GB2612319A - System and method for detecting anomalous system behaviour - Google Patents

Abstract

A method and apparatus for predicting failure of an engineering asset based on real-time data. The system comprises a plurality of sensors for measuring data on the engineering asset. The method of predicting failure comprises receiving a data record comprising data on the engineering asset collected from the plurality of sensors at time t S122; generating, using a trained machine learning algorithm, a probability Pf that the received data record indicates that the asset is in a faulty state S127; determining what number of data records received in a look-back time Lt are indicative of the asset being in a faulty state, wherein the look-back time Lt is a time period occurring before the time t at which the data record was collected; and predicting a probability of the engineering asset failing during a horizon time Ht S136, wherein the horizon time Ht is a time period after time t at which the data record was collected. The predicting step implements a Bayes forecasting model to predict the probability of failure based on the generated probability Pf and the number of data records which were determined to be faulty S135 within the look-back time Lt. Further aspects disclose receiving multiple data records and grouping the records into a plurality of envelopes S130.

Description

SYSTEM AND METHOD FOR DETECTING ANOMALOUS SYSTEM BEHAVIOUR

TECHNICAL FIELD

The present invention relates generally to a system and method for detecting anomalous system behaviour, for example in complex engineering assets including those found in maritime vessels.

BACKGROUND

Prognostics and Health Management (PHM) is an active and well subscribed to field of study. Companies managing complex distributed or safety-critical engineering assets want to understand the status of their portfolio in order to maximise efficiency and minimise downtime. Approaches such as health scores, anomaly detection, failure prediction and remaining useful life (RUL) calculation are applied across various domains in order to advise maintenance and repair schedules.

The published literature in this field will typically use synthetic data or carefully collected lab-based data to demonstrate the proficiency of a technique, however the use of "real-world" applications is less common. The challenges of developing PHM algorithms to run in a real-world environment are discussed and postulated, but tangible examples are not widespread. For example, see as a reference "Prognostics and Health Management for Maintenance Practitioners-Review, Implementation and Tools Evaluation" by Atamuradov et al. published in International Journal of Prognostics and Health Management in 2017.

Most known PHM applications apply to engineered systems operating in and on a well-defined single-state and without consideration of prior usage and age. Especially, but not exclusively, in the maritime domain, engineering systems' operating envelopes are manifold by design and operating environment, and can include rapid, anomalous transient events (the term may be used interchangeably with "transients"). These transients can pose problems for classic machine learning (ML) algorithms, in that during these events a system may exhibit behaviour consistent with impending failure before recovering to "healthy" -2 -behaviour, leading to a false-positive which results in wasted repair efforts and a loss of operator confidence in the prognostic algorithm. This problem is further compounded by the fact that such transient events can cause low-level damage to a device, resulting in an increased risk of failure at a later date.

ML algorithms performing anomaly-based fault detection are rarely capable of estimating the remaining useful life (RUL) of a system. However, it is of enormous benefit to a maintainer to have such an estimate and a probability value attached to the estimate in order to permit maximum efficiency in task planning and supply chain optimisation. Furthermore, existing non-deep learning ML anomaly-based detectors, especially in the maritime domain, tend to trigger far in advance of system failure, provoking excessively early intervention by a maintainer, denying the operator the economic benefits of extracting maximum use from an asset.

It is an example aim of the present invention to at least partially overcome or avoid one or more problems of the prior art, whether identified herein or elsewhere, or to at least provide an alternative to existing systems and related methodologies.

SUMMARY

According to a first aspect of the present invention, there is provided a method for a computer-implemented method for predicting failure of an engineering asset based on real-time data, the method comprising: receiving a data record comprising data on the engineering asset collected from a plurality of sensors at time t; generating, using a trained machine learning algorithm, a probability PF that the received data record indicates that the engineering asset is in a faulty state; determining what number of data records received in a look-back time Lt are indicative of the engineering asset being in a faulty state, wherein the look-back time Lt is a time period occurring before the time t at which the data record was collected; and predicting a probability of the engineering asset failing during a horizon time Ht, wherein the horizon time Ht is a time period after time t at which the data record was collected; wherein the predicting step comprises -3 -implementing a Bayesian model to predict the probability of failure within the horizon time Fir based on the generated probability PF and the number of data records which were determined to be faulty within the look-back time b. According to another aspect of the invention, there is also provided a system for predicting failure of an engineering asset based on real-time data, the system comprising: a plurality of sensors for measuring data on the engineering asset and a processor which is configured to receive a data record comprising data on the engineering asset collected from the plurality of sensors at time t; generate, using a trained machine learning algorithm, a probability PF that the received data record indicates that the engineering asset is in a faulty state; determine what number of data records received in a look-back time Lt are indicative of the engineering asset being in a faulty state, wherein the look-back time Lt is a time period occurring before the time t at which the data record was collected; and predict a probability of the engineering asset failing during a horizon time Fit, wherein the horizon time Flt is a time period after time t at which the data record was collected; wherein predicting comprises implementing a Bayesian model to predict the probability of failure with the horizon time Ht based on the generated probability PF and the number of data records which were determined to be faulty within the look-back time Lt The invention may be considered to be a prognostics and health management (PHM) application which provides a probability of failure forecast for users including operators and maintenance staff. The invention may be used for monitoring any complex engineering asset whose operation may be described by a data set including a plurality of physical parameters which are measured by the plurality of sensors. Merely as examples, such engineering assets may be within a marine vessel, for example a diesel engine or a high voltage switchboard. Such engineering assets may be in other systems, e.g. railway rolling stock bogies, nuclear reactor coolant pumps, gas turbine engines for passenger aircraft or oil rig drill bearings.

As set out above, an output of the trained machine learning algorithm is the probability of the received data record being faulty (PF). The trained machine learning algorithm may also be used to generate the probability of each of the -4 -data records received within the look-back time being faulty. From these generated probabilities a binary fault value for each of the multiple data records (i.e. the received data record and the data records within the look-back time) may be determined. For example, the fault value for each of the multiple data records may be equal to 1 when a probability that the data record is faulty exceeds a probability threshold and 0 otherwise. The fault value Fi,t for each record i at time t may be defined as: (0' Pp > Prp (equation 1) o where Pp is the probability of the data record being faulty and Pth is the probability threshold.

Owing to the potential for transients (as defined in the background section above) leading to an erroneous classification of the status of a data record, a rolling average of the faulty state value Fi,t over a period (or "envelope") of time (duration eL) may be computed. The method may thus comprise grouping the received data record with multiple previously received data records to form an envelope of data records. Determining what number of data records received in the look-back time are indicative of the engineering asset being in a faulty state may comprise determining what number of envelopes received in the look-back time are indicative of the engineering asset being in a faulty state.

Each envelope may have an envelope length (eL) and may contain a sequence of all data records recorded during a time interval which is equal in length to the envelope length eL. The time interval may be measured back in time from the recordal time t of the final (or most recent) data record of the sequence of data records. There may be the same or a different number of data records in each data envelope. When each of the data records is captured at a set, uniform frequency, there will be the same number of data records within each envelope. However, the data records may be recorded at irregular time intervals and in such an instance, the envelopes may contain different numbers of data records.

The use of a rolling average (i.e. envelopes) has the effect of "smoothing" out the impact that a rapid transient/anomalous record may have on the determination of whether the underlying monitored engineering asset truly is in a faulty or "non-faulty" state. This rolling average is hereafter referred to as the sick rate. The sick rate of each data envelope may be equal to the sum of each fault value for the multiple data records within the data envelope, divided by the number of data records within the data envelope. In other words, the sick rate may be a normalised sum of the fault values. The sick rate SR may be calculated as defined below: EL, Ftt SR = ' (equation 2) where Ftt is the binary fault value for each record i within the envelope and N is the number of records within the envelope.

From these generated sick rates, a binary sick value for each of the multiple data envelopes may be determined. A data envelope (recorded in the jth position of a sequence of envelopes and terminating at a time t) may be determined to be indicative of running in a faulty state when a sick rate of the data envelope equals or exceeds a sick rate threshold and a data envelope may be indicative of running in a non-faulty state when a sick rate of the data envelope is lower than the sick rate threshold. An envelope which is determined to be faulty state may be given a sick value of 1 and an envelope which is determined to be non-faulty state may be given a sick value of 0. The sick value Sj.t may be defined as: (0'5R < S"it Sj't ti, SR > S"it (equation 3) where SR is the sick rate and &fit is the sick rate threshold.

As set out above, the Bayesian model may be implemented to predict the probability of failure within the horizon time Fit based on the generated probability PF and the number of data records which were determined to be faulty or "non-"faulty" within the look-back time Lt. For example, the Bayesian model may be defined as: P(FritinLt) P(nLt) P(711,t1FHt) x F(Fftr) -6 - (equation 4) where FH, is the failure of the engineering asset within the time period which follows the moment of evaluation t and which is equal in length to the horizon time Hi, nu is the number of faulty records or envelopes within the look-back time Li prior to the moment of evaluation t, P(FHtInci) is the probability of failure of the engineering asset within the horizon time Hi given n records or envelopes classed as faulty in time Li, P(nulFHt) is the probability of n records or envelopes being classed as faulty in look back time Li given that a failure is known to have occurred within the horizon time Hi, P(FHt) is the probability of failure of the engineering asset in any time period equivalent in length to the horizon time Hi and P(nu) is the probability of n records or envelopes being classed as faulty in any time equivalent in length to the look-back time Li. In other words, for any given record, a probability function which predicts the probability of the engineering asset failing completely in the horizon time from that record may be determined as a product of the probability of n records/envelopes being classed as faulty in the look-back time and the probability of the engineering asset failing in any time of length equivalent to the horizon time divided by the probability of n records being classed as faulty in any time of length equivalent to the look-back time.

As set out above, the output of the trained machine learning algorithm is the probability (PF) of the received data record indicating that the engineering asset is in a faulty state. The number of data records/envelopes (nu) in the look-back time Lt which are indicative of the engineering asset being in a faulty state is also determined. In addition to the probability (PF) and the determined number of data records/envelopes, the Bayesian model may also be built based on knowledge of how many faults the engineering asset may manifest per unit time and/or how many records (or envelopes) may be classified as pertaining to a faulty state per unit time. This information may be obtained from a look-up table or other database which stores this information. The various probabilities which are input into the Bayesian model, e.g. P(nulFHt), P(FHt) and P(nu) may be calculated using known techniques from the probability (Pr), the number of faulty data records/envelopes nu, the faults per unit time and the number of faulty records per unit time. This may be done using any suitable technique, for example -7 -a frequentist approach (i.e. consulting a stored historical database and determining the frequency of events) may be suitable. As an example, P(FH) may be determined by calcuiating the total number of records for which there is a failure within the following horizon time, and dividing by the total number of records available The predicted probability may be compared to a failure threshold and when the predicted probability is equal to or exceeds the failure threshold, an alert may be output to a user. The alert may be an audio or visual cue, or a combination of both.

The Bayes forecasting model may comprise a plurality of sub-models, one for each of a plurality of different look-back times (Lo, Li, L2, M. The method may comprises, for each sub-model: determining what number of data records (nLo, nu, flu,... nu) received in the corresponding look-back time (Loy Li, L2y... Li) are indicative of the engineering asset being in a faulty state.

Before the implementation of the method and system described above, an arbitrary machine learning algorithm needs to be trained before use. The data records within the training data correspond to the data records which are received for prediction purposes. In other words, the training data comprises data records having data collected from at least the same set of sensors. The training data may be stored on a database which may be separate to the system which implements the prediction. Indeed, the training may be carried out on a separate system and the trained algorithm transferred to the prediction system when required.

The method may comprise training the machine learning algorithm by receiving multiple data records for at least one engineering asset which corresponds to the engineering asset for which failure is to be predicted classifying each of the multiple data records as either indicative of at least one engineering asset running in an acceptable state or running in a faulty state. Classifying an engineering asset as running in an "acceptable" running state may be considered as a "non-faulty" running state and may also be considered as an alternative to the known method of classification as a "healthy" state. The multiple data records may comprise data previously collected from the plurality of sensors -8 -at a sequence of times prior to failure of at least one engineering asset which may be identical to the engineering asset being monitored.

The training may be implemented separately to the prediction and thus according to another aspect of the invention, there is provided a computer-implemented method for training a machine learning algorithm to predict failure of an engineering asset based on real-time data, the method of training comprising: receiving multiple data records for at least one engineering asset which corresponds to the engineering asset for which failure is to be predicted, wherein the multiple data records comprise data previously collected from the plurality of sensors at a sequence of times prior to failure of the least one engineering asset; and classifying each of the multiple data records as either indicative of the at least one engineering asset running in an acceptable state or running in a faulty state.

Classifying each of the multiple data records as indicative of running in a faulty state may comprise obtaining a minimum remaining running time, MRRT, which is a time period before failure of the at least one engineering asset; and classifying each of the multiple data records which is within the minimum remaining running time as indicative of running in a faulty state. The MRRT may be obtained by an input from an operator who has knowledge of the engineering asset or may be input automatically by the management system. The MRRT may be determined based on cost/benefit calculations. Such a determination, in the context of a maritime vessel, might be performed by an insurer or maritime engineer. The MRRT may be any appropriate time period, e.g. measured in seconds, hours or even days, depending on the system.

Classifying each of the multiple data records as indicative of running in an acceptable state may comprise obtaining a minimum running time, MRT, which is a time period after the at least one similar engineering asset has been started; and classifying each of the multiple data records which is after the minimum running time and before the minimum remaining running time as indicative of running in an acceptable state. In other words, the "acceptable" running data is sampled from the data records between the MRT and the MRRT and thus this training uses acceptable running sampling (ARS). The MRT may be obtained by an input from an operator who has knowledge of the engineering asset or may be input automatically by the management system. The MRT may be any appropriate time period, e.g. measured in seconds, hours or even days, depending on the system. Although data received in the MRT may not be included in the training data, data received in the MRT for an engineering asset for which failure is being monitored is likely to be included when making the predictions.

As an alternative, or in addition to training using the acceptable running sampling (ARS) described above, time envelopes (as described above) may also fi) be used. The method may comprise training the machine learning algorithm by receiving multiple data records for at least one engineering asset which corresponds to the engineering asset for which failure is to be predicted, wherein the multiple data records comprise data previously collected from the plurality of sensors at a sequence of times prior to failure of the least one engineering asset; grouping the multiple data records into a plurality of data envelopes, wherein each envelope has an envelope length (eL) and contains a sequence of all data records recorded during a time interval of length eL; and targeting correct classification of each data record within each data envelope as indicative of running in a faulty state or running in a not-faulty state.

This training may also be implemented separately to the prediction and thus according to another aspect of the invention, there is provided a computer-implemented method for training a machine learning algorithm to predict failure of an engineering asset based on real-time data, the method of training comprising: receiving multiple data records for at least one engineering asset which corresponds to the engineering asset for which failure is to be predicted, wherein the multiple data records comprise data previously collected from the plurality of sensors at a sequence of times prior to failure of the least one engineering asset; grouping the multiple data records into a plurality of data envelopes, wherein each envelope have an envelope length (eL) and contains a sequence of all data records recorded during a time interval of length eL; and targeting correct classification of each data record within each data envelope as indicative of running in a faulty state or running in a not-faulty state.

-10 -As set out above, there may be the same or a different number of data records in each data envelope. Targeting correct classification may comprise calculating a sick rate for each envelope, for example using equation 2 above. Targeting correct classification may comprise calculating a sick value for each envelope, for example using equation 3 above.

Thus, according to another aspect of the invention, there is provided a method for a computer-implemented method for detecting anomalous behaviour of an engineering asset based on real-time data, the method comprising: receiving a data record comprising data on the engineering asset collected from o a plurality of sensors at time t; and classifying, using a trained machine learning algorithm, the data record as faulty or not-faulty, wherein the machine learning algorithm has been trained as described above using one or both of acceptable running sampling and time envelopes. Classifying may further comprise grouping the received data record with a set of previously received data records to form a data envelope, classifying the data envelope as faulty or not-faulty and applying the classification of the data envelope to the data record.

According to another aspect of the invention, there is a non-transitory computer-readable medium comprising processor control code which when running on a system causes the system to carry out the method described above.

The various aspects described above may be considered to include a novel three-fold approach using one or more of acceptable running sampling, time envelopes and a Bayesian forecasting approach. The outputs may be presented as a decision aid in a graphical user interface (GUI) that allows the user to identify suitable maintenance actions. The application is designed such that the maximum acceptable use can be extracted from an engineering asset without excessively early intervention It will be appreciated that any one or more features described in relation to an aspect of the invention may replace, or be used in combination with, any one or more features described in relation to another aspect of the invention, unless such replacement or combination would be understood by the skilled person as mutually exclusive, after a reading of this disclosure. In particular, any features described in relation to apparatus-like aspects may be used in combination with, or in place of, any features described in relation to method-like aspects. They may also be carried out in any appropriate order.

BRIEF DESCRIPTION OF THE DRAWINGS

For a better understanding of the invention, and to show how embodiments of the same may be carried into effect, reference will now be made, by way of example, to the accompanying diagrammatic figures in which: Figure 1 is a schematic illustration of a system for detecting anomalous system behaviour; o Figure 2a is a flowchart setting out the steps carried out by the system of Figure 1 when training; Figure 2b is a flowchart setting out the steps carried out by the system of Figure 1 to predict anomalous behaviour; Figure 3 plots a plurality of data measurements over time including measurements taken during healthy operation and failure; Figure 4 illustrates how faulty data may be sampled from the data shown in Figure 3; and Figure 5 illustrates an alternative technique for sampling faulty data.

zo DETAILED DESCRIPTION

Figure 1 schematically shows an example embodiment comprising a management system 10 for monitoring an engineering asset. The engineering asset may be on a marine vessel, for example chilled water (CHW) plants, transmission, starboard shaft, port shaft, one or more CHW circulation pumps, gas turbine, converters, DBRs, propulsion motor, high voltage (HV) switch board, low voltage (LV) switchboard, HV filters, HV/LV transformers, LV filters, diesel generators, emergency generators, heating, ventilation and air conditioning (HVAC) system, pressurised air, fresh water system, salt water (SW) system. It will be appreciated that the management system may also be used on other -12 -systems, e.g. railway rolling stock bogies, nuclear reactor coolant pumps, gas turbine engines for passage aircraft or oil rig drill bearings.

The management system 10 is connected to a plurality of sensors, including a temperature sensor 50, a pressure sensor 52, an electrical current sensor 54, a voltage sensor 56 and other sensor(s) 58, e.g. a sensor to measure the mass of any fluids in the system. It will be appreciated that the depicted sensors are merely illustrative, and any suitable sensors may be included in the overall system. The combination of sensors may record a plurality of physical parameters, e.g. temperatures, pressures, salinity, chemical concentrations, fluid flow rates, electrical currents, voltages or any other physical parameter, which alone or together with digital parameters describing the configuration of the system, e.g. valve positions in a piping assembly, are indicative of whether the engineering asset is operating in a healthy or faulty state. For example, for a chilled water plant, the sensor data may include readings such as temperature and pressure of the water and the coolant, from before and after the pump, and pump mechanical power, current rate etc. Some of the internal detail of the display system 10 is shown in Figure 1. There are standard components of whichever hardware solution is deployed, including for example a display 20, a processor 30, memory 40 and an interface 42 for connecting with the sensors 50, 52, 54, 56, 58. The display may include a touchscreen 22 or may be any other suitable display on which information may be displayed to a user 70. The management system 10 is also connected to a database 60 which may be located remotely (i.e. in a different location) from the management system, including for example in the cloud. The database 60 comprises training data 62 which is used to train the management system as described in more detail below. It will be appreciated that there may be other standard components which are omitted for clarity.

The processor 30 may process the received sensor data and training data in any appropriate way. As illustrated, the specific processing may be completed in dedicated modules, e.g. a machine learning engine (or module) 32. Terms such as 'component', 'module', 'processor' or 'unit' used herein may include, but are not limited to, a hardware device, such as circuitry in the form of discrete or -13 -integrated components, graphical processing units (GPUs), a Field Programmable Gate Array (FPGA) or Application Specific Integrated Circuit (ASIC), which performs certain tasks or provides the associated functionality. In some embodiments, the described elements may be configured to reside on a tangible, persistent, addressable storage medium and may be configured to execute on one or more processors. These functional elements may in some embodiments include, by way of example, components, such as software components, object-oriented software components, class components and task components, processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays, and variables. Although the example embodiments have been described with reference to the components, modules and units discussed herein, such functional elements may be combined into fewer elements or separated into additional elements.

Figures 2a and 2b illustrate the steps which may be carried out by the system shown in Figure 1. It will be appreciated that, unless indicated otherwise, the following steps can be carried out in any order, and not just the one described below. Furthermore, some of the steps could be carried out simultaneously or some steps could be omitted.

Figure 2a shows the first phase of the method which relates to the training phase in which supervised learning may be carried out as explained below. Training data is received at step S100. The training data may comprise a plurality of data records from the sensor(s) monitoring an asset identical to the subject system. Any given record within the training data may be labelled based on its ordinal position in the sequence of available data (e.g. first, second, third etc corresponding to i= 1, 2, 3 etc.) and the time t at which it is recorded. The data may have been recorded from the sensors at a fixed frequency or may be irregular and thus the time between records may or may not be constant. The data is recorded for engineering assets which are ideally identical to the one which is to be monitored. The training data also records the failure time of the engineering asset. The training data may also include recorded data from the very same system which is being monitored.

-14 -Figure 3 illustrates an example of a time series of training data. As shown in Figure 3, the data begins when the machine (or engineering asset) is started and continues until and after the machine has failed. When training classical (i.e. non-deep learning) machine learning binary classification algorithms for PHM using supervised learning, it is necessary to provide the algorithms with examples of both known healthy running behaviour for a system and faulty running. The selection of data conforming to these requirements in the existing literature is typically based on the assumption that a system's behaviour when new (or newly repaired) will be "healthy" whereas immediately prior to a failure the system will be faulty.

Returning to Figure 2a, as shown at step S102, an optional step is to obtain the minimum running time (MRT) within the received training data. The MRT will depend on the nature of the engineering asset being monitored but may be defined as the minimum time period for the complex engineering asset to decay from its original pristine condition after being started. For example, the MRT may thus be any appropriate time, e.g. measured in seconds, hours or even days, depending on the system. This minimum running time may be obtained by reference to an operator's (e.g. marine engineering officers') knowledge of the engineering asset, may be input automatically by the management system or may come from the general knowledge of the certifying authorities, e.g. the ship insurers. The data may then be sampled from a period of time after the Minimum Running Time (MRT) as illustrated in Figure 3.

Returning to Figure 2a, as shown at step S104, another optional step is to obtain the Minimum Remaining Running Time (MRRT). This step may be done simultaneously with step S102 as illustrated or the steps may be carried out in any order. As illustrated in Figure 3, the MRRT is a relatively brief period prior to system failure and it is only in this period that human intervention is required. The MRRT may be determined based on risk/cost/benefit calculations weighing the costs of replacement parts and equipment and working hours against the costs incurred by the failure of an engineering asset. For example, like the MRT, the MRRT may be any appropriate time, e.g. measured in seconds, hours or even days, depending on the system.

-15 -When the MRT and MRRT are obtained, as shown in Figure 2a, at step S106, the received training data may be sampled as "acceptable" running data and "faulty" running data. "Acceptable" running data may be defined as data records having occurred after the initial MRT has elapsed until immediately before the MRRT. Figure 3 illustrates the period in which "acceptable" data is sampled and this period may be sampled heavily from the available data. Faulty data is sampled during the MRRT, i.e. from the time immediately prior to failure as shown in Figure 4.

"Acceptable" running data may be considered as an alternative to including all the data records which are indicative of truly "healthy" running. This may alleviate issues with known classifiers in which "faulty" behaviour is erroneously detected very early in a system's life cycle, prompting unnecessary human intervention, and provoking enormous error in RUL estimations (as with so much remaining life, the variations in the behaviour of a vessel and the conditions a system is exposed to can be enormous). In the present proposal, it is assumed that a system will accrue wear and damage throughout its life yet continue to function and retain economic value.

Figures 3 and 4 also illustrate that the data records may be grouped into a plurality of data envelopes each having an envelope length (eL). As shown in Figure 3 each of the data records is captured at a set, uniform frequency and thus there are the same number (merely as an illustration -four in this example) of data records within each envelope. However, as set out above, the data records may be recorded at irregular time intervals and in such an instance, the envelopes may contain different numbers of data records. Thus, returning to Figure 2a, an optional step S108 may be to select data envelopes within the sampled data. Selecting data envelopes may comprise selecting the envelope length (eL).

For each data record 1, if a probability (Per) of the record being faulty at a time t exceeds or equals a critical level (which may be referred to as a probability threshold, Ph), the record may be assigned a 'fault' value (Ft) of one.

Alternatively, if a probability (Per) of the record being faulty at a time t is lower than -16 -the probability threshold, Pill, the fault value is zero, as defined below by equation 1: fO, Pp < Pth tt, PF > Pth equation (1) The notation above in which the fault values are assigned both i and t subscript values reflects the fact that any given record can be labelled based on its ordinal position in the sequence of available data (i) and the time t at which it is recorded.

For each data envelope, a value which is termed the sick rate SR may be calculated from the fault values for the data records. The sick rate SR may be 10 calculated as sum of the fault values Fa for each record within the envelope divided by the number of records (N) within the envelope as defined by equation 2: ELFit SR - L equation (2) where eL is the envelope length as defined above.

In other words, the fault values for each record within a period of time (eL) preceding the most recent record at time t are summed together. It should also be noted that in this formulation, the sick rate can be applied both in the case of regular data supply (i.e. data being recorded for all sensor channels simultaneously at a set frequency), and in the case of irregular data supply. This is because it does not assume that the time between records is consistent (allowing for N to fluctuate for a given value of eL).

For each envelope j, if the sick rate equals or exceeds a critical threshold (which may be referred to as a sick rate threshold Sera), the system may be determined to be faulty at the time t. An envelope which is determined to be in faulty state may be given a sick value of 1 and an envelope which is determined to be in a non-faulty state may be given a sick value of 0. The sick value is determined by equation 3: (0,5R < S "Et equation (3)

_

Sbt - SR > S "rt -17 -The use of the probability and sick rate thresholds described above, provides new options for training the underlying machine learning algorithm. For example, as shown in step S110, arbitrary values for the probability threshold (Prh) and the sick rate threshold (Serif) may be set.

The next step S112 is then to train the machine learning algorithm to target correct classification of the envelopes and the records within the envelopes as faulty or not-faulty using the received training data (i.e. to generate values for PF which, via equations 1, 2 and 3 lead to correct classification of the state of the engineering asset). The training may be done in any standard way, for example -up by evaluating an appropriate loss function. Examples of suitable loss functions include area under receiver operating or precision/recall curves In this case, effort is directed primarily at optimising the machine learning classification algorithm to produce values of PF which lead to correct classification of the state of a data envelope given set values of for the probability threshold (Pm) and the sick rate threshold (Serif).

Another option for training the machine learning algorithm may be to obtain as in step 5114 a pre-trained algorithm (e.g. one trained as in steps 5110 and S112 -or perhaps using a different training method). The next step S116 is then to train the machine learning algorithm to target the optimum values of the probability threshold (Pm) and the sick rate threshold (Sr) which best fit the training data to correctly classify the envelopes and hence the records within the envelopes as faulty or not-faulty. Again, the training may be done in any standard way, for example by simple brute force searching of the two-dimensional space described by these variables, and evaluating an appropriate loss function across the search space. In this case effort is directed primarily at optimising the values of for the probability threshold (Pm) and the sick rate threshold (Salt) given a pre-trained machine learning classification algorithm.

Both training options lead to a trained algorithm being output at step S118. The trained algorithm may be used as described in Figure 2b.

Figure 2b illustrates the application of the output trained algorithm to monitor the performance of an engineering asset in real-time. As shown at step -18 -S120, the trained algorithm is obtained. A new data record is also received at step 5122. The new data record may be a large corpus of sensor data which corresponds to the data records used to train the trained algorithm to correctly classify whether the engineering asset is in a "faulty" or "non-faulty" state. The new data record may or may not be derived from the same engineering asset as the training data, however the engineering asset should be essentially the same (and preferably identical) and capable of generating data equivalent in type and dimensionality to the training data.

The management system may forecast the probability of the asset (or system) failing within a prescribed time period. This prescribed time period may be termed the horizon time (HO and may be set (or simply obtained) in step S124. Figure 5 illustrates a time series of data records and illustrates the horizon time in this example occurs after another five data records are obtained. The horizon time may be set based on the engineering asset being monitored and the requirements of the end user. A cost/benefit calculation similar to that used to produce the MRRT may be used to determine a desirable horizon time. Example time frames may range from minutes (to provide warning for immediate engineering intervention) to months (to provide adequate warning for the repositioning of engineering supplies/acquisition of the long-lead items require for a repair/scheduling of major maintenance). The applicant has commonly used values of 24 hours in previous work.

In addition to obtaining a horizon time which looks into the future, a look-back time (Lt) may also be set in step S126 of Figure 2b. The look-back time (Lt) may be for the same time period as the horizon time (HO, e.g. five previous time records as in the example shown in Figure 5. Alternatively, the horizon time (Ht) and the look-back time (Lt) may have differing lengths. Steps S126 and S128 are depicted in Figure 2b as occurring at the same time but it will be appreciated that they may be carried out sequentially or at any suitable time in the method.

The management system may then use the trained algorithm at step S127 to determine a first probability (PF) as to whether the new data record (recorded at time t) indicates that the asset which is being managed is "faulty" or "not-faulty".

-19 -As shown at step S128, equation 1 may then be applied to determine if the data record is indicative of the engineering system being in a faulty state (Fu has a value 1) or a non-faulty state (Fit has a value 0).

Optionally, the new data record may then be grouped with at least some of the previously received data records in the look-back time (b) to form a data envelope (labelled j and ending at time t) as described above. Each data envelope includes all records available within the envelope length time eL prior to the time t. The envelope length time eL is normally shorter than the length of the look-back time (b) and other envelopes which occur before the envelope including the new data record may also be defined by grouping earlier data records. Once the records have been grouped, the sick rate (SR) of the (or each) envelope is calculated using equation 2 (step 5132). The sick rate (SR) is used in equation 3 to determine the sick value (Sit) for the (or each) envelope. The sick value may be considered to be an overall classification of the system being in a faulty state (Sit has a value 1) or a non-faulty state (Sit has a value 0).

As shown in step S135, the number of faulty data records or envelopes (nu) within the look-back time may then be determined. Using the look-back time (b), the horizon time (Hi) and the number of faulty records or envelopes (nu) within the look-back time, the probability of the system failing with the horizon time (Fir) after the moment of calculation (t) may then be forecast as in step S136.

The probability may be forecast by using a Bayes equation such as that shown in equation 4, which is repeated below: P (Flit 13(riLtiFt) x P(F Ht) intt) P @ILO (Equation 4) where FHt is the failure of the system within the time period which follows the moment of evaluation t and which is equal in length to the horizon time Fit, nu is the number of faulty records or envelopes within the look-back time L1 prior to the moment of evaluation t, P(FHtinu) is the probability of failure of the engineering asset within the horizon time Fir after the moment of evaluation t given n records/envelopes classed as faulty in the look-back time Lt prior the moment of evaluation t, P(nLtIFHt) is the probability of n records/envelopes being classed as -20 -faulty in look back time Li prior to the moment of evaluation t given that a failure is known to have occurred within time Ht after the moment of evaluation t, P(FHO is the probability of failure in any time period equivalent in length to the horizon time Ht and P(nu) is the probability of n records/envelopes classed as failure in any time period equivalent in length to the look-back time L. In other words, for any given record, the probability of the system failing completely in the horizon time from that record is determined as a product of the probability of n records/envelopes being classed as faulty in the look-back time measured from the time of the new data record and the probability of failure happening in any time period of length equivalent to the horizon time divided by the probability of n records/envelopes being classed as faulty in any time of length period equivalent to the look-back time. This formulation may permit the use of existing probabilistic computing techniques to determine the probability function, for example as described in the textbook "Practical Probabilistic Programming" by Avi Pfeffer, published by Manning Publications or by using standard software packages such as Edward, PyMCs and PyStan. It is also noted that a background to Bayesian optimisation techniques is found in "Bayesian Approach to Global Optimization: Theory and Applications" by Jonas Mockus published by Springer Science & Business Media.

Once the probability of failure has been determined, the management system may determine compare the calculated probability with a pre-set threshold determined by a user (step S134). If the calculated probability is greater than the threshold, an alert may be output to the user (step S136). For example, a visual alert may be output on the display screen or an audible alert may be output. The output may be presented as a decision aid in a graphical user interface on the display screen. If the probability of system failure within the horizon time is below the threshold set by the user, the system continues to collect and evaluate more data.

Equation 4 may be extended to produce a plurality of hierarchical models comprising sub-models in which the number of records/envelopes (nLo, nw) classed as faulty through multiple look-back time windows (Lo, Li, L2,... Lj) of varying length are used to construct prior distributions. This may provide a more -21 -accurate prediction. Other adaptations may be made, e.g. to include more background or more detail. The adaptability of the Bayesian forecasting is a useful tool in improving the predictions.

Additional data may also be used to generate other sub-models. For example, total system run-time since last repair (survival analytics) may be used to generate sub-models. For marine vessels, metrics pertaining to the current state of the sea (or other water body) on which the marine vessel is operating may be used to generate sub-models. These additional sub-models may be brought into the overall hierarchical model to provide increasingly sophisticated forecasts for the probability of failure within the horizon time.

The three approaches described above (determining an acceptable running sample, use of time envelopes and Bayesian forecasting) may be combined to produce a fault detection algorithm with the ability to provide the user with an estimate of the probability of failure in the following horizon time.

Appropriately trained and configured copies of the algorithm may work simultaneously on multiple systems and specific sub-systems. For example, a marine vessel may comprise a plurality of chilled water plants and four copies of the algorithm may be deployed, either having been trained on data accumulated on the specific chilled water plant they are to monitor, or having been trained on data pooled from several chilled water plants.

When training, the training data may have been sampled as "faulty" or "acceptable" using the MRT and MRRT described above and/or grouped in time envelopes. As set above, the training may be done by evaluating an appropriate loss function. The applicant has also recognised that more elegant solutions, such as the application of Bayesian optimisation functions may also be used and may be highly beneficial in terms of reduced computational overhead.

Attention is directed to all papers and documents which are filed concurrently with or previous to this specification in connection with this application and which are open to public inspection with this specification, and the contents of all such papers and documents are incorporated herein by reference. -22 -

Although a few preferred embodiments have been shown and described, it will be appreciated by those skilled in the art that various changes and modifications might be made without departing from the scope of the invention, as defined in the appended claims. The invention is not restricted to the details of the foregoing embodiment(s). The invention extends to any novel one, or any novel combination, of the features disclosed in this specification (including any accompanying claims, abstract and drawings), or to any novel one, or any novel combination, of the steps of any method or process so disclosed.

All of the features disclosed in this specification (including any accompanying claims, abstract and drawings), and/or all of the steps of any method or process so disclosed, may be combined in any combination, except combinations where at least some of such features and/or steps are mutually exclusive. Each feature disclosed in this specification (including any accompanying claims, abstract and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise. Thus, unless expressly stated otherwise, each feature disclosed is one example only of a generic series of equivalent or similar features.

Claims (1)

1. -23 -CLAIMS 1. A computer-implemented method for predicting failure of an engineering 5 asset based on real-time data, the method comprising: receiving a data record comprising data on the engineering asset collected from a plurality of sensors at time t; generating, using a trained machine learning algorithm, a probability Pp that the received data record indicates that the engineering asset is in a faulty 10 state; determining what number of data records received in a look-back time Lt are indicative of the engineering asset being in a faulty state, wherein the look-back time Lt is a time period occurring before the time t at which the data record was collected; and predicting a probability of the engineering asset failing during a horizon time Ht, wherein the horizon time Ht is a time period after time t at which the data record was collected; wherein the predicting step comprises implementing a Bayes forecasting model to predict the probability of failure within the horizon time Ht based on the generated probability PF and the number of data records which were determined to be faulty within the look-back time Lt 2. The method of claim 1, further comprising generating a fault value Pt for each data record i at time t, from where Pp is the probability of the data record being faulty and Pm is a probability threshold.3. The method of claim 1 or claim 2, further comprising grouping the received data record with multiple previously received data 30 records to form an envelope of data records, wherein each envelope has an -24 -envelope length (e0 and contains a sequence of all data records recorded during a time interval equal to the envelope length.4. The method of any one of the preceding claims wherein determining what number of data records received in the look-back time are indicative of the engineering asset being in a faulty state comprises determining what number of envelopes received in the look-back time are indicative of the engineering asset being in a faulty state.5. The method of claim 3 or claim 4, comprising determining that an envelope is indicative of the engineering asset being in a faulty state by calculating a sick rate of the data envelope from a normalised sum of each fault value for each data record within the envelope; comparing the sick rate to a sick rate threshold; and when a sick rate of the data envelope equals or exceeds the sick rate threshold determining that the data envelope is indicative of running in a faulty state.6. The method of any one of the preceding claims, wherein the Bayes forecasting model is defined as: P(nuIFFID x P(FH) p(FH,Ina) = P(TILt) where Fryt is the failure of the system within the time period equal to Htfollowing the moment of evaluation t, nu is the number of faulty records or envelopes within the look-back time Lt prior to the moment of evaluation t, P(FHtlnu) is the probability of failure within a time period time Ht after the moment of evaluation t given n records or envelopes classed as faulty in the time Lt prior the moment of evaluation t, P(rulFHt) is the probability of n records or envelopes classed as faulty in look back time Lt prior to the moment of evaluation t given that a failure is known to have occurred within time Htafter the moment of evaluation t, P(FHt) is the probability of failure in any time equivalent in length to the horizon time Ht -25 -and P(nu) is the probability of n records or envelopes classed as faulty in any time equivalent in length to the look-back time Lt 7. The method of any one of the preceding claims, wherein the Bayes forecasting model comprises a plurality of sub-models, one for each of a plurality of different look-back times (Lo, Li, L2,... Lj) and the method comprises, for each sub-model: determining what number of data records or envelopes (nLo, nu, nL2,...nu) received in the corresponding look-back time (Lo, Li, LZ... L1) are indicative of the engineering asset being in a faulty state.8. The method of any one of the preceding claims, further comprising training the machine learning algorithm by receiving multiple data records for at least one engineering asset which corresponds to the engineering asset for which failure is to be predicted, wherein the multiple data records comprise data previously collected from the plurality of sensors at a sequence of times prior to failure of the least one engineering asset; and classifying each of the multiple data records as either indicative of the at least one engineering asset running in an acceptable state or running in a faulty state.9. The method of claim 8, wherein classifying each of the multiple data records as indicative of running in a faulty state comprises obtaining a minimum remaining running time, MRRT, which is a time period before failure of the at least one engineering asset; and classifying each of the multiple data records which is within the minimum remaining running time as indicative of running in a faulty state. -26 -10. The method of claim 9, wherein classifying each of the multiple data records as indicative of running in an acceptable state comprises obtaining a minimum running time, MRT, which is a time period after the at least one similar engineering asset has been started; and classifying each of the multiple data records which is after the minimum running time and before the minimum remaining running time as indicative of running in an acceptable state.11. The method of any one of the preceding claims, further comprising training the machine learning algorithm by receiving multiple data records for at least one engineering asset which corresponds to the engineering asset for which failure is to be predicted, wherein the multiple data records comprise data previously collected from the plurality of sensors at a sequence of times prior to failure of the least one engineering asset; grouping the multiple data records into a plurality of data envelopes each having an envelope length (eL); and targeting correct classification of each data record within each data envelope as indicative of running in a faulty state or running in a not-faulty state.12. A computer-implemented method for training a machine learning algorithm to predict failure of an engineering asset based on real-time data, the method of training comprising: receiving multiple data records for at least one engineering asset which corresponds to the engineering asset for which failure is to be predicted, wherein the multiple data records comprise data previously collected from the plurality of sensors at a sequence of times prior to failure of the least one engineering asset; grouping the multiple data records into a plurality of data envelopes each having an envelope length (eL); and targeting correct classification of each data record within each data envelope as indicative of running in a faulty state or running in a not-faulty state. -27 -13. The method of claim 11 or claim 12, wherein targeting correct classification comprises determining that a data envelope is indicative of running in a faulty state when a sick rate of the data envelope equals or exceeds a sick rate threshold and that a data envelope is indicative of running in a non-faulty state when a sick rate of the data envelope is lower than the sick rate threshold; wherein the sick rate of each data envelope is equal to the sum of each fault value for the multiple data records within the data envelope, divided by the 10 number of data records within the data envelope, and wherein. the fault value for each of the multiple data records is set to 0 or 1 and the fault value for each of the multiple data records is equal to 1 when a probability that the data record is faulty exceeds a probability threshold and 0 otherwise.14. A computer-readable medium comprising processor control code which when running on a system causes the system to carry out the method of any one of claims 1 to 13.15. A system for predicting failure of an engineering asset based on real-time data, the system comprising: a plurality of sensors for measuring data on the engineering asset and a processor which is configured to receive a data record comprising data on the engineering asset collected from the plurality of sensors at time t; generate, using a trained machine learning algorithm, a probability PF that the received data record indicates that the engineering asset is in a faulty state; determine what number of data records received in a look-back time Lt are indicative of the engineering asset being in a faulty state, wherein the look-back time Lt is a time period occurring before the time t at which the data record was collected; and -28 -predict a probability of the engineering asset failing during a horizon time Ht, wherein the horizon time Ht is a time period after time t at which the data record was collected; wherein prediction implements a Bayes forecasting model to predict the probability of failure based on the generated probability PF and the number of data records which were determined to be faulty within the look-back time b.COCO C\I