Computing device and method

Abstract

A method performed by a computing device is provided. The method comprises receiving, from one or more pressure sensors configured to measure gas pressure at one or more respective locations in a Gas Distribution Network (GDN), measured pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a time period. The method comprises using a trained machine learning (ML) model to score the measured pressure data over the time period in accordance with a degree to which the measured pressure data for the time period deviates from normal pressure data at the one or more pressure sensors as characterised by the ML model. The ML model is trained to characterise normal pressure data at the one or more pressure sensors based on previous pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a previous time period. The method comprises comparing the score of the measured pressure data over the time period to a threshold score. The method comprises determining that the score of the measured pressure data over the time period has exceeded the threshold score. The method comprises, in response, outputting an alert indication indicating that an anomaly may have occurred, or may occur, in the GDN.

Description

[0001] P131560GB 1

[0002] COMPUTING DEVICE AND METHOD

[0003] BACKGROUND

[0004] Field of Disclosure

[0005] The present technique relates to a computing device and method.

[0006] The present application claims Paris Convention priority from GB patent application number GB 2503634.4, filed on 12 March 2025, the contents of which are hereby incorporated by reference in their entirety.

[0007] Description of Related Art

[0008] The “background” description provided herein is for the purpose of generally presenting the context of the disclosure. Work of the presently named inventors, to the extent it is described in this background section, as well as aspects of the description which may not otherwise qualify as prior art at the time of filing, are neither expressly or impliedly admitted as prior art against the present invention.

[0009] Gases may be distributed by a network of pipes which convey the gas under pressure from a source to one or more consumer units. One example of such a network is gas distribution network (GDN) which may be used to provide combustible gas for consumption by industrial, commercial or domestic consumers.

[0010] GDNs are typically formed from a network of pipelines through which the gas passes under pressure from a source to reach the consumers. The gas pressure in the network is set by one or more pressure control stations known as district governor stations (hereinafter referred to as “governor stations”) which receive the gas from a source. It is generally desirable to set the pressure in the network to achieve a balance between: safety, maintaining consumer service levels and gas leakage in the GDN. A gas supply pressure which is too low may be dangerous to consumers. For example, a low gas supply pressure may lead to incomplete combustion and consequent formation of toxic carbon monoxide or may lead to consumer units simply not functioning. There is therefore a statutory requirement that the gas supply pressure in GDNs should not fall below a minimum value. Conversely, GDNs may be prone to leakage which is of both environmental and financial concern to gas suppliers. Generally, leakage increases with the gas supply pressure. A gas supplier may therefore wish to impose a maximum gas supply pressure in the GDN to reduce the financial loss and environmental impact of gas leakage. Furthermore, for GDNs which supply bio-methane gas, it becomes more difficult to feed the bio methane-gas into the GDN if the pressure exceeds a predetermined threshold or if a pressure ratio between the GDN pressure and a bio-methane planet outlet pressure exceeds a predetermined threshold. Excess bio-methane by be burnt by flaring, thus resulting in environmental damage. Therefore, higher or lower than expected pressures in GDNs can cause technical problems.

[0011] Furthermore, a higher demand for gas results in a decrease in the pressure at pressure low-points in the GDN as consumers of the gas supply consume gas. An increased gas supply demand decreases the gas supply pressure in the GDN and vice versa. Consequently, gas suppliers are required to supply gas at a high enough pressure which takes into consideration potential pressure drops due to increased demand.P131560GB 2

[0012] The pressure of a gas at a point in a GDN is determined by a plurality of factors including: an outlet pressure at one or more governor stations in the GDN, the distance the point is away from the one or more governor stations and a demand for the gas in the GDN. Typically, the pressure of the gas is measured at one or more pressure low points in the GDN using digital gas pressure data loggers which may be alternatively referred to as “low-point loggers”. A low-point is a point in the GDN which has a low (possibly minimum) pressure. The GDN may have a number of low-points.

[0013] In some GDNs, gas supply pressures are set manually at governor stations. The gas supply pressure is typically set at a high value to prepare for a worst case scenario. For example, the pressure may be set at a pressure high enough such that the pressure at the low-points is expected to remain above the statutory minimum requirement even if the gas demand is expected to be the highest gas demand of any day. In some GDNs, governor stations are configured to automatically change a gas supply pressure based on a “clock”. For example, the governor stations may supply gas at one pressure during the day and at another pressure during the night. In some GDNs, governor stations are configured to alter a gas supply pressure based on predetermined pressure profiles.

[0014] A method which utilises machine learning to autonomously and accurately control the gas pressure to satisfy a gas pressure condition at one or more points in a GDN has been proposed in [1], the contents of which are hereby incorporated by reference in their entirety.

[0015] An additional problem in GDNs is the formation of anomalies. For example, one or more components in the GDN (such as governor stations) may fail to function properly, leaks may develop in one or more of the pipes of the GDN, or water may ingress into the pipes etc. Such anomalies can lead to improper performance of the GDN. For example, the gas pressure in the GDN can be affected by such anomalies which can lead to, for example, lower or higher than expected pressures. As explained above, higher pressures may result in increased gas leakage in pipes or an increased difficulty to feed bio-methane into a GDN, thus leading to environmental damage, while lower pressures may cause incomplete combustion, thus creating toxic carbon monoxide which is dangerous consumers. Both higher and lower than expected pressures may cause consumer units (such as boilers) not to function properly, thus adversely affecting consumer quality of experience.

[0016] The detection of such anomalies in GDNs represents a technical challenge. Typically, anomalies in GDNs are detected manually by, for example, consumers, other members of the public or network operational teams. In one example, a downstream consumer may observe and report a lower than expected boiler pressure. In other example, members of the public may notice and report a gas leak. Such reports may prompt network operational teams to review recent pressure profiles in the network to determine if there is unexpected pressure behaviour and thus whether there is an anomaly in the GDN. In other example, network operational teams may determine component failure or leaks in the GDN as a result of routine maintenance carried out on the GDN. The manual detection of anomalies means that the anomalies may go undetected for a significant amount of time. This may lead to increased pollution (due to leakage or flaring), increased formation of carbon monoxide (due to low pressure) and poor consumer experience (e.g. lower than expected pressure in consumer units). Furthermore, regulators may prescribe Key Performance Indicators (KPI) such as that critical reports must be responded to within one hour. Such KPIs may be difficult to meet when anomalies are manually detected.P131560GB 3

[0017] There is therefore a desire to improve anomaly detection in GDNs.

[0018] Previously disclosed arrangements include [2], [3], [4], [5] and [6],

[0019] SUMMARY OF DISCLOSURE

[0020] Various aspects and features of the present disclosure are defined in the appended claims.

[0021] It is to be understood that both the foregoing general description and the following detailed description are exemplary, but are not restrictive, of the present technology. The described embodiments, together with further advantages, will be best understood by reference to the following detailed description taken in conjunction with the accompanying drawings.

[0022] BRIEF DESCRIPTION OF THE DRAWINGS

[0023] A more complete appreciation of the disclosure and many of the attendant advantages thereof will be readily obtained as the same becomes better understood by reference to the following detailed description when considered in connection with the accompanying drawings wherein like reference numerals designate identical or corresponding parts throughout the several views, and wherein: Figure 1 illustrates a simplified Gas Distribution Network (GDN);

[0024] Figure 2 is a schematic diagram illustrating a section of a GDN which serves a plurality of consumers with gas along a route;

[0025] Figure 3 is a graph of gas pressure against time for a 24-hour period for the section of a GDN for a constant governor pressure;

[0026] Figure 4 illustrates a method performed by a computing device in accordance with example embodiments;

[0027] Figure 5 illustrates an example of a pressure profile graph and anomaly score graph in accordance with example embodiments;

[0028] Figure 6 illustrates an example of a pressure profile graph and anomaly score graph in accordance with example embodiments.;

[0029] Figure 7 schematically illustrates an example of a computing device communicating with a GDN in accordance with example embodiments;

[0030] Figure 8 schematically illustrates a computing device in accordance with example embodiments; Figure 9 schematically illustrates an example of an autoencoder demonstrating the passage of a data via an encoder neural network through a low dimensional bottleneck before reconstructing the data in a decoder neural network; and

[0031] Figure 10 is a graph illustrating gradual changes in pressure overtime.

[0032] DESCRIPTION OF EXAMPLE EMBODIMENTS

[0033] In the following description, a number of specific details are presented in order to provide a thorough understanding of the embodiments of the present invention. It will be apparent, however, to a person skilled in the art that these specific details need not be employed to practice the present invention. Conversely, specific details known to the person skilled in the art are omitted for the purposes of clarity where appropriate.

[0034] An example of a GDN, which may be configured in accordance with example embodiments, is schematically illustrated in Figure 1. In particular, Figure 1 represents a simplified representation in which a gas source 14 supplies gas to a GDN. In particular, the gas source 14 supplies gas to a governor station 12 through a gas distribution pipe 10. The governor station 12 controls a pressure of gas from the gas distribution pipe 10 to a pipe network 2. The gas pressure in the pipe network 2 is at a lower pressure than the gas pressure in the gas distribution pipe 10. TheP131560GB 4

[0035] pipe network 2 supplies a plurality of consumers 16 with the gas received from the gas source 14 via the governor station 12. A plurality of low point loggers 6 are used to measure a gas pressure at low-points in the pipe network 2.

[0036] A consumer of gas may be an industrial, commercial or domestic consumer or the like. The term “gas distribution network” is used herein to refer to a network of pipes and one or more governor stations for distributing gas to one or more consumers.

[0037] The gas source 14 is a representation of a source of gas which may be a standalone container of gas, or may represent gas received from one or more other gas networks. For example, the National Grid System is a gas network serving high pressure gas which is delivered to GDNs throughout the UK. The gas source 14 may also be a source of bio gas (such as bio-methane) generated from a source such as a farm or dedicated plant. It will be appreciated that although a single gas source 14 is shown in Figure 1, a GDN may be supplied by a plurality of gas sources.

[0038] Gas pressure is typically highest at the point of entry into the GDN and lowest at extremities of the GDN as a result of gas leakage and gas usage by consumers. For example, the National Grid may supply the gas source 14 of a GDN. Gas moves through the pipe network 2 driven by the pressure and the pressure drops as a result of gas usage consumers 16 and leakage.

[0039] The governor station 12 (or “governor”) in a GDN typically receives the gas from the higher pressure gas source 14 and contains pressure control means to lower the pressure of the gas received from the gas source 14. Consequently, the pressure of gas arriving at the governor station 12 is higher than the pressure of gas leaving the governor station 12. In one example, gas arrives at the governor station 12 with a pressure of about 1 to 2 bar and leaves the governor station with a pressure of up to about 50 mbar. The pressure of gas received by the plurality of consumers 16 will typically be lower than the pressure leaving the governor station 12 due to usage of gas by the consumers 16 and due to gas leakage. Hence, one or more points exist in a GDN for which the gas pressure may be low or minimal (referred to as “low-points”). Low-point loggers 6 are typically placed at some or all of these locations to monitor the gas pressure there as shown in Figure 1. For example, the low-points may be located using models calibrated by measurements of pressure throughout the GDN, and the low-points loggers 6 are placed at the low-points.

[0040] Figure 2 is a schematic diagram illustrating a section of the GDN shown in Figure 1 which serves a plurality of consumer 16 with gas. As shown in Figure 2, the governor station 12 receives medium pressure gas through the gas distribution supply pipe 10. As mentioned above, the source of the medium pressure gas supply is not limited and may be, for example, a connection to another gas network. The governor station 12 alters the medium pressure gas supply to a pressure Pd. The governor station 12 may alter the gas pressure through the use of pressure control means such as a pilot valve and an actuator. The gas at pressure Pd leaves the governor station and enters a pipe network 2 of the GDN to supply the gas to the plurality of consumers 16. The low-point logger 6 is typically disposed at a point near the edge of the GDN which is likely to have a low or minimal gas pressure. In Figure 2, a pressure Pi is measured at the low point 6. The governor station 12 is may be configured manually with settings to reach the pressure Pd. The gas pressure Pd is conventionally set manually. As will be appreciated, gas consumption typically varies throughout the year commensurate with environmental conditions. For example during winter in northern Europe, the weather is typically colder and so gas consumption willP131560GB 5

[0041] increase. Accordingly, the gas pressure Pd at the governor station 12 is set manually with different pressures between summer and winter. The pressure Pd set by the governor station 12 is required to ensure that the gas pressure Pi at the low-point 6 is above a minimum required by consumers 16 to operate gas burning devices. However, the pressure in the GDN will vary as a function of demand for gas from the consumers 16 connected to the GDN. This necessitates setting the pressure Pd at the governor station 12 to a value which delivers the minimum pressure at the low point 6 when consumer demand is highest. As a result, when a demand for gas is lower, the pressure set by the governor station 12 is higher than it needs to be, which can increase an amount of gas leakage from the GDN.

[0042] Figure 3 provides an example illustration of a need to set the gas pressure at the governor station 12 to a maximum when the demand is greatest which can result in too much pressure in the GDN at other times.

[0043] Figure 3 provides an illustration of a graphical plot of gas pressure against time for a 24-hour period for the GDN of Figure 2. Figure 3 provides a graphical plot of network pressure against time throughout a day illustrating an example of a relationship between a pressure measured at the governor station 12 (herein after referred to as the “governor pressure 30, Pd”) and a pressure measured at a low-point of the GDN (hereinafter referred to as the “low-point pressure 32, P”). The governor pressure 30 may also be referred to as the “gas supply pressure” herein. As will be appreciated from Figure 3, the governor pressure 30 is constant in time for a recorded 24 hours. This is because the governor pressure 30 in this example is set manually at the governor station 12. However, the low-point pressure 32 is variable in time over the 24 hours. The low-point pressure 32 may drop due to an increased consumer demand, for example. In Figure 3, the low-point pressure 32 falls in the early morning hours. This is likely as a result of cold temperatures typical of the early morning hours and an increased consumer demand as household heating systems are turned on. An excess pressure 22 representing a difference between a minimum customer pressure 20 and the low-point pressure 32 is shown. The minimum customer pressure 20 may be a minimum statutory pressure or a pressure required to meet consumer service levels, for example. A high excess pressure is undesirable because a higher pressure can increase a likelihood of a higher gas leakage than is necessary to meet the minimum customer pressure 20. A method which utilises machine learning to autonomously and accurately control the gas pressure to satisfy a gas pressure condition at one or more points in a GDN has been proposed in [1], The method disclosed in [1] can operate by controlling the gas pressure at pressure control stations (such as governor stations) such that the gas pressure at one or more points downstream from the pressure control station (e.g. at pressure sensors such as low-point loggers) satisfy a gas pressure condition (e.g. the gas pressure does not fall below a minimum). Nevertheless, there may exist anomalies in the GDN which cause unexpected performance. The detection of anomalies in GDNs represents a technical challenge.

[0044] Examples of anomalies in GDNs include one or more components in the GDN (such as governor stations) may fail to function properly, leaks may develop in one or more of the pipes of the GDN, or substances (such as water) ingress into the pipes etc.

[0045] In respect of substance ingress, water or other substances may find their way into pipes of other infrastructure of the GDN. This may cause a complete or partial blockage in the GDN prevents or reduces the flow of gas beyond the point of ingression.P131560GB 6

[0046] In respect of component failure, GDNs comprise various components which enable the safe and effective delivery of gas to consumers. Generally, these components have a working lifespan which necessitates a service schedule to ensure they maintain operational capability. Nevertheless, components may fail outside of this servicing schedule. Failure can take different modes, such as catastrophic failure or gradual failure. In each case, gas supply through the GDN can be affected.

[0047] The present inventors have recognised there may be unexpected variations in gas pressure due to anomalies in the GDN and that such pressure variations can be used to detect the presence of anomalies in the GDN. Figure 4 illustrates a method performed by a computing device in accordance with example embodiments.

[0048] In step S2, the method comprises receiving, from one or more pressure sensors configured to measure gas pressure at one or more respective locations in a GDN, measured pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a time period.

[0049] The locations of the pressure sensors may be anywhere in the GDN. As an example, one or more of the pressure sensors may be located at governor stations of the GDN, and are thus configured to measure the outlet pressure of the governor stations. One or more of the pressure sensors may be located at pressure “low-points” in the GDN, for example.

[0050] In step S4, the method comprises using a trained machine learning (ML) model to score the measured pressure data over the time period in accordance with a degree to which the measured pressure data for the time period deviates from normal pressure data at the one or more pressure sensors as characterised by the ML model.

[0051] Although reference is made herein to “an ML model”, the ML model may comprise a plurality of ML models. In some embodiments, a plurality of ML models may be configured to score the measured pressure data over the time period and produce a weighted average of the plurality of outputs into a final score.

[0052] The ML model is trained to characterise normal pressure data at the one or more pressure sensors based on previous pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a previous time period.

[0053] The score may be referred to herein as an “anomaly score”.

[0054] “Normal” pressure data may be considered as pressure data which would be measured by the one or more pressure sensors if there were no anomalies in the GDN. Therefore, the previous pressure data used to train the ML model to characterise normal pressure variation is preferably measured when there are no known anomalies in the GDN. For example, the previous pressure data is measured when there are no known leaks in the GDN, there are no known faults in components of the GDN which would cause unexpected pressure behaviour and there is no known substance ingression in the GDN. By training the ML model based on previous pressure data measured (which may be measured when there are no known anomalies in the GDN), the ML model can characterise normal pressure data and thus gain an understanding of what pressure data at the one or more pressure sensors should look like in the absence of anomalies in the GDN. Accordingly, the concept of “normal” pressure data as characterised by the MLP131560GB 7

[0055] reflects what the ML model considers to be normal pressure data based on the previous pressure data.

[0056] A maintenance check may be carried out on the GDN prior to measuring the previous pressure data to reduce the risk that any anomalies are present.

[0057] In some embodiments, the ML model is continuously or periodically trained to characterise normal pressure data based on measured pressure data from the one or more pressure sensors. Thus, the ML model’s understanding of what normal pressure data looks like, is continuously or periodically evolving.

[0058] A characteristic of pressure data may include, for example, characteristic variation in the pressure data over time, characteristic maximum pressure measured by a pressure sensor, characteristic minimum pressure measured by a pressure sensor, characteristic average pressure measured by a pressure sensor, a characteristic pattern in the pressure data over time, and / or a rate of change in the pressure data over time.

[0059] In a particular example, the ML model may learn based on the previous pressure data that a characteristic normal maximum pressure for a pressure sensor is X bar because, for example, the previous pressure data does not exceed X bar for the pressure sensor. In a particular example, the measured pressure is above this maximum characteristic normal maximum pressure, the greater the anomaly score.

[0060] In some embodiments, the greater the degree to which the measured pressure data deviates from normal pressure data as characterised by the ML model, the greater the score. Similarly, the lower the degree to which the measured pressure data deviates from normal pressure data as characterised by the ML model (or the closer the measured pressure data matches the normal pressure data as characterised by the ML model), the lower the score.

[0061] In some embodiments, the deviation of the measured pressure data from the normal pressure data is a deviation in the variation of the measured pressure data with time from a variation of the normal pressure data with time.

[0062] In some embodiments, using the trained ML model to score the measured pressure data over the time period comprises scoring the measured pressure data over the time period separately according to a plurality of types of deviation from the normal pressure data at the one or more pressure sensors.

[0063] Examples of types of deviation from the normal pressure data may comprise one or more of: 1 ) Deviation of the measured pressure data from normal pressure data which is characteristic for a first period, and

[0064] 2) Deviation of the measured pressure data from normal pressure data which is characteristic for a second period, where the second period is longer than the first period.

[0065] The first time period may be a day or a week for example. The second time period may be a month, for example. Therefore, normal pressure data which is characteristic for the first period is normal pressure data which is characteristic of a shorter time period and normal pressure dataP131560GB 8

[0066] which is characteristic for the second period is normal pressure data which is characteristic of a longer time period. Thus, the deviation type 1) is a measure of shorter term pressure deviation whereas deviation type 2) is a measure of longer term pressure deviation. For example, a given fluctuation in pressure data may represent a large short term deviation but a small long term variation (e.g. pressures do not usually fluctuate by that amount on a daily level but do on a monthly level).

[0067] In some embodiments, for deviation type 1) and type 2), the deviation in the measured pressure data is a deviation in an average measured pressure data, where the averaged measured pressure data is an average of the measured pressure data over an interval. The time period over which the measured pressure data is scored may comprise a plurality of intervals. For example, the measured pressure data may be averaged every 6-minutes, and a score is generated for every 6-minutes (representing the deviation of the average pressure from a normal average pressure characterised by the ML model), for the duration of the time period.

[0068] For deviation type 1), the deviation of the measured pressure data from normal pressure data which is characteristic for a first period may be a deviation measured over a time period substantially the same as (or the same as), the first period. For example, the normal pressure data characteristic for a first period may be normal pressure data characteristic for 24 hours and the deviation in the measured pressure data may be the deviation in the measured pressure data over the last 24 hours.

[0069] For deviation type 2), the deviation of the measured pressure data from normal pressure data which is characteristic for a second period may be a deviation measured over a time period substantially the same as (or the same as), the second period. For example, the normal pressure data characteristic for a second period may be normal pressure data characteristic for a month and the deviation in the measured pressure data may be the deviation in the measured pressure data over the last month.

[0070] The deviation types may comprise:

[0071] 3) Deviation in the maximum and / or minimum pressure measured by each pressure sensor in an interval. The time period over which the measured pressure data is scored comprises a one or more (e.g. a plurality of) intervals.

[0072] Deviation type 3) may be represented by the difference between the maximum and minimum pressure over an interval (such as a 6-minute interval). For example, a score is generated every 6-minutes for the duration of the time period, and the score for every 6-minute interval represents a deviation of a difference between the maximum and minimum pressure measured by a pressure sensor during the 6-minute interval from a normal difference between the maximum and minimum pressure for a 6-minute interval as characterised by the ML model. The period over which the ML model is trained may be a rolling window. For example, the ML model may be trained to characterise a normal difference between the maximum and minimum pressure for a 6-minute interval over a rolling period of 7 to 14 days in the past. A high anomaly score may be represented by a maximum and / or minimum pressure for which is changing rapidly and increasing in magnitude, for example.

[0073] In some embodiments, deviation type 1) may comprise a plurality of deviation sub-types. For example, deviation sub-type 1a) may represent deviation in the measured pressure data fromP131560GB 9

[0074] normal pressure data which is characteristic for a first period, where the normal pressure data which is characteristic for the first period is a typical pressure profile pattern for the one or more pressure sensors for the first period (e.g. a day). For example, network pressures are typically raised in the morning and evening when demand increases to ensure continuity of supply. This results in a repeating pattern of two daily peaks for governor pressures. Deviation sub-type 1a) may represent the deviation of the measured data from this pattern. Alternatively or additionally, deviation type 1) may also comprise deviation type 1b) which represents deviation in the measured pressure data from normal pressure data which is characteristic for a first period, where the normal pressure data is pressure data measured for a first period (e.g. a day) in the recent past by the one or more pressure sensors (e.g. the last 24 hours). Alternatively, or additionally, deviation sub-type 1c) may represent deviation in the measured pressure data from normal pressure data which is characteristic for a first period, where the normal pressure data is characterised based on pressure data measured by one or more other pressure sensors in the GDN during the same period as the measured pressure data from the one or more pressure sensors. Deviation type 1) may comprise any combination of one or more of deviation sub-type 1a), 1b) and 1c).

[0075] In some embodiments, at least one ML model may be trained for each deviation type and / or each deviation sub-type. For example, in one embodiment, a separate M L model is trained for deviation type 1a), 1b) 1c), 2) and 3). In this embodiment, the score may be a weighted average of the score produced by one or more of the ML models. In some embodiments, a separate ML model is trained for each of deviation sub-types 1a), 1b) and 1c) and each ML model produces a separate score. The overall score for deviation type 1) may be represented by a weighted average of the deviation sub-types.

[0076] In some embodiments, where each deviation type uses one or more ML models, the overall score may be a weighted average of the score of all deviation types.

[0077] It will be appreciated that the deviation types described above are examples, and other deviation types may be used.

[0078] In step S6, the method comprises comparing the score of the measured pressure data over the time period to a threshold score.

[0079] In some embodiments, comparing the score of the measured pressure data over the time period to a threshold score comprises comparing the score of the measured pressure data over the time period to a plurality of threshold scores. Each of the plurality of threshold scores corresponds to a respective one of the plurality types of deviation from the normal pressure data. In the case of the first deviation type, the score may be a weighted average of the deviation sub-types as discussed above.

[0080] The threshold scores may be preconfigured by a user, for example.

[0081] In step S8, the method comprises determining that the score of the measured pressure data over the time period has exceeded the threshold score. In response to the determining that the score of the measured pressure data over the time period has exceeded the threshold score, the method proceeds to step S10.P131560GB 10

[0082] In some embodiments, determining that the score of the measured pressure data over the time period has exceeded the threshold score comprises determining that at least one of the plurality of threshold scores is exceeded, and the method proceeds to step S10 in response to the determination that at least one of the plurality of threshold scores is exceeded. In some embodiments, the alert indication is only output if more than one (for example, all) of the threshold scores are exceeded.

[0083] In some embodiments, the threshold score which is exceeded provides an indication of a type of the anomaly. For example, a significant deviation in the maximum pressure measured by a sensor in the time period (e.g. the measured pressure falls significant below the characteristic minimum pressure) may indicate that a governor station has failed.

[0084] In step S10, the method comprises outputting an alert indication indicating that an anomaly may have occurred, or may occur, in the GDN.

[0085] In some embodiments, the alert indication explicitly indicates that an anomaly may have occurred (i.e. already occurred). An example of such an alert indication may be: “Analysis of pressure data for this GDN indicates that an anomaly may have occurred in the GDN. Please perform a maintenance check on the GDN”. Certain deviations from normal pressure data as characterised by the ML model may be indicative that an anomaly has already occurred.

[0086] In some embodiments, the alert indication explicitly indicates that an anomaly may occur in the GDN (i.e. may occur in the future). An example of such an alert indication may be: “Analysis of pressure data for this GDN indicates that an anomaly may occur in the GDN. Please perform a maintenance check on the GDN”. Certain deviations from normal pressure data as characterised by the ML model may be indicative that an anomaly may occur in the future. In other words, such deviations may be a precursor to an anomaly occurring and are thus indicative that an anomaly is likely to occur.

[0087] In some embodiments, the alert indication does not explicitly identify whether an anomaly has already occurred or whether an anomaly will occur. An example of such an alert indication may be: “Analysis of pressure data for this GDN indicates that an anomaly may have occurred, or may occur, in the GDN. Please perform a maintenance check on the GDN”.

[0088] In some embodiments, the outputting of the alert indication comprises transmitting the alert indication. In some embodiments, alert indication comprises the measured pressure data. In some embodiments, the alert indication indicates the threshold score which was exceeded. In some embodiments, alert indication indicates a type of the anomaly.

[0089] The alert indication may be transmitted, for example, to an operator of the GDN. This enables the operator to quickly co-ordinate an anomaly correction process. In embodiments where the alert indication does not indicate a type of the anomaly, the operator may launch an investigation into the GDN to identify the type of anomaly. In embodiments where the alert indication indicates the type of anomaly, the operator may send one or more personnel to correct the anomaly. For example, one or more personnel may be sent to correct a leak because the alert indication indicated a leak was present in the GDN. For example, the operator may send out a team to correct a leak in the GDN.P131560GB 11

[0090] The alert indication may be transmitted as a wireless transmission. Examples of transmitting the alert indication by wireless transmission include radio transmission such as Wifi, 5G or radio transmission according to any other wireless communications standard. The alert indication may be broadcast or transmitted to dedicated receivers. The alert indication may be transmitted to computing devices controller by network personnel, for example, who are responsible for checking the GDN for anomalies.

[0091] The anomaly type may be, for example: substance ingress (such as water ingress) in the GDN, a leak in the GDN, or a failure of a component of the GDN (such as failure of a governor station of the GDN, for example). It will be appreciated that these are examples and an anomaly in a GDN may be regarded more generally as a fault or problem in the GDN which causes deviation from expected performance.

[0092] By training an ML model based on previously measured pressure data, the ML model can characterise (and thus gain an understanding of) normal pressure data. By using the trained ML model to score the measured pressure data over the time period in accordance with a degree to which the measured pressure data for the time period deviates from normal pressure data as characterised by the ML model, and outputting an alert indication in response to determining that a threshold score has been exceeded, anomaly detection can be performed more quickly than by manual checks following consumer pressure reports. This means that anomalies can be more quickly corrected in the GDN. Therefore, there is reduced cost to the consumer and network operator. There may also be reduced environmental damage due to leaks in the GDN. The ability to meet KPIs in respect of anomaly response times can also be increased.

[0093] Figure 5 illustrates an example of a pressure profile graph and an anomaly score graph for a GDN in accordance with example embodiments. In particular, Figure 5 shows a pressure graph 520 with three pressure profiles (a first pressure profile 502, a second pressure profile 504, and a third pressure profile 506) for three respective governor stations at different locations of a GDN and an anomaly score graph 530 showing an anomaly score 508 for the third pressure profile 506.

[0094] The first pressure profile 502 represents measured pressure data measured by a first pressure sensor configured to measure an outlet pressure of a first governor station in the GDN. In particular, the first pressure profile 502 shows the average pressure measured by the first pressure sensor in 6 minute intervals between November and December of a given year.

[0095] Similarly, the second pressure profile 504 represents measured pressure data measured by a second pressure sensor configured to measure an outlet pressure of a second governor station in the GDN. In particular, the second pressure profile 504 shows the average pressure measured by the second pressure sensor in 6 minute intervals between November and December of the given year.

[0096] Similarly, the third pressure profile 506 represents measured pressure data measured by a third pressure sensor configured to measure an outlet pressure of a third governor station in the GDN. In particular, the third pressure profile 506 shows the average pressure measured by the second pressure sensor in 6 minute intervals between November and December of the given year. It will be appreciated that an interval of less than or greater than 6 minutes may be used for averaging the pressure data.P131560GB 12

[0097] As shown in Figure 5, the anomaly score graph 530 illustrates an anomaly score 508 for the third pressure profile 506. As will be appreciated from Figure 5, the pattern of pressure variation in the third pressure profile is relatively consistent between the start of November and 21 December. This pattern may therefore be regarded as representing normal pressure data for the third pressure sensor and thus expresses characteristics of normal pressure data measured by the third pressure sensor. In accordance with example embodiments, an ML model may be trained to characterise normal pressure data for the third sensor based on the measured pressure data from the third sensor during a period before 21 December. For example, the ML model may characterise normal pressure data for the third sensor as having an average pressure which oscillates between approximately 24 and 30 mbar with a particular frequency. Since the average pressure is approximately normal until 21 December, the anomaly score 508 remains low until 21 December as shown.

[0098] After 21 December, the third pressure profile 506 exhibits abnormal pressure data. For example, there are spikes in the third pressure profile 506 reaching above 50 mbar. Therefore, after 21 December, there is an increase in the anomaly score, thus indicating an increased likelihood that there is an anomaly in the GDN, or a that an anomaly in the GDN will occur.

[0099] Figure 6 illustrates an example of a pressure profile graph and an anomaly score graph with multiple anomaly score thresholds for a GDN in accordance with example embodiments.

[0100] In Figure 6, a pressure graph 630 is shown for a pressure sensor in a GDN. The pressure graph 630 shows an average pressure profile 634 representing the average pressure measured by the pressure sensor in 6 minute intervals between 24 Feb and 2 March, a minimum pressure profile 636 representing the minimum pressure measured by the pressure sensor in 6 minute intervals between 24 Feb and 2 March and a maximum pressure profile 632 representing the maximum pressure measured by the pressure sensor in 6 minute intervals between 24 Feb and 2 March. In accordance with example embodiments, the ML model may score measured pressure data separately according to a plurality of types of deviation from the normal data at the one or more pressure sensors. In Figure 6, there are three different types of deviation from the normal pressure data corresponding to deviation types 1), 2) and 3) explained above.

[0101] In Figure 6, there is shown a first anomaly score 608 for deviation type 1) between 24 Feb and 2 March, a second anomaly score 612 for deviation type 2) between 24 Feb and 2 March and a third anomaly score 610 for deviation type 3) between 24 Feb and 2 March.

[0102] In Figure 6, a higher anomaly score means the higher the likelihood that an anomaly has occurred or will occur in the GDN. In the anomaly graph 620 shown in Figure 6, the highest shown anomaly score is 10 and the lowest shown anomaly score is -6.

[0103] Also shown in Figure 6 is a first anomaly threshold score 604 for deviation type 1), a second anomaly threshold score 602 for deviation type 2) and a third anomaly threshold score 606 for deviation type 3). In accordance with example embodiments, the first anomaly score 608 is compared with the first anomaly threshold score 604 to determine if the first anomaly score 608 exceeds the first anomaly threshold score 604, the second anomaly score 612 is compared with the second anomaly threshold score 602 to determine if the second anomaly score 612 exceeds the second anomaly threshold score 602, and the third anomaly score 610 is compared with the third anomaly threshold score 606 to determine if the third anomaly score 610 exceeds the third anomaly threshold score 606. In the example shown in Figure 6, none of the threshold scores isP131560GB 13

[0104] exceeded. It is therefore determined that an anomaly has not occurred in the GDN. In some embodiments, the first anomaly score 608 is a weighted average of the anomaly score according to each of deviation sub-type 1a), 1b) and 1c).

[0105] As shown in Figure 6, the first anomaly score 608 and the second anomaly threshold score 602 remain relative constant between 24 Feb and 2 March. This is indicative that average pressure does not vary significantly from normal average pressure as characterised by the ML model. However, the third anomaly score 610 increases around the end of February I start of March, indicating that the difference between the maximum and minimum pressure data is deviating from the normal difference between the maximum and minimum pressure data as characterised by the ML model.

[0106] Combining Pressure Data from Different Pressure Sensors

[0107] In some embodiments, the ML model may be trained to characterise normal pressure data for at least one of the pressure sensors based on previous pressure data measured over a time period by the at least one pressure sensor and based on previous pressure data measured over the time period from at least one other pressure sensor. For example, there may be two pressure sensors in a GDN configured to measure the outlet pressure of two governor stations in the GDN. In some embodiments, the ML model is trained to characterise normal pressure variation for at least one of the governor stations based on previous pressure data received over a time period from the pressure sensor at each governor station. For example, the ML model may detect a similar pressure oscillation pattern in the pressure data from both pressure sensors and determines that this oscillation pattern is a characteristic of normal pressure data at the sensors. By training the ML model in this way, deviations in pressure data from pressure sensors at different locations in the GDN can be used to determine whether or not an anomaly may have occurred, or may occur, in the GDN. For example, if the pressured data measured at one governor station has a significantly different characteristic to pressure data measured at another governor station, this may be an indication that an anomaly has occurred, or is likely to occur, in the GDN.

[0108] Determining Anomaly Type

[0109] In some embodiments, the ML model may be additionally trained based on anomalous pressure training data to characterise anomalous pressure data. The anomalous pressure training data may comprise a plurality of gas pressure measurements performed by each of the one or more pressure sensors over a time period while a known anomaly was present in the GDN.

[0110] In some embodiments, the anomalous pressure training data may comprise a plurality of sets of anomalous pressure training data where each set of anomalous pressure training data comprises anomalous pressure training data measured when a different anomaly type was present in the GDN. For example, a first set of anomalous pressure training data may be measured when a substance ingression was known to exist in the GDN, a second set of anomalous pressure training data may be measured when a leak was known to be present in the GDN, a third set of anomalous pressure training data may be measured when there was a known component failure in the GDN (e.g. failure of a governor station). For example, by checking the GDN in response to the alert indication or otherwise, an operator may determine that a specific type of anomaly was present in the GDN. The pressure data for when the specific type of anomaly was present can then be used as anomalous pressure training data. In some embodiments, an investigation may be conducted after an alert indication has been transmitted to find out whether or not a true anomaly was present in the GDN. For example, the network operators and / or consumers of the GDN may be asked whether a true anomaly was present. If a true anomaly was present, then pressure dataP131560GB 14

[0111] for when the specific type of anomaly was present can then be used as anomalous pressure training data.

[0112] Thus, the ML algorithm may be trained based on the anomalous pressure training data to characterise anomalous pressure data for a plurality of anomaly types.

[0113] Therefore, the computing device may be configured to determine the anomaly type based on a comparison of the measured pressure data and the anomalous pressure data for each anomaly type as characterised by the ML model. For example, the computing device may score the measured pressure data in accordance with a degree to which the measured pressure data matches the anomalous pressure data for each anomaly type as characterised by the ML model. Then, the computing device may determine that the anomaly type which may have occurred, or which may occur, in the GDN is the anomaly type for the anomalous pressure data as characterised by the ML model which most closely matches the measured pressure data (e.g. the anomaly type with the highest score). The anomaly type may be transmitted in the alert indication. In some embodiments, the anomaly type is transmitted in a separate signal to the alert indication. Thus example embodiments can provide a way of detecting the anomaly type. This means that anomaly correction can be more quickly implemented since the anomaly type is known without manual investigation. In some embodiments, the alert indication may indicate the anomaly type.

[0114] In respect of substance ingress, water or other substances may find their way into pipes of other infrastructure of the GDN. This may cause a complete or partial blockage in the GDN prevents or reduces the flow of gas beyond the point of ingression.

[0115] In respect of component failure, GDNs comprise various components which enable the safe and effective delivery of gas to consumers. Generally, these components have a working lifespan which necessitates a service schedule to ensure they maintain operational capability. Nevertheless, components may fail outside of this servicing schedule. Failure can take different modes, such as catastrophic failure or gradual failure. In each case, gas supply through the GDN can be affected.

[0116] As explained above, an example of an anomaly type is substance ingress (such as water ingress). The present inventors have recognised that substance ingress (and in particular water ingress) can cause characteristic pressure data in a GDN. For example, substance ingression causes a restriction in gas flow which may cause an increase in pressure upstream of the restriction and / or characteristic fluctuations in gas pressure. Furthermore, water ingression (and liquid ingression more generally) can cause oscillations in pressure over time with characteristic magnitude and frequency. In particular, the water may settle at a physical low point in the GDN. This will, in turn, cause a constriction to the gas flow, but because of the dynamic nature of water, the constriction can vary in size as the gas pushes the water. This in turn causes characteristic oscillations in the gas pressure as water is pushed by the gas and returns to a position where it settles under gravity, he ML model can learn such characteristics in pressure data caused by substance ingression and identify these characteristics in measured pressure data to detect a substance ingression.

[0117] As explained above, another example of an anomaly type is the failure of GDN components. The present inventors have recognised that component failure (and in particular failure of a governor station) cause characteristic pressure data in a GDN. For example, a failure of a governor station may lead to sharp drop in the outlet pressure (e.g. a predetermined drop in outlet pressure) of the governor station over a short time period (e.g. 3 to 24 hours). The ML model can learn suchP131560GB 15

[0118] characteristics in pressure data caused by component failure and identify these characteristics in measured pressure data to detect component failure.

[0119] Although embodiments have been described above where the anomalous pressure training data comprises a plurality of gas pressure measurements performed by each of the one or more pressure sensors over a time period while a known anomaly was present in the GDN, it will be appreciated that the above embodiments are equally applicable when anomalous pressure training data comprises a plurality of gas pressure measurements performed by each of the one or more pressure sensors over a time period preceding a time during which a known anomaly was present in the GDN. In such embodiments, the anomalous pressure data as characterised by the ML model comprise characteristics of pressure data which precedes (and is therefore a precursor to) an anomaly occurring. For example, certain pressure patterns may indicate that an anomaly is about to occur in the GDN.

[0120] Computing Device

[0121] Figure 7 is a schematic diagram illustrating an example of a computing device 202 communicating with a GDN. As shown in Figure 7, the computing device 202 is configured to receive measured pressure data from one or more pressure sensors. In particular, the computing device 202 is configured to receive measured pressure data from a pressure sensor configured to measure an outlet pressure of a governor station via a controller 206 of the governor station as shown. Similarly, in the example of Figure 7, the computing device 202 is configured to receive measured pressure data from a low pressure data logger 214 which is an example of a pressure sensor. A schematic representation of the computing device 202 is shown in Figure 8. As shown in Figure 8, the computing device 202 comprises circuitry such as controller circuitry 206 and communications circuitry 204. The communications circuitry 206 may comprise transmitter circuitry and receiver circuitry. Therefore, the communications circuitry may 206 may be configured to transmit and receive signals. The transmitter circuitry and the receiver circuitry may include radio frequency filters and amplifiers as well as signal processing components and devices in order to transmit and receive radio signals in accordance for example with the 5G / NR standard, Wi-Fi or any other wireless communications standard. The controller circuitry 206 may be configured to control the communications circuitry 206. The controller circuitry 206 may be, for example, a microprocessor, a CPU, or a dedicated chipset, etc., configured to carry out instructions which are stored on a computer readable medium, such as a non-volatile memory. The processing steps described herein may be carried out by, for example, a microprocessor in conjunction with a random access memory, operating according to instructions stored on a computer readable medium. The transmitters, the receivers and the controllers are described separately for ease of explanation. However, it will be appreciated that the functionality of these elements can be provided in various different ways, for example using one or more suitably programmed programmable computer(s), or one or more suitably configured application-specific integrated circuit(s) I circuitry I chip(s) I chipset(s).

[0122] References herein to a computing device 202 performing an operation, or being configured to perform an operation, should be understood as also disclosing that the circuitry of the computing device 202 is configured to perform that operation. For example, the controller circuitry 206 may be configured in combination with the communications circuitry 204 to perform that operation. ML Model to Characterise Normal Pressure DataP131560GB 16

[0123] In some embodiments, the ML model may implement one or more of a supervised learning model, an autoencoder model and a reinforcement learning model, though it will be appreciated, that other Al models may be used.

[0124]

[0125] In some embodiments, the ML model may implement a supervised ML model.

[0126] The supervised learning model is trained using labelled training data to learn a function that maps inputs (typically provided as feature vectors) to outputs (i.e. labels). The labelled training data comprises pairs of inputs and corresponding output labels. The output labels are typically provided by an operator to indicate the desired output for each input. The supervised learning model processes the training data to produce an inferred function that can be used to map new (i.e. unseen) inputs to a label.

[0127] The input data (during training and / or inference) may comprise various types of data, such as numerical values, images, video, text, or audio. Raw input data may be pre-processed to obtain an appropriate feature vector used as input to the model - for example, features of an image or text input may be extracted to obtain a corresponding feature vector. It will be appreciated that the type of input data and techniques for pre-processing of the data (if required) may be selected based on the specific task the supervised learning model is used for.

[0128] Once prepared, the labelled training data set is used to train the supervised learning model. During training the model adjusts its internal parameters (e.g. weights) so as to optimize (e.g. minimize) an error function, aiming to minimize the discrepancy between the model’s predicted outputs and the labels provided as part of the training data. In some cases, the error function may include a regularization penalty to reduce overfitting of the model to the training data set. The supervised learning model may use one or more machine learning algorithms in order to learn a mapping between its inputs and outputs. Example suitable learning algorithms include linear regression, logistic regression, artificial neural networks, decision trees, support vector machines (SVM), random forests, and the K-nearest neighbour algorithm.

[0129] In accordance with example embodiments, the input training data to the supervised learning model may comprise normal pressure data (i.e. pressure data from the one or more sensors when no known anomalies were present in the GDN) and the output training data from the supervised learning model may be characteristics of the normal pressure data.

[0130] Once trained, the supervised learning model may be used for inference - i.e. for predicting outputs for previously unseen input data. The supervised learning model may perform classification and / or regression tasks. In a classification task, the supervised learning model predicts discrete class labels for input data, and / or assigns the input data into predetermined categories. In a regression task, the supervised learning model predicts labels that are continuous values.

[0131] For example, in accordance with example embodiments, the input data to the supervised learning model may comprise normal pressure data (i.e. pressure data from the one or more sensors when no known anomalies were present in the GDN) and the output data from the supervised learning model may be characteristics of the normal pressure data.P131560GB 17

[0132] In some cases, limited amounts of labelled data may be available for training of the model (e.g. because labelling of the data is expensive or impractical). In such cases, the supervised learning model may be extended to further use unlabelled data and / or to generate labelled data. Considering using unlabelled data, the training data may comprise both labelled and unlabelled training data, and semi-supervised learning may be used to learn a mapping between the model’s inputs and outputs. For example, a graph-based method such as Laplacian regularization may be used to extend a SVM algorithm to Laplacian SVM in order to perform semi-supervised learning on the partially labelled training data.

[0133] Considering generating labelled data, an active learning model may be used in which the model actively queries an information source (such as a user, or operator) to label data points with the desired outputs. Labels are typically requested for only a subset of the training data set thus reducing the amount of labelling required as compared to fully supervised learning. The model may choose the examples for which labels are requested - for example, the model may request labels for data points that would most change the current model, or that would most reduce the model's generalization error. Semi-supervised learning algorithms may then be used to train the model based on the partially labelled data set.

[0134] Autoencoders

[0135] In some embodiments, the ML model may implement Autoencoding.

[0136] An autoencoder is a type of an unsupervised machine learning model that uses one or more artificial neural networks to learn an efficient representation of unlabelled input data. The autoencoder may be used to encode various types of data, such as images, video, text, audio, or positioning measurements.

[0137] The autoencoder may comprise an encoder neural network that encodes input data into a reduced representation (also called a “latent space”), and a decoder neural network that aims to recreate the input data from the encoded reduced representation. The latent space is typically of a lower-dimension than the input data - thus, the latent space generated by the encoder typically provides a more efficient, compressed representation of the input data that requires less memory storage than the original input data.

[0138] The encoder neural network may comprise one or more layers that transform input data into a reduced representation. The encoder neural network receives input data, and the final layer of the encoder neural network outputs a reduced representation of the input data, i.e. a latent space (also termed a “bottleneck layer”).

[0139] The decoder neural network comprises one or more layers that transform data from the latent space into output data of the same dimensionality as the data input to the encoder. The decoder aims to reconstruct the data originally input to the encoder neural network from the latent space representation of the data.

[0140] The encoder and / or decoder neural networks typically comprise a plurality of hidden layers. For example, an encoder may comprise a plurality of hidden layers that progressively extract further reduced representations of the input data. Using deeper neural networks (i.e. with a higher number of hidden layers) for the encoder and / or the decoder may improve performance of the autoencoder, and in some cases may reduce the amount of training data that is required.P131560GB 18

[0141] The encoder and decoder neural networks are typically trained together. During training the autoencoder may adjust its internal parameters (e.g. weights and biases of the encoder and decoder neural networks) so as to optimize (e.g. minimize) a loss / error function, aiming to minimize discrepancy between the data input to the encoder and the output reconstructed data generated by the decoder. It will be appreciated that the specific loss function, and algorithm used to optimize the function may vary depending on the nature of the autoencoder model, and its intended application. In an example, a mean squared error loss function optimized using gradient descent may be used. In some cases, a sparse autoencoder may be used in order to promote sparsity of the latent representation (as compared to the input) and to prevent the autoencoder from learning the identity function - for example, a sparse autoencoder may be implemented by modifying the loss function to include a sparsity regularization penalty.

[0142] In some cases, the autoencoder may be a Variational Autoencoder (VAE). The VAE is a specific type of auto-encoder in which a probability model is imposed on the encoded representation by the training process (in that deviations from the probability model are penalised by the training process). The VAE may be used for generative artificial intelligence applications to generate new output data which exhibits similar characteristics to the input encoded data by sampling from the learned latent space.

[0143] In accordance with example embodiments, the input data may comprise normal pressure data (i.e. pressure data from the one or more sensors when no known anomalies were present in the GDN) and the output data may be characteristics of the normal pressure data.

[0144] Reinforcement learning

[0145] In some embodiments, the ML model may implement reinforcement learning (RL).

[0146] Reinforcement learning is a type of machine learning directed to training an artificial intelligence agent to take actions in an environment that maximize the notion of a cumulative reward. During reinforcement learning, the agent interacts with the environment, and learns from the results of its actions, thus allowing the agent to progressively improve its decision-making.

[0147] An RL model typically comprises an action-reward feedback loop. The feedback loop comprises: an environment, state, agent, policy, action, and reward. The environment is the system with which the agent interacts and in which the agent operates - for example, the environment may be a virtual environment of a video game. The state represents the current conditions in the environment. The agent receives the state as an input and takes an action which may affect the environment and change the state of the environment. The agent takes the action based on its policy which is a mapping from states of the environment to actions of the agent. The policy may be deterministic or stochastic. The reward represents feedback from the environment to the action taken by the agent. The reward provides an indication (typically in the form of a numerical value) of the desirability of the result of the agent’s action. The reward may comprise positive signals to reward desirable behaviour of the agent and / or negative signals to penalize undesirable behaviour of the agent.

[0148] Through multiple iterations of action-reward feedback loop, the agent aims to maximise the total cumulative reward it receives, thus learning how to take optimal actions in the environment. The reinforcement learning process thus allows the agent to learn an optimal policy that maximizes the cumulative reward. The cumulative award may be estimated using a value function whichP131560GB 19

[0149] estimates the expected return starting from a given state or from a given state and action. Using the cumulative reward in the reinforcement learning process allows the agent to consider longterm effects of its policy.

[0150] A reinforcement learning algorithm may be used to refine the agent’s policy and the value function over iterations of the action-reward feedback loop. The learning algorithm may rely on a model of the environment (e.g. based on Markov Decision Processes (MDPs)) or be model-free. Example suitable model-free reinforcement learning algorithms include Q-learning, SARSA (State-Action-Reward-State-Action), Deep Q-Networks (DQNs), or Deep Deterministic Policy Gradient (DDPG).

[0151] It will be appreciated that the agent will typically engage in both exploration and exploitation of the environment in which it operates. In exploration, the agent takes typically random actions to gather information about the environment and identify potentially desirable actions (i.e. actions that maximise cumulative reward). In exploitation, the agent takes actions that are expected to maximise reward (e.g. by selecting the action based on the agent’s latest policy). Various techniques may be used to control the proportion of explorative and exploitative actions taken by the agent - for example, a predetermined probability of taking an explorative action in a given iteration of the feedback loop may be set (and optionally reduced over time to allow the agent to shifts more towards exploitation over time to maximise cumulative reward in view of diminishing returns for further exploration).

[0152] In some cases, the RL model may be configured to learn from feedback provided by a user. Utilising user feedback in this way may allow the agent to improve its choice of actions and better align with user preferences. For example, reinforcement learning from human feedback (RLHF) techniques may be used. RLHF includes training a reward model based on user feedback and using this model for determining the reward in the reinforcement learning process described above. The user feedback may be received in various forms depending on the specific reinforcement learning problem being solved - for example, the feedback may be received in the form of a user ranking of instances of the agent’s actions. RLHF thus allows incorporating user feedback into the reinforcement learning process. RLHF approaches may be advantageous where it is easier for a user than for an algorithm to assess the quality of the machine learning model’s output (e.g. for generative artificial intelligence RL models).

[0153] In accordance with example embodiments, the input data may comprise normal pressure data (i.e. pressure data from the one or more sensors when no known anomalies were present in the GDN) and the output data may be characteristics of the normal pressure data.

[0154] Detailed Examples

[0155] As explained above, deviation type 1) may comprise a plurality of deviation sub-types where a separate ML model is used to score each deviation sub-type. For example, deviation sub-type 1a) may represent deviation in the measured pressure data from normal pressure data which is characteristic for a first period, where the normal pressure data which is characteristic for the first period is a typical pressure profile pattern for the one or more pressure sensors for the first period (e.g. a day). For example, network pressures are typically raised in the morning and evening when demand increases to ensure continuity of supply. This results in a repeating pattern of two daily peaks for governor pressures. Deviation sub-type 1a) may represent the deviation of the measured pressure data from this pattern. Deviation type 1b) may representP131560GB 20

[0156] deviation in the measured pressure data from normal pressure data which is characteristic for a first period, where the normal pressure data is pressure data measured for a first period (e.g. a day) in the recent past by the one or more pressure sensors (e.g. the last 24 hours). Deviation sub-type 1c) may represent deviation in the measured pressure data from normal pressure data which is characteristic for a first period, where the normal pressure data is characterised based on pressure data measured by one or more other pressure sensors in the GDN during the same period as the measured pressure data from the one or more pressure sensors.

[0157] In one example, each ML model creates an anomaly score based on taking measured pressure data for a governor over the preceding 24 hours (the shortest period of time with repeating “seasonal” behaviour comprising 240 records based on 6-minute sampling). Then, the ML model for sub-type 1a) generates a first score by comparing the measured pressure data from a governor to a typical pressure profile pattern over a day for the governor. The ML model for subtype 1b) generates a second score by comparing the measured pressure data from the governor with measured pressure data from a day in the recent past of the governor and the ML model for sub-type 1c) generates a third score by comparing the measured pressure data from the governor to the pressure data from one or more other governors in the GDN for the same day.

[0158] Each ML model may be applied globally; that is, these techniques and their common elements may be universally applied to every GDN and governor (as opposed to individual ML models being developed per GDN or governor).

[0159] The comparison may be performed in three stages:

[0160] 1. Reduction in the dimensionality of the measured pressure data using a dimensionality reduction algorithm known as an autoencoder

[0161] 2. Calculation of a distance between

[0162] a. The measured pressure profile and a reconstruction of the low dimensional encoded representation

[0163] b. The encoded representation for the current 24-hour period and a prior 24-hour period

[0164] c. The encoded representation for the current governor and one or more other governors on the same GDN for the current 24 hour period.

[0165] 3. Z-scoring each distance for a) b) and c).

[0166] Z-scoring modifies a value by subtracting the mean and dividing by the standard deviation. The resulting values have a mean of zero and standard deviation of one. The magnitude of the value corresponds to how unusual the value is and provides a consistent basis for comparison across anomaly scores.

[0167] A weighted average of the Z-scores can then be created to produce a final view of the 24-hour period. An anomaly is considered to be presented if the weighted average exceeds a threshold. The above-approach has the following advantages:P131560GB 21

[0168] • The approach enables the detection of multiple different signals in pressure data that indicate changes in expected behaviour

[0169] • The models are lightweight (in terms of compute) and fast to run (typically under a minute during development), meaning it will not be expensive to run multiple times a day • The models are global (i.e. not specific to a single network) and so can be applied to new networks with relatively little historic data as long as demand and pressure control is similar to networks already seen

[0170] • A basic rationale can be provided indicating the anomaly detected was flagged due to abnormal behaviour in the previous 24 hours

[0171] • The sensitivity of the ML models can be tuned by changing a threshold associated with the anomaly score.

[0172] Data

[0173] The measured pressure data received from the one or more pressure sensors may be represented by a time series. The time series data may be processed into a tabular data set. For example, create a pressure dataset may be created where each row is the pressure data from a single pressure sensor for a given day, and each column is a 6-minute window covering a 24 period (resulting in 240 columns). Each 6 minute window is treated as an input feature in a tabular dataset that can be ingested by the model.

[0174] Once processed into this tabular form, the raw data may then be scaled using the min and max values seen during training in that 6 minute window to bring all values into the range of zero to one. It will be appreciated that other methods of scaling the data (such as “standard scaling”) may be used.

[0175] Dimensionality Reduction

[0176] Due to the large number of records involved in each 24 hour period, it is beneficial to perform dimensionality reduction on the pressure data prior to comparing these time periods.

[0177] Dimensionality reduction methods can be used to filter out noise from the data and retain the key information needed to make the comparison.

[0178] The ML model used to perform dimensionality reduction may be an autoencoder. The autoencoder may be trained to pass the input data through an “information bottleneck” (i.e. a low dimensionality representation) and then recreate the input data as closely as possible. This is achieved by using two models in tandem - an “encoder”, that takes in the original data and outputs a low dimensional representation, and a “decoder”, that takes the low dimensional representation and outputs an approximation of the real data.

[0179] Once trained, these models perform well at reconstructing data that they have seen frequently in the training data (such as “normal” pressure patterns). A useful feature of autoencoders is that the lower dimensional representation of infrequently seen pressure patterns will often be very different from the representations of more commonly seen patterns, thereby facilitating the detection of anomalies.

[0180] The autoencoder may be constructed with a relatively simple structure with one hidden layer on the encoder and decoder to limit computational costs. The activation functions and number ofP131560GB 22

[0181] nodes in each layer were chosen through experimentation and can be configured during training to enable continued experimentation.

[0182] An example of an autoencoder demonstrating the passage of a data via an encoder neural network through a low dimensional bottleneck before reconstructing the data in a decoder neural network is shown in Figure 9.

[0183] Deviation Sub-Type 1a)

[0184] As explained above, deviation sub-type 1a) may represent deviation in the measured pressure data from normal pressure data which is characteristic for a first period, where the normal pressure data which is characteristic for the first period is a typical pressure profile pattern for the one or more pressure sensors for the first period (e.g. a day). In a particular example, the deviation is represented by the distance between the measured pressure profile pattern and a reconstruction of the pressure profile as reconstructed after passing through the autoencoder. Large errors in the reconstruction of the data by the autoencoder can indicate that a pressure profile is anomalous. This is because anomalous data points will be very infrequent in the training data, and therefore the autoencoder will not have been optimised to minimise reconstruction error for these pressure profiles. To calculate the reconstruction error, the measured pressure profile and reconstructed pressure profile are compared and the difference calculated. During inference, the raw score is then z-scored to express how unusual the observed reconstruction error is compared to a typical pressure profile pattern.

[0185] This metric is quick to calculate as it utilises information that has already been modelled.

[0186] Deviation Sub-Type 1b)

[0187] As explained above, deviation type 1b) may represent deviation in the measured pressure data from normal pressure data which is characteristic for a first period, where the normal pressure data is pressure data measured for a first period (e.g. a day) in the recent past by the one or more pressure sensors (e.g. the last 24 hours). In an example, the deviation may be measured using the distance in low dimensional space from the measured pressure profile to one in the past for the same governor station. Different distance metrics may be used to quantify the deviation. A particular beneficial metric is the Manhattan distance

[0188] As with the reconstruction error, the raw distance metrics are z-scored using the mean and standard deviation observed in historical data to express how unusual the observed distance is. The model may be modified to follow a “stateless” design in which the mean and standard deviation are continuously calculated from a defined reference period loaded for every inference. The time between the current observation and the past point being used for comparison is configurable and experimentation suggested that the weighted combination of multiple periods would be beneficial. Additionally, the reference period is configurable and can be a defined period or expand to include as much historical data as is available. In practice, when an expanding reference period is used the duration will be defined by the duration of data loaded.

[0189] Deviation Sub-'P131560GB 23

[0190] As explained above, deviation sub-type 1c) may represent deviation in the measured pressure data from normal pressure data which is characteristic for a first period, where the normal pressure data is characterised based on pressure data measured by one or more other pressure sensors in the GDN during the same period (e.g. a day) as the measured pressure data from the one or more pressure sensors. Deviations in behaviour compared to other governors in a network are implemented in a similar way to comparing a governor to its past as explained above. However, for deviation sub-type 1c), a past pressure data for a governor may be switched for the current pressure data for another governor in the GDN. The raw distance metric between governors may be z-scored with the mean and standard deviation of the distance seen between the governors during a reference period and then the score may be aggregated across all governors in the network.

[0191] The relatedness between governors will be different depending on the governors being considered. Governors closely connected in the network will likely have similar pressure patterns and this relatedness will be evident by a very stable distance in the low dimensional space. By z-scoring the distance to each governor before aggregating, the impact of a change in the relationship between closely-related governors will be larger (and therefore so will the anomaly score) than for governors that are unrelated and have very variable distances in the low dimensional space.

[0192] Aggregation across all governors enables anomalies at specific governors to be identified rather than being limited to pairs of governors. For example, if the distance between governors A and B is unusually large it is unclear whether A or B are anomalous. However, if the distance between governors A and C is also unusually large and the distance between B and C is in the normal range it is likely that governor A has an anomaly.

[0193] Overall Deviation for Deviation Type 1

[0194] The models described above may produce an overall anomaly score by taking the sum of weighted values from each individual anomaly score. The weights may be chosen based on performance using a set of labelled examples from the training data. These weights can also be tuned during field trials on the basis of which alerts users are interested in.

[0195] In the event of missing data it is possible some components of the anomaly score may be missing. For example, missing historical data would impact the temporal and interstation distance anomaly score components but not the reconstruction error score. If one of the scores cannot be generated the weights for the remaining anomaly score components may be rebalanced to compensate.

[0196] Deviation type 3)

[0197] As explained above, deviation in the maximum and / or minimum pressure measured by each pressure sensor in an interval. The time period over which the measured pressure data is scored comprises a one or more (e.g. a plurality of) intervals.

[0198] The min-max spread is related to the range in pressures observed within an interval (such as a 6 minute window) at a governor station. The min-max spread may be calculated directly from the recorded data without dimensionality reduction. The anomaly score to detect changes in theP131560GB 24

[0199] min-max spread identifies governors for which the variability of pressure in the network is changing rapidly and is increasing in magnitude.

[0200] During development of the model for deviation type 3), it was noted that min-max pressure patterns differ significantly between governors, which may be caused by turbulence in the pipes at the location of the sensor. The min-max spread may also be related to the mean pressure for many stations, with higher mean pressures relating to greater min-max spreads.

[0201] To overcome these issues the min-max spread may be first normalised by the mean pressure for the interval and then the model is implemented in a rolling fashion such that the min-max spread in the interval is compared, by z-scoring, to a configurable period in the past. This configurable period may be a rolling period. For example, the model was trained to characterise a normal max-min spread for a 6-minute interval over a period of 7 to 14 days in the past.

[0202] The use of this modelling approach includes the following advantages:

[0203] • This is a light lightweight approach which means models do not have to be governor specific and a relatively small amount of training data is required (e.g. 7 to 14 days) • The model is easy and cheap to run inference with which helps to reduce costs Deviation Type 2)

[0204] As explained above, deviation type 2) may represent the deviation of the measured pressure data from normal pressure data which is characteristic for a second period, where the second period is longer than the first period. The second period may be a month or year, for example, and therefore the deviation type 2) represents gradual deviations in pressure behaviour over time.

[0205] As pressures will naturally change over the course of a year, the anomaly score to detect longer term changes in pressure may operate selecting a low point in the pressure during each day, and then calculating the trend in these pressures over a configurable number of days.

[0206] Exploratory data analysis indicated that the mean pressure during the hours of 10pm, midnight and 2am had reasonably low seasonality and were chosen for use in the anomaly score creating three scores based on pressures observed at 10pm-11pm, 12am-1am, and 2am-3am. As shown in Figure 10, the mean pressure during selected hours of the day is shown across the year to understand the seasonality in the pressure. Line 914 represents the variation of mean pressure during each month at 0 hr; Line 912 represents the variation of mean pressure during each month at 1 hr; Line 908 represents the variation of mean pressure during each month at 2 hr; Line 904 represents the variation of mean pressure during each month at 8 hr; Line 902 represents the variation of mean pressure during each month at 17 hr; Line 906 represents the variation of mean pressure during each month at 22 hr; and Line 910 represents the variation of mean pressure during each month at 23 hr. From month 5 (May) to month 10 (October) pressures are low and stable for all times of the day. From month 11 (November) to month 4 (April), pressures at 8am and 5pm are higher indicating the seasonality in demand at these times. Pressures between 10pm and 2am are less impacted by the time of year.

[0207] The gradient in the pressure was calculated for each 90-day period and z-scored based on how likely it would be observed by chance. The final anomaly score was then calculated by combining the three scores based on the data from 10pm, 12am, and 2am. This approach mayP131560GB 25

[0208] be tuned based on performance using a set of manually labelled examples from the training data.

[0209] This approach includes the following advantages:

[0210] • Models are extremely lightweight

[0211] • Can be applied to new networks (after 90 days of data has been collected)

[0212] • Require no retraining, keeping costs low.

[0213] Alert generation and explainability

[0214] Each anomaly score may produce values that are centred around zero due to the application of z-scoring. It is assumed that only the highest scores are anomalous. The interpretation of a high score is:

[0215] • Deviation type 1) - General anomaly: Given the combination of three separate scores for each deviation sub-type one or more of the following must be true

[0216] - The autoencoder has struggled to represent the measured pressure pattern and the reconstruction error is larger than typically observed

[0217] - The pressure pattern observed for the current day is more dissimilar to the pattern observed 3 hours, 2 days, or 7 days ago than is typically observed

[0218] - The pressure pattern observed for this governor is more dissimilar to other governors on the network than is typically observed

[0219] • Deviation Type 2) - Gradual change: The slope in the overnight pressures for this governor is higher or more positive than would be expected by chance

[0220] • Deviation Type 3) - Minmax: The difference between the high and low pressures observed at this governor is greater than has typically been observed

[0221] A consequence of how the scores are calculated is the possibility of negative values. As with high scores, low negative values suggest the pattern observed is unusual but due to less than typical differences. Specifically:

[0222] • Deviation Type 1) - General anomaly: Given the combination of three separate scores for each deviation sub-type one or more of the following must be true

[0223] - The autoencoder has been able to represent the measured pattern very well and the reconstruction error is smaller than is typically observed.

[0224] - The pressure pattern observed for the current day is more similar to the pattern observed 3 hours, 2 days, or 7 days ago than is typically observed

[0225] - The pressure pattern observed for this governor is more similar to other governors on the network than is typically observed

[0226] • Deviation Type 2) - Minmax: The difference between the high and low pressures observed for this governor is less than has typically been observedP131560GB 26

[0227] • Deviation Type 3) - Gradual change: The slope in the overnight pressures for this governor is lower or more negative than would be expected by chance

[0228] An alert indication may be output if the threshold anomaly score for at least one of Deviation Type 1), 2) or 3) is exceeded.

[0229] In some embodiments, the alert indication may comprise the anomaly score which caused the alert indication to be output. For example, brief text description may be presented to the user to explain why the alert was generated. Examples of the descriptions to be generated might include, “abnormal pressure pattern seen in previous 24 hours”, “change in pressure variability compared to 7 days ago” or “increase in pressure observed over previous 90-days.” Described embodiments may be implemented in any suitable form including hardware, software, firmware or any combination of these. Described embodiments may optionally be implemented at least partly as computer software running on one or more data processors and / or digital signal processors. The elements and components of any embodiment may be physically, functionally and logically implemented in any suitable way. Indeed the functionality may be implemented in a single unit, in a plurality of units or as part of other functional units. As such, the disclosed embodiments may be implemented in a single unit or may be physically and functionally distributed between different units, circuitry and / or processors.

[0230] Although the present disclosure has been described in connection with some embodiments, it is not intended to be limited to the specific form set forth herein. Additionally, although a feature may appear to be described in connection with particular embodiments, one skilled in the art would recognise that various features of the described embodiments may be combined in any manner suitable to implement the technique.

[0231] Further examples of feature combinations taught by the present disclosure are set out in the following numbered paragraphs:

[0232] Paragraph 1. A method performed by a computing device, the method comprising receiving, from one or more pressure sensors configured to measure gas pressure at one or more respective locations in a Gas Distribution Network (GDN), measured pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a time period,

[0233] using a trained machine learning (ML) model to score the measured pressure data over the time period in accordance with a degree to which the measured pressure data for the time period deviates from normal pressure data at the one or more pressure sensors as characterised by the ML model, wherein the ML model is trained to characterise normal pressure data at the one or more pressure sensors based on previous pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a previous time period, comparing the score of the measured pressure data over the time period to a threshold score,

[0234] determining that the score of the measured pressure data over the time period has exceeded the threshold score, and in response,

[0235] outputting an alert indication indicating that an anomaly may have occurred, or may occur, in the GDN.P131560GB 27

[0236] Paragraph 2. A method according to paragraph 1, wherein the outputting the alert indication comprises transmitting the alert indication.

[0237] Paragraph 3. A method according to paragraph 1 or paragraph 2, wherein

[0238] using the trained ML model to score the measured pressure data over the time period comprises scoring the measured pressure data over the time period separately according to a plurality of types of deviation from the normal pressure data,

[0239] comparing the score of the measured pressure data over the time period to a threshold score comprises comparing the score of the measured pressure data over the time period to a plurality of threshold scores, wherein each of the plurality of threshold scores corresponds to a respective one of the plurality types of deviation from the normal pressure data, determining that the score of the measured pressure data over the time period has exceeded the threshold score comprises determining that at least one of the plurality of threshold scores is exceeded, and the alert indication is output in response to the determination that at least one of the plurality of threshold scores is exceeded.

[0240] Paragraph 4. A method according to paragraph 3, wherein the plurality of types of deviation comprise

[0241] a first deviation type representing the deviation of the measured pressure data from normal pressure data which is characteristic for a first period, and

[0242] a second deviation type representing the deviation of the measured pressure data from normal pressure data which is characteristic for a second period, the second period being longer than the first period.

[0243] Paragraph 5. A method according to paragraph 4, wherein the first period is a day or a week and the second period is a month or a year.

[0244] Paragraph 6. A method according to any of paragraphs 3 to 5, wherein the plurality of types of deviation comprise

[0245] a third deviation type representing a deviation in the maximum and / or minimum pressure measured by each pressure sensor in an interval, wherein the time period over which the measured pressure data is scored comprises one or more intervals.

[0246] Paragraph 7. A method according to paragraph 6, wherein the interval is 6 minutes.

[0247] Paragraph 8. A method according to any of paragraphs 3 to 7, wherein the threshold score which is exceeded indicates a type of the anomaly.

[0248] Paragraph 9. A method according to any preceding paragraph, wherein the alert indication comprises the measured pressure data.

[0249] Paragraph 10. A method according to any preceding paragraph, wherein the alert indication indicates a type of the anomaly.

[0250] Paragraph 11. A method according to paragraph 10, the type of anomaly comprises one or more of: substance ingress in the GDN, a leak in the GDN, a failure of a component of the GDN. Paragraph 12. A method according to paragraph 11, wherein the failure of the component of the GDN comprises a failure of a governor station in the GDN.P131560GB 28

[0251] Paragraph 13. A method according to any preceding paragraph, wherein the method comprises receiving, from the one or more pressure sensors, the previous pressure data, and training the ML model to characterise normal pressure data at the one or more pressure sensors based on the previous measured pressure data.

[0252] Paragraph 14. A method according to any preceding paragraph, wherein the method comprises receiving, from the one or more pressure sensors, additional measured pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a time period subsequent the time period over which the plurality of gas measurements in the measured pressure data were performed, and

[0253] updating the ML model to characterise normal pressure data at the one or more pressure sensors based on the additional measured pressure data.

[0254] Paragraph 15. A method according to any preceding paragraph, wherein one or more of the pressure sensors are configured to measure the gas pressure at one or more respective governor stations of the GDN.

[0255] Paragraph 16. A method according to any preceding paragraph, wherein the ML model is trained to characterise normal pressure data for at least one of the pressure sensors based on the previous pressure data measured over the time period by the at least one pressure sensor and based on

[0256] previous pressure data measured over the time period by at least one of the other pressure sensors.

[0257] Paragraph 17. A method according to any preceding paragraph, wherein the ML model is additionally trained to characterise anomalous pressure data for a plurality of anomaly types based on a plurality of sets of anomalous pressure training data, each set of anomalous training pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a time period when the respective anomaly type was present in the GDN, and the method comprises

[0258] determining, based on a comparison of the measured pressure data and the anomalous pressure data for each of the plurality of anomaly types as characterised by the ML model, the type of the anomaly which may have occurred, or which may occur, in the GDN.

[0259] Paragraph 18. A computing device, the computing device comprising circuitry configured to receive, from the one or more pressure sensors configured to measure gas pressure at one or more respective locations in a Gas Distribution Network (GDN), measured pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a time period,

[0260] use a trained machine learning (ML) model to score the measured pressure data over the time period in accordance with a degree to which the measured pressure data for the time period deviates from normal pressure data at the one or more pressure sensors as characterised by the ML model, wherein the ML model is trained to characterise normal pressure data at the one or more pressure sensors based on previous pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a previous time period,

[0261] compare the score of the measured pressure data over the time period to a threshold score,P131560GB 29

[0262] determine that the score of the measured pressure data over the time period has exceeded the threshold score, and in response,

[0263] output an alert indication indicating that an anomaly may have occurred, or may occur, in the GDN.

[0264] Paragraph 19. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any of paragraphs 1 to 17.

[0265] REFERENCES

[0266] [1] WO 2023 / 218196, “Computing Device, Pressure Control Station, System and Methods for Controlling Fluid Pressure in a Fluid Distribution Network”, published 16 November 2023.

[0267] [2] SIF Beta Project Intelligent Gas Grid (IGG) 1stAnnual Progress Report, 1stJuly 2023 -30thJune 2024, https: / / smarter.enerqynetworks.org / proiects / 10063754 / , published 27 November 2024.

[0268] [3] SIF Beta Project Registration, https: / / smarter.enerqynetworks.org / proiects / 10063754 / , published 23 August 2023.

[0269] [4] CN116625438B, “Gas pipe network safety on-line monitoring system and method thereof”, published 1 December 2023.

[0270] [5] CN116772944B, “Intelligent monitoring system and method for gas distribution station”, published 1 December 2023.

[0271] [6] https: / / www.gov.uk / government / publications / examining-patent-applications-relating-to-artificial-intelligence-ai-inventions / scenarios-applying-the-guidelines-for-examining-patent-applications-for-ai, retrieved 4 March 2025.

Claims

P131560GB 30CLAIMSWhat is claimed is:

1. A method performed by a computing device, the method comprisingreceiving, from one or more pressure sensors configured to measure gas pressure at one or more respective locations in a Gas Distribution Network (GDN), measured pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a time period,using a trained machine learning (ML) model to score the measured pressure data over the time period in accordance with a degree to which the measured pressure data for the time period deviates from normal pressure data at the one or more pressure sensors as characterised by the ML model, wherein the ML model is trained to characterise normal pressure data at the one or more pressure sensors based on previous pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a previous time period, comparing the score of the measured pressure data over the time period to a threshold score,determining that the score of the measured pressure data over the time period has exceeded the threshold score, and in response,outputting an alert indication indicating that an anomaly may have occurred, or may occur, in the GDN.

2. A method according to claim 1, wherein the outputting the alert indication comprises transmitting the alert indication.

3. A method according to claim 1, whereinusing the trained ML model to score the measured pressure data over the time period comprises scoring the measured pressure data over the time period separately according to a plurality of types of deviation from the normal pressure data,comparing the score of the measured pressure data over the time period to a threshold score comprises comparing the score of the measured pressure data over the time period to a plurality of threshold scores, wherein each of the plurality of threshold scores corresponds to a respective one of the plurality types of deviation from the normal pressure data, determining that the score of the measured pressure data over the time period has exceeded the threshold score comprises determining that at least one of the plurality of threshold scores is exceeded, and the alert indication is output in response to the determination that at least one of the plurality of threshold scores is exceeded.

4. A method according to claim 3, wherein the plurality of types of deviation comprise a first deviation type representing the deviation of the measured pressure data from normal pressure data which is characteristic for a first period, anda second deviation type representing the deviation of the measured pressure data from normal pressure data which is characteristic for a second period, the second period being longer than the first period.

5. A method according to claim 4, wherein the first period is a day or a week and the second period is a month or a year.P131560GB 316. A method according to claim 3, wherein the plurality of types of deviation comprisea third deviation type representing a deviation in the maximum and / or minimum pressure measured by each pressure sensor in an interval, wherein the time period over which the measured pressure data is scored comprises one or more intervals.

7. A method according to claim 6, wherein the interval is 6 minutes.

8. A method according to claim 3, wherein the threshold score which is exceeded indicates a type of the anomaly.

9. A method according to claim 1, wherein the alert indication comprises the measured pressure data.

10. A method according to claim 1 , wherein the alert indication indicates a type of the anomaly.

11. A method according to claim 10, the type of anomaly comprises one or more of: substance ingress in the GDN, a leak in the GDN, a failure of a component of the GDN.

12. A method according to claim 11, wherein the failure of the component of the GDN comprises a failure of a governor station in the GDN.

13. A method according to claim 1, wherein the method comprisesreceiving, from the one or more pressure sensors, the previous pressure data, and training the ML model to characterise normal pressure data at the one or more pressure sensors based on the previous measured pressure data.

14. A method according to claim 1 , wherein the method comprisesreceiving, from the one or more pressure sensors, additional measured pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a time period subsequent the time period over which the plurality of gas measurements in the measured pressure data were performed, andupdating the ML model to characterise normal pressure data at the one or more pressure sensors based on the additional measured pressure data.

15. A method according to claim 1, wherein one or more of the pressure sensors are configured to measure the gas pressure at one or more respective governor stations of the GDN.

16. A method according to claim 1, wherein the ML model is trained to characterise normal pressure data for at least one of the pressure sensors based on the previous pressure data measured over the time period by the at least one pressure sensor and based onprevious pressure data measured over the time period by at least one of the other pressure sensors.

17. A method according to claim 1, wherein the ML model is additionally trained to characterise anomalous pressure data for a plurality of anomaly types based on a plurality of sets of anomalous pressure training data, each set of anomalous training pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a time period when the respective anomaly type was present in the GDN, and the method comprisesP131560GB 32determining, based on a comparison of the measured pressure data and the anomalous pressure data for each of the plurality of anomaly types as characterised by the ML model, the type of the anomaly which may have occurred, or which may occur, in the GDN.

18. A computing device, the computing device comprising circuitry configured to receive, from the one or more pressure sensors configured to measure gas pressure at one or more respective locations in a Gas Distribution Network (GDN), measured pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a time period,use a trained machine learning (ML) model to score the measured pressure data over the time period in accordance with a degree to which the measured pressure data for the time period deviates from normal pressure data at the one or more pressure sensors as characterised by the ML model, wherein the ML model is trained to characterise normal pressure data at the one or more pressure sensors based on previous pressure data comprising a plurality of gas pressure measurements performed by each pressure sensor over a previous time period,compare the score of the measured pressure data over the time period to a threshold score,determine that the score of the measured pressure data over the time period has exceeded the threshold score, and in response,output an alert indication indicating that an anomaly may have occurred, or may occur, in the GDN.

19. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of claim 1.

Images