 18 relates to a zone Z of the displayed image. This advantageously makes it possible to take into account knowledge of the operation of the industrial equipment 12, or the experience of the user, in particular the thresholds not to be crossed by certain operating parameters, or the zones of good operation.

For example, on the , the user wants to make a negative correction on an elliptical shaped zone Z1 (normal operation inside the ellipse) and a positive correction on a rectangular shaped zone Z2 (abnormal operation outside the rectangle).

In step f), the calculation of the corrected model Fp comprises a summation of the initial model F and a penalty function Pen having as variables at least said coordinates of the image, advantageously all the operating parameters C1 to CN. The penalty function Pen is adapted to carry out in the corrected image 36 a correction of elliptical shape Fp in the zone Z1 and rectangular in the zone Z2.

The corrected model is shown in the (which is therefore the sum of the images shown in figures 4 and 9).

Alternatively or in addition, the correction has a band shape in a Z3 zone, as shown in the .

This type of zone correction can be obtained using functions, or a sum of functions known per se to those skilled in the art.

For example, an elliptic correction comes for example from the choice of the Euclidean distance in the Pen function defined previously.

For a rectangular correction, we choose for example the distance induced by the l-p norm:

, with the parameter p sufficiently high.

For a vertical band correction, the distance is for example calculated with only C1 as the operating parameter, unlike in the previous cases.

For example, the correction can be "abrupt" (with a drop when leaving the ellipse, rectangle or band), or "soft" (with a transition zone whose extension can be advantageously configured). This is adjustable via the ord parameter: the higher it is, the more abrupt the correction.

On the , we have represented a penalty function resulting from the combination of the penalty functions represented on the (resulting from one-off corrections) and on the (corrections by zones). The method according to the invention makes it possible to introduce several corrections, for example of different nature.

Iteration of steps b) to g)

As we have seen, in steps e) and f), the user can introduce one or more corrections, one-off or by zone.

Steps b) to g) are optionally iterated over time, so that the user can, at each iteration, reintroduce one or more corrections that are cumulative with the previous ones. The new corrections may come from a new operating point or a new zone whose anomaly score the user wants to correct, or may come from a change in the user's opinion, taking into account his acquired experience, on a previously corrected operating point or zone. The corrected model Fp obtained at step f) of an iteration becomes the initial model F of the following iteration.

The process therefore advantageously allows several simultaneous or successive corrections to be made over time, and also allows previous corrections to be returned to and modified.

In step i), the last corrected model Fp is used.

Effect of parameter γ - choice of this parameter

As seen above, the preselected parameter γ represents, in the case of a point correction desired by the user, a distance for which the correction at point X is only a fraction, for example half, of what it is at correction point Y.

There represents the penalty function Pen(X) obtained from the point corrections 28 shown in the , using γ=0.25. The shows the corrected model Fp(X), which is the sum of the initial model F(X) shown in the and the penalty function Pen(X) shown in the .

Figures 13 and 14 show respectively the penalty function Pen(X) and corrected model Fp(X) obtained with γ=0.05. It can be seen that the corrections do not diffuse in the corrected image ( ) and have little effect. In particular, the corrections do not propagate to operating points close to those that the user has chosen to label on the .

Figures 15 and 16 show respectively the penalty function Pen(X) and corrected model Fp(X) obtained with γ=1.0. Figures 17 and 18 show respectively the penalty function Pen(X) and corrected model Fp(X) obtained with γ=5.0. In both cases, the corrections are spread too widely and affect not only the point cloud located in the vicinity of the points labeled by the user, but other point clouds. In the case γ=5.0, the information is spread so much in the corrected image that the corrected model Fp(X) is no longer discriminating (no longer allows to distinguish abnormal points from normal points), particularly in the left part of image 36 shown in .

The parameter γ is for example stored in memory 22 and has been advantageously chosen so that the corrections diffuse sufficiently, but not too much.

According to a particular embodiment, the method comprises a step f1) of determining the preselected parameter γ.

Step f1) of determining the preselected parameter γ

For example, we start from the plurality of operating points Y _i represented on the , having respectively as coordinates the values of the operating parameters.

For increasing values of a parameter of rank k , and for each of said operating points Y _i , a distance D _k (Y _i ) is calculated between the operating point Y _i and its k ^th closest neighbor in the plurality of operating points. For each of the values of the parameter of rank k , an average distance D _k is calculated equal to the arithmetic mean of the distances D _k (Y _i ) obtained for the operating points Y _i :

, Or

N is the number of operating points Y _i , and

k is an integer that can take values from 1 to N-1.

D _k is represented as a function ofkon the graph located at the top of the .

Then, we calculate a first derivative D _k ' of D _k as a function of k . For example: .

D _k ' is represented on the graph in the middle of the .

In Figure 19, a second derivative is also represented on the graph below.

If the first derivative D _k ' has at least one first peak, we choose a value k0 of the parameter of rank k located just before this peak, and the preselected parameter γ is then defined as the average distance D _k0 associated with this value k0 of the parameter of rank k . We have:

γ = D _k0 , k0 being the parameter of rank k located just before the first peak of D _k '.

In the example, we see on the , in the graph vertically in the middle, that the function D _k ' presents several successive peaks P1, P2, P3… To be just before the first peak P1, it is necessary to choose for the rank parameter the value k0 = 15. We obtain γ = D ₁₅ = 0.25.

It was observed that the parameter γ thus obtained advantageously allows the diffusion of point corrections in the cloud of points located in the vicinity of the correction point, without going beyond.

In the example of the , the second derivative D _k '' is not used and is advantageously not calculated, because the first derivative D _k ' has at least a first peak P1, due to the fact that the plurality of operating points Y _i comprises several point clouds distinct from each other ( ).

If, instead of starting from the plurality of operating points Y _i represented on the , we start from the plurality of points Y _i represented on the , we obtain the graphs D _k , D _k ' and D _k '' represented on the .

We then note that the curve D _k ' does not have a first peak that can be isolated. This is due to the fact that the points Y _i on the form a single cloud and not several.

In this case, when the first derivative D _k ' does not have a peak, we calculate the second derivative D _k '' and we choose a value k1 of the parameter of rank k for which the second derivative D _k '' becomes stable, and the preselected parameter γ is then defined as the average distance D _k1 associated with this value k1 of the parameter of rank k . We have:

γ = D _k1 , k1 being the parameter of rank k for which the second derivative D _k '' becomes stable.

In the example of the , we see that the function D _k '' stabilizes quite quickly after a value k1 = 6. We obtain γ = D ₁₆ = 0.50.

It was observed that the parameter γ thus obtained advantageously allows the point corrections to be diffused towards significant points in the vicinity of the correction point, without “contaminating” the set of points Y _i .

Thanks to the characteristics described above, the method is more reliable, because the anomaly scores are progressively made more reliable by the one-off or zone-specific corrections requested by the user. The method is also easy to implement, because it allows one or more simultaneous corrections of the anomaly scores to be made, and possibly successive corrections to be made that accumulate, without carrying out a complete re-parameterization of the initial model. The control of the industrial equipment 12 is therefore improved.

According to the particular embodiments described above, the method makes it possible to label as “normal” or “abnormal” an actual operating point, or simply a displayed point. The correction can be punctual (with diffusion) or immediately concern a given area of the displayed image. The correction is positive (towards “abnormality” in the example) or negative (towards “normality”) and its value (between -1 and +1 in the example) can be modulated according to the user’s degree of certainty.

For area corrections, by means of an appropriate shape, such as elliptical, rectangular or strip, the user can easily translate his experience or knowledge of the operation of industrial equipment into correction form 12.

For a one-time correction (from a correction point), the value of the parameter γ makes it possible to control the diffusion of the correction from the correction point to its neighbors. The parameter γ can be preselected, or advantageously determined by a method based on the plurality of historical operating points available.

Claims (10)

Method for controlling industrial equipment (12) to detect operating anomalies of the industrial equipment (12), implementing a control system (16) adapted to be used by at least one user (18), the method comprising the following steps:
a) obtaining, in a memory (22) of the control system (16), an initial model (F) suitable for calculating anomaly scores from a plurality of operating parameters (C1, …CN) representative of operating states (X) of the industrial equipment (12), the anomaly scores being respectively representative of probabilities that the operating states (X) are abnormal, the initial model (F) resulting from an automatic learning process,
b) obtaining a plurality of values (30) of the operating parameters (C1, …CN) defining at least one operating point (M) of the industrial equipment (12),
c) calculation, using the initial model (F), of the anomaly score associated with the operating point (M),
d) displaying, on a screen (24) of the control system (16), at least one image (32) showing anomaly scores calculated by the initial model, as a function of at least two (C1, C2) of the operating parameters (C1…CN) forming two coordinates of the image (32), the image (32) showing the anomaly score of the operating point (M),
e) determining, by the user (18), at least one desired correction (28) in the image (32) displayed on the screen (24), the desired correction (28) relating to at least one correction point (Y), or at least one correction zone (Z), and entering by the user (18) the desired correction (28) into the control system (16),
f) from the initial model (F) and the desired correction (28), obtaining a corrected model (Fp) suitable for calculating, from the plurality of operating parameters (C1…CN), corrected anomaly scores,
g) displaying, on the screen (24), a corrected image (36) showing the corrected anomaly scores as a function of the two coordinates,
steps b) to g) being optionally iterated over time, the corrected model (Fp) of one iteration becoming the initial model (F) of a following iteration,
h) obtaining a plurality of values (38) of the operating parameters (C1…CN) defining an operating point (M) of the industrial equipment (12), and
i) calculation of the anomaly score (Fp(M)) associated with said operating point (M) using the corrected model (Fp) obtained during the last of the iterations. Method according to claim 1, wherein step a) comprises a machine learning sub-step to obtain the initial model (F). Method according to claim 1 or 2, in which, in step h), and optionally in step b), measurements are carried out by a plurality of sensors (14) of the industrial equipment (12) to obtain the values (30, 38) of the operating parameters (C1…CN). Method according to any one of claims 1 to 3, further comprising a step j) of development by the user (18) or a machine, from said anomaly score (Fp(M)) obtained in step i), of an administration decision (40) of the industrial equipment (12) intended to modify the operating state of the industrial equipment (12). A method according to any one of claims 1 to 4, wherein:
- in step e), the desired correction (28) relates to an area (21, 22) of the displayed image (32), and
- in step f), the calculation of the corrected model (Fp) comprises a summation of the initial model (F) and a penalty function (Pen) having as variables at least said coordinates of the image (32), the penalty function (Pen) being adapted to carry out in the corrected image (36) a correction of predefined shape, for example spherical, elliptical, rectangular, or by band. A method according to any one of claims 1 to 4, wherein:
- in step e), the desired correction (28) relates to at least one correction point (Y) belonging to the displayed image (32), and
- in step f), the calculation of the corrected model (Fp) comprises a summation of the initial model (F) and a penalty function (Pen) having as variables at least said coordinates of the image (32), the penalty function (Pen) comprising a term associated with the correction point (Y), the term being, in absolute value, maximum for said correction point (Y) and decreasing as a function of a distance (d(X,Y)) from said correction point (Y). A method according to claim 6, wherein said term is proportional to a Gaussian function having said distance (d(X,Y)) as a variable. A method according to claim 6 or 7, wherein said term comprises a preselected parameter (γ) for defining a propagation of the desired correction (28) in the corrected image (36), the term being of the form , Or :
X is a current point having as coordinates the operating parameters (C1…CN), and where the penalty function is calculated,
Y is the correction point,
a is a coefficient,
K is a decreasing function, with K(0) = 1,
d(X,Y) is the distance between the correction point Y and the current point X, and
γ is the preselected parameter. Method according to claim 8, comprising a step f1) of determining the preselected parameter (γ) from a data set (40) comprising a plurality of operating points (Y _i ) having respectively as coordinates values of the operating parameters (C1…CN), the determination comprising the following sub-steps:
- for increasing values of a parameter of rank k , and for each of said operating points (Y _i ), calculation of a distance (D _k (Y _i )) between said operating point (M) and its kth closest ^neighbor in the plurality of operating points (Yi),
- for each of the values of the parameter of rank k , calculation of an average distance (D _k ) equal to an average of the distances (D _k (Y _i )) obtained for the operating points (Y _i ),
- calculation of a first derivative (D _k ') of the mean distance with respect to the parameter of rank k and, if the first derivative (D _k ') has at least a first peak (P1), choice of a value (k0) of the parameter of rank k located just before the first peak (P1), the preselected parameter (γ) then being defined as the mean distance (D _k ) associated with said value (k0) of the parameter of rank k ,
- if the first derivative (D _k ') does not have a first peak (P1), calculation of a second derivative (D _k '') of the mean distance (D _k ) with respect to the parameter of rank k and choice of a value (k1) of the parameter of rank k for which the second derivative (D _k '') becomes stable, the preselected parameter (γ) then being defined as the mean distance (D _k ) associated with said value (k1) of the parameter of rank k . Installation (10) comprising:
- industrial equipment (12),
- sensors (14) adapted to provide a plurality of operating parameters (C1…CN) representative of operating states (X) of the industrial equipment (12), and
- a control system (16) adapted to detect operating anomalies of the industrial equipment (12), the control system (16) comprising at least a central unit (20), a memory (22), a screen (24) and a keyboard (26), the memory (22) comprising instructions which when read by the central unit (20) lead the installation (10) to carry out steps a) to d) and f) to i) of a method according to any one of claims 1 to 9, the screen (24) being adapted to display the image (32) showing anomaly scores or the corrected image (36).