EP4323846A1

EP4323846A1 - Method for monitoring the quality of screwing or drilling operations including unsupervised machine learning

Info

Publication number: EP4323846A1
Application number: EP22722836.8A
Authority: EP
Inventors: Philippe Leray; Mathieu RITOU; Mahmoud FERHAT
Original assignee: Universite de Nantes; Georges Renault SAS
Current assignee: Universite de Nantes; Georges Renault SAS
Priority date: 2021-04-16
Filing date: 2022-04-14
Publication date: 2024-02-21
Also published as: CN117355803A; FR3122013A1; US20240198506A1; WO2022219153A1

Abstract

The present invention relates to a method for monitoring the quality of screwing or drilling operations performed by way of a tool, said method comprising unsupervised machine learning of a statistical model.

Description

DESCRIPTION

Title: Method for controlling the quality of screwing or drilling operations including unsupervised automatic learning

1. Field of the invention

The field of the invention is that of controlling the quality of operations carried out in industrial production, such as for example screwing or drilling operations, by the implementation of intelligent control processes implementing a model of the type machine learning.

2. Prior Art

In many industrial fields, particularly in the automotive industry, components are assembled by screwing.

Screwdriving consists in driving a screw in rotation until a tightening criterion, which can for example be a tightening torque or a tightening angle, reaches a predetermined threshold. The torque developed by a screwdriver follows the resistive torque of the screw which, in general, is a linear function of the angle.

The functional objective of a screw is to install a certain tension in the body of the screw, this tension making it possible to carry out the assembly of two components.

However, it can happen that a screwing does not take place correctly and that the tension installed in the screw at the end of a screwing is not satisfactory so that the assembly is defective.

This can happen in particular following poor engagement of the screw in the thread (crossed thread). The torque applied by the screwdriver is then observed to rise to the desired torque level, giving the impression that the tightening has been carried out properly, while the assembly tension in the screw is insufficient or even zero. The tightening is then considered positively good even though it is not correct; this is a false positive.

Other faults can occur during tightening with similar consequences: seizure in the threads, stick slip, etc.

In many industrial fields, particularly in aeronautics, components are drilled or counter-drilled.

Drilling consists of making a hole in a component.

Counter-drilling consists of drilling several components pressed against each other before assembling them together.

In the aerospace industry, counter drilling is widely used to make holes in components to allow the installation of screws or rivets allowing then to assemble these components to produce aircraft structures.

These holes are more and more frequently made in stacks of materials of different nature: aluminum alloy, titanium, carbon fiber, etc.

The drilling systems make it possible to measure in real time the axial force and the cutting torque supported by the drill.

These values follow an evolution according to the depth of penetration of the drill bit into the material, which reflects the variations in hardness of the different materials present in the stack.

The torques and axial forces are rather constant, in the form of a leveling off, when the tip of the drill progresses through a thickness of homogeneous material. On the contrary, they evolve in a strong way when the tip of the drill passes from one material to another.

In principle, these changes in axial torque and axial force are repeatable from one drilling to the next. A drift in these evolutions can reflect premature wear of the drill bit or jamming of chips with consequences on the quality of the hole (geometry, surface condition, burr at the entrance or exit of the hole, flaking of the carbon fiber).

The production and quality control departments of these industries therefore implement processes aimed at detecting defects by analyzing the physical data (or quantities) recorded during each screwing or drilling. This data is commonly referred to as screwing or drilling report or screwing or drilling results. These processes are based on "machine leaming" or neural network techniques and make it possible to issue an alert in real time to operators or method technicians in order to be able to isolate or correct non-compliant production operations, i.e. ie whose quality is not satisfactory.

Supervised learning model type operation quality control techniques are known. The implementation of these techniques includes two phases, namely a learning phase and a production control phase.

Learning is based on the training of a statistical model, from results of screwing or drilling operations, to define groupings (or "cluster" in English) of results. Prior to this learning, an operator must label, one by one, the results of operations taken into account. This labeling phase consists, for the operator, in analyzing each operation result to qualify it as a good or bad result. After the operator has labeled all the results taken into consideration, he triggers automatic learning of the supervised model so as to make it able to determine the label of new results.

During the production control phase, as soon as a new operation is carried out, its result is analyzed by the control system, which is able to automatically assign it a label.

Such a technique is advantageous in that it makes it possible in production to inform automatically, that is to say without the intervention of a control technician, whether the result of a new operation is good or bad. .

However, supervised learning is constraining. Indeed, for a model to be reliable, several hundred operation results must be taken into consideration. This assumes that the method technician reviews all of these operation results and status on their compliance (i.e. good or bad character) prior to learning. This long and therefore costly task.

Moreover, for a supervised model to be efficient, the set of operation results intended to train it must contain a number of results of each type (good and bad) likely to be encountered during the training. use of the model to classify the results. However, during the commissioning of such a model, the results of operations available are those emitted during the first production operations and these results are for the most part good results.

Moreover, even if a large number of results of different types is provided in the learning phase, these supervised models are generally designed to classify the new results entering production mode into the types identified during the learning. However, as some types may be unknown or missing during the training phase, the model cannot detect the emergence of a new type of result not encountered before.

Finally, most of these models require a sufficient training time to fulfill their objective with considerable precision. Moreover, as the accuracy is strongly proportional to the complexity of the model and is crucial in an industrial context, the ability to interpret the results is often deteriorated.

3. Objects of the invention

The aim of the invention is in particular to provide an effective solution to at least some of these various problems.

In particular, according to at least one embodiment, an objective of the invention is to improve quality control techniques for machine-learning type operations.

In particular, the object of the invention, according to at least one embodiment, is to provide such a technique which makes it possible to be initialized when learning from a set of results of operations containing the vast majority of results compliant with quality requirements and few defects, i.e. a large majority of good results and few bad results.

Another object of the invention is, according to at least one embodiment, to provide such a technique requiring only little work on the part of the technician.

Another objective of the invention is, according to at least one embodiment, to provide such a technique capable of detecting the appearance of results of an unknown type during production.

Another objective of the invention is, according to at least one embodiment, to provide such a technique capable of allowing an update of the model when results of an unknown type arise.

Another objective of the invention is, according to at least one embodiment, to provide such a technique which makes it possible to label groups (clusters) and to characterize the physical nature of the fault encountered during the operations belonging to these groups.

Another objective of the invention is, according to at least one embodiment, to provide such a technique which is simple to implement and/or robust and/or versatile and or scalable.

4. Presentation of the invention

For this, the invention proposes a method for controlling the quality of screwing or drilling operations carried out by means of a tool, said method comprising: a step of automatically learning a model, said step of learning automatic initial comprising: a step of collecting initial results of screwing or drilling operations recorded during a plurality of screwing or drilling operations; a step of obtaining, by means of said model, at least one statistical representation of at least part of the initial results; a step of labeling said at least one statistical representation with at least one of said following labels: labeled statistical representation of compliant results; labeled statistical representation of non-compliant results; a step of automatic control in production of the quality of screwing or drilling operations, said step of automatic control in production being implemented at the end of each operation and comprising: a step of collecting the new result of the screwing or drilling operation recorded during the operation in question; a step of automatically assigning said new result of the operation considered to the at least one statistical representation; a step for issuing an alert if the new operation result is assigned to a non-compliant statistical representation.

According to the invention, said learning of the model is of the unsupervised type.

Thus, the invention relates to a technique for quality control of results of screwing or drilling operations including automatic learning of the unsupervised type. This technique makes it possible to simplify and reduce the time required for learning, or more precisely, if necessary, for the pre-processing of the results of operations, necessary for the automatic control of the results of operations. Indeed, rather than labeling all of the operation results as in a technique of the prior art with supervised learning, the operator must only label the statistical representations to which the results are assigned.

According to one possible characteristic, said production control step includes a step for rejecting new results of operations that cannot be assigned to one of said statistical representations, and a step for recording rejected results as exception results.

It is thus possible to discard the results of operations whose likelihood of belonging to the normal laws already established is low.

In this case, a method according to the invention preferably comprises a step of issuing an alert when a new result obtained at the end of a screwing or drilling operation is identified as an exceptional result during said rejection step. According to one possible characteristic, said step of obtaining at least one statistical representation implements a statistical model of the Gaussian mixture model type, said at least one statistical representation being a multivariate normal distribution.

According to another variant, said step of obtaining at least one statistical representation implements a statistical model of K-means type, said at least one statistical representation being an average vector.

According to a possible variant, said statistical model of the Gaussian mixture type calculates, during said step of obtaining a statistical representation, a plurality of initial multivariate normal laws each representative of a group of initial results of tightening operations or drilling, each of said initial multivariate normal laws possessing a weight, said step of rejecting comprising a step of calculating the probability density of each new result of screwing or drilling operation obtained during said step of automatic control in production and a step of comparing this density with a predetermined global rejection threshold depending on the weight of each of said initial multivariate normal laws.

According to a possible variant, said weight of an initial multivariate normal law is representative of the number of results of screwing or drilling operations assigned to said multivariate normal law relative to the total number of results of screwing or drilling operations taken into account.

According to a possible variant, said classification step is followed by a step of updating the weight of all of said initial multivariate normal distributions.

According to a possible variant, said classification step is followed by a step of updating said predetermined global rejection threshold.

According to a possible variant, a method according to the invention comprises a step of updating said learning, said step of updating being implemented during said step of automatic control in production and comprising a step of updating the obtaining a statistical representation of the initial results in the form of at least one multivariate normal law, said updating step taking into account the new operation results identified as exceptional results to generate at least one new multivariate normal law.

According to a possible variant, during said step of updating said learning, said statistical model of Gaussian mixture type calculates new multivariate normal laws, from the exceptional results, each representative of a new group of results d operations, said step of updating said learning comprising a step of calculating the new weight of said initial and new multivariate normal laws.

According to a possible variant, said step of updating said learning generates new multivariate normal laws, from the new results of operations and from the initial results of operations, each representative of a new group of results of tightening operations or drilling, said new multivariate normal laws being substituted for the initial multivariate normal laws.

According to one possible characteristic, a method according to the invention comprises a step of counting the number of new operation results identified as exceptional results, said step of updating said learning being implemented when the number of results of operation identified as exception results reaches a predetermined threshold.

According to one possible characteristic, each operation result is a series of data, said series of data being the subject of a preprocessing consisting in carrying out on a series of data a series of predetermined calculations each resulting in an extracted characteristic, said extracted characteristics being taken into consideration for carrying out said step of obtaining a statistical representation of the initial results in the form of at least one multivariate normal law. According to one possible characteristic, said extracted characteristics are taken into consideration by said statistical model of the Gaussian mixture type to generate said multivariate normal laws each representative of a group of results of screwing or drilling operations.

According to a possible characteristic, which said data series belong to the group comprising: torque as a function of drilling angle or depth or time; an angle as a function of time; an intensity, in particular of a motor driving a screwing tool in rotation or driving a cutting tool in rotation or translation as a function of an angle or of a drilling depth or of time; a force as a function of an angle or a drilling depth or time.

According to one possible characteristic, a method according to the invention comprises a step of selecting, in each of said series, a portion of data of interest, said portion of data of interest being the subject of said pre-processing.

The invention also relates to a device for controlling the quality of screwing or drilling operations carried out by means of a tool, said device comprising: means for automatically learning a model, said initial automatic learning means comprising: means for collecting initial results of screwing or drilling operations recorded during a plurality of screwing or drilling operations;

Means for obtaining, by means of said model, at least one statistical representation of at least part of the initial results; means for labeling said at least one statistical representation with at least one of said following labels: labeled statistical representation of compliant results; labeled statistical representation of non-compliant results; means for automatic control in production of the quality of screwing or drilling operations, said means for automatic control in production being able to be implemented at the end of each operation and comprising: means for collecting the new result the screwing or drilling operation recorded during the operation in question; means for automatically assigning said new result of the operation considered to the at least one statistical representation; means for issuing an alert if the new operation result is assigned to a non-compliant statistical representation; said learning of said model being of the unsupervised type.

According to one possible characteristic, a device according to the invention comprises means for rejecting new results of operations which cannot be assigned in production to one of said statistical representations, and means for recording the results rejected as results of exception.

According to one possible characteristic, a device according to the invention comprises means for transmitting an alert when a new result obtained at the end of a screwing or drilling operation is identified as an exceptional result by said means. of rejection.

According to one possible characteristic, said means for obtaining at least one statistical representation implement a statistical model of the Gaussian mixture model type, said at least one statistical representation being a multivariate normal law.

According to one possible characteristic, said means for obtaining at least one statistical representation implement a statistical model of the K-means type, said at least one statistical representation being an average vector.

According to one possible characteristic, said statistical model of the Gaussian mixture type is able to calculate a plurality of initial multivariate normal laws each representative of a group of initial results of screwing or drilling operations, each of said initial multivariate normal laws having a weight, said rejection means comprising means for calculating the probability density of each new result of a screwing or drilling operation obtained in production and means for comparing this density with a predetermined overall rejection threshold depending on the weight of each of said initial multivariate normal distributions.

According to one possible characteristic, said weight of an initial multivariate normal law is representative of the number of results of screwing or drilling operations assigned to said multivariate normal law relative to the total number of results of screwing or drilling operations taken into account. According to one possible characteristic, a device according to the invention comprises means for updating the weight of all of said initial multivariate normal laws.

According to one possible characteristic, a device according to the invention comprises comprising means for updating said predetermined global rejection threshold.

According to one possible characteristic, a device according to the invention comprises means for updating the at least one statistical representation of the initial results in the form of at least one multivariate normal law obtained by said learning means, said updating taking into account the new operation results identified as exceptional results to generate at least one new multivariate normal law.

According to one possible characteristic, said updating means using said statistical model of the Gaussian mixture type to calculate new multivariate normal laws, from the exception results, each representative of a new group of results of operations, said means for updating said learning comprising means for calculating the new weight of said initial and new multivariate normal laws.

According to one possible characteristic, said updating means generate new multivariate normal laws, from the new results of operations and from the initial results of operations, each representative of a new group of results of tightening or drilling, said new multivariate normal laws being substituted for the initial multivariate normal laws.

According to one possible characteristic, a device according to the invention comprises means for counting the number of new operation results identified as exceptional results, said updating means being implemented when the number of operation results identified as exception results reaches a predetermined threshold.

According to one possible characteristic, each operation result is a series of data, said device comprising means for pre-processing said series of data consisting in carrying out on a series of data a series of predetermined calculations each resulting in an extracted characteristic, said extracted characteristics being taken into consideration by said means for obtaining a statistical representation of the initial results in the form of at least one multivariate normal distribution. According to one possible characteristic, said extracted characteristics are taken into consideration by said statistical model of the Gaussian mixture type to generate said multivariate normal laws each representative of a group of results of screwing or drilling operations.

According to one possible characteristic, said series of data belong to the group comprising: a torque as a function of an angle or of a drilling depth or of time; an angle as a function of time; an intensity, in particular of a motor driving a screwing tool in rotation or driving a cutting tool in rotation or translation as a function of an angle or of a drilling depth or of time; a force as a function of an angle or a drilling depth or time.

According to one possible characteristic, a device according to the invention comprises means for selecting, in each of said series, a portion of data of interest, said portion of data of interest being taken into account by said pre-processing means.

The invention also relates to a computer program product, comprising program code instructions for implementing a method according to any one of the variants set out above, when said program is executed on a computer.

The invention also relates to a non-transitory, computer-readable storage medium storing a computer program product according to claim 35.

5. Description of the Ligurians

Other characteristics and advantages of the invention will appear on reading the following description of particular embodiments, given by way of simple illustrative and non-limiting example, and the appended drawings, among which:

[Fig 1] Figure 1 illustrates an example of an operation result table;

[Fig 2] Figure 2 illustrates an example of a tightening curve;

[Fig 3] Figure 3 illustrates the functional blocks of an example of a control device according to the invention; [Fig 4] Figure 4 illustrates learning a method according to the invention;

[Fig 5] Figure 5 illustrates the use in production of a method according to the invention;

[Fig 6] Figure 6 illustrates a first variant of updating the learning of a method according to the invention;

[Fig. 7] FIG. 7 illustrates a first variant of updating the learning of a method according to the invention;

[Fig 8] Figure 8 illustrates an example of the architecture of a quality control system according to the invention.

6. Description of particular embodiments 6.1. Quality control device

The invention relates to a technique for controlling the quality of screwing or drilling operations using an unsupervised statistical model of the unsupervised machine learning type.

Screwing or drilling operations are carried out with a work device conventionally comprising a tool, such as a screwdriver or a drill, and a controller equipped with an HMI allowing the tool to be controlled.

The tool and/or the controller make it possible in particular to record, or during the performance of each operation, an operation result.

During a tightening operation, various physical quantities can be measured by the tool and/or the controller to monitor the progress of the operation in progress and detect its completion by the achievement of objective thresholds by some of the quantities physics measured.

In the example described here, it will be considered that the physical quantities measured during a tightening operation include the tightening angle and the tightening torque.

During a drilling operation, the physical quantities measured are in particular, for example, the torque and the axial thrust on the tool generated by the cutting tool.

During a tightening operation, control means integrated into the tool and/or the controller make it possible to measure and record physical quantities on the basis of a predetermined acquisition frequency, these physical quantities possibly include one or more of the following quantities: the tightening angle, the tightening torque or the intensity of the tool motor.

The tool and/or the controller can then for example record, according to the rate of its microprocessor, couplets comprising a tightening angle value and a tightening torque value. A tightening angle and tightening torque couplet corresponds to the result of an angle and tightening torque measurement at the same time.

During a tightening operation, these torques are recorded in the form of a table of values and can lead to the construction of a tightening curve that an operator can view on a monitor.

Figure 1 illustrates a table of values in which the torques measured during a tightening operation are recorded. Figure 2 illustrates an example of a tightening curve.

The result of a tightening is therefore in this example a table of values listing the tightening angle/tightening torque couplets recorded during a tightening operation. Such a table therefore groups couplets which constitute a series of data, also called series of temporal data.

In the context of drilling, the physical quantities measured generally include the torque and the axial force generated by the cutting tool as a function of the depth reached by the tip of the cutting tool (drill).

The result of a drilling is therefore in this example a table of values listing the drilling depth/torque couplets and a table of values listing the drilling depth/axial force couplets recorded during a tightening operation. Each table therefore groups couplets which constitute a series of data, also called series of temporal data.

An operation result is therefore one (in the case of a tightening) or several (in the case of a drilling) series of data recorded during the performance of the operation.

This operation result is then analyzed by a quality control device 100 which makes it possible to check whether the operation carried out complies or does not comply with the expected level of quality.

A quality control device according to the invention comprises: functional means for collecting results of operations 10; functional means (optional) for selecting a portion of interest in each operation result 11; functional means for pre-processing operation results or portions of interest 12; functional means for generating at least one statistical representation (multivariate normal laws or mean vectors) representative of groupings of operation results 13; functional means for labeling the statistical representations 14; functional means (optional) for rejecting non-assignable operation results 15; functional means for assigning the results of operations 16; functional means for transmitting an alert signal 17; functional means for updating the statistical representations 18.

Such a device comprises in particular a processing unit equipped for example with a microprocessor, a random access memory, a read only memory containing a computer program comprising program code instructions for the execution of a method according to the invention.

These functional means are not described here in more detail. They typically integrate hardware components and software bricks. Their respective functionalities will be described in more detail later with the description of the control method according to the invention.

6.2. Quality control process

A work device is implemented to perform successively on a workstation, a plurality of operations of the same type. For example, a screwing operation of one type corresponds to the tightening of a given assembly, such as that of a screw fixing a cylinder head on an engine, which is repeated a plurality of times successively on a chain of production.

When it is put into service, a quality control device according to the invention must be trained to control the quality of the type of operation for which it must make it possible to ensure quality control.

For this, a quality control method according to the invention comprises an initial learning step of the control model, the learning being of the unsupervised type. The preferred embodiment uses a model of the Gaussian mixture type leading to the generation of statistical representations of the results of operation in the form of multivariate normal laws.

6.2.1. Initial model training

The initial learning step 200 consists, from a plurality of initial results of operations, in obtaining at least one statistical representation of the initial results in the form of at least one initial multivariate normal law, each law being representative of an initial group of results. Each multivariate normal distribution is then labeled as being representative of compliant results or of non-compliant results.

In production, each new operation result is assigned to one of the multivariate normal distributions in such a way that it is assigned a compliant or non-compliant result label.

The initial learning step 200 of the model comprises several steps. i. Collection

The learning firstly comprises a step 20 of collecting a plurality of initial results of operations.

The initial results of operations are the results of operations obtained by carrying out a plurality of operations of a type whose quality is subsequently to be controlled by the implementation of the quality control method. An operation result is, for example, one or more tables of values of several measured physical quantities constituting a series of data, or time series, as indicated previously, such as for example the table shown in Figure 1. ii. Selection of portions of interest

It is possible to train the model on all the data sets of the initial results. It is however more efficient to select from the series of data of each result of operation, a portion of interest, corresponding to a part of an operation in which it is appropriate to be of particular interest. The fact of selecting a portion of interest makes it possible to reduce the number of data taken into consideration by concentrating on a large portion and thus to simplify the model for example to reduce the calculation times. The learning can therefore comprise an optional step 21 of selecting a portion of interest in the data series of each initial result of operation.

In the context of a tightening operation, an operation result table can include a certain number of markers corresponding to characteristic instants of a tightening operation. In this case, the markers taken into consideration in this example are: the angle counting threshold pair; the motor stopping torque of the tightening device.

The angle counting threshold torque corresponds to the value of the tightening torque from which the value of the tightening angle is measured.

The motor stopping torque is the tightening torque at the moment when the tightening device stops to mark the end of the tightening operation.

The markers are mentioned here in a line of the table in the form of the number 4 (it could be any other character) in the column corresponding to the reaching of the angle counting threshold torque and the number 2 (it could be any other character different from the previous marker) in the column corresponding to the reaching of the engine stopping torque.

The step of selecting a portion of interest therefore consists, for each operation result, in extracting from the data series the data located between the markers. iii. Pre-processing: feature extraction

The learning then comprises a step 22 of pre-processing the data series (or portions of interest of data series).

Such a pre-processing may for example consist in carrying out, on each series of data or portion of interest, a series of predetermined calculations each resulting in an extracted characteristic.

A series of calculations can for example come from a library such as the tsfresh library supported by Python. Other libraries could be used such as SciPy for example.

The series of a number d of predetermined calculations carried out on each series of data or portion of interest (resulting from an operation result) gives a vector Xi comprising d extracted features. Assuming that the number of data series (or portions of interest of data series) taken into consideration is equal to N, a matrix X of data of dimensions Nxd is obtained at the end of the preprocessing. These extracted feature vectors are preferably saved for later use, for example to update the model, as will be explained in more detail below. iv. Obtaining a set of statistical representations of the initial operation results in the form of initial multivariate normal law(s)

The learning then comprises a step 23 of obtaining a statistical representation of the initial results in the form of at least one initial multivariate normal law. Each multivariate normal distribution is representative of a group (or cluster) of initial operation results. In this embodiment, this step implements a statistical model of the Gaussian mixture model (GMM) type.

The GMM model determines, from the matrix X of extracted features, Kmax proposals for mixing initial multivariate normal distributions. Each mixture contains K multivariate normal laws, K varying from 1 to Kmax.

Each multivariate normal distribution of a mixture of multivariate normal distributions is representative of a cluster of initial operation results taken into consideration to carry out the learning.

Each multivariate normal distribution is characterized by: pk: mean vector of each multivariate normal distribution åk: covariance matrix of each multivariate normal distribution nk: weight of each Gaussian of the mixture of the multivariate normal distributions The weight of a multivariate normal distribution in a mixture of multivariate normal distributions is representative of the number of operation results assigned to said multivariate normal distribution with respect to the total number of operation results taken into account and the sum of the weights of all the multivariate normal distributions of a mixture is equal to 1. Among the Kmax multivariate normal law mixture proposals (or set of statistical representations), it is then determined which offers the best compromise between precision and simplicity of the model. For this, the Bayesian Information Criterion (BIC) is calculated using the formulas below:

BIC(K) = -2 log (C\Q _K ) + Y _k log(N)

K ^* =argminBIC(K)

With :

K Number of normal laws considered in the different mixing proposals.

X Matrix of features extracted from the results used for learning = {nk, \ik,åk} with kE{\,..,K} P parameters of the mixture of multivariate normal distributions nk Weight of each Gaussian of the mixture of the normal distribution multivariate Yk = K+ Ad + Ad(d-1)/2 Number of parameters of each proposition N Number of operation results used for learning K* optimal number of initial multivariate normal distributions.

At the end of the initial learning, the GMM provides a mixture of a number K* of initial multivariate normal distributions in which each initial multivariate normal distribution is representative of a cluster of initial operation results.

These elements are not further detailed here since they are also known in the state of the prior art. v. Labeling of initial groups of operation results

The learning continues with the implementation of a step 25 of labeling the multivariate normal distributions.

During the labeling, each multivariate normal distribution is assigned a label, for example one of the following labels: labeled multivariate normal distribution of conforming results; labeled multivariate normal distributions of nonconforming results.

For this, the system provides the method technician, for each multivariate normal law, with an operation result having led to the generation of this law, in the form of an operation result curve, a result table of operation or otherwise, so that it can determine whether that result is compliant or not. The normal law multivariate is labeled accordingly. Thus, in this unsupervised method, the technician must only label one operation result per cluster rather than all of the operation results.

A compliant or non-compliant label is thus associated, by an operator via an HMI interface, with each multivariate normal law. The nature of the fault encountered for the operations having led to each multivariate normal law can also be entered (stick slip, seizure, etc.).

The labeling of the multivariate normal distributions marks the end of the initial training of the model.

When the initial learning of the model is complete, the quality control process can be implemented during production to check, in real time, at the end of each operation, whether or not it complies with the level of requirement. . vi. Determination of an overall rejection threshold

It may happen, during production, that new results of operations collected cannot be assigned to any of the multivariate normal laws generated during learning.

These operating results that can be assigned to any existing multivariate normal law may result from a drift in the production process and affect quality requirements.

They may correspond to the appearance of a new fault for which it is not only necessary to generate an alert but also to prepare monitoring.

Model learning is an opportunity to establish a rejection mechanism to detect, in production, the results of operations that cannot be assigned to the multivariate normal laws generated during learning.

It is known to check whether an operation result X _t can be assigned to a multivariate normal law by comparing its probability density / _¾ .( _έ ) with a threshold th _k specific to said law.

With: f _k {Xd = p{Xi\k) d Number of predetermined calculations å _k Covariance matrix for the Gaussian considered t is chosen between 1 and 10% , the higher t is, the more the assignment of a result to a normal distribution implies that this result is centered on the normal distribution. If this normal law is representative of compliance with the quality requirements or of a given physical defect, then choosing a high t amounts to retaining as belonging to this law the physical results close to the quality requirement or close to the physical defect. .

Whether : < ^th k

Then, it is considered that the result of operation does not belong to the said normal law.

In the context of the present invention, the objective of the rejection mechanism is to determine whether an operation result X _t is likely to belong no longer to a single normal distribution but to a mixture of several normal distributions.

It is therefore envisaged to establish a global threshold specific to the mixture of normal distributions with which the global probability density f(X _t ) of an operation result X _t will be compared. The aim in establishing a global rejection threshold is firstly to allow verification of whether a new operation result belongs to one of the multivariate normal laws in a single control operation and secondly to make the rejection mechanism more reliable by tightening the assignment of new results to low-order normal laws and by facilitating the assignment of new results to high-order normal laws.

Thus, the weight of the different normal laws is taken into account to calculate the global threshold during step 24 for determining the global rejection threshold.

The global rejection threshold thg is then calculated as follows:

The use of this global threshold is explained below.

6.2.2. Quality control in production During each operation carried out in production with the tool, the tool and/or the controller conventionally records the result of the operation. This operation result is transmitted for checking to the quality control device so that a compliant or non-compliant label is assigned to each operation result. i. Collection of a new operation result

Production quality control 300 therefore includes a step 30 of collecting the new operation result in the form of a series of data. ii. Selection of a portion of interest

This collection step 30 is optionally followed by a step 31 of selecting a portion of interest in the data series of the new operation result. This step is implemented when it was during the initial training of the model. iii. Pretreatment

The series of data, or the portion of interest, is then preprocessed in a preprocessing step 32 equivalent to the preprocessing step implemented during the initial learning.

At the end of the pre-processing step, the pre-processing means generate a vector K _j of extracted characteristics which is recorded for possible use during subsequent steps. This vector can be used in the rejection mechanism, in the classification if it is not an exception result, or when updating the model. Xi is a vector of features extracted from an initial operation result while l^ is a vector of features extracted from a new operation result recorded in production. iv. Exception result rejection

Quality control in production continues with the (optional) implementation of a step 33 for rejecting the new operation result if it cannot be assigned to any of the labeled multivariate normal laws, and if necessary a step 34 saving the rejected result as an exception result.

This rejection step 33 consists in rejecting any new result of operation which is known not to be assignable to any of the multivariate normal laws generated during the learning phase. The step of rejecting exception results includes: calculating the global probability density f{Yi) of the new operation result Kr the comparison of the probability density f{Y ) of the new result Y _L with the global rejection threshold thg, the rejection of the new operation result if its probability density f{Yi) is lower than the global rejection threshold thg, : f(Yù < thg the recording of the new operation result as an exception result if it is rejected. v. Assignment to an existing multivariate normal distribution of a new non-rejected operation result

Classification

If the new operation result is not rejected during the rejection step, the quality control in production continues by implementing a step 35 for automatically assigning the new operation result considered to one of the initial labeled multivariate normal distributions.

This step consists of determining the multivariate normal distribution to which the new operation result can be attributed. The multivariate normal distribution to which the new affected operation result is determined using the following relationship:

Ci = argmaxp k

Updated Gaussian weight and global rejection threshold

Each time a new operation result is assigned to one of the multivariate normal distributions, the weight of the multivariate normal distributions is updated according to the formula:

With, if k different from the index of the cluster receiving the new result then: p(fc|¾ = 0 The global rejection threshold thg is updated accordingly. vi. Alert in the event of rejection or assignment to a non-compliant multivariate normal distribution

If the new operation result considered is assigned to a multivariate normal law that does not comply with the quality requirements, then the quality control process in production continues by implementing a step 36 for issuing an alert ( for example visual and/or sound) to the attention of the method technician and/or the operator working to carry out the operations to warn him that the screwing or drilling operation requires a corrective action. vii. Quality control in recurring production

The quality control process in production is implemented each time a new operation is carried out by an operator. viii. Counting Exception Results and Updating Learning

The quality control process in production preferably comprises: a step 37 of counting the number of exception results recorded; a step 38 of comparing the number of exception results with a predetermined counting threshold; a step 39 for triggering a learning update when the counting threshold is reached.

6.2.3. Learning Update

The method includes a learning update step 400, 400'. This step is implemented during production quality control.

The learning update is implemented throughout the implementation of the quality control process in production, for example each time the number of exception results collected is sufficient. It could also be triggered manually or even according to a given temporal frequency.

The learning update can be carried out according to two variants i. Updating Learning with Exception Findings Only Updating 400 learning can be performed considering only exception findings.

This update consists of: 40- in obtaining, in a step for obtaining an update, one or more new multivariate normal laws representative of new groups of results from the exception results previously collected and preprocessed during the rejection step, c that is to say from the characteristics extracted previously recorded from these results; the method for obtaining these new multivariate normal laws uses, as before, the GMM and the calculation of the Bayesian information criterion to determine a mixture of optimum multivariate normal laws.

41- labeling the normal laws obtained by an operator assigning them, as explained previously, one of the following labels: multivariate normal laws labeled with conforming results; labeled multivariate normal distributions of nonconforming results.

The weights of the initial multivariate normal distributions obtained during the initial learning and of the new multivariate normal distributions are then updated, as well as the global rejection threshold by implementing the following steps:

42- to calculate the new number of operation results taken into account: /V = N<P ^rev > + N< ^new >)

43- to calculate the new weight of the initial multivariate normal distributions: 77l

= n _h <P ^rev >.N ⁽ P ^rev >/N

44- to calculate the weight of the new multivariate normal distributions: 77l =

_Nk < ^new >/N

45- to calculate the new value of the global rejection threshold thg.

With o N ⁽ P ^rev > Number of operation results assigned to the multivariate normal laws resulting from initial learning o N(new) Number of operation results belonging to the multivariate normal laws obtained during the update.

These operations, which are similar to those implemented during the initial training, are not described in more detail here. The new multivariate normal laws resulting from step 40 are associated with the initial multivariate normal laws of the mixture obtained during the initial learning.

The new model thus obtained is applied when implementing the quality control process in production after this update.

With each new learning update, new multivariate normal laws are generated from the exception results which are associated with the previous normal laws including the initial laws and the laws generated during the previous updates. ii. Updating learning with exception results and results assigned to previous multivariate normal distributions

The 400' learning update can be carried out by taking into account the exceptional results and the results assigned to existing multivariate normal laws resulting from the initial learning.

This update consists of:

40'- to obtain, in a step of updating obtaining, new multivariate normal laws from the exception results previously collected and preprocessed during the rejection step and from the results assigned to the multivariate normal laws resulting initial learning, that is to say from the characteristics extracted previously recorded from all the results;

41 '- to label the normal laws obtained by an operator assigning them, as explained previously, one of the following labels: multivariate normal laws labeled with conforming results; labeled multivariate normal distributions of nonconforming results.

42' - to calculate the new number of operation results taken into account: N= N ⁽ P ^æv > + N< ^new >)

43' - to calculate the new weight of the new multivariate normal distributions; 44' - to calculate the new value of the global rejection threshold thg.

The new multivariate normal distributions are substituted for the multivariate normal distributions resulting from the initial learning. At each new learning update, the new multivariate normal laws are generated from the new operation results and the previous results including the initial results and the previous operation results iii. Advantages and disadvantages of the two learning update solutions The first type of update has the advantage of being less computationally heavy than the second type of update. It also keeps the clusters obtained during the initialization and can thus allow the highlighting of a possible drift of the new results vis-à-vis these initial clusters.

The second type of update is heavier in terms of calculation but it offers the advantage of greater precision given that the multivariate normal laws obtained during the initialization are re-estimated taking into account not only the initial results but also the results that have been attributed to the clusters associated with these laws during the classification phase in production.

6.3. Variant of K-means

In a variant, the statistical model implemented for the grouping of the results of operations into clusters could be different from that of the GMM type.

A k-means type model could for example be used. Indeed, the objective of the k-means is to partition the set of results into a number of K clusters, such that the sum of the squares of the distances of each point with respect to the center of the cluster to which it has been assigned is minimal.

Similar to the Gaussian mixture model, this learning involves several cluster set calculations, each of its calculations presupposes a number of clusters chosen between 1 and a maximum Kmax.

Then, among the different sets of clusters, the one that offers the best compromise between accuracy and simplicity is selected.

The calculation of the Silhouette criterion (Sht(K)) allows this evaluation.

With : K Number of cluster considered

Xi vector of the characteristics extracted from the result i Shte{X _j Silhouette index associated with the vector/^ a*: the average of the intra-cluster distance in which X _t is affected bi: the distance between X _t and the second-most cluster close.

The rejection mechanism will be characterized by:

Calculate a numeric threshold for each cluster

The calculation of this numerical threshold is based on the dispersion of the points which were assigned to it

For each result i assigned to a cluster k, the square of the distance Df _k is calculated.

D _k = å”=i df _j where d _j is the distance between result i and the ^jth nearest result (which belongs to the same cluster as that of result i).

The number n is the number of n nearest neighbors of the result i and is fixed by the method technician.

The numerical threshold th _k is fixed by the 100(1-a)% percentile of the squared distance distribution on the ^kth cluster. a is similar to the first kind risk on the control charts and is set by the method technician.

For a new result p, if Dp _k > th _k vfc then the new result is considered to be of unknown type.

6.4. Example of system architecture

FIG. 8 illustrates a complete work and quality control system 500 which may for example comprise: a tool (50 for example a screwdriver or a drill); a controller 51 able to communicate with the tool 50; a man-machine interface (HMI) 52 comprising for example the form of a PC; a server 53 able to communicate with the controller and with GIHM. During the initial learning, the tool makes it possible to carry out a plurality of operations and the controller exports the result of each operation to the server.

To perform learning, the server collects the initial operation results exported by the controller. The server carries out the initial learning while IΊHM is used at the end of learning by the method technician to carry out the labeling of the multivariate normal laws.

During quality control in production, the tool makes it possible to carry out successive operations. Each operation result is sent by the tool to the server. The server either rejects the new operation result as an exception result or assigns it to a conforming or nonconforming multivariate normal distribution. If the operation result is rejected as an exception result or assigned to a non-compliant multivariate normal law, the server generates the emission of an alert signal, for example by IΊHM.

The technician controls by HMI the activation of the learning update. The update is performed by the server. During this update, the technician labels the new multivariate normal laws using IΊHM.

In a variant, during quality control in production, the tool makes it possible to carry out successive operations. Each operation result is exported by the tool to the server. The server rejects the new operation result as an exception result or assigns it to a conforming or nonconforming multivariate normal. If the operation result is rejected or assigned to a non-conforming multivariate normal law, the server commands the emission of an alert signal. The server orders, according to a predetermined frequency or the achievement of a given number of exceptional results, the activation of the learning update. The update is performed by the server. During this update, the technician labels the new multivariate normal laws using IΊHM.

The different functional means of the quality control device according to the invention are installed in the different equipment of these architectures according to the needs.

6.5. Variant

In general, the present invention relates to a method for controlling the quality of screwing or drilling operations carried out by means of a tool, said method including automatic learning of an unsupervised statistical model.

More generally still, the invention relates to a method for controlling the quality of any type of industrial operation, said method comprising automatic learning of an unsupervised statistical model.

Claims

1. Method for controlling the quality of screwing or drilling operations carried out by means of a tool, said method comprising: a step of automatic learning of a model, said step of initial automatic learning comprising: a step collection of initial results of screwing or drilling operations recorded during a plurality of screwing or drilling operations; a step of obtaining, by means of said model, at least one statistical representation of at least part of the initial results; a step of labeling said at least one statistical representation with at least one of said following labels: labeled statistical representation of compliant results; labeled statistical representation of non-compliant results; a step of automatic control in production of the quality of screwing or drilling operations, said step of automatic control in production being implemented at the end of each operation and comprising: a step of collecting the new result of the screwing or drilling operation recorded during the operation in question; a step of automatically assigning said new result of the operation considered to the at least one statistical representation; a step for issuing an alert if the new operation result is assigned to a non-compliant statistical representation; characterized in that said learning of said model is of the unsupervised type.

2. Method according to claim 1, in which said production control step comprises a step of rejecting new results of operations which cannot be assigned to one of said statistical representations, and a step of recording the rejected results as exceptional results.

3. Method according to claim 2 comprising a step of issuing an alert when a new result obtained at the end of a screwing or drilling operation is identified as an exception result during said step of rejecting .

4. Method according to claim 1 to 3, in which said step of obtaining at least one statistical representation implements a statistical model of the Gaussian mixture model type, said at least one statistical representation being a multivariate normal law.

5. Method according to claim 1 to 3, in which said step of obtaining at least one statistical representation implements a K-means type statistical model, said at least one statistical representation being an average vector.

6. Method according to claims 2 and 4 wherein said statistical model of the mixture of Gaussian type calculates, during said step of obtaining a statistical representation, a plurality of initial multivariate normal laws each representative of a group of results initial screwing or drilling operations, each of said initial multivariate normal laws having a weight, said step of rejecting comprising a step of calculating the probability density of each new screwing or drilling operation result obtained during said step of automatic control in production and a step of comparing this density with a predetermined global rejection threshold depending on the weight of each of said initial multivariate normal laws.

7. Method according to claim 6, in which said weight of an initial multivariate normal law is representative of the number of results of screwing or drilling operations assigned to said multivariate normal law relative to the total number of results of screwing operations or drilling taken into account.

8. Method according to claim 6 or 7 wherein said classification step is followed by a step of updating the weight of all of said initial multivariate normal distributions.

9. Method according to any one of claims 6 to 8 wherein said classification step is followed by a step of updating said predetermined global rejection threshold.

10. Method according to any one of claims 6 to 9 comprising a step of updating said learning, said updating step being implemented during said step of automatic control in production and comprising a step of updating obtaining a statistical representation of the initial results in the form of at least one multivariate normal law, said updating step taking into account the new operation results identified as exceptional results to generate at least one new law multivariate normal.

11. Method according to claim 10, in which during said step of updating said learning, said statistical model of the Gaussian mixture type calculates new normal distributions, based on the exception results, each representative of a new group of results of operations, said step of updating said learning comprising a step of calculating the new weight of said initial and new multivariate normal laws.

12. Method according to claim 10, in which said step of updating said learning generates new multivariate normal laws, from the new results of operations and from the initial results of operations, each representative of a new group of results of screwing or drilling operations, said new multivariate normal laws being substituted for the initial multivariate normal laws.

13. Method according to any one of claims 10 to 12 comprising a step of counting the number of new operation results identified as exceptional results, said step of updating said learning being implemented when the number of results operations identified as exception results reach a predetermined threshold.

14. Method according to any one of claims 1 to 13 wherein each operation result is a series of data, said series of data being the subject of a pre-processing consisting in carrying out on a series of data a series of predetermined calculations each leading to an extracted characteristic, said extracted characteristics being taken into consideration for carrying out said step of obtaining a statistical representation of the initial results in the form of at least one multivariate normal law.

15. Method according to claims 6 and 14, in which said extracted characteristics are taken into consideration by said statistical model of the Gaussian mixture type to generate said multivariate normal laws, each representative of a group of results of screwing or drilling operations.

16. Method according to claim 14 or 15 wherein said series of data belong to the group comprising: a torque as a function of an angle or of a drilling depth or of time; an angle as a function of time; an intensity, in particular of a motor driving a screwing tool in rotation or driving a cutting tool in rotation or translation as a function of an angle or of a drilling depth or of time; a force as a function of an angle or a drilling depth or time.

17. Method according to any one of claims 14 to 16 comprising a step of selecting, in each of said series, a portion of data of interest, said portion of data of interest being the subject of said pre-processing.

18. Device for controlling the quality of screwing or drilling operations carried out by means of a tool, said device comprising: means for automatic learning of a model, said means for initial automatic learning comprising: means collection of initial results of screwing or drilling operations recorded during a plurality of screwing or drilling operations;

Means for obtaining, by means of said model, at least one statistical representation of at least part of the initial results; means for labeling said at least one statistical representation with at least one of said following labels: labeled statistical representation of compliant results; labeled statistical representation of non-compliant results; means for automatic control in production of the quality of screwing or drilling operations, said means for automatic control in production being able to be implemented at the end of each operation and comprising: means for collecting the new result the screwing or drilling operation recorded during the operation in question; means for automatically assigning said new result of the operation considered to at least one statistical representation; means for issuing an alert if the new operation result is assigned to a non-compliant statistical representation; characterized in that said learning of said model is of the unsupervised type.

19. Device according to claim 18, comprising means for rejecting new results of operations which cannot be assigned in production to one of said statistical representations, and means for recording the rejected results as exception results.

20. Device according to claim 19 comprising means for transmitting an alert when a new result obtained at the end of a screwing or drilling operation is identified as an exception result by said rejection means.

21. Device according to any one of claims 18 to 20, in which said means for obtaining at least one statistical representation implement a statistical model of the Gaussian mixture model type, said at least one statistical representation being a multivariate normal law.

22. Device according to any one of claims 18 to 20, in which said means for obtaining at least one statistical representation implement a statistical model of K-means type, said at least one statistical representation being an average vector .

23. Device according to claims 19 and 21, in which said statistical model of the Gaussian mixture type is capable of calculating a plurality of initial multivariate normal laws each representative of a group of initial results of screwing or drilling operations, each of said initial multivariate normal laws possessing a weight, said rejection means comprising means for calculating the probability density of each new result of a screwing or drilling operation obtained in production and means for comparing this density with a rejection threshold global predetermined depending on the weight of each of said initial multivariate normal distributions.

24. Device according to claim 23, in which said weight of an initial multivariate normal law is representative of the number of results of screwing or drilling operations assigned to said multivariate normal law relative to the total number of results of screwing operations. or drilling taken into account.

25. Device according to claim 23 or 24 comprising means for updating the weight of all of said initial multivariate normal laws.

26. Device according to any one of claims 23 to 25 comprising means for updating said predetermined global rejection threshold.

27. Device according to any one of claims 23 to 26 comprising means for updating the at least one statistical representation of the initial results in the form of at least one multivariate normal law obtained by said learning means, said updating means taking into account the new operation results identified as exceptional results to generate at least one new multivariate normal law.

28. Device according to claim 27, in which said updating means using said statistical model of the Gaussian mixture type to calculate new multivariate normal laws, from the exception results, each representative of a new group of results of operations, said means for updating said learning comprising means for calculating the new weight of said initial and new multivariate normal laws.

29. Device according to claim 27, in which said updating means generate new multivariate normal laws, from the new results of operations and from the initial results of operations, each representative of a new group of results of operations. screwing or drilling, said new multivariate normal laws being substituted for the initial multivariate normal laws.

30. Device according to any one of claims 27 to 29 comprising means for counting the number of new operation results identified as exceptional results, said updating means being implemented when the number of results of operation identified as exception results reaches a predetermined threshold.

31. Device according to any one of claims 18 to 30, in which each operation result is a series of data, said device comprising means for pre-processing said series of data consisting in carrying out on a series of data a series of predetermined calculations resulting each with an extracted characteristic, said extracted characteristics being taken into consideration by said means for obtaining a statistical representation of the initial results in the form of at least one multivariate normal law.

32. Device according to claims 23 and 31 wherein said extracted characteristics are taken into consideration by said statistical model of the Gaussian mixture type to generate said multivariate normal laws each representative of a group of results of screwing or drilling operations.

33. Device according to Claim 31 or 32, in which the said series of data belong to the group comprising: a torque as a function of an angle or of a drilling depth or of time; an angle as a function of time; an intensity, in particular of a motor driving a screwing tool in rotation or driving a cutting tool in rotation or translation as a function of an angle or of a drilling depth or of time; a force as a function of an angle or a drilling depth or time.

34. Device according to any one of claims 31 to 33 comprising means for selecting, in each of said series, a portion of data of interest, said portion of data of interest being taken into account by said pre-processing means. .

35. Computer program product, comprising program code instructions for implementing a method according to any one of claims 1 to 17, when said program is executed on a computer.

36. A computer-readable, non-transitory storage medium storing a computer program product according to claim 35.