Method, model, program, storage medium and system for predicting maintenance state

Abstract

The present disclosure particularly relates to a computer-implemented method for predicting a maintenance state of a real-world device. The method includes providing a data set. The data set includes data describing historical real-world maintenance events and attributes of devices of the same type as real-world devices. The method further includes training a neural network based on the data set to predict parameters of an MCDA classification model. The MCDA classification model is configured to take as input at least one temporal measurement of maintenance-related physical and / or functional data of the real-world device, and to output a prediction of a maintenance status of the real-world device.

Description

Technical Field

[0001] The present disclosure relates to the field of computer programs and systems, and more particularly to methods, systems and programs for predicting the maintenance status of real-world devices. Background Art

[0002] Maintenance is a key activity in many technical fields (e.g., industry, energy, transportation) and has a significant impact on the reliability of many devices. Any unexpected downtime of a machine, device or device may interrupt the production process and may lead to safety failures, energy interruptions, accidents. For example, an aircraft part that does not undergo proper maintenance at the right time may lead to significant consequences. Therefore, developing a well-implemented and efficient maintenance strategy has become crucial in many technical fields.

[0003] Maintenance strategies have evolved from reactive maintenance to preventive maintenance. Reactive maintenance is performed only to restore the operating state of the device after a failure occurs. Preventive maintenance is performed based on time or process iterations according to a planned schedule to prevent failures, so unnecessary maintenance may be performed.

[0004] Preventive maintenance is a newer paradigm that performs maintenance only after analytical models predict certain failures or degradation. This approach allows maintenance to be performed as infrequently as possible to prevent unexpected reactive maintenance without doing too much preventive maintenance. Emerging technologies have made preventive maintenance more accessible and enhanced its potential to detect, isolate, and identify precursor and incipient failures of mechanical devices and components, monitor and predict the progression of failures, and provide decision support or automation to develop maintenance schedules.

[0005] In pursuit of effective preventive maintenance, various technologies have been developed and can be categorized into four main types: model-based, data-based, hybrid, and vibration and sound analysis technologies.

[0006] Model-based techniques use mathematical models such as Weibull analysis to predict component failures. Although they are simple and easy to apply, these models are often insufficient because they usually assume that failures are random processes with a limited number of input variables. This assumption does not always apply to complex industrial systems, for example. On the other hand, data-based techniques use historical data to train models that predict failures. These techniques are able to handle complex systems. However, the interpretation and verification of their predictions can pose significant challenges. Hybrid techniques attempt to strike a balance by combining elements from model-based and data-based techniques. This provides a good compromise between accuracy and interpretability. However, they may still be limited in terms of accuracy and ability to handle complexity. Finally, vibration and sound analysis techniques focus on analyzing the vibrations or sounds generated by the machine to predict possible failures. These techniques typically require specialized sensors and a deep understanding of signal analysis. In addition, these techniques may encounter difficulties in systems with multiple failure criteria.

[0007] Against this background, there is a need for improved solutions for predicting the maintenance status of real-world devices. Summary of the invention

[0008] Therefore, a computer-implemented method for predicting a maintenance state of a real-world device is presented. The method includes providing a data set. The data set includes data describing historical real-world maintenance events and attributes of a device of the same type as the real-world device. The method also includes training a neural network based on the data set to predict parameters of an MCDA classification model. The MCDA classification model is configured to take as input at least one time measurement of physical and / or functional data related to maintenance of the real-world device and output a prediction of the maintenance state of the real-world device.

[0009] The method may include one or more of the following:

[0010] The MCDA classification model is the NCS model;

[0011] The neural network is based on a sigmoid activation function, the sigmoid activation function is used to implement a comparison rule in the MCDA classification model, at least one parameter of the MCDA classification model is a sigmoid function used to implement the comparison rule; and / or

[0012] The maintenance status is one of: normal operation, maintenance recommended, and maintenance required.

[0013] Also provided is an MCDA classification model obtainable according to the method.

[0014] An MCDA classification model is also provided, the MCDA classification model being configured to take as input measurements of physical and / or functional data related to maintenance of a real-world device and output a prediction of a maintenance state of the real-world device. At least one parameter of the MCDA classification model is a sigmoid function that implements a comparison rule.

[0015] A method of using one of the above two models is provided. The method of using includes providing at least one time measurement of physical and / or functional data related to maintenance of the real-world device. The method of using also includes applying the MCDA classification model to the provided at least one time measurement, thereby outputting a prediction of the maintenance state of the real-world device.

[0016] The method of use may include one or more of the following:

[0017] - said at least one time measurement comprises at least one real-time measurement;

[0018] - the MCDA classification model is applied in real time;

[0019] - said at least one time measurement originates from at least one physical sensor of said device and / or attached to said device;

[0020] - The method of use also includes:

[0021] o comparing one or more predictions of the MCDA classification model to one or more real-world maintenance conditions of the device; and

[0022] o if the comparison results in a discrepancy, updating the MCDA classification model based on one or more real-world maintenance conditions of the device; and / or

[0023] The method of use further comprises performing maintenance on the device based on the prediction results of the MCDA classification model.

[0024] A computer program is also provided, the computer program comprising instructions for performing the method and / or the use of the method.

[0025] A computer-readable storage medium is also provided, on which a computer program and / or one or both of the above two models are recorded.

[0026] A computer system is also provided, comprising a processor coupled to a memory, wherein the memory has recorded thereon a computer program and / or one or both of the above two models.

[0027] A device is also provided, comprising a data storage medium on which a computer program and / or one or both of the above two models are recorded.

[0028] The device may form or be used as a non-transitory computer-readable medium, for example on a SaaS (Software as a Service) or other server or cloud-based platform, etc. The device may alternatively include a processor coupled to a data storage medium. Thus, the device may form a computer system in whole or in part (e.g., the device is a subsystem of the entire system). The system may also include a graphical user interface coupled to the processor. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Non-limiting examples will now be described with reference to the accompanying drawings, in which:

[0030] - Figures 1 to 4 A method is shown; and

[0031] - Figure 5 An example of a computer system is shown. DETAILED DESCRIPTION

[0032] A computer-implemented method for predicting a maintenance state of a real-world device is presented. The method includes providing a data set. The data set includes data describing historical real-world maintenance events and attributes of a device of the same type as the real-world device. The method also includes training a neural network based on the data set to predict parameters of an MCDA classification model. The MCDA classification model is configured to take as input at least one time measurement of physical and / or functional data related to maintenance of the real-world device and output a prediction of the maintenance state of the real-world device.

[0033] This constitutes an improved solution for predicting the maintenance status of real-world devices.

[0034] Notably, the method learns an MCDA (Multi-Criteria Decision Assistance) classification model configured to take as input at least one time measurement of physical and / or functional data related to maintenance of a real-world device and output a prediction of a maintenance state of the real-world device. In other words, the method learns an MCDA classification model for predicting the maintenance state of the real-world device based on the relevant measured maintenance-related data. Thus, the method applies the MCDA classification paradigm to the problem of predicting the maintenance state of a real-world device, which is a new and unconventional approach. Thus, the prediction of the maintenance state benefits from the capabilities of the MCDA classification model.

[0035] Furthermore, the method uses a neural network to learn the MCDA model, which is an unconventional approach. In fact, learning the MCDA model, or rather its parameters, is usually done by conventional methods such as mixed integer programming (MIP) and logic formulas. The proposed method instead learns a neural network on the training data. This neural network is trained to / learn to predict the parameters of the MCDA model. This allows for a lot of customization, as the neural network can be trained on specific data, so that the MCDA model is customized for a specific maintenance use case. This distinguishes the proposed method from more generic preventive maintenance strategies, thereby providing a better customized solution.

[0036] Furthermore, the method and the MCDA classification model inferred by the method by training a neural network to predict its parameters can be used to predict the maintenance state of a real-world device (e.g., based on the method of use). In other words, the MCDA classification model, once inferred, can be fed as input physical measurements (e.g., measured by sensors) describing physical and / or functional data about the real-world device and, upon receiving these data, predict the maintenance state of the device (e.g., "normal operation" (i.e., no maintenance required), "maintenance recommended", or "maintenance required"). In other words, the model predicts (and thus indirectly measures) the internal functional state of the device (in terms of whether operation is normal or maintenance must be completed) based on the physical measurements of the maintenance-related physical and / or functional data. The maintenance state can be output to the user of the device (e.g., the owner of a heat pump, an aircraft pilot) or transmitted to a company or organization responsible for the maintenance of the device or the manufacturer of the device, which in any case allows the necessary maintenance physical actions to be performed to return the device to its normal operating state.

[0037] The proposed method and its examples also offer the following advantages:

[0038] -Preventive maintenance is an area that has seen significant evolution over the years, transitioning from model-based techniques that use predefined models to predict maintenance status to data-based techniques that learn directly from historical or real-time data. There are also hybrid techniques that combine elements of the two aforementioned as well as techniques based on vibration and sound analysis that focus on detecting specific faults. However, despite these advances, there is still room for improvement, as demonstrated by the proposed new approach, which uses neural networks to train multi-criteria decision-making assistance models, in particular MCDA classification models such as NCS models.

[0039] - Model-based techniques, while offering high interpretability, generally have moderate accuracy due to their simple nature. Additionally, they have low ability to handle complexity and low sensitivity to data, which limits their usefulness in complex situations or when data changes rapidly. The proposed method overcomes these limitations by training a classification model using a neural network, allowing it to capture complex patterns in the maintained data and improve the accuracy of predictions.

[0040] -Data-based techniques can handle complexity and provide high accuracy, but they often have interpretability issues because they are treated as "black boxes". In fact, it can be challenging to understand how they reach their conclusions. The proposed method overcomes this shortcoming by combining a rule-based model (classification model) with a neural network. The classification model provides interpretability because it is based on defined rules, while the neural network provides the ability to handle complex patterns.

[0041] -Hybrid techniques combine model-based and data-based techniques in an attempt to exploit the benefits of each. However, they may still be limited in accuracy and ability to handle complexity. Similarly, vibration and sound analysis are specialized techniques that are very useful for detecting certain types of faults but less useful for other types of maintenance issues.

[0042] - The proposed approach provides increased flexibility by being able to process various forms of functional data, including previous maintenance plans and real-time sensor qualitative and quantitative data. This is in contrast to some traditional approaches that may require specific data types or be limited in their ability to process real-time data.

[0043] - In addition to flexibility, the proposed approach allows for extensive customization. Neural networks can be trained on company or industry specific data, allowing the classification model to be tailored to the specific needs of the application. This distinguishes the proposed approach from more generic preventive maintenance strategies, providing a better customized solution.

[0044] - A significant advantage of the proposed method is the incorporation of sigmoid functions for some parameters of the MCDA (e.g. NCS) model, as opposed to relying solely on step functions. This allows for smoother transitions between maintenance states, which can lead to better prevention performance. In other words, a neural network is trained to determine the parameters of a model that has been enhanced with at least one parameter that is presented as a sigmoid function, unlike the model's traditional step function. This makes the model more accurate and flexible.

[0045] - This proposed method relates to a computerized method for predicting the maintenance status and remaining life of a component. Said method relies on a multi-criteria decision-making assistance model, in particular an NCS model, calculated by a previously trained neural network. This method exploits the analysis of functional data, including previous maintenance plans and real-time sensor data.

[0046] - By using the proposed method, maintenance status can be predicted, which can range from normal operation to urgent maintenance needs or to durability improvement. Therefore, the proposed method provides optimization of maintenance resources, improvement of safety and extension of the life of the device;

[0047] - In summary, the proposed method provides a more accurate, flexible and customizable approach for preventive maintenance, which offers significant benefits over existing methods in the field of preventive maintenance.

[0048] The method will now be discussed further. The use of the method will also be discussed further below.

[0049] The method (which trains a neural network to predict MCDA parameters) is used to predict the maintenance state of a real-world product. In particular, the method infers an MCDA classification model that is configured to perform the prediction based on at least one time measurement of physical and / or functional data related to the maintenance of the real-world device. Therefore, the output of the method, once configured (i.e., once its parameters have been predicted by the trained neural network), is an MCDA model. The uses of the method may be part of the method, in which case they correspond to those uses of the method performed after performing the steps of the method, wherein the parameters of the MCDA model have been predicted by the neural network trained by the method. In other words, the method and the uses of the method may be included in the same computer-implemented process. Alternatively, the methods may be performed independently, for example, by different participants.

[0050] Since the method trains a neural network, the method is a method of machine learning. As is known from the field of machine learning itself, the processing of an input by a neural network includes applying an operation to the input, the operation being defined by data including weight values. Therefore, learning a neural network includes determining the values ​​of weights based on a data set configured for such learning, which data set may be referred to as a learning data set or a training data set. To this end, the data set includes data segments that each form a corresponding training sample. The training samples represent the diversity of situations in which the neural network is to be used after learning. Any training data set herein may include a plurality of training samples higher than 1000, 10000, 100000 or 1000000. In the context of the present disclosure, "learning / training a neural network based on a data set" means that the data set is a learning / training data set of a neural network, based on which the values ​​of weights (also referred to as "parameters") are set.

[0051] In the context of the proposed method, a training dataset is a provided dataset that includes data describing historical real-world maintenance events and attributes of devices of the same type as the real-world devices. Prior to training, the method includes providing a training dataset. This and its data are now discussed.

[0052] The training data set includes (e.g., consists of) data describing historical real-world maintenance events and attributes of devices of the same type as the real-world device. "Device" means any mechanical part or any component or any equipment, such as the following non-limiting examples: an aircraft part (e.g., an aircraft engine), an engine (e.g., an aircraft engine or a wind turbine engine), a heating device (e.g., a heat pump unit), a part of an electrical grid, a part of a nuclear power plant, or a manufacturing machine. The real-world device contemplated in the present disclosure is a device that requires maintenance. Such a device may include one or more sensors attached to the device or built into the device and configured to obtain time measurements of physical and / or functional data related to maintenance of the real-world device. These physical and / or functional data may include any data related to the maintenance status / correct operation of the device, that is, when these data or a portion thereof deviates from a normal value (i.e., corresponding to normal operation / operation), it indicates recommended maintenance (e.g., if the inconsistency with the normal value exceeds a certain threshold) or required maintenance (e.g., if the inconsistency with the normal value is higher than the certain threshold). The claimed method considers a type of device (e.g., aircraft engines, e.g., aircraft engines of the same type, such as a certain type of aircraft and / or specific to a certain company) and infers MCDA parameters for predicting the maintenance state of this type of real-world device. Thus, the data of the training data set relate to the same type of device (e.g., all aircraft engines, e.g., aircraft engines of the same type, such as a certain type of aircraft and / or specific to a certain company) of the real-world device.

[0053] Data describing historical real-world maintenance events and attributes specifies, for each device involved in the training data set, any data representing the history of maintenance events for the device (e.g., a history of maintenance states at each associated time / timestamp) and the values ​​of data about physical and / or functional attributes of the device. These training data may, for example, be data samples, each indicating a specific maintenance state at a specific time, given the measured values ​​of these attributes of the device at that time and / or previous times. For example, these data may include log files for several devices (but still of the same type, as described above).

[0054] Log files are historical records of component maintenance. They contain key information about past operations and maintenance activities. These records include details of when the device was operating normally, when maintenance was recommended, or when emergency maintenance was required. This historical data serves as a rich learning resource for neural networks, enabling them to understand patterns and draw valuable insights. Log files connect data about the physical and / or functional properties of the device (e.g., in the form of IoT sensor data streams) with the corresponding maintenance status of the device. For example, a log file may include, for example, physical and / or functional property data (such as IoT sensor data) tagged with maintenance status at different times, so as to train a neural network in a supervised manner to connect input physical property data (such as IoT sensor data) with maintenance status. In other words, each training data sample may correspond to a corresponding timestamp and may include the log file maintenance status at that timestamp and the measured IoT sensor data at that timestamp.

[0055] This allows the neural network to predict the parameters of the MCDA model so that the MCDA model can take the IoT sensor data stream as input and infer the corresponding maintenance status because its parameters have been learned / inferred on the data connecting the maintenance status with the IoT sensor data stream (i.e., log files).

[0056] IoT (Internet of Things) sensor data streams are data from various IoT sensors installed on the device. These are data that can be used in real-time operation (online phase), such as in the use of the method. These sensors monitor different operating parameters, such as temperature, pressure, vibration or other component-specific metrics. The data streams from these sensors provide continuous and up-to-date insights into the condition of the components. These real-time data can then be processed and analyzed by the MCDA model (once inferred by the method) to predict the maintenance status.

[0057] Providing the training data set may include forming the training data set by retrieving (e.g., downloading) at least some (e.g., all) data (e.g., log files, e.g., resulting from measurement and maintenance operations performed on the real-world device) from one or more (e.g., remote) memories or servers. Alternatively or additionally, providing the training data set may include measuring and / or acquiring at least some data (e.g., log files) from the real-world device by any suitable method and / or actually synthesizing such data. Alternatively or additionally, providing the training data may include simulating the training data or at least a portion thereof, e.g., using Microsoft Azure as discussed below.

[0058] In addition to providing a training data set, the method further includes: training a neural network based on the training data set to predict parameters of the MCDA classification model.

[0059] MCDA stands for "Multi-Criteria Decision Aiding". MCDA is a paradigm that aims to develop a decision support model explicitly based on the construction of a set of criteria that reflect the relevant aspects of the decision-making problem. These n criteria (N = {1, 2 ..., n}, n ≥ 2) evaluate a set of alternatives A = {a, b, c, ...} under consideration with respect to different viewpoints. The purpose of the MCDA method is to support the decision maker (DM) by providing a method and framework for making decisions about the decision situation under consideration. The decision problems considered in MCDA are of different types, and three reference problems are encountered in practice (discussed in the following reference, B. Roy. Multicriteria Methodology for Decision Aiding. Kluwer Academic, Dordrecht, 1996, which is incorporated herein by reference):

[0060] · Selection problem: It involves choosing an alternative or a subset of alternatives. A common MCDA selection problem is the supplier selection problem (discussed in the following references, M. Khalilzadeh, A. Karami and Alborz Hajikhani. The multi-

[0061] Objective supplier selection problem with fuzzy parameters and solving the order allocation problem with coverage. Journal of Modelling in Management, 15: 705-725, 2020, which is incorporated herein by reference). In practice, suppliers are evaluated based on several criteria such as cost, quality, and on-time delivery. The goal is to select a supplier (the best supplier) from a list of candidates.

[0062] Ranking problem: It consists in sorting a set of alternatives from best to worst according to the DM’s preferences. The result of the ranking method can be a partial or complete ranking of the set of alternatives. An interesting example of a ranking problem is to sort the company’s products according to several criteria (discussed in the following references, Ana Paula Henriques de and C. Medeiros. A model for selecting a strategic information system using the fitradeoff. Mathematical Problems in Engineering, 2016: 1-7, 2016,

[0063] which is incorporated herein by reference) for selection of information systems for evaluation.

[0064] Classification problem: It consists in assigning each alternative to a category selected from a set of predefined and ordered categories. The result of the classification method is the distribution of alternatives in different categories. For example, a committee of doctors makes a decision on whether a patient is admitted for surgery by considering several criteria for assessing the patient's physical health status (discussed in the following references, O. Sobrie, MEA Lazouni, S. Mahmoudi, V.

[0065] Mousseau and M. Pirlot. A new decision support model for preanesthetic evaluation. Computer Methods and Programs in Biomedicine, 133: 183-193, 2016, which is incorporated herein by reference).

[0066] The modeled MCDA (i.e., once its parameters are predicted by the trained neural network) is configured to predict the maintenance state of a real-world device (i.e., a device of the same type as those that were the source of the training data set) based on one or more input time measurements of physical and / or functional data related to maintenance of the real-world device (i.e., a device of the same type as those that were the source of the training data set (e.g., based on an input IoT data stream about the device). In the present disclosure, a maintenance state may be one of: normal operation, maintenance recommended, and maintenance required. This means that any maintenance state herein (i.e., in the training data or predicted by the MCDA classification model) is one of normal operation, maintenance recommended, and maintenance required.

[0067] In the case of the method, the MCDA model is an MCDA classification model, i.e. it solves a classification problem. The MCDA model can be an NCS model. NCS stands for "non-compensatory classification". NCS corresponds to a generalization and formal description of the Electre Tri procedure (discussed in the following reference: Salvatore Greco, José Figueira, Slowinski Roman, and Bernard Roy. Electre methods: Main features and recent developments. 06 2010, which is incorporated herein by reference). One of its characteristics is to consider alternative evaluations from an ordered perspective in order to avoid compensation and to be able to meaningfully process qualitative data. The NCS model may be intended to classify a device into one of three predefined ordered maintenance states: "normal operation", "recommended maintenance" and "maintenance required". Let A be a set of devices evaluated according to n criteria. Device a∈A is represented as a vector (a1,…,a n ), where a i is the evaluation of device a according to criterion i. Each criterion i has an overall pre-ranking, and a preference relation ≥ i can be implemented in the model to compare the devices. Embodiments of the method use two boundaries to define the maintenance state. These boundaries are defined as constraint profiles (vectors of n values), one for each criterion:

[0068] · Separate maintenance statuses "Normal Operation" and "Maintenance Recommended" Separate maintenance statuses "Maintenance Recommended" and "Maintenance Required"

[0069] Because the maintenance state is ordered, there is dominance between restriction profiles:

[0070]

[0071] A device will be assigned to maintenance if it outperforms the lower profile on a sufficiently strong subset of criteria, but not when comparing the component to the upper profile. These are called "comparison rules" of the NCS model, a concept known in the art of NCS models.

[0072] The NCS model can be implemented according to the MR classification method as follows. A positive weight is associated with each criterion that sums to 1, and:

[0073] If and only if The device is assigned the maintenance state "Normal Operation"

[0074] If and only if and The device is assigned the maintenance status "Maintenance Recommended" when

[0075] If and only if The device is assigned the maintenance status "Maintenance Required" when

[0076] The method infers an MCDA model by predicting its parameters using a neural network. That is, once the neural network has been trained, the method may include applying the neural network, or using the parameters predicted by the network at the end of its training (e.g., once the convergence criteria of the training have been reached) to set the parameters of the MCDA model to the parameters predicted by the neural network. Prior to this, the method includes training the neural network. The training includes the following iterations: feeding the neural network with training samples of the training data set; applying the MCDA model or its allocation rules (the concept of allocation rules for MCDA models is known) with the current values ​​of its parameters (i.e., as predicted by the neural network in its current training state); and if the result of applying the MCDA model or its allocation rules is not satisfactory, modifying the parameters / weights of the neural network, and so on until a satisfactory result is achieved. For example, this may include repeatedly feeding the neural network with IoT data of a log file, and evaluating whether the resulting predicted parameters lead to a correct prediction of the maintenance status associated with the IoT data in the log file, until a satisfactory result is achieved.

[0077] For example, when the MCDA model is an NCS model, the training uses machine learning techniques to learn MR classification parameters from a training data set, such as consisting of previous maintenance interventions (log files). The neural network represents the assignment rules of the NCS model. This constitutes a neural representation of the MR classification parameters, which are therefore the parameters inferred by the trained neural network.

[0078] The neural network may be based on a sigmoid activation function for implementing a comparison rule in the MCDA classification model. In this case, at least one parameter of the MCDA classification model is a sigmoid function implementing the comparison rule. For example, if the MCDA model is an NCS model, the assignment rule of the MCDA model corresponds to an inequality / condition (also called a "comparison rule") of the type a≥b. Therefore, the NCS model includes one or more parameters to structurally represent these conditions / inequalities, and these parameters include one or more sigmoid functions: one sigmoid function for each inequality / condition of the type a≥b. In order to learn such parameters, the neural network may be based on one or more sigmoid activation functions (e.g., one function for each condition, e.g., each layer of the neural network).

[0079] Typically, in the NCS paradigm, these conditions are structurally represented by step functions. In contrast, the proposed method can structurally represent these conditions using sigmoid functions as described above. Figure 1 and Figure 2A step function and a sigmoid function are shown to illustrate the difference between the two. A neural network can use a sigmoid function as an activation function, which is a differentiable activation function. This allows the use of gradient descent during training. In contrast, a step function is not differentiable. The method can modify each sigmoid function To more closely resemble its corresponding step function. This is achieved by multiplying the difference between a and b by a large constant M. This multiplication is used to flatten the S-shaped function vertically. By fine-tuning the value of M, the embodiment of the method can control the steepness of the S-shaped curve around the threshold, making it sharp or smooth as needed. However, balance must be maintained because the extremely high value of M may bring challenges during the gradient-based optimization process. For each given condition a≥b, the adapted S-shaped function (that is, the parameters that form the NCS model) is then defined as:

[0080]

[0081] This ensures that the output of the function is approximately 0 and 1 for values ​​significantly below or above the threshold, respectively.

[0082] In order to infer these sigmoid function parameters of the classification model for the condition a≥b of the type, the neural network can be based on the sigmoid activation function, as described above. This can be achieved as follows to train a neural network model made based on the principle of NCS model learning. Such a model treats each component as a separate input. Therefore, in an embodiment, the architecture of the neural network includes n independent single layers, where each layer corresponds to a specific evaluation a i Therefore, each of these individual layers compares the input value to a series of trainable, positive, and sequentially ordered restriction profiles using the corresponding sigmoid activation function as described above. As a result, each i th The layer has two unique activation sigmoid functions. These functions are used to activate the i With each restriction profile b 1 and b 2 The proximity of the input value a i Mapping to fractions between 0 and 1 and This effectively classifies the input.

[0083] In these embodiments, the subsequent stage of the training process involves computing two weighted sums, expressed as:

[0084]

[0085] These sums cover the performance of each component relative to the limiting profile b 1 and b 2 It is important to note that the described implementation sets the uniform weights w iApplied to each sum to maintain consistency.

[0086] By design, this sequence of calculations produces 2 ordered fractions c 1 (a)≤Sc 2 (a) that fall within the range of 0 to 1. These embodiments then compare these scores to a trainable threshold λ in the final layer to determine the appropriate maintenance state.

[0087] The maintenance status assignment rules are as follows:

[0088] If and only if Sc 1 When (a) < λ, the component is assigned the maintenance state “normal operation”

[0089] If and only if Sc 2 (a) <λ and Sc 1 When (a)≥λ, the component is assigned the maintenance state “Maintenance Recommended”: and

[0090] If and only if Sc 2 When (a)≥λ, the component is assigned the maintenance status "maintenance required"

[0091] To implement this approach, each score is compared to λ using a sigmoid function. Similarly, for the initial layer, this generates two additional sigmoid functions.

[0092] These embodiments apply a SoftMax function coupled with a Cross-Entropy Loss function to classify components. This is done by using the formula To achieve, where t i is the true label, p i Indicates i th The complete neural network architecture according to these implementations is shown in Figure 3 Shown in.

[0093] These embodiments may use a gradient descent variant called "Adam", a well-respected algorithm in the art. This choice was driven by several attractive properties of Adam, including its adaptive learning rate, impressive accuracy, and fast execution time. Adam is therefore an ideal fit for the currently proposed method, balancing efficiency and effectiveness in training neural networks.

[0094] These embodiments may use the following optimization (borrowed and adapted from reference Loshchilov, Ilya & Hutter, Frank. (2017). Fixing Weight Decay Regularization in Adam., which is incorporated herein by reference):

[0095]

[0096] Several hyperparameters may be necessary for the Adam optimizer to function. These hyperparameters may seriously affect the efficiency, speed, and effectiveness of the solution determined by the method (as discussed in the following references, KrzysztofMartyn, Kadziński, Deep preference learning for multiple criteria decision analysis, European Journal of Operational Research, Volume 305, Issue 2, 2023, Pages 781-805, ISSN 0377-2217).

[0097] The learning rate, denoted as α, determines the size of the adjustments made to the parameters during each optimization step. If this rate is set too low, the learning process may be slow, risking premature stopping at a local optimum. Conversely, a high rate may overlook the optimum and lead to convergence failure.

[0098] Momentum factors β1 and β2 evaluate the impact of past parameter improvements on the current step. Momentum promotes faster and more efficient optimization by applying insights gained from earlier stages of the learning process, smoothing the path towards a stable optimization that is less susceptible to perturbations during training.

[0099] • Factor ∈, used as a small denominator value, is introduced to ensure stability in the calculation.

[0100] Weight decay factor, denoted as w τ , which also helps in the optimization process.

[0101] In addition to the parameters involved in the Adam optimizer, in an embodiment, the neural network may also feature its own set of hyperparameters. These may include: M and M2, which help approximate the step function; and the number of epochs, which indicates the duration of model training. The exact values ​​of these parameters are a matter of embodiment.

[0102] An MCDA classification model obtainable (e.g., directly obtained) according to the method is also proposed. In other words, for a given device type, the MCDA classification model has parameters having the same values ​​as the parameters to be inferred / predicted by a neural network trained by the method on a training data set related to the device type. For example, the MCDA classification model may have the values ​​of its parameters being the values ​​of the parameters generated directly by the method on such a training data set, i.e., the values ​​of those parameters inferred by a neural network trained according to the method. Thus, the proposed MCDA classification model is configured to take as input measured values ​​of physical and / or functional data related to maintenance of a real-world device and output a prediction of the maintenance state of the real-world device. The proposed MCDA classification model may be an NCS model. The proposed MCDA classification model may include one or more parameters, each parameter being a sigmoid function that implements a corresponding comparison rule (i.e., structurally defining a condition / inequality as described above).

[0103] A method for using the MCDA classification model is also provided and is now discussed.

[0104] The method of use includes providing at least one time measurement of physical and / or functional data related to the maintenance of a real-world device. The at least one time measurement originates from at least one physical sensor of the device and / or attached to the device. Providing the at least one time measurement may include receiving one or more IoT sensor data streams from one or more IoT physical sensors measuring these data by a computer system executing the method of use. These one or more IoT sensors may each be an IoT sensor of the device (i.e., integrated therein) or an external IoT sensor attached to the device. Providing the at least one time measurement may include measuring physical and / or functional data by one or more physical sensors (e.g., IoT sensors). This may be done automatically, the sensor being configured to provide measurements, for example, at regular time intervals. Therefore, providing at least one time measurement may include sending the measurements (e.g., in operation when the measurements are available, or in one-time or grouped, e.g., with an intermediate storage stage) to a computer executing the method of use, and receiving the measured data as described above by the computer.

[0105] The method of use then comprises applying the MCDA classification model to the at least one time measurement provided. The application of the MCDA classification model results in the MCDA classification model outputting a prediction of the maintenance state of the real-world device, as it has been designed for this purpose (i.e., its parameters have been inferred). It should be understood that the device on which the MCDA classification model is used in the method of use is of the same type as the device involved in the training data set based on which the model parameters are inferred.

[0106] The provided measurements may be real-time measurements, such as a real-time IoT sensor data stream. The MCDA classification model may be applied in real-time and continuously as the real-time measurements are received, such as in real-time and continuously as the real-time IoT sensor data stream is received. This amounts to the steps of the method being iteratively usable, wherein the provision of the measurements is done repeatedly and in real-time (e.g., at regular time intervals, such as short time intervals, to produce, for example, a real-time stream of IoT sensor data), and the application of the MCDA classification model is then performed iteratively and in real-time on these received measurements (e.g., as they are received, or at short time intervals as the measurements are received) to predict the maintenance status continuously and in real-time as the IoT sensor data stream is received.

[0107] In an embodiment, the method may optionally use advanced stream analytics techniques to process the IoT sensor data stream in real time before feeding it to the classification model. The goal is to identify patterns and detect any anomalies that may indicate potential maintenance needs. The processed data can then be fed into the NCS model for preventive maintenance decisions. If so implemented, the corresponding pre-processing is done directly on the provided training data set or during training to infer the NCS model, which is suitable for taking the processed data as input instead of directly taking the IoT data stream.

[0108] The method of use may include displaying the maintenance status of the device on a computer screen of the device and / or of a computer (e.g., a computer of the device manufacturer or maintenance company) connected (e.g., by wireless connection) to the device. Alternatively or additionally, the maintenance status may be transmitted to and stored on a computer or storage medium (or database) of the device manufacturer or maintenance company when it is predicted (e.g., in the form of a log file) and / or may be stored on a storage medium (e.g., a database) of the device itself (e.g., in the form of a log file). Alternatively or additionally, the method may include outputting a visual and / or sound alarm, for example, on the device itself and / or on a computer of the device manufacturer or maintenance company whenever the status is "maintenance required" (and optionally, the status is for each item for which maintenance is recommended).

[0109] The method of use may also include performing maintenance on the device based on the prediction results of the MCDA classification model. When the predicted maintenance status is maintenance required or maintenance recommended, as described above, this can be done upon receipt of the alarm, and may then include physically performing the necessary physical technical actions to return the device to normal operation.

[0110] The method of use may also include comparing (e.g., for one or more devices) one or more prediction results of the MCDA classification model with one or more real-world maintenance states of the device. The real-world maintenance states may, for example, result from a professional's conclusions about maintenance (e.g., in parallel with the model's predictions). In this case, the method of use also includes: if the result of the comparison is that there are differences (i.e., at least some of the maintenance states output by the MCDA model do not correspond to the real-world maintenance states), updating the MCDA classification model based on the one or more real-world maintenance states of the device. The updating may include fine-tuning some parameters of the model and / or retraining the neural network to refine the parameters, for example, by adding data corresponding to the real-world maintenance states to the training data, such as physical and / or functional data associated with these states.

[0111] Figure 4 A flow chart of an embodiment of a method and a method for using the method in which the update of the NCS model is completed is shown. The figure provides an overview of the proposed innovative process for preventive maintenance. A neural network (NN) trained on a log file is used to learn a maintenance classification model - the NCS model. The model classifies components into three maintenance states: normal operation, recommended maintenance, and required maintenance. In real-time operation, data from IoT sensors is processed and analyzed and then passed through the NCS model for preventive maintenance decisions. The results can be stored in a database and visualized for easy interpretation. In order to keep the system adaptive and accurate, the method can update the NCS model regularly as described above. This is done by comparing the model's predictions with the actual component status, updating the log file accordingly, and retraining the NN. The method is dynamic, adaptive, and is designed to optimize maintenance operations, thereby minimizing unforeseen equipment downtime.

[0112] In production, the proposed preventive maintenance system operates in real time, continuously analyzing data from IoT sensors to predict maintenance status.

[0113] The NCS model, trained with a neural network using log files, is the core of the implementation. When a real-time data stream is received from IoT sensors, the data is passed through the NCS model. The model then classifies the maintenance status of the component into one of three categories: normal operation, maintenance recommended, or maintenance required.

[0114] IoT sensor data streams provide up-to-date information about various operating parameters of components. Advanced streaming analytics techniques can be used to process this data in real time. The goal is to identify patterns and detect any anomalies that may indicate a potential need for maintenance. The processed data is then fed into the NCS model for preventive maintenance decisions.

[0115] Once the NCS model has made its predictions, these predictions can be stored in a database as described above for record keeping and further analysis. The method can also include visualizing the results in a user-friendly format. This makes it easy for maintenance personnel to interpret the predictions and make informed decisions about necessary maintenance actions. The visualization tool can show component status in an at-a-glance manner, alert operators to potential problems, and even track the performance of the preventive maintenance system over time.

[0116] The model update process depends on the comparison of the predicted maintenance state with the actual maintenance state. After running the model on the real-time data stream, the update may include performing an evaluation by comparing the predicted results with the observed actual maintenance state. The first step may involve aligning the predicted maintenance states with their real counterparts for direct comparison. This allows any differences to be found, whether they are false alarms (situations where maintenance is predicted but not required) or missed detections (situations where maintenance is not predicted). Upon completion of the evaluation, the method of use may then update the log files. They can enrich the newly collected information: sensor data, predicted maintenance states, and actual maintenance states. This additional data enhances the maintenance database and helps to refine future model training. These updated log files can then be used to retrain the neural network. This continuous learning process enables the NCS model to gradually evolve and adapt to changes and new trends observed in the maintenance data. This cycle of evaluation, updating, and retraining can be repeated regularly to ensure continuous improvement of the proposed preventive maintenance system.

[0117] Examples of methods and uses of methods are now discussed where the device is an aircraft engine.

[0118] In the context of preventive maintenance production of IoT, a complex infrastructure is required to monitor and measure the condition of machines, especially aircraft engines. More precisely, data-based predictions depend on the availability of a statistically significant number of telemetry and maintenance execution records that reached potential engine failures. From these, equipment degradation models can be learned, enabling predictions based on historical and newly collected data. Finding a real-world dataset containing records that reached failures is practically impossible due to its commercially sensitive nature.

[0119] For this purpose, an open source simulator from Microsoft Azure can be used, allowing such simulation to be possible. Based on theoretical physics equations, the temperature (in degrees Celsius (°C)), pressure (in bar), speed (in revolutions per minute (RPM)) and pressure state (in bar) of the engine can be simulated over time, as well as environmental conditions based on the same standards. These form the physical and / or functional data discussed above for the aircraft. Arbitrarily large data can be generated, and a real-time telemetry stream can be provided, allowing interactive testing.

[0120] An example of the results obtained for a telemetry stream is provided below.

[0121]

[0122] A maintenance log file is also obtained which records malfunctions of the engine.

[0123] Timestamp grade 2017-06-06 21:49:29 Maintenance required 2017-06-08 06:18:10 Maintenance required 2017-06-08 20:16:42 Maintenance required … …

[0124] Each training data sample may correspond to a corresponding timestamp and may include a log file maintenance state at the timestamp and measured data of the telemetry stream at the timestamp.

[0125] Let's now discuss the training in this example. The initial work in the training process involves making this simulation compatible with the implementation of the model, for which it is necessary to obtain a monotonically increasing criterion. To achieve this, three features related to rotation speed, temperature and pressure will be created: the absolute difference of the three states of the machine and the three states of the environment will be taken, for example, for temperature; it will be |ambient temperature - machine temperature|. Finally, the last feature representing the operating duration of the machine will be used, which allows to measure how long the machine has been operating.

[0126] It should also be noted that in the first seconds (about 10 seconds) when the machine starts running, the temperature, pressure and rotation speed are significantly lower until it starts working, which is why these measurements will be removed.

[0127] Therefore, the modeling includes four criteria for the three categories of normal operation, recommended maintenance, and required maintenance. To label the training set, a maintenance file is used, which indicates when the engine failed and the associated time. According to this time reference, the telemetry stream is labeled as follows:

[0128] A machine is always marked as functioning normally if no failures occurred during execution.

[0129] If the machine fails after period p, then t1=p / 2 and t2=3p / 4 are calculated, representing two time markers, the first at 50% of the total time before the failure and the second at 75%. The marking is then performed from 0 to t1 in case of normal operation, from t1 to t2 in case of recommended maintenance, and between t2 and the last day of period p in case of required emergency maintenance.

[0130] This preprocessing provides training and test sets for the neural network. The input data includes:

[0131] · List of 4 values ​​for each of the 4 monotonic criteria: temperature, rotation speed, pressure and operating duration. Since the criteria must be maximized and fall between 0 and 1, it is sufficient to normalize the values ​​and define the best category as requiring maintenance. In practice, machine states of similar environments are sought, where smaller absolute differences indicate better machine functionality, while higher values ​​present more risks. Regarding the criterion of operating duration, a machine at the beginning of the service presents less risk than one at the end of the service. Normalization between 0 and 1 is performed by the standard method: Normalized value = (old value - minimum value) /

[0132] (maximum value - minimum value).

[0133] · A label or category associated with the previous standard value. The neural network will then be trained using the training set as described previously, allowing for preliminary classification of the test set.

[0134] Let's discuss inference in this example. To do this, let's revisit the telemetry data stream generated previously to see what's going on. It should be noted that the telemetry stream can contain a very large number of measurements, in this case tens of thousands. As mentioned above for training, the first 10 measurements are discarded while the machine reaches its cruising rhythm. Knowing that the machine on the second maintenance table did not fail, it ran for 2 minutes at a rate of 1 measurement per second, for a total of 120 measurements. Given the initial running measurements of the discarded machine, note measurements 11 and 12: The process begins by establishing the absolute value of the difference between the machine's measurement and the environmental measurement, and the elapsed time:

[0135] state temperature pressure speed Duration State 1 23.05 1062.97 70.76 3 State 2 23.26 1064.34 71.49 4

[0136] It was also measured that during the machine's operating interval, the maximum and minimum states of the first three criteria were:

[0137] value temperature pressure speed Minimum 22.98 1060.32 69.52 Maximum 23.54 1069.26 72.89

[0138] This produces the normalized values:

[0139] state temperature pressure speed Duration State 1 0.125 0.2964 0.3679 0.025 State 2 0.5 0.4497 0.5846 0.0333

[0140] Therefore, if the values ​​from the previous table are input into the neural network, normal operation will be returned for each state of the machine.

[0141] The update of the model in this example is discussed below. The goal is to have a scalable model, meaning that it should be adaptable if given new reference data that becomes increasingly accurate. After a few weeks of operation, a machine failure is detected while the model is predicting maintenance recommendations, so the model must be updated to correct this error. The idea is to add these new reference data to the learning set and retrain the neural network, initializing the network variables with the old values.

[0142] The method is computer-implemented. This means that the steps (or substantially all steps) of the method are performed by at least one computer or any system. Therefore, the steps of the method are performed by a computer, possibly fully automatically or semi-automatically. In an example, the triggering of at least some steps of the method may be performed by user-computer interaction. The level of user-computer interaction required may depend on the level of automation foreseen and balanced with the needs of the implementing user. In an example, the level may be user-defined and / or predefined.

[0143] A typical example of a computer implementation of the method is to perform the method using a system suitable for this purpose. The system may include a processor coupled to a memory and a graphical user interface (GUI), the memory having recorded thereon a computer program including instructions for performing the method. The memory may also store a database. The memory is any hardware suitable for such storage, and may include several physically distinct parts (e.g., one for the program and possibly one for the database).

[0144] Figure 5 An example of a system is shown, where the system is a client computer system, such as a user's workstation.

[0145] The client computer of the example includes a central processing unit (CPU) 1010 connected to an internal communication bus BUS1000 and a random access memory (RAM) 1070 also connected to the bus. The client computer is also provided with a graphics processing unit (GPU) 1110, which is associated with a video random access memory 1100 connected to the BUS. The video RAM 1100 is also referred to as a frame buffer in the art. A mass storage device controller 1020 manages access to a mass storage device such as a hard disk drive 1030. Mass storage devices suitable for tangibly embodying computer program instructions and data include all forms of non-volatile memory, including, for example, semiconductor memory devices such as EPROM, EEPROM, and flash memory devices; magnetic disks such as internal hard disks and removable disks; magneto-optical disks. Any of the foregoing may be supplemented or incorporated by a specially designed ASIC (Application Specific Integrated Circuit). A network adapter 1050 manages access to a network 1060. The client computer may also include a tactile device 1090, such as a cursor control device, a keyboard, and the like. A cursor control device is used in the client computer to allow the user to selectively position the cursor at any desired location on the display 1080. In addition, the cursor control device allows the user to select various commands and input control signals. The cursor control device includes a plurality of signal generating devices for inputting control signals to the system. Typically, the cursor control device may be a mouse, the buttons of which are used to generate the signals. Alternatively or additionally, the client computer system may include a sensitive board and / or a sensitive screen.

[0146] The computer program may include instructions executable by a computer, the instructions including means for causing the above-mentioned system to perform the method. The program may be recorded on any data storage medium, including the memory of the system. The program may be implemented, for example, in a digital electronic circuit, or in computer hardware, firmware, software, or a combination thereof. The program may be implemented as a device, such as a product tangibly embodied in a machine-readable storage device for execution by a programmable processor. The method steps may be performed by a programmable processor executing an instruction program to perform the functions of the method by operating on input data and generating output. Therefore, the processor may be programmable and coupled to receive data and instructions from a data storage system, at least one input device, and at least one output device, and to send data and instructions to the data storage system, at least one input device, and at least one output device. The application program may be implemented in a high-level program or object-oriented programming language, or in an assembly or machine language if necessary. In any case, the language may be a compiled or interpreted language. The program may be a fully installed program or an update program. The application of the program on the system results in any case of instructions for executing the method. The computer program may alternatively be stored and executed on a server in a cloud computing environment, which communicates with one or more clients via a network. In this case, the processing unit executes instructions included by the program, so that the method is performed on the cloud computing environment.

Claims

1. A computer-implemented method for predicting a maintenance status of a real-world device, the method comprising: providing a data set comprising data describing historical real-world maintenance events and attributes of a device of the same type as the real-world device; as well as Based on the data set, a neural network is trained to predict parameters of an MCDA classification model, the MCDA classification model being configured to take as input at least one time measurement of physical and / or functional data related to maintenance of the real-world device and to output a prediction of a maintenance state of the real-world device.

2. The method according to claim 1, wherein: The MCDA classification model is the NCS model.

3. A method according to claim 1 or 2, wherein, The neural network is based on a sigmoid activation function, which is used to implement a comparison rule in the MCDA classification model, and at least one parameter of the MCDA classification model is a sigmoid function used to implement the comparison rule.

4. The method according to any one of claims 1 to 3, wherein: The maintenance status is one of: normal operation, maintenance recommended, and maintenance required.

5. An MCDA classification model, which can be obtained according to the method according to any one of claims 1 to 4.

6. An MCDA classification model, the MCDA classification model being configured to take as input time measurements of physical and / or functional data related to maintenance of a real-world device and output a prediction of a maintenance state of the real-world device, at least one parameter of the MCDA classification model being a sigmoid function for implementing a comparison rule, wherein Optionally, the MCDA classification model can be obtained according to the method of claim 3.

7. A method for using the MCDA classification model according to claim 5 or 6, comprising: providing at least one temporal measurement of physical and / or functional data related to maintenance of the real-world device; as well as The MCDA classification model is applied to the provided at least one time measurement to output a prediction of a maintenance status of the real-world device.

8. The method according to claim 7, wherein: The at least one time measurement includes at least one real-time measurement.

9. The method according to claim 8, wherein: The MCDA classification model is applied in real time.

10. The method according to any one of claims 7 to 9, wherein: The at least one time measurement originates from at least one physical sensor of the device and / or attached to the device.

11. The method according to any one of claims 7 to 10, wherein: The method further comprises: comparing one or more predictions of the MCDA classification model to one or more real-world maintenance conditions of the device; and If a result of the comparison is that there is a difference, the MCDA classification model is updated based on the one or more real-world maintenance conditions of the device.

12. The method according to any one of claims 7 to 11, wherein: The method further comprises: Based on the prediction result of the MCDA classification model, maintenance of the device is performed.

13. A computer program comprising instructions which, when executed by a computer system, cause the computer to perform the method of any one of claims 1 to 4 and / or the method of any one of claims 7 to 12.

14. A computer-readable data storage medium having recorded thereon the computer program of claim 13 and / or the MCDA classification model of claim 5 and / or the MCDA classification model of claim 6.

15. A computer system comprising a processor coupled to a memory, wherein the computer program of claim 13 and / or the MCDA classification model of claim 5 and / or the MCDA classification model of claim 6 are recorded on the memory.

Images