Prediction of maintenance status of real-world device

Abstract

To provide a computer-implemented method for predicting a maintenance status of a real-world device, a multi-criteria decision aid (MCDA) sorting model, a usage method, a computer program, a computer readable data storage medium, and a computer system.SOLUTION: The method includes a step of providing a dataset. The dataset includes data describing past real-world maintenance events, as well as characteristics of a device of the same type as a real-world device. The method further includes a step of training a neural network based on the dataset to predict parameters of an MCDA sorting model. The MCDA sorting model takes as input at least one time measurement of maintenance-related physical and / or functional data of the real-world device, and outputs the prediction of a maintenance status of the real-world device.SELECTED DRAWING: Figure 4

Description

Technical Field

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

Background Art

[0002] Maintenance is an important activity that has a significant impact on the reliability of numerous devices in many technical fields (such as industry, energy, transportation, etc.). Unplanned downtime of machines, devices, and equipment can interrupt the production process and lead to a lack of safety, energy supply interruption, and accidents. For example, parts of an airplane that have not received appropriate maintenance at the appropriate time can have dramatic consequences. Therefore, in many technical fields, it has become important to develop well-implemented and efficient maintenance strategies.

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

[0004] Predictive maintenance is a newer paradigm in which maintenance is only carried out after an analysis model has predicted a specific failure or degradation. With this approach, the frequency of maintenance can be reduced as much as possible to prevent unplanned corrective maintenance without overdoing preventive maintenance. New technologies have made predictive maintenance more accessible and are likely to provide decision-making support and automation for detecting, isolating, and identifying precursors and incipient failures of machinery and components, monitoring and predicting the progression of failures, and formulating maintenance schedules.

[0005] To pursue effective predictive maintenance, various methods have been developed and classified into four categories: model-based, database, hybrid, and vibration / acoustic analysis methods.

[0006] Model-based techniques use mathematical models such as Weibull analysis to predict component failures. Although these models are simple and easy to apply, they are often insufficient because they generally assume that failures are probabilistic processes with a limited number of input variables. This assumption does not always hold true in complex industrial systems. On the other hand, database techniques use past data to train models for predicting failures. These techniques can handle complex systems. However, there are significant challenges in interpreting and validating their predictions. Hybrid techniques attempt to balance by combining elements of both model-based and database techniques. These offer a good compromise between accuracy and interpretability. However, they may still have limitations in terms of handling accuracy and complexity. Finally, vibration / acoustic analysis techniques focus on analyzing vibrations and sounds generated by machines to predict the likelihood of failure. Such techniques often require special sensors and a deep understanding of signal analysis. Additionally, these techniques may struggle in systems with multiple failure criteria.

[0007] Against this backdrop, an improved solution is needed to predict the maintenance status of real-world devices.

Summary of the Invention

[0008] Therefore, a computer-implemented method for predicting the maintenance status of real-world devices is proposed. This method includes the step of providing a dataset. The dataset includes past real-world maintenance events and data describing the characteristics of devices of the same type as the real-world devices. The method further includes the step of training a neural network based on the dataset to predict the parameters of the MCDA sorting model. The MCDA sorting model is configured to take at least one time measurement of physical and / or functional data related to the maintenance of a real-world device as an input and output a prediction of the maintenance status of the real-world device.

[0009] The method may include one or more of the following. · The MCDA sorting model is an NCS model. · The neural network is based on a sigmoid activation function for implementing the comparison rules of the MCDA sorting model, and at least one parameter of the MCDA sorting model is a sigmoid function for implementing the comparison rules. And / or, · The maintenance status is any one of "normal operation", "maintenance advice", and "require maintenance".

[0010] Furthermore, an MCDA sorting model obtained according to this method is provided.

[0011] Furthermore, an MCDA sorting model is provided that is configured to take a time measurement of physical and / or functional data related to the maintenance of a real-world device as an input and output a prediction of the maintenance status of the real-world device. At least one parameter of the MCDA sorting model is a sigmoid function for implementing the comparison rules.

[0012] A method of using one of the two models described above is provided. The method of use includes the step of providing at least one time measurement of at least one of the physical and / or functional data related to the maintenance of a real-world device. The method of use further includes the step of outputting a prediction of the maintenance status of a real-world device by applying an MCDA sorting model to the at least one time measurement provided.

[0013] The method of use may include one or more of the following. · At least one time measurement consists of at least one real-time measurement. · The MCDA sorting model is applied in real time. · At least one time measurement is derived from at least one physical sensor of the device and / or a sensor attached to the device. · The method of use further includes the following. The step of comparing one or more predictions of the MCDA sorting model with one or more actual maintenance statuses of the device. And If the result of the comparison is inconsistent, the step of updating the MCDA sorting model based on one or more real-world maintenance statuses of the device. And / or · The method of use further includes the step of performing maintenance on the device based on the prediction of the MCDA sorting model.

[0014] A computer program including the method and / or instructions for performing the method of use is further provided.

[0015] Furthermore, a computer-readable recording medium on which the computer program and / or one or both of the two models described above are recorded is provided.

[0016] Furthermore, a computer system including a processor coupled to a memory is provided, and the memory has recorded therein the computer program and / or one or both of the two provided models.

[0017] Furthermore, an apparatus is provided that includes a computer program and / or a data storage medium on which one or both of the above two models are recorded.

[0018] The apparatus can form, or function as, a non-transitory computer-readable medium in, for example, a server such as SaaS (Software as a Service) or a cloud-based platform. Alternatively, the apparatus may include a processor coupled to the data storage medium. Thus, the device can form a computer system, wholly or in part (e.g., the device is a subsystem of the overall system). The system can further include a graphical user interface coupled to the processor.

Brief Description of the Drawings

[0019]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Modes for Carrying Out the Invention

[0020] A computer-implemented method for predicting the maintenance status of real-world devices is proposed. This method includes the step of providing a dataset. The dataset includes past real-world maintenance events and data describing the characteristics of devices of the same type as the real-world devices. The method further includes the step of training a neural network based on the dataset to predict the parameters of an MCDA sorting model. The MCDA sorting model is configured to take as input at least one time measurement of physical and / or functional data related to the maintenance of a real-world device and output a prediction of the maintenance status of the real-world device.

[0021] This constitutes an improved solution for predicting the maintenance status of real-world devices.

[0022] Notably, this method learns an MCDA (Multi-Criteria Decision Aid) sorting model configured to take as input at least one time measurement of physical and / or functional data related to the maintenance of a real-world device and output a prediction of the maintenance status of the real-world device. In other words, this method learns an MCDA sorting model for predicting the maintenance status of real-world devices based on relevant measured maintenance-related data. This method thus applies the MCDA sorting paradigm to the problem of predicting the maintenance status of real-world devices, which is a novel approach not previously seen. The prediction of the maintenance status thus benefits from the power of the MCDA sorting model.

[0023] Furthermore, the method learns the MCDA model using a neural network, which is an unprecedented approach. In fact, the learning of the MCDA model (or rather its parameters) is usually carried out by conventional methods such as mixed integer programming (MIP) or logical formulation. The proposed method, instead, learns a neural network from the training data. This neural network is trained to predict the parameters of the MCDA model. This enables the neural network to be trained with specific data and the MCDA model to be customized for specific maintenance use cases, allowing for extensive customization. As a result, the proposed method is distinguished from more general predictive maintenance strategies and provides a more customized solution.

[0024] Furthermore, the method, and the MCDA sorting model inferred by the method through parameter prediction by learning the neural network, can be used to predict the maintenance status of real-world devices (e.g., according to this usage method). In other words, once the MCDA sorting model is inferred, it can be given physical measurements (e.g., those measured by sensors) that describe the physical and / or functional data regarding the real device as input, and upon receiving these data, it predicts the maintenance status 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 infers) the internal functional state of the device (in terms of whether the operation is normal or maintenance is required) based on physical measurements of maintenance-related physical and / or functional data. The maintenance status is output to the user of the device (e.g., the owner of a heat pump, the pilot of an airplane), or sent to a company or organization responsible for the maintenance of the device, or the manufacturer of the device, and in either case, the physical actions of maintenance necessary to return the device to its normal operating state can be carried out.

[0025] The proposed method and its examples also have the following advantages: · Predictive maintenance is a field that has evolved significantly over the years, transitioning from model-based techniques that use pre-defined models to predict maintenance status to database techniques that learn directly from past and real-time data. There are also hybrid techniques that combine the previous two elements and vibration and acoustic analysis-based techniques focused on detecting specific faults. However, despite such progress, there is still room for improvement, as shown by the proposal of a new approach that uses neural networks to train multi-criteria decision support models, specifically MCDA sorting models such as the NCS model.

[0026] · Model-based techniques offer high interpretability, but due to their simple nature, they generally have medium accuracy. Furthermore, their low ability to handle complexity and low sensitivity to data limit their usefulness in complex situations or when data changes rapidly. The proposed approach overcomes these limitations by using neural networks to train sorting models, enabling it to capture complex patterns in maintenance data and improve prediction accuracy.

[0027] · Database techniques can handle complexity and provide high accuracy, but they are often regarded as "black boxes" and thus often have problems with interpretability. In fact, it is difficult to understand how they reach their conclusions. The proposed approach overcomes this drawback by combining a rule-based model (sorting model) with a neural network. The sorting model is easy to interpret as it is based on defined rules, and the neural network has the ability to handle complex patterns.

[0028] · Hybrid technology combines model-based technology and database technology, aiming to take advantage of their respective merits. However, there may still be limitations in terms of the ability to handle accuracy and complexity. Similarly, vibration and acoustic analysis is also a special technology that is very useful for detecting certain types of faults but not very helpful for other types of maintenance problems.

[0029] · The proposed approach can handle various forms of functional data such as past maintenance plans and qualitative and quantitative data from real-time sensors, thus improving flexibility. This is in contrast to conventional approaches that require specific data types or have limitations in the ability to handle real-time data.

[0030] · In addition to flexibility, the proposed approach enables a wide range of customization. Since neural networks can be trained with enterprise- or industry-specific data, the screening model can be customized according to the specific needs of the application. As a result, the proposed approach provides a more customized solution, different from more general predictive maintenance strategies.

[0031] · One of the remarkable advantages of the proposed approach is to incorporate a sigmoid function into part of the parameters of the MCDA (such as NCS) model instead of relying only on step functions. This makes the transition between maintenance states smoother and leads to an improvement in prediction performance. In other words, the neural network is trained to determine the parameters of the model enhanced with at least one parameter represented as a sigmoid function, different from the conventional step function of the model. As a result, this model becomes more accurate and flexible.

[0032] · This approach relates to a computerized method for predicting the maintenance status and remaining life of components. This method relies on a multi-criteria decision-making support model, particularly the NCS model, calculated by a pre-trained neural network. In this approach, the analysis of functional data including past maintenance plans and real-time sensor data is utilized.

[0033] · By using the proposed approach, the maintenance status can be predicted. The maintenance status can cover a wide range from normal operation to the need for emergency maintenance or improvement of durability. Thus, the proposed approach realizes the optimization of maintenance resources, the improvement of safety, and the extension of the equipment life.

[0034] · In summary, the proposed approach provides a more accurate, flexible, and customizable method for predictive maintenance, bringing greater advantages than existing methods in the field of predictive maintenance.

[0035] This method will be further described. The usage method will be described later.

[0036] This method (training a neural network to predict MCDA parameters) is for predicting the maintenance status of real-world products. Specifically, this method infers an MCDA sorting model configured to perform this prediction based on at least one temporal measurement of physical and / or functional data related to the maintenance of a real device. Thus, the output of this method is an MCDA model once configured (i.e., after its parameters have been predicted by a trained neural network). This usage method may be part of this method, in which case, after performing the steps of this method, it corresponds to performing that of this usage method which is an MCDA model whose parameters are predicted by the neural network learned by this method. In other words, this method and this usage method may be included in the same computer-implemented process. Alternatively, these methods may be performed independently, for example, by different actors.

[0037] The method is a machine learning technique such that the method trains a neural network. As is known per se from the field of machine learning, the processing of an input by a neural network involves applying operations to the input, the operations being defined by data including weight values. Thus, the learning of a neural network involves determining the values of the weights based on a data set configured for such learning, such a data set being sometimes referred to as a learning data set or a training data set. Therefore, the data set includes pieces of data that form respective training samples. The training samples represent the diversity of the situations in which the neural network after learning will be used. Any training data set herein may be composed of a number of training samples greater 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 the learning / training data set of the neural network and the values of the weights (also called "parameters") are set based thereon.

[0038] In the context of the proposed method, the training data set is a provided data set that includes data describing the characteristics of past real-world maintenance events and devices of the same type as the real-world devices. Before training, the method provides the training data set. This and its data are described next.

[0039] The training dataset includes data that describes the characteristics of past real-world maintenance events and devices of the same type as real-world devices (e.g., composed of these). "Device" means any mechanical part, component, device, such as non-limiting examples below: aircraft parts (such as aircraft engines), engines (such as aircraft engines and wind turbine engines), heating devices (such as heat pump units), part of a power transmission network, part of a nuclear power plant, manufacturing machinery, etc. The real devices envisioned in this disclosure are devices that require maintenance. Such devices can include one or more sensors attached to or built into the device and configured to acquire time measurements of maintenance-related physical and / or functional data of the real-world device. These physical data and / or functional data can consist of any data related to the maintenance status / proper functioning of the device. That is, when these data or a part thereof deviate from normal values (i.e., values corresponding to normal function / operation) (e.g., when the discrepancy from the normal value exceeds a certain threshold), it indicates that maintenance is required or (e.g., when the discrepancy from the normal value exceeds the said certain threshold) shows that maintenance is necessary. The claimed method takes into account one type of device (e.g., aircraft engine, e.g., of the same type, e.g., of a specific type of aircraft and / or company-specific) and infers MCDA parameters for predicting the maintenance status of real-world devices of this type. Therefore, the data in the training dataset pertains to devices of the same kind as this real-world device (e.g., all aircraft engines, e.g., of a specific kind of aircraft and / or company-specific, etc., of the same kind).

[0040] Data describing past real-world maintenance events and characteristics, for each device included in the training dataset, specifies, along with the values of data regarding the physical and / or functional characteristics of the device, any data representing the history of the maintenance events of the device (e.g., a history of maintenance situations with their respective associated timestamps / timestamp). These training data are, for example, data samples where measurements of these characteristics of the device at that time and / or at an earlier time indicate a particular maintenance situation at a particular time. These data include, for example, log files of multiple devices (of the same type as described above).

[0041] Log files are records of the maintenance history of components. They contain important information about past operations and maintenance activities. These records include details such as when the equipment was operating normally, when maintenance was recommended, and when emergency maintenance was required. These historical data serve as rich learning resources for neural networks, enabling them to understand patterns and draw valuable insights. Log files connect data regarding the physical and / or functional characteristics of the device (e.g., in the form of a data stream from IoT sensors) with the corresponding maintenance situation of the device. For example, log files can be used to train a neural network in a supervised manner to connect input physical characteristic data such as IoT sensor data with the maintenance situation. For example, they may consist of physical characteristic data and / or functional characteristic data such as IoT sensor data labeled with the maintenance situation at different times. In other words, each training data sample can correspond to a respective timestamp and, together with the measured IoT sensor data at this timestamp, constitute the log file maintenance situation at this timestamp.

[0042] As a result, the neural network can predict the parameters of the MCDA model, and the MCDA model can take in the IoT sensor data stream as input and infer the corresponding maintenance status because the parameters are learned / inferred on the data connecting the maintenance status and the IoT sensor data stream (i.e., the log file).

[0043] The data stream of IoT (Internet of Things) sensors is data from various IoT sensors attached to the device. These are data that can potentially be used in real-time operations (online phase), such as in the present method of use. These sensors monitor various operating parameters such as temperature, pressure, vibration, and other component-specific metrics. The data stream from these sensors provides continuous and up-to-date insights into the state of the component. This real-time data can be processed and analyzed to predict the maintenance status through the MCDA model (inferred by the present method).

[0044] Providing the training data set may include forming the training data set by obtaining (e.g., downloading) at least a part (e.g., all) of the data (e.g., log files derived from measurements and maintenance operations performed on real-world devices, etc.) from one or more (e.g., remote) memories or servers. Alternatively or additionally, providing the training data set may consist of measuring and / or obtaining at least a part of the data (e.g., log files) from real-world devices and / or realistically synthesizing these data in any suitable way. Alternatively or additionally, the step of providing the training data may include, for example, simulating the training data or at least a part thereof using Microsoft Azure as described later in this specification.

[0045] In addition to providing the training dataset, the method trains a neural network that predicts the parameters of the MCDA sorting model based on the training dataset.

[0046] MCDA is an abbreviation for "Multi-criteria decision-aiding". MCDA is a paradigm aimed at explicitly developing a decision-making support model based on the construction of a set of criteria that reflect relevant aspects of a decision-making problem. These n criteria (N = {1, 2, …, n}, where n ≥ 2) evaluate a set of alternatives (A = {a, b, c, …}) under consideration from various perspectives. The purpose of the MCDA approach is to assist the decision maker (DM) by providing a method and framework for making decisions regarding the considered decision-making situation. There are various types of decision-making problems considered in MCDA, and in practice, three reference problems are encountered (Reference: B. Roy. Multicriteria Methodology for Decision Aiding. Kluwer Academic, Dordrecht, 1996, which is incorporated herein by reference).

[0047] · Selection problem: The problem of selecting an alternative or a subset of alternatives. A familiar MCDA selection problem is the supplier selection problem (M. Khalilzadeh, A. Karami, and Alborz Hajikhani. The multi-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 fact, suppliers are evaluated according to several criteria such as cost, quality, and on-time delivery. The aim is to select a supplier (the best supplier) from the candidate list.

[0048] · Ranking problem: The problem of ordering a set of alternatives from best to worst according to the preference of the DM. The result of the ranking method can be a partial ranking or a complete ranking of the set of options. An interesting example of a ranking problem is the selection of an information system evaluated according to several criteria in an enterprise (Reference: Ana Paula Henriques de Gusmao and C. Medeiros. A model for selecting a strategic information system using the fitrade off. Mathematical Problems in Engineering, 2016:1 - 7, 2016, which is incorporated herein by reference).

[0049] · Sorting problem: The problem of assigning each option to a category selected from a predefined set of ordered categories. The result of the sorting method is the assignment of options between different categories. For example, a committee of physicians decides on the feasibility of surgery considering several criteria for evaluating the patient's physical health (Reference: O. Sobrie, M. E. A. Lazouni, S. Mahmoudi, V. 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).

[0050] The modeled MCDA is configured to predict the maintenance status of real-world devices (i.e., of the same type as those from which the training dataset is derived) (i.e., once its parameters are predicted by a neural network that has been trained) based on one or more input time-measured maintenance-related physical data and / or functional data of the real-world devices (e.g., based on an input IoT data stream regarding the device). In the present disclosure, the maintenance status is any of normal operation, maintenance recommendation, and maintenance required. That is, the maintenance status referred to herein (i.e., as predicted by the training data or the MCDA sorting model) is any of normal operation, maintenance recommendation, and maintenance required.

[0051] In the case of this method, the MCDA model is an MCDA sorting model, i.e., one that solves a sorting problem. The MCDA model may be an NCS model. NCS is an abbreviation for "non-compensatory sorting". NCS corresponds to a generalization and formal description of the Electre Tri procedure (References: Salvatore Greco, Jose Figueira, Slowinski Roman, Bernard Roy, Electre methods: Main features and recent developments. 06 2010, this document is incorporated herein by reference). One of its features is to explain alternative evaluation from the perspective of order, which avoids correction and enables meaningful handling of qualitative data. The NCS model aims to classify a device into one of three predefined maintenance statuses: "normal operation", "maintenance recommendation", and "maintenance required". Assuming an evaluation based on a set of devices. Device a ∈ A is a vector (a1,…, an) is expressed as, where ai is the evaluation of device a with respect to criterion i. Each criterion i has a total order that can implement a preference relation ≧i within the model for comparing devices. In the implementation of this method, two boundaries are used to demarcate the maintenance status. These boundaries are defined as limit profiles, and a vector of values is set for each criterion, one by one.

Number

Number

Number

[0052] If a device is superior to the lower limit profile in a sufficiently strong subset of criteria, the device is assigned to a maintenance status, but this is not the case when comparing the components with the upper limit profile. These are called the "comparison rules" of the NCS model and are concepts known in the field of the NCS model.

[0053] The NCS model can be implemented as follows according to the MR sorting method. Each criterion is associated with a positive weight whose sum is 1: The device is assigned to the maintenance status of "normal operation" only when the following conditions are met.

Number

Number

[0054] This method infers the MCDA model by predicting parameters using a neural network. That is, once the neural network is trained, this method may consist of applying the neural network or setting the parameters of the MCDA model as those predicted by the network at the end of its training (e.g., after reaching the training convergence criterion). Before that, this method trains the neural network. The training includes supplying the training samples of the training dataset to the neural network, applying the MCDA model, or its assignment rules (the concept of the assignment rules of the MCDA model is known) with the current values of its parameters (i.e., the values predicted by the neural network in its current training state), and if the result of the application of the MCDA model or its assignment rules is unsatisfactory, modifying the parameters / weights of the neural network and repeating this until a satisfactory result is obtained. For example, this includes repeatedly supplying IoT data from a log file to the neural network 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 obtained.

[0055] For example, when the MCDA model is the NCS model, the training uses machine learning techniques to learn the MR sort parameters from a training dataset composed, for example, 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 sort parameters and becomes the parameters inferred by the once-trained neural network.

[0056] The neural network can be based on the sigmoid activation function to implement the comparison rules of the MCDA sorting model. In this case, at least one parameter of the MCDA sorting model is the sigmoid function that implements the comparison rules. For example, when the MCDA model is the NCS model, the assignment rules of the MCDA model correspond to inequalities / conditions of the form a ≥ b (also called "comparison rules"). Therefore, the NCS model is composed of one or more parameters for structurally representing these conditions / inequalities, and these parameters are composed of one or more sigmoid functions (one sigmoid function for each a ≥ b type of inequality / condition). To learn such parameters, the neural network may be based on one or more sigmoid activation functions (for example, one for each condition, for example, for each layer of the neural network).

[0057] Normally, in the NCS paradigm, these conditions are structurally represented by step functions. On the contrary, the proposed method can structurally represent these conditions using sigmoid functions as described above. Figures 1 and 2 are diagrams for explaining the difference between the step function and the sigmoid function. The neural network can use the sigmoid function, which is a differentiable activation function, as the activation function. This allows the use of the gradient descent method during training. On the contrary, the step function is non-differentiable. The present method is for each sigmoid function

Number

[0058] This ensures that the output of the function is approximated and becomes a value significantly below or above the threshold respectively.

[0059] To infer the parameters of such a sigmoid function for the condition a ≥ b, the neural network can be based on the sigmoid activation function as described above. This can be implemented as follows to train a neural network model created based on the principles of NCS model learning. This model treats each component as an individual input. Thus, in the embodiment, the architecture of the neural network consists of n independent single layers where each layer corresponds to a specific evaluation a i Each of these layers, as described above, uses its respective sigmoid activation function to compare the input value to a series of trainable, positive, sequentially ordered limit profiles. As a result, the i-th layer hosts two unique activation sigmoid functions. These functions map the input value a i from 0 to 1, [Number] and map it to a score between, and effectively classify the input based on the proximity of b1 and b2 to each limit profile.

[0060] In these implementations, in the subsequent phase of the training process, a weighted sum formulated as follows is calculated.

Number

[0061] These sums include the scores of each component for the limit profiles b1 and b2. Importantly, to maintain consistency, each implementation applies a uniform weight wi to its respective sum.

[0062] By design, this calculation order results in two ordered sets of scores

Number

[0063] The rules for assigning maintenance situations are as follows. · A component is assigned the maintenance situation "Normal Operation" only if

Number

Number

Number

[0064] To execute this method, each score is compared using the sigmoid function. Similar to the initial layer, this further generates two sigmoid functions.

[0065] These implementations apply the softmax function combined with the cross-entropy loss function to classify the components. This is achieved using the formula

Number

[0066] In these implementations, the gradient descent method known as "Adam", which is a highly regarded algorithm in this field, can be used. This choice is due to the attractive characteristics of Adam (such as adaptive learning rate, excellent accuracy, fast execution time, etc.). Adam thus ideally fits the currently proposed method while balancing efficiency and effectiveness in the training of neural networks.

[0067] These implementations can use the following optimizations (see reference Loshchilov, Ilya & Hutter, Frank. (2017). Fixing Weight Decay Regularization in Adam., which is incorporated herein by reference).

Number

[0068] The Adam optimizer's functionality requires several hyperparameters. These hyperparameters can have a decisive impact on the efficiency, speed, and effectiveness of the solution determined by this method (reference: Krzysztof Martyn, Milosz Kadzinski, Deep preference learning for multiple criteria decision analysis, European Journal of Operational Research, volume 305, Issue 2, 2023, 781 - 805, ISSN0377 - 2217).

[0069] · The learning rate, denoted as α, determines the magnitude of the adjustment applied to the parameters during each optimization step. Setting this rate too low may slow down the learning process and risk premature stopping at a local optimum. Conversely, a high rate may miss the optimum and fail to converge.

[0070] · The momentum factors β1 and β2 evaluate the impact of past parameter improvements on the current step. By applying insights obtained from the initial stages of the learning process, momentum facilitates faster and more efficient optimization, smoothing the path towards a stable optimization direction that is less susceptible to the effects of perturbations during training.

[0071] · The factor ε functions as a small denominator value introduced to ensure computational stability.

[0072] · w τ The weight decay coefficient, represented as such, also contributes to the optimization process.

[0073] In addition to the parameters related to the Adam optimizer, hyperparameters can also be set independently in the implementation of neural networks. These include M and M2, which are useful for approximating the step function, and the number of epochs indicating the learning period of the model. The exact values of these parameters are implementation - specific issues.

[0074] Also proposed is an MCDA sorting model that can be obtained according to this method (for example, obtained directly). In other words, for a given device type, the MCDA sorting model has parameters with the same values as those that would be inferred / predicted by a neural network trained by this method on a training dataset related to this device type. For example, the MCDA sorting model can have, as the values of its parameters, those directly obtained from this method on such a training dataset, that is, those inferred by a neural network once trained according to this method. Thus, the proposed MCDA sorting model is configured to take in as input the time measurement of the physical and / or functional data related to the maintenance of real-world devices and output a prediction of the maintenance status of real-world devices. The proposed MCDA sorting model may be an NCS model. The proposed MCDA sorting model may be composed of one or more parameters that are sigmoid functions each implementing its respective comparison rule (that is, structurally defining conditions / inequalities as described above).

[0075] A method of using the MCDA sorting model is also provided. This will be discussed now.

[0076] This method of use includes the step of 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 is derived from at least one physical sensor of the device and / or a sensor attached to the device. The providing of the at least one time measurement may include, by a computer system executing this method of use, receiving one or more IoT sensor data streams from one or more IoT physical sensors that measure these data. Each of these one or more IoT sensors may be an IoT sensor of the device (i.e., integrated therein) or an external IoT sensor attached to the device. Providing at least one time measurement may include the step of measuring physical data and / or functional data by one or more physical sensors (e.g., IoT sensors). This may be done automatically, for example, the sensors are set to provide measurements at regular time intervals. Next, providing at least one time measurement may include the step of transmitting the measurements (e.g., as they become available, either in real time, or all at once, or in groups, e.g., with an intermediate storage step) to the computer executing this method of use and the step of receiving, by this computer, the data measured as described above.

[0077] This method of use then includes the step of applying an MCDA sorting model to the at least one measurement provided. When the MCDA sorting model is applied, since the MCDA sorting model is designed for this purpose (i.e., its parameters are inferred), it will output a prediction of the maintenance status of the real-world device. It should be understood that the device for which the MCDA sorting model is used in this method of use is of the same type as the devices included in the training data set for which the parameters of the model were inferred.

[0078] The measurement(s) provided may be real-time measurement(s), e.g., a real-time data stream of IoT sensors. The MCDA sorting model can be applied in real-time and continuously along with the reception of real-time measurement(s), e.g., along with the reception of a real-time data stream of IoT sensors. This may be repeated for the steps of this method of use, where the provision of the measurement is done iteratively and in real-time (e.g., at regular time intervals, e.g., short time intervals, to provide a real-time stream of IoT sensor data), and the application of the MCDA sorting model is then done iteratively and in real-time on these received measurements (e.g., along their reception, or at short regular time intervals along the reception of the measurement) to predict the maintenance status continuously and in real-time along with the reception of the data stream of the IoT sensors.

[0079] Before feeding the IoT sensor data stream into the sorting model, the method can optionally, in implementation, use advanced stream analysis techniques to process this IoT data in real-time. The purpose is to identify patterns and detect anomalies that may indicate a potential need for maintenance. The processed data is fed into the NCS model for predictive maintenance decision-making. If this is implemented, the corresponding preprocessing is done directly on the provided training data set or during training, and an NCS model adapted to take this processed data as input instead of directly inputting the IoT data stream is inferred.

[0080] This method of use may include the step of displaying the maintenance status of the device on the computer screen of the device and / or on the computer screen of a computer (such as the computer of the device manufacturer or the maintenance company) connected to the device (for example, via a wireless connection) at the time of its prediction. Alternatively or additionally, the maintenance status may be transmitted to the computer or storage medium (or database) of the device manufacturer or maintenance company at the time of its prediction, stored therein (for example, in the form of a log file), and / or stored in the storage medium (for example, database) of the device itself (for example, in the form of a log file). Alternatively or additionally, the method may consist of outputting visual and / or audible warnings, for example, on the device body and / or on the computer of the device manufacturer or maintenance company, each time the status is "maintenance required" (and optionally, each time the status is "maintenance required").

[0081] This method of use may further include the step of performing maintenance on the device based on the prediction of the MCDA sorting model. This may be done, as described above, when a warning is received when the predicted maintenance status is maintenance required or a maintenance recommendation, and then may include the step of physically performing the physical and technical operations necessary to return the device to normal operation.

[0082] This method of use may further include the step of comparing one or more predictions of an MCDA sorting model (for example, for one or more devices) with one or more actual maintenance situations of the device. The actual maintenance situation may, for example, be derived from the conclusions of an expert regarding maintenance (for example, in parallel with the predictions of the model). In this case, if a discrepancy occurs as a result of the comparison (i.e., at least some of the maintenance situations output by the MCDA model do not match the real-world ones), this method of use further includes the step of updating the MCDA sorting model based on one or more real-world maintenance situations of the device. The update may include fine-tuning some of the model's parameters and / or retraining the neural network to refine the parameters, for example, by adding data corresponding to the real-world maintenance situations, such as physical data and / or functional data related to these statuses, to the training data.

[0083] Figure 4 is a flowchart showing an example of a method of use in which the present method and the NCS model are updated. This figure shows an overview of an innovative process proposed for predictive maintenance. A neural network (NN) trained on a log file is used to learn a maintenance classification model (NCS model). In this model, components are classified into the following three maintenance situations: normal operation, maintenance recommendation, and maintenance required. In real-time operation, data from IoT sensors is processed and analyzed and passed to the NCS model for predictive maintenance decisions. The results are stored in a database and visualized for easy interpretation. To maintain the adaptability and accuracy of the system, this method may periodically update the NCS model as described above. This is done by comparing the predictions of the model with the actual status of the components, updating the log file accordingly, and retraining the NN. This approach is dynamic and adaptable and is designed to optimize maintenance operations and minimize unexpected downtime of the device.

[0084] In the actual operation, the predictive maintenance system operates in real time, continuously analyzes data from IoT sensors, and predicts the maintenance status.

[0085] The NCS model trained with a neural network using log files is the core of the implementation. When receiving a live data stream from IoT sensors, this data is passed to the NCS model. Next, this model classifies the maintenance status of the component into one of three categories: "normal operation", "maintenance recommendation", and "require maintenance".

[0086] The data stream of IoT sensors provides up-to-date information on various operating parameters of the component. In order to process this data in real time, advanced stream analysis techniques can be used. The purpose is to identify patterns and detect anomalies that may indicate a potential need for maintenance. The processed data is input into the NCS model for predictive maintenance decision-making.

[0087] After the NCS model makes a prediction, it can be saved in the database as described above for record-keeping and further analysis. This method can also visualize the results in a user-friendly format. Thereby, the maintenance staff can easily interpret the prediction and make an information-based decision about the necessary maintenance actions. The visualization tool can also show the status of the component at a glance, warn the operator of potential problems, and track the performance of the predictive maintenance system over time.

[0088] The model update process depends on the comparison between the predicted maintenance situation and the actual maintenance situation. After operating the model on the real-time data stream, this update includes the step of performing an evaluation by contrasting the predicted results with the actually observed maintenance situation. The first step may be to align the predicted maintenance situation with the actual corresponding one for direct comparison. This enables the discovery of any discrepancies, whether they are false alarms (cases where maintenance was predicted but not necessary) or misses (cases where necessary maintenance was not predicted). When this evaluation is complete, the present method of use may update the log file. It may also be enhanced with newly collected information such as sensor data, predicted maintenance situation, and actual maintenance situation. This additional data helps reinforce the maintenance database and improve future model training. These updated log files are utilized for the retraining of the neural network. Through this continuous learning process, the NCS model can consistently evolve and adapt to the variations and new trends observed in the maintenance data. This cycle of evaluation, update, and retraining may be periodically repeated to ensure the continuous improvement of the proposed predictive maintenance system.

[0089] Examples of this method and the present method of use will be described.

[0090] In the context of predictive maintenance manufacturing for IoT, a complex infrastructure is required to monitor and measure the state of machines, particularly aircraft engines. More precisely, data-based prognosis prediction depends on whether a statistically significant amount of telemetry and maintenance execution records are available until the possibility of engine failure occurs. A degradation model of the device can be learned from these and predictions can be made based on past and newly collected data. Finding a real-world dataset that includes records until failure is virtually impossible due to its commercial confidentiality.

[0091] Therefore, by using the open-source simulator provided by Microsoft Azure, this simulation can be enabled. Based on theoretical physical equations, it is possible to simulate over time the engine's temperature (in degrees Celsius [°C]), pressure (in bar), rotational speed (in RPM), pressure state (in bar), and environmental state based on the same criteria. These form the physical and / or functional data of the aircraft described above. Since arbitrarily large data can be generated and a real-time telemetry stream can be provided, interactive tests can be carried out.

[0092] Examples of the results obtained from the telemetry stream are shown below. [Table 1]

[0093] Maintenance log files recording engine failures are also obtained. [Table 2]

[0094] Each training data sample corresponds to its respective timestamp and is composed of the measurement data of the telemetry stream at this timestamp together with the log file maintenance status at this timestamp.

[0095] The training in this example will be described. In the initial stage of training, it is necessary to adapt this simulation to the implementation of the model, and for this purpose, it is necessary to obtain a monotonically increasing criterion. To achieve this, three features related to rotational speed, temperature, and pressure are created. The absolute value of the difference between the three states of the machine and the three states of the environment is taken. For example, for temperature, it is |ambient temperature - machine temperature|. Finally, a function representing the operating time of the machine is used, and the operating time of the machine can be confirmed by measurement.

[0096] Also, for several seconds (about 10 seconds) until the machine starts operating, since the temperature, pressure, and rotational speed are extremely low, these measurements are excluded.

[0097] Therefore, the modeling includes four criteria for three categories: normal operation, maintenance advice, and maintenance required. To label the training set, a maintenance file indicating when the engine failed and the associated time is used. From this time-based criterion, the telemetry stream is labeled as follows. · If the machine does not fail during operation, it is always labeled as "normal operation". · If the machine fails after a period p, t1 = p / 2 and t2 = 3p / 4 are calculated (these represent two time markers. The first marker indicates 50% of the total time until failure, and the second marker indicates 75%). Then, from 0 to t1, it is labeled as normal operation, from t1 to t2 as maintenance required, and from t2 to the final day as urgent maintenance required.

[0098] This preprocessing yields both the training set and the test set for the neural network. The input data is as follows. · For four monotonic criteria of temperature, rotational speed, pressure, and operation time, a list of four values each. Since the criteria are to be maximized and must fall between 0 and 1, it is sufficient to normalize the values and define the best category as the required maintenance. In fact, a state of the machine similar to the environment is required, and the smaller the absolute value of the difference, the better the machine's function, but the larger the value, the higher the risk. For the criterion regarding the operation period, the machine at the start of operation has a lower risk than the machine at the end of operation. The normalization between 0 and 1 is performed in a standard way (normalized value = (old value - minimum value) / (maximum value - minimum value)). · The label, or category, associated with the previous criterion value. The neural network is trained with the training set as described above, enabling preliminary classification of the test set.

[0099] Regarding the inference, it will be explained with this example. For this purpose, review the previously generated telemetry data stream again and observe a specific case. Note that the telemetry stream may contain a very large number of measurements. As described in the above training, the first 10 measurements are discarded and used for the machine to reach a cruising rhythm. The machine did not fail in the second maintenance table, measured once per second, and operated for 2 minutes until 120 measurements were made. If the initial operation measurements of the machine are discarded, attention is directed to measurements 11 and 12. The process starts by establishing the absolute value of the difference between the machine's measurements and the environmental measurements over time.

Table 3

[0100] Also, within the operating interval of the machine, the maximum and minimum states of the first three criteria are measured.

Table 4

[0101] As a result, a normalized value is obtained.

Table 5

[0102] Therefore, when the values in the previous table are input into the neural network, normal operation will be returned for each state of the machine.

[0103] Next, the model update in this example will be described. The goal is to have a scalable model, that is, it should be adaptable and become more accurate when new reference data is provided. After several weeks of operation, it was found that the machine had failed while the model was predicting maintenance advice, so it was necessary to update the model to correct this error. The idea is to add these new reference data to the learning set, retrain the neural network, and initialize the variables of the network with the old values.

[0104] This method is computer-implemented. This means that the steps (or substantially all steps) of this method are executed by at least one computer or any system. Therefore, the steps of this method are executed by a computer, in some cases completely automatically or semi-automatically. In the example, at least some of the triggers for the steps of this method are executed by the interaction between the user and the computer. The required level of user-computer interaction may depend on the expected level of automation and the need to fulfill the user's wishes. In the example, this level can be user-defined and / or pre-defined.

[0105] A typical example of the computer implementation of the method is to execute this method on a system adapted for this purpose. The system can comprise a processor coupled to a memory and a graphical user interface (GUI), and the memory has recorded thereon a computer program containing instructions for executing this method. The memory can also store a database. The memory is hardware adapted for such storage and, in some cases, is composed of a plurality of physically different parts (for example, one for the program and one for the database).

[0106] Figure 5 shows an example of a system, which is a client computer system, for example, the user's workstation.

[0107] The client computer of this embodiment is composed of a central processing unit (CPU) 1010 connected to an internal communication bus and a random access memory (RAM) 1070 also connected to the bus. The client computer further includes a graphical processing unit (GPU) 1110 related to a VRAM 1100 connected to the BUS. The video RAM 1100 is also known as a frame buffer in the art. The mass storage controller 1020 manages access to mass memory devices such as a hard drive 1030. Examples of mass memory devices suitable for embodying computer program instructions and data include semiconductor memory devices such as EPROM, EEPROM, and flash memory devices, all forms of non-volatile memory, magnetic disks such as built-in hard disks and removable disks, and magneto-optical disks. Any of the above may be complemented or incorporated by a specially designed ASIC (application-specific integrated circuit). The network adapter 1050 manages access to the network 1060. The client computer may also include tactile devices 1090 such as a cursor control device and a keyboard. A cursor control device is used in the client computer so that a cursor can be selectively placed at any position on the display 1080. Also, various commands can be selected and control signals can be input by the cursor operation device. The cursor control device includes a number of signal generators for inputting control signals to the system. Generally, the cursor control device is a mouse, and the buttons of the mouse are used to generate signals. Alternatively or additionally, the client computer system can include a touch pad and / or a touch screen.

[0108] A computer program may include instructions executable by a computer, the instructions including means for causing the system to perform the method. The program is recordable on any data storage medium, including the system's memory. The program can be implemented, for example, in digital electronic circuitry, computer hardware, firmware, software, or combinations thereof. The program can be implemented as a product tangibly embodied in a machine-readable storage device for execution by an apparatus, e.g., a programmable processor. The steps of the method may be performed by a programmable processor executing a program of instructions that operate on input data to produce output, thereby performing the functions of the method. Accordingly, the processor may be programmable to receive data and instructions from a data storage system, at least one input device, and at least one output device, and to send and receive data and instructions, and may be coupled. An application program can be implemented in a high-level procedural or object-oriented programming language, or in assembly or machine language as desired. In any case, the language may be a compiled or interpreted language. This program may be a full-installation program or an update program. When this program is applied to the system, in any case, it is instructed to perform the method. Alternatively, the computer program may be stored and executed on a server in a cloud computing environment, the server communicating with one or more clients via a network. In such a case, the method is performed on the cloud computing environment by a processing device executing instructions configured by the program.

Claims

1. 1. A computer-implemented method for predicting a maintenance status of a real-world product, comprising: providing a dataset including data describing past real-world maintenance events and characteristics of devices of the same type as the real-world device; training a neural network to predict parameters of an MCDA sorting model based on the dataset, the MCDA sorting model being configured to take as input at least one time 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; 23. A computer-implemented method comprising:

2. The MCDA sorting model is an NCS model. The method of claim 1.

3. the neural network is based on a sigmoid activation function for implementing a comparison rule in the MCDA sorting model; At least one parameter of the MCDA sorting model is a sigmoid function that implements a comparison rule. The method according to claim 1 or 2.

4. The maintenance status is one of normal operation, maintenance recommendation, and maintenance required.

4. The method according to any one of claims 1 to 3.

5. 5. An MCDA sorting model obtained according to the method of any one of claims 1 to 4.

6. An MCDA sorting model configured to take as input time measurements of maintenance-related physical and / or functional data of a real-world device and to output a prediction of a maintenance status of the real-world device, wherein at least one parameter of said MCDA sorting model is a sigmoid function implementing a comparison rule, and optionally, the MCDA sorting model is obtainable according to the method of claim 3. MCDA sorting model.

7. A method of using the MCDA sorting model according to claim 5 or 6, comprising the steps of: providing at least one time measurement of maintenance-related physical and / or functional data of a real-world device; applying the MCDA sorting model to at least one provided time measurement, thereby outputting a prediction of a maintenance status of a real-world device; The method of use has:

8. The at least one time measurement comprises at least one real-time measurement.

8. The method of claim 7.

9. MCDA sorting model is applied in real time 9. The method of claim 8.

10. The at least one time measurement is derived from at least one physical sensor of the device and / or a physical sensor attached to the device.

10. Use according to any one of claims 7 to 9.

11. A method of use according to any one of claims 7 to 10, comprising: comparing one or more predictions of the MCDA sorting model to one or more actual maintenance situations of a device; if the comparison results in a discrepancy, updating an MCDA sorting model based on one or more real-world maintenance conditions of the device; The method of use has:

12. A method of use according to any one of claims 7 to 11, comprising: performing maintenance on the device based on the predictions of the MCDA sorting model. The method of use has:

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

14. A computer readable data storage medium having recorded thereon the computer program according to claim 13 and / or the MCDA sorting model according to claim 5 and / or the MCDA sorting model according to claim 6.

15. A computer system including a processor coupled to a memory, the memory storing the computer program of claim 13 and / or the MCDA sorting model of claim 5 and / or the MCDA sorting model of claim 6.

Images