Automated method for monitoring rotating components of a rotating machine by domain adaptation

Abstract

The invention relates to a method (100) for automatically monitoring a plurality of rotating components of a rotating machine based on a target database comprising a plurality of time signals generated by each rotating component from a distribution and based on a source database comprising a plurality of time signals from a distribution S different from the distribution T generated by a source rotating component of a source rotating machine and associated with a category of operation, the monitoring being performed by means of an adaptive deep learning model making it possible to adapt the source distribution to the target distribution, the deep learning module being trained by minimizing a cost function related to a Gaussian kernel function with a parameter σ; σ being calculated at each epoch based on a distribution difference weighted by a constant static value estimated based on a Pascal triangle.

Description

Technical Field

[0001] The technical field of this invention is the field of distributed domain adaptation and deep learning transfer.

[0002] This invention relates to an automated method for monitoring rotating components of rotating machinery. Background Technology

[0003] In many industrial sectors, the diagnosis and monitoring of various mechanical components (such as engines and their different rotating parts (bearings, gears, shafts, fans, etc.)) are essential to determine their operating condition or health status and thus plan maintenance operations to minimize downtime. Therefore, reliable diagnostic and monitoring systems or methods enable the early detection and identification of damage to prevent its spread to other mechanical components and to plan appropriate maintenance based on the health status of the monitored components. Consequently, mechanical monitoring is a significant challenge in the mechanical engineering industry, particularly in the aerospace sector.

[0004] Typically, monitoring of rotating components in rotating machinery is achieved by analyzing vibration signals generated by these components and collected by vibratory acoustic sensors (such as accelerometers). This method is commonly used to determine the operational status of aircraft engines and their rotating components. During production or maintenance, high-frequency vibration signals are collected while the machinery is running to detect weak signals (called signatures) that characterize damage to mechanical components, thereby preventing engine failure. Monitoring via vibration signal analysis is one of the most widely used methods due to its non-invasive nature and the rich diagnostic information it provides.

[0005] Typically, vibration diagnostic methods are based on signal processing techniques that utilize vibration signals. The signal is input into a processing task that includes multiple processing methods (source separation, filtering, denoising, etc.) aimed at extracting or refining the "vibration features" of interest and relating them to mechanical components. Therefore, vibration analysis involves inferring the health status of rotating mechanical components from their vibration characteristics; this inference requires prior kinematic knowledge and monitoring expertise. Subsequently, appropriate indices can be constructed to quantify damage and facilitate decision-making.

[0006] Artificial intelligence-based monitoring and diagnostic methods are rapidly developing. The main goal of these methods is to replace the human knowledge required in traditional methods with machine learning based on large amounts of data.

[0007] Against this backdrop, intelligent fault diagnosis (IFD) is emerging as an auxiliary or alternative solution to the previously discussed signal processing methods. In recent years, with the rapid development of deep learning, intelligent fault diagnosis has received considerable attention.

[0008] The success of deep learning models depends on both the choice of model architecture and the representativeness and richness of the training data. For example, in bearing monitoring, signals from healthy bearings with various fault types are needed to train the model, enabling it to analyze new bearings and classify their conditions. The success of IFD methods relies on a common assumption: sufficient labeled data to form a reliable diagnostic model. However, in the aerospace field, due to the rarity of faults, it is difficult to collect enough labeled data. Therefore, unlabeled data from real machinery cannot train a diagnostic model to provide accurate results. Furthermore, the difference in the acquisition context between the training signals and the signals to be classified in existing databases can affect the performance of the learning model, as two signals acquired in two different contexts originate from two different probability distributions.

[0009] Domain adaptation is a tool for reusing knowledge learned from a set of signals acquired from rotating machinery in a first context (this set of data is called the source domain) by transferring the data to the analysis of relevant signals acquired for the same type of rotating machinery in a different context (this set of data is called the target domain). The source domain represents a set of data that has already been acquired and labeled, i.e., its operating category is known.

[0010] To reuse the source database for diagnostic tasks on the target database, deep transfer learning models are used due to their efficiency and ability to reduce the large distributional differences between the source and target domains. These deep transfer learning models take into account the parameters to be set, particularly the model's cost function. On the other hand, the health status of the monitored mechanical components is unknown. Therefore, the data from the target domain is naturally unlabeled. Thus, the main challenge lies in being able to adjust the parameters of the deep learning model without relying on labels of the rotating machinery's operating state to determine whether the model has correctly predicted the current state of the rotating machinery. Summary of the Invention

[0011] This invention provides a solution to the problems discussed above, enabling automatic monitoring of rotating parts with unknown operating categories by adjusting parameters of the learning model that are independent of the target database, based on a deep transfer learning model using a source and target database.

[0012] A first aspect of the invention relates to a method for automatically monitoring at least one rotating component of rotating machinery from an unmarked database and a source database, the unmarked database being referred to as a target database, the target database including at least one time signal originating from a distribution T, the time signal being generated by the rotating component, and the source database including a plurality of time signals originating from a distribution S different from the distribution T, each time signal in the source database being generated by a source rotating component of the source rotating machinery and associated with an operating category in a set of operating categories, the set of operating categories including at least a nominal operating category and a fault operating category, characterized in that the method includes the following steps:

[0013] A first neural network is trained to associate a run category from a set of run categories with a non-stationary time signal originating from a distribution S. The first neural network includes a first part for feature extraction and a second part for classification. The first neural network is trained based on source data.

[0014] A so-called adaptive artificial neural network is trained for M training epochs to obtain a trained artificial neural network capable of associating one of a set of running categories with a temporal signal derived from a joint distribution of S and T. This adaptive artificial neural network includes a first part for feature extraction corresponding to a first part of a first neural network, and a second part for adapting distribution S to distribution T and for classification. The artificial neural network is trained simultaneously on both a source database and a target database, and training is performed by minimizing a cost function, which includes:

[0015] The first term corresponds to the error between the category obtained by the neural network for each signal in the source database and the category associated with that signal in the source database;

[0016] The second term is calculated based on at least one maximum mean difference between the functions of the source database and the functions of the target database, the maximum mean difference being based on parameters. The parameter is calculated using at least one Gaussian kernel function. This parameter is estimated based on Pascal's triangle. It is related to the variance of the Gaussian kernel function and depends on each training epoch. The Gaussian kernel function of order is based on the size of The last row of Pascal's triangle is determined.

[0017] An adaptive artificial neural network trained on each signal in the target database is used to associate a running category with the signal.

[0018] Target domain D T Defined as ;in X It is the space of signal characteristic descriptors. It is a marginal probability distribution, and .

[0019] Source Domain D S Defined as ;in X It is the space of signal feature descriptors. It is a marginal probability distribution, and .

[0020] "Distribution S is different from distribution T" means .

[0021] Specifically, adaptation from distribution S to distribution T is achieved by minimizing the second term of the cost function, which is computed based on a source database from distribution S and a target database from distribution T. The goal of the adaptation domain is to reduce the distributional discrepancies between the source and target databases (during or after training) so that the classification of the target database is nearly identical to that of the source database. Both databases deal with the same topic (e.g., bearing failures in rotating machinery), but each database pertains to different operating conditions of the rotating machinery (speed, load, torque, etc.). In particular, Maximum Mean Discrepancy (MMD) is one of the most efficient functions for minimizing the discrepancies between distributions, and thus performing adaptation across multiple domains, for example, by minimizing the maximum discrepancy between the distributions.

[0022] Therefore, before training the adaptive artificial neural network, the first part of the adaptive artificial neural network is identical to the first part of the first neural network. The architecture and parameters of the first part of the artificial neural network are the same as those of the first part of the first neural network. The architecture of the neural network (or a part of it) corresponds to the number of layers, neurons, and their arrangement. In other words, the first part of the artificial neural network and the first part of the first neural network are the same.

[0023] The present invention advantageously makes it possible to: obtain data from a labeled source database, by means of parameters dependent on Pascal's triangle. The signal from an unlabeled target database (where the state of the rotating machinery is unknown under the condition of acquiring a signal of distribution T) is classified into one of a set of categories without adjustment for missing labels in the target database. The signal from the source database has a different distribution than the signal from the target database. In fact, since the cost function of the neural network is partially minimized by means of a Gaussian kernel function, the last row of Pascal's triangle advantageously allows a Gaussian kernel function of the same size as that last row to be represented as a triangle (e.g., the size is chosen according to the source and target databases), and thus its parameters can be determined without any additional external adjustments. Therefore, the present invention makes it possible to automatically monitor the rotating parts of a rotating machinery using signals from a distribution T without associated labels, without expert intervention, thus saving time in maintaining the rotating machinery.

[0024] To reduce the distributional differences among multiple databases (both between and within each database), a parameter calculation method was introduced. The Gaussian kernel (and therefore the cost function) is used. Therefore, using the same kernel every time to reduce this difference is not very efficient, because the difference changes every time the model adjusts its weights (i.e., in each training epoch). Therefore, the Gaussian kernel must be defined during each iteration to update this reduction paradigm. On the other hand, the parameter defining the Gaussian kernel is its variance, or equivalently its standard deviation (in this case, determined by...). express, This parameter varies accordingly, depending on the differences in the distribution.

[0025] Therefore, due to the parameters It varies with each training epoch, making the distributions more similar in each epoch, so the estimate of this parameter is very robust to the variance of the differences between the distributions.

[0026] Advantageously, the second term of the cost function, and in particular the maximum mean difference, is a dynamic value that changes in each training epoch.

[0027] Advantageously, the operator does not need to try, experiment, or adjust to find the optimal variance (or equivalent standard deviation) of the Gaussian kernel.

[0028] In addition to the features just discussed in the preceding paragraphs, the method according to the first aspect of the invention may have one or more of the following additional features, which are considered individually or in any technically possible combination:

[0029] The first part of this adaptive artificial neural network includes N c1 The number of layers, N c1It is a natural number greater than or equal to 1, and the second part of this adaptive artificial neural network includes N. c2 The number of layers, N c2 The second term of the cost function is a natural number greater than or equal to 1. It is computed on one or more layers belonging to the second part of the neural network, and it lies before the last layer of that second part. Specifically, the maximum mean difference is computed on each of the layers in the second part (before the last layer). Therefore, the second term of the cost function includes the sum of the computed maximum mean differences. The last layer is a decision layer used to classify the input signal into a running category after the distributions in the one or more layers preceding the last layer have become more similar.

[0030] For multiple n T Monitoring is performed on rotating mechanical components, and the target database includes... It represents multiple signals.

[0031] The source database contains multiple signals from This indicates that the second term of the cost function is determined according to the function in each period. To calculate the signal set This indicates that the m-th layer of the adaptive artificial neural network is related to the input. The output of each signal Having length and signal set This indicates that the m-th layer of the adaptive artificial neural network is related to the input. The output of each signal Having length , m is an integer included in the set Ncc, which includes each stratum number from the second part, and the maximum mean difference is calculated based on the set Ncc, where:

[0032]

[0033] in:

[0034] ,in Is with sets Related Gaussian kernel function variance

[0035] ,in Is with sets Related Gaussian kernel function variance

[0036] ,in Is with sets and set Related Gaussian kernel function variance

[0037] P is a natural number greater than or equal to 1, and each vector Depend on This means that each vector Depend on This means that each vector Depend on Representation, and each matrix Depend on This indicates that the parameter equals set Advantageously, the Gaussian kernel function The quantity P makes improving the function This enables the computational accuracy of the training cost function of the adaptive artificial neural network. Specifically, the distributions of S and T change at each training epoch. Therefore, at each epoch, the parameters used to compute the second term (and including its three components) of the cost function are... They also change to correspond to the new distribution. Therefore, these three components change in each period to adapt to the changes in distributions S and T.

[0038] The training of this adaptive artificial neural network is performed over M epochs, and for each of the M epochs, parameters are estimated in the following sub-steps. :

[0039] For each layer :

[0040] Sure , It is a set The length of each signal and the set The maximum length between each signal,

[0041] Each signal Resampling to size Two-dimensional image and each signal Resampling to size Two-dimensional image ,

[0042] Determine in the following sub-steps Variance of the Gaussian kernel function :

[0043] Constructing a matrix The matrix express A Gaussian kernel of order 1, the matrix equal: , express The last row of Pascal's triangle of order 1, and yes transpose,

[0044] According to the formula Calculate variance , equal matrix The maximum value,

[0045] Calculate the vector according to the following sub-steps :

[0046] according to To calculate ,

[0047] according to To calculate ,

[0048] according to To calculate ,

[0049] Calculate according to the following formula Advantageously, to save computation time, the signal is resampled and converted into an image. Furthermore, parameters are estimated using the source and target databases and Pascal's triangle. This eliminates the need to label the target database to be categorized. The amount , and It is calculated for each training period because The amount , and It depends on the difference in the distribution of changes in each training period. It is a static reference value (in other words, the value does not change with training periods) and is used to evaluate the changes after each training period, and thus to evaluate each adjustment of the parameters (or weights) of the adaptive artificial neural network.

[0050] According to one embodiment:

[0051] Each coefficient equal ,

[0052] Each coefficient equal ,

[0053] Each coefficient equal .

[0054] P equals 5. The higher the value of P, the more accurate the parameter estimation. However, increasing the order has drawbacks from the perspective of available computational resources. Advantageously, P equals 5, making it easier to... A trade-off between accurate estimation and reduced computation time becomes possible.

[0055] A third aspect of the invention relates to a computer program product comprising instructions that, when executed on a computer, cause the computer to perform the steps of the method according to the first aspect of the invention and the method according to the second aspect of the invention.

[0056] A better understanding of the invention and its various applications will be gained after reading the following description and examining the accompanying drawings. Attached Figure Description

[0057] These accompanying drawings are provided to illustrate the purpose of the invention and do not limit the purpose of the invention in any way.

[0058] Figure 1 A block diagram of a method for monitoring rotating components of rotating machinery according to the present invention is shown.

[0059] Figure 2 This indicates that the time signal is resampled into a two-dimensional image.

[0060] Figure 3 This represents a Gaussian kernel with dimensions of 28×28. Detailed Implementation

[0061] These accompanying drawings are provided to illustrate the purpose of the invention and do not limit the purpose of the invention in any way.

[0062] Figure 1 A schematic representation of a block diagram of a method 100 for automatically monitoring at least one rotating component of rotating machinery based on a “target” database and using a source database, the “target” database comprising at least one time signal from a distribution T and generated by the rotating component, and the source database comprising multiple time signals from a distribution S different from the distribution T.

[0063] Rotating machinery is, for example, an engine, and more preferably, an aircraft turbojet engine.

[0064] For example, the rotating parts of rotating machinery are one or more mechanical shafts, one or more bearings, one or more fans, one or more turbines, or one or more compressors.

[0065] According to one embodiment, for multiple (n) rotating machinesT (Number) rotating components are monitored, and the target database includes (number) rotating components. Represents multiple signals, n T It is a natural number greater than 1.

[0066] n T The rotating parts are of the same type, therefore n T A rotating component is, for example, a bearing.

[0067] Each time signal By n T One of the rotating components is generated by a rotating component.

[0068] Each time signal can be a non-stationary time signal. A "non-stationary time signal" refers to a physical time signal whose frequency content changes with time.

[0069] In the remainder of this document, the terms “time signal” or “signal” will be used interchangeably.

[0070] In the remainder of this document, the terms “target database signal” or “target signal” will be used interchangeably.

[0071] Each target signal is preferably generated by the vibration of a so-called target rotating component of a so-called target rotating machine.

[0072] For example, each target signal is measured using a sensor that may be located on the rotating machinery of the target; such a sensor is a vibration-acoustic sensor. This vibration-acoustic sensor could be, for example, an accelerometer, a strain gauge, or a microphone.

[0073] Each target signal includes L T There are a number of points, where the number L is... T It can depend on the sensor's sampling frequency.

[0074] According to a preferred embodiment, each target signal is a vibration signal.

[0075] Multiple target signals are from It means that n T It is a non-zero natural number representing the number of target signals in the target database.

[0076] Each non-stationary time signal from distribution T is a signal generated in background T' by the vibration of rotating parts of rotating machinery.

[0077] Target domain D T Defined as ;in X It is the space of signal feature descriptors. It is a marginal probability distribution, and .

[0078] Background (e.g., values ​​of parameters related to rotating machinery) corresponds to the operating conditions of the rotating machinery. Parameters related to rotating machinery can be the rotational speed of the rotating machinery, the load, the temperature of the rotating machinery, the type of rotating machinery, or the location of vibration-acoustic sensors in the rotating machinery.

[0079] Multiple source signals from It means that n s It is a non-zero natural number representing the number of source signals in the source database. The source database includes n from the source rotating machinery. s n generated by the rotating component of the source s One signal.

[0080] Each time signal in the source database is generated by the source rotating part of the source rotating machinery and is associated with one of a set of operating categories. Preferably, each signal in the source database is generated by the vibration of the source rotating component.

[0081] The database in which each signal is associated with one of the categories in a set is called "labeled". Therefore, the source database is labeled.

[0082] In the following text, the terms “source database signal” or “source signal” will be used interchangeably.

[0083] For example, a sensor, possibly located on the rotating machinery at the source, can be used to measure each source signal; this sensor could be, for example, a vibration-acoustic sensor. The vibration-acoustic sensor could be, for example, an accelerometer, a strain gauge, or a microphone.

[0084] Each source rotating component has the same type as each other rotating component; that is, if each rotating component is a bearing, then each source rotating component is a bearing.

[0085] Each source signal includes L S There are a number of points, where the number L is... S It can depend on the sensor's sampling frequency.

[0086] Each source rotating machinery has the same type as each target rotating machinery; that is, if each target rotating machinery is an aircraft engine, then each source rotating machinery is also an aircraft engine.

[0087] Natural number n s It can be equal to or different from the natural number n T .

[0088] If the value of a single parameter related to the rotating machinery changes during the acquisition process, then the two signals originate from the same distribution.

[0089] Each non-stationary time signal originating from the S-distribution is a signal generated in the background S' by the vibration of rotating parts of rotating machinery.

[0090] Source domain D S Defined as ;in X It is the space of signal feature descriptors. It is a marginal probability distribution, and .

[0091] Background S' and background T' are different.

[0092] If two signals are acquired in two different backgrounds, at least two parameters related to the rotating machinery will have different values, and the two signals will have different distributions.

[0093] When two signals are acquired in two different backgrounds, the domains of these two signals are considered to have different distributions. Therefore, the domain D s and D T They are considered to have different distributions.

[0094] Domain D s and D T "Having different distributions" means: and (and therefore) ).

[0095] For example, a set of signals, including a first signal, a second signal, and a third signal, is measured by sensors attached to the same rotating machinery. Each signal is measured on the same type of rotating machinery, at the same temperature, and at the same location on the machinery; the only parameter affecting its value is the rotational speed of the machinery, denoted by N². The first signal is... The second signal was measured at revolutions per minute (rpm). The third signal is measured below. The measurements were taken against the same background, and the first, second, and third signals formed part of the same distribution.

[0096] For example, the first signal and the second signal are measured by sensors connected to each other. The first signal is the rotational speed. The measurement was taken at a load of 1 HP, and the second signal was measured at the rotational speed. Furthermore, the measurements were taken under a load of 3 HP. When the first and second signals were acquired, the two parameters related to the rotating machinery had different values, therefore the first and second signals did not originate from the same distribution.

[0097] This group of operating categories is a group of operating categories for rotating parts.

[0098] For example, a set of operating categories for rotating components includes at least the following categories: nominal operating category and fault operating category.

[0099] For example, a set of operating categories for rotating components includes the following categories: nominal operating category, operating category with a first fault, operating category with a second fault, and operating category with a third fault. For example, the first fault could be a wear fault. For example, the second fault could be a scaling fault.

[0100] According to one embodiment, method 100 may include a first step 101 of supervising the training of a first artificial neural network to obtain a trained artificial neural network that is capable of providing categories included in a set of running categories based on signals belonging to distribution S.

[0101] Supervised training (or supervised learning) enables the training of an artificial neural network for a predefined task by updating its parameters in a manner that minimizes the cost function corresponding to the error between the output data provided by the artificial neural network and the actual output data. That is, what the artificial neural network should provide as output to complete a predefined task based on certain input data.

[0102] The artificial neural network preferably includes a first part for feature extraction and a second part for classification.

[0103] The first part preferably comprises a set of convolutional artificial neuron layers and a set of so-called pooling artificial neuron layers, each pooling artificial neuron layer preceding a convolutional artificial neuron layer and each pooling artificial neuron layer following a convolutional artificial neuron layer. For example, the first part performs the extraction of health indicators from rotating machinery and is trained to appropriately (by means of a cost function) extract features of these health indicators.

[0104] The first part includes N c1 A number of layers, N c1 It is a natural number that is greater than or equal to 1, and preferably equal to 4.

[0105] Supervised training of the first artificial neural network includes updating the parameters of the first artificial neural network to minimize the cost function corresponding to the error between the prediction of the category provided by the artificial neural network based on the source signals of the source database and the category associated with the source signals of the source database.

[0106] The cost function can be, for example, the root mean square function or the cross-entropy function.

[0107] For example, a stochastic gradient descent algorithm with back-propagation through time (BPTT) can be used to minimize the cost function.

[0108] The first neural network is trained on B periods, where B is a natural number greater than 1.

[0109] The method 100 includes step 102 of training a so-called adaptive artificial neural network to obtain a trained artificial neural network capable of associating a class from a set of classes with a non-stationary time signal originating from a joint distribution of S and T. The non-stationary time signal originating from the joint distribution of S and T is a signal that may originate from either distribution S or distribution T.

[0110] An adaptive artificial neural network is trained simultaneously on both the source and target databases.

[0111] Adaptive artificial neural networks consist of a first part for feature extraction and a second part for so-called domain adaptation and classification.

[0112] Preferably, before training the adaptive artificial neural network begins, the first part of the adaptive artificial neural network is identical to the first part of the first neural network. Therefore, the first part of the adaptive artificial neural network includes N. c1 A number of layers, N c1 It is a natural number that is greater than or equal to 1, and preferably equal to 4.

[0113] Preferably, the second part of the adaptive artificial neural network includes a set of so-called "fully connected" layers.

[0114] Preferably, the second part of the adaptive artificial neural network includes N c2 A number of layers, for example, N c2 It is a natural number greater than or equal to 1, and preferably equal to 4 or 5.

[0115] In the following text, the notation This represents the m-th layer of an adaptive artificial neural network for the input. The output, and This represents the m-th layer of an adaptive artificial neural network for the input. The output of , where m is included in the interval The natural numbers in the array.

[0116] gather equal .

[0117] gather Each element The number of points is The signal, in which It is a non-zero natural number.

[0118] gather equal .

[0119] gather Each element The number of points is The signal, in which It is a non-zero natural number.

[0120] Integer and They can be equal or different.

[0121] In the following text, the terms "number of points in the signal" and "length of the signal" will be expressed using the same term.

[0122] An adaptive artificial neural network is trained by minimizing the cost function associated with the adaptive artificial neural network in each of M periods, where M is an integer greater than or equal to 1.

[0123] In the following text, the expressions “cost function related to adaptive artificial neural networks” and “adaptive cost function” will be the same.

[0124] For example, a stochastic gradient descent algorithm with backpropagation over time (BPTT) can be used to minimize the adaptive cost function.

[0125] The cost function includes at least the first and second terms.

[0126] The first term of the cost function is combined with the class prediction provided by the adaptive artificial neural network based on the source signals from the source database and a set of function categories {y}. i} i≥1 The error between the categories associated with the source signals in the source database and the error in the source database is proportional.

[0127] For the source and target databases input into the adaptive artificial neural network, the second term of the cost function is calculated based on the sum of each maximum mean difference (MMD) computed at the output of at least one layer in the second part.

[0128] The second item is calculated using the following formula:

[0129]

[0130] Index m belongs to set Ncc, which includes each stratum number of the selected second part, and the maximum mean difference is calculated based on this set Ncc. Ncc can include a single stratum number or be greater than 1 and strictly less than 1. Multiple layer numbers.

[0131] The formula is as follows:

[0132]

[0133] This is the Gaussian kernel function.

[0134] The sum of multiple Gaussian kernel functions is considered to be a single Gaussian kernel function.

[0135] P is a natural number greater than or equal to 1, and P represents the number of Gaussian kernel functions considered when calculating the second term.

[0136] Notice item ,in equal: ,in Is with sets Related Gaussian kernel function The variance.

[0137] Notice item ,in equal: ,in Is with sets Related Gaussian kernel function The variance.

[0138] Notice item ,in equal: ,in Is with sets and set Related Gaussian kernel function The variance.

[0139] therefore, .

[0140] Each vector Depend on This means that each vector Depend on This means that each vector Depend on express.

[0141] In the following text, for each layer Each matrix Depend on Representation, and set Depend on This indicates. Therefore, It is a function The parameters.

[0142] This function It is the sum of multiple Gaussian kernel functions.

[0143] In each of the M number of epochs of training the adaptive artificial neural network, the cost function is minimized, and thus the first and second terms of that function are minimized.

[0144] Before minimizing the second term of the cost function, the parameters are estimated for each of the M periods. .

[0145] Determine parameters This includes several sub-steps as described below.

[0146] For each layer Determine parameters The first sub-step is to determine the representation length. and length The length of the maximum length between Sub-steps.

[0147] The second step of estimation method 100 is to convert each signal Resampled two-dimensional image to grayscale and each signal Resampled two-dimensional image to grayscale The steps.

[0148] image Each column includes equal to The number of pixels, and Each row includes equal to The number of pixels.

[0149] Vertical image Each column includes equal to The number of pixels, and the image Each row includes equal to The number of pixels.

[0150] Figure 2 This represents the steps involved in performing two-dimensional resampling on a signal of length M.

[0151] Determine parameters The second sub-step is to construct the matrix. Sub-steps. Matrix Indicates according to The last row of the Pascal's triangle of order 1 has a size of 1. And the variance is Gaussian kernel.

[0152] matrix Included Vectors of columns and rows Build, and This indicates that The last row of Pascal's triangle of order 1.

[0153] matrix equal , yes The transpose of .

[0154] matrix Indicates the size is The reproduction of the Gaussian kernel, where the center value of the matrix is ​​the maximum value, and the values ​​of the matrix gradually decrease from the center, thus reaching the maximum value of the matrix. The place where the value is equal to 1.

[0155] For example, for The last row of a 5th-order Pascal's triangle consists of the following vectors: .

[0156] Therefore, for , yes and .

[0157] therefore, yes .

[0158] exist Figure 3 The matrix represents a Gaussian kernel of size 28×28. The image.

[0159] Determine parameters The third sub-step is to compute the matrix The variance of the Gaussian kernel The sub-steps. This matrix The Gaussian kernel for the following formula is: When the exponent term approaches 1, the function... It is the largest, and therefore the function yes .

[0160] Notice It is a matrix The maximum value, therefore, The result is:

[0161] .

[0162] It does not depend on the training period, therefore, It can be calculated only for the first period and then reused for each of the M periods, or it can be recalculated for each period.

[0163] Determine parameters The fourth sub-step is to compute the matrix. Sub-steps.

[0164] According to the formula To calculate .

[0165] According to the formula To calculate .

[0166] According to the formula To calculate .

[0167] Preferably, .matrix Each coefficient equal .

[0168] Preferably, .matrix Each coefficient equal .

[0169] Preferably, .matrix Each coefficient equal .

[0170] According to one embodiment, P equals 1. Therefore, in this embodiment, , ,and .

[0171] Preferably, P equals 5. Setting P to 5 makes it possible to achieve a balance between the computation time of the adaptive cost function and the satisfactory accuracy of estimating the differences in each distribution.

[0172] Therefore, this matrix Depending on each training period, the set and set Modify the process for each training period out of M periods.

[0173] Used for estimating matrices Methods include those based on variance sum matrix To calculate the matrix by multiplication The steps.

[0174] Preferably, .

[0175] Therefore, in the preceding sub-steps, the set equal to... parameters Make an estimate.

[0176] Method 100 includes step 103 of associating a running category with the signal using an adaptive artificial neural network trained on each signal in the target database.

Claims

1. A method (100) for automatically monitoring at least one rotating component of a rotating machine from a no-label database, called target database, comprising at least one time signal originating from a distribution T, said time signal being produced by said rotating component, and from a source database comprising a plurality of time signals originating from a distribution S different from said distribution T, each time signal of said source database being produced by a source rotating component of a source rotating machine and being associated with an operating class of a set of operating classes, said set of operating classes comprising at least a nominal operating class and a fault operating class, characterized in that, Monitoring is performed on n T rotating components of a rotating machine, and the target database includes a plurality of signals represented by , the plurality of signals of the source database being represented by , n s representing the number of signals in the source database, the method comprising the steps of: 101 : training a first trained neural network, said first neural network being able to associate a run class from said set of run classes to a non-stationary time signal originating from a distribution S, said first neural network comprising a first part for feature extraction and a second part for classification, said first neural network being trained based on source data, 102: training an adaptive artificial neural network according to M number of training epochs to obtain a trained artificial neural network able to associate a class of the set of operating classes with a time signal originating from a S and T joint distribution, the adaptive artificial neural network comprising a first part for feature extraction corresponding to a first part of the first neural network and a second part for adapting the distribution S to the distribution T and for classification, the first part of the adaptive artificial neural network comprising N c1 a number of layers, said N c1 being a natural number greater than or equal to 1 and the second part of the adaptive artificial neural network comprising N c2 a number of layers, said N c2 being a natural number greater than or equal to 1, the artificial neural network being trained simultaneously on the source database and the target database, the training being performed by minimizing a cost function comprising: a first term, said first term corresponding to an error between a class obtained by said neural network for each signal of said source database and said class associated to said signal of said source database; a second term of said cost function is computed on one or more layers belonging to a second part of said neural network and said second term of said cost function is located before the last layer of said second part, said maximum mean discrepancy being computed according to at least one Gaussian kernel function of order said parameter being related to the variance of said Gaussian kernel function and depending on each training epoch, said parameter being estimated according to the Pascal triangle, said Gaussian kernel function of order being determined according to the last row of said Pascal triangle of size said second term of said cost function being computed at each epoch according to the function a set of signals representing the output of the m-th layer of an adaptive artificial neural network for an input each signal having a length and a set of signals representing the output of the m-th layer of said adaptive artificial neural network for an input each signal having a length m being an integer included in the set Ncc comprising each layer number of said second part, the maximum mean discrepancy being computed according to said set Ncc, wherein: wherein: wherein is the variance of the Gaussian kernel function related to the set of the Gaussian kernel function wherein is the variance of the Gaussian kernel function related to the set of the Gaussian kernel function wherein is the Gaussian kernel function related to the set and the set of variances of the Gaussian kernel function P denotes a Gaussian kernel function the number of vectors, and is a natural number greater than or equal to 1, each vector is represented by each vector is represented by each vector is represented by each matrix is represented by the parameter is equal to the set ; 103 : using said adaptive artificial neural network trained on each signal in said target database in order to associate a run class to said signal; wherein the step of training the adaptive artificial neural network (102) is performed according to a number M of epochs, and for each epoch of the number M of epochs, the parameters are estimated in the following sub-steps : For each layer : determining , is the maximum length between each signal of the set and each signal of the set , resample each signal to a two-dimensional image of size and resample each signal to a two-dimensional image of size and resample each signal to a two-dimensional image of size , In the following sub-steps are determined variance of the step-gaussian kernel function : constructing a matrix , said matrix represents a Gaussian kernel of order, said matrix is equal to: , represents the last row of the Pascal triangle of order, while is the transpose of According to the formula The variance is calculated , is equal to the maximum value of the matrix , The vector is calculated in the following substeps : According to come , According to come , According to come , The following formula is used to calculate .

2. The method (100) according to claim 1, characterized in that: each coefficient is equal to , each coefficient is equal to , each coefficient is equal to .

3. The method (100) according to any one of claims 1 to 2, characterized in that, P is equal to 5.

4. A computer program product comprising instructions which, when the program is executed on a computer, cause the computer to carry out the steps of the method (100) according to any one of claims 1 to 3.

Images