Method for determining a health indicator of a facility.

Abstract

A method for determining a health indicator for an installation comprising a plurality of sensors, the method comprising the steps: a) obtaining a plurality of data from the plurality of sensors; b) storing the plurality of data, the plurality of data comprising at least first context data and second context data; c) determining an actual value of the installation's efficiency from the stored information; d) training a learning model of an estimated value of the installation's efficiency from the first context data and supervised learning; e) determining an estimated value of the installation's efficiency from the trained learning model and the second context data; f) determining a health indicator for the installation, being equal to the difference between the actual value of the installation's efficiency and the estimated value of the installation's efficiency.

Description

Title of the invention: Method for determining a health indicator of an installation.

[0001] The present invention relates to a method for determining a health indicator of an installation such as an air separation installation.

[0002] .

[0003] Supervised learning methods are used to enable better estimation of needs and lowering of the costs of industrial facilities.

[0004] With the collection of data from sensors in these installations, it becomes possible to improve production and distribution processes through these new methods, made possible by greater computing power.

[0005] Such learning methods are described in: “ARTIFICIAL INTELLIGENCE BASED PERFORMANCE PREDICTION SYSTEM, RAJNAYAK MAMTA AGGARWAL, 2018”.

[0006] Monitoring the health status of installations, particularly industrial installations, is a key point, as it determines the costs of maintenance and therefore replacement of the various equipment.

[0007] To detect such a need for replacement, an indicator called a "health indicator" is used. This indicator is calculated from data from sensors.

[0008] This type of indicator provides a good estimate of the age of the installation.

[0009] This is therefore a good way to estimate the remaining useful life.

[0010] When trying to determine which formula could be a good indicator of health, two solutions exist: - use a large amount of historical data to isolate which data may represent the health of the asset under study. This has been used in several previous works such as: "Sanz-Bobi MA et al., 2018", but this is difficult to apply due to a lack of historical data concerning past deteriorations; - use the knowledge of experts, which requires a work of capitalizing on knowledge (cf. "Knut Oien et al., 2018").

[0011] Moreover, the available dataset is not large enough.

[0012] Therefore, it is necessary to have an effective health indicator, which is also adapted to the targeted degradation.

[0013] The present invention aims to effectively remedy all or part of these drawbacks by proposing a method for determining a health indicator of a installation such as an air separation system comprising a plurality of sensors, the process comprising the following steps: - a) obtain a plurality of data relating to the state and / or activity of the installation and / or the context of the installation, the data being obtained from the plurality of sensors; - b) store the plurality of data in the form of a plurality of time series, including real or decimal numbers, the plurality of data comprising at least first context data relating to outside temperature or outside pressure and second context data relating to outside temperature or outside pressure; - c) determine a real efficiency value of the installation, based on the stored information; - d) train a learning model of an estimated value of the efficiency of the installation, from the first context data and supervised learning such as regression; - e) determine an estimated value of the installation's efficiency, from the trained learning model and the second context data; - f) determine a health indicator of the installation, being equal to the difference between the actual value of the efficiency of the installation and the estimated value of the efficiency of the installation.

[0014] Such a process allows a better estimation of the effect of the context on an installation in particular by abstracting from any parasitic effect and thus proposing a high-performance health indicator.

[0015] The invention will be better understood upon reading the following description and examining the accompanying figure.

[0016] [Fig-1] Fig. 1 is a representation of the isothermal efficiency of a compressor depending on the weather.

[0017] The invention relates to a method for determining a health indicator of an installation such as an air separation installation comprising a plurality of sensors, the method comprising the following steps: - a) obtain a plurality of data relating to the state and / or activity of the installation and / or the context of the installation, the data being obtained from the plurality of sensors; - b) store the plurality of data in the form of a plurality of time series, including real or decimal numbers, the plurality of data comprising at least some initial contextual data, including data relating to the outside temperature or outside pressure, and some second contextual data, particularly relating to outside temperature or outside pressure; - c) determine a real efficiency value of the installation, based on the stored information; - d) train a learning model of an estimated value of the efficiency of the installation, from the first context data and supervised learning such as regression; - e) determine an estimated value of the installation's efficiency, from the trained learning model and the second context data; - f) determine a health indicator of the installation, being equal to the difference between the actual value of the efficiency of the installation and the estimated value of the efficiency of the installation.

[0018] In one embodiment, the method is implemented to determine a health indicator of an installation such as an air separation installation comprising a compressor.

[0019] The compressor has a very long lifespan (at least 30 to 40 years), and may experience, during its use, a decrease in efficiency due to aging.

[0020] From a theoretical point of view, one would like to have a formula that takes all the context data and subtracts it from the health indicator. However, this is generally not the case, as the only thing available is a partial health indicator, which does not cover all possible failures / wear / aging.

[0021] In order to obtain the remaining useful life of such a compressor, it is proposed to use an indicator which takes into account the isothermal efficiency.

[0022] Compressors are the main components of the ASU (Air Separation Unit) from an energy perspective. Furthermore, they are subject to degradation over time and must be regularly monitored, as must proper use cases.

[0023] Compressor performance is considered overall. Under the same operating conditions (load, cooling temperature and pressure ratio), electrical consumption should be the same unless it is affected by failures.

[0024] The minimum power required to compress a gas from a pressure A to a pressure B is defined by the isothermal power. The isothermal power is the reference, and comparison with the actual power provides an indication of the compressor's efficiency.

[0025] The isothermal efficiency of the compressor is defined as: w r F «si®

[0026] With: nactuai = isothermal efficiency (%) ; wactuai = electrical consumption (kW); X = ideal gas constant (=8.317'0.00001240079); Tw = cooling water temperature (K); Qm = compressed gas flow (Nm3 / h); Pout / Pin = pressure ratio (abs).

[0027] The application of this calculation of the isothermal efficiency, to the compressor is illustrated [Fig.1], with the isothermal efficiency on the ordinate and the number of days on the abscissa.

[0028] It can be observed, over the ten years of data as represented in [Fig. 1], that annual seasonality has an impact on efficiency. This highlights the fact that this efficiency depends on the context.

[0029] Although we observe a downward trend, it is impossible to draw any conclusions, as the contexts vary. It is therefore necessary to normalize the efficiency values ​​to allow for comparison at different points in the life of an asset such as a compressor.

[0030] Thus, to compensate for this probable bias, it is proposed to compare the value calculated using the above formula with another value determined by a machine learning algorithm.

[0031] Indeed, a regressor can learn the actual efficiency based solely on the context (external temperature and pressure, imposed flow rate). If the difference remains the same, then there is no trend in efficiency. Such an indicator proves that the compressor remains as efficient as if it were new and therefore does not require any specific maintenance.

[0032] In an example embodiment, a machine learning regressor is trained on the first 20 percentages of the dataset as a function of time. The inputs, which are those of the context, are: cooling water temperature atmospheric pressure flow ratio between outlet and inlet pressures.

[0033] Once the context is fixed, the effectiveness should depend only on the health of the asset.

[0034] Cross-validations could also have been carried out to ensure stability of the model and a good balance between bias and variance, but this would not be a realistic test because the system could use future data to learn about the past data. To solve this problem, it is proposed to divide the data according to time, and to use MAE / RMSE as the main cost function for performance evaluation. [Tables 1] Training Size / Validation Ratio Performance on Training Data (MAE / RMSE) Performance on Test Data (MAE / RMSE) 20 / 80 0.0016 / 0.024 0.0076 / 0.010 40 / 60 0.0024 / 0.0035 0.0078 / 0.010 60 / 40 0.0019 / 0.0024 0.0072 / 0.010 80 / 20 0.0021 / 0.0032 0.0068 / 0.0084

[0035] It appears that only 20% for the training portion is sufficient, as the inference gathers relevant information about the model. Indeed, more than a year's worth of data has already been captured, which seems to allow for generalization of the knowledge.

[0036] If we compare the algorithm's ability to predict learned values, we obtain good values, around 0.0016 of MAE.

[0037] If we compare the algorithm's prediction on the remaining data (the remaining 80%) which it has never seen with the actual values, the results are around 0.0076 of MAE.

[0038] There is absolutely no drift in efficiency, even if we abstract from the context.

[0039] The value remains absolutely stable, the relative deviation having an average of 1% and a standard deviation of 0.9%.

[0040] There is no trend (around 4.10-8, a decrease in Es compared to the approach without normalization, and an average of 0.9993, which is very close to 1 because there is no variation, and the standard deviation of 0.014), i.e., no variation clear enough to conclude that the targeted compressor exhibits wear due to aging over the ten years of measurement.

[0041] If aging were to occur, we would have seen a decrease in performance, i.e. an average ratio greater than 1.

[0042]

[0043] As a final statistical evaluation of compressor stability, we propose to divide into N parts the time series created from the isothermal efficiency and to evaluate separately the median of the difference between the predicted and actual isothermal efficiency value.

Claims

[Claim 1] Demands Method for determining a health indicator for an installation such as an air separation system comprising a plurality of sensors, the method comprising the following steps: - a) obtain a plurality of data relating to the state and / or activity of the installation and / or the context of the installation, the data being obtained from the plurality of sensors; - b) store the plurality of data in the form of a plurality of time series, including real or decimal numbers, the plurality of data comprising at least first context data relating to outside temperature or outside pressure and second context data relating to outside temperature or outside pressure; - c) determine a real efficiency value of the installation, based on the stored information; - d) train a learning model of an estimated value of the efficiency of the installation, from the first context data and supervised learning such as regression; - e) determine an estimated value of the installation's efficiency, from the trained learning model and the second context data; - f) determine a health indicator of the installation, being equal to the difference between the actual value of the efficiency of the installation and the estimated value of the efficiency of the installation.

Images