GB2618987A - System for detecting abnormal operating states of a heating system - Google Patents

Abstract

Detection of anomalous operating states of a temperature control system (e.g. a central heating system, fig 1) is performed using a prediction model. The model is trained using training samples based on operational data of the system 702. The training samples contain values for training features derived from operational data which are processed using an anomaly detection model, which outputs an anomaly score that is indicative of an anomalous outlier. An anomaly indication is associated with each anomaly score. The prediction model is then trained 710 on the processed training samples and anomaly indications so that it may predict a corresponding anomaly indication for a test sample of feature values. The prediction model is then used to identify an anomalous operating state of a temperature control system being monitored based on further operational data received from the monitored temperature control system.

Description

System for detecting abnormal operating states of a heating system The present invention relates to systems and methods for detecting abnormal operational states and faults in a heating system (or other temperature control system), in particular a boiler.

Faults in domestic heating equipment, in particular boilers, often cause great inconvenience and cost to homeowners. Typically, problems are not recognised until it is too late, at which point the homeowner is left without heating / hot water.

Although early detection of developing faults would be beneficial, homeowners do not have the expertise to monitor operation of their heating systems. In modern boilers, the mode of operation of the boiler and indications of certain faults may be available through a control panel. However, users are often not aware of the mode or fault because they do not closely monitor their boiler and a faulty boiler may still produce hot water for some time. Fault diagnostics and maintenance generally require an engineer to attend, which is costly.

Embodiments of the present invention seek to provide approaches for detection of faults without necessarily requiring in situ inspection by an engineer, and in some cases for early detection of developing faults before they result in complete system failure, to enable timely intervention.

Accordingly, in a first aspect of the invention, there is provided a computer-implemented method for detecting an anomalous operating state of a temperature control system for controlling a temperature in an environment, the method comprising: receiving operational data indicating operational characteristics of one or more temperature control systems; generating a set of training samples based on the operational data, each training sample comprising values for a plurality of training features derived from operational data associated with a given temperature control system; processing training samples of the set of training samples using an anomaly detection model, the anomaly detection model arranged to output for each training sample an anomaly score that is indicative of whether the training sample represents an anomalous outlier in the set of training samples; associating with each processed training sample an anomaly indication based on the anomaly score assigned to the training sample by the anomaly detection model; training a prediction model on the processed training samples and the associated anomaly indications to obtain a trained prediction model that is arranged to predict a corresponding anomaly indication for a test sample of feature values; and applying the prediction model to identify an anomalous operating state of a temperature control system being monitored, based on further operational data received from the monitored temperature control system.

The operational data used for training may typically be received from a population of temperature control systems, e.g. over some period of time. However, it would also be possible to train a prediction model specific to just one temperature control system.

The applying step may comprise receiving the further operational data from the temperature control system being monitored and deriving one or more test samples of feature values from the further operational data; applying the trained prediction model to the test samples to obtain predicted anomaly indications for the test samples; and identifying an anomalous operating state of the monitored temperature control system based on the predicted anomaly indications. Test samples correspond to training samples, e.g. they typically include the same features as the training samples (or at least a subset of those features e.g. where the prediction model only uses a subset of the features).

The step of applying the prediction model may be performed for a population of monitored temperature control systems. A fault alert or notification may then be generated for the monitored system (or any of them) if an anomalous operating condition is detected.

A training or test sample provides information On the form of the feature values) on the operating state of a temperature control system at a particular time. Thus, each training or test sample preferably comprises values for a plurality of features derived from operational data associated with a given temperature control system corresponding to a particular measurement time. For example, a sample may correspond to a particular time instant or measurement time interval (e.g. determined based on timestamps in the operational data). This may therefore result in generating a times series of training samples (or equivalently test samples) for a given temperature control system, characterising the operation of the system at different times.

The application of the prediction model may be performed as pad of a monitoring operation for monitoring operational data received from one or more boilers over time (and formed into a time series of test samples). Monitoring may be performed live, e.g. as data is received, or at intervals, e.g. to process received data for a past monitoring interval.

The operational data (used to generate the training and/or test samples) preferably comprises one or more of: sensor data from one or more sensors measuring operating characteristics of a temperature control system; and control data associated with a temperature control system. Operational data may include respective time series of sensor measurements or other control/operational values for different sensors / sources, each measurement or other value associated with measurement time information (e.g. a timestamp).

The features of the training and test samples may comprise one or both of: values of the received operational data or sensor data, and derived values computed from the received operational or sensor data. Thus, features may include the raw operational data as received (e.g. sensor measurements) and/or may include derived features, where a derived feature can be calculated from one or more values of the operational data.

The operational data (used to generate the training and/or test samples) may comprises a demand signal indicating a request for activation of a temperature control function, preferably a heating function, of the temperature control system. The sample features may then comprise a time duration since the demand signal transitioned to (has remained in) a given state, preferably an ON state, the ON state indicating that activation of the temperature control function is being requested. -4 -

Preferably, the operational data (used to generate the training and/or test samples) comprises temperature measurements obtained using one or more temperature sensors associated with a given temperature control system.

In preferred embodiments, the temperature control system is a heating system comprising a heating appliance such as a boiler for heating water that is circulated through a water circulation system to provide space heating and/or for supplying heated water to hot water outlets, the operational data comprising one or more of: measurement data indicating a temperature at a supply flow pipe supplying heated water from the heating appliance into the circulation system; measurement data indicating a temperature at a return flow pipe returning water from the circulation system to the heating appliance; and measurement data indicating a temperature at a hot water supply pipe for supplying heated water to hot water outlets. Note in some embodiments a different fluid may be substituted for water.

The supply flow temperature may correspond to the temperature of the supply flow or specifically of the heated water, or may be indicative or representative of the supply flow temperature. Similarly, the return flow temperature may correspond to, or be indicative or representative of, the return flow temperature (specifically of returning water). In particular, the supply flow temperature may be one of: the temperature of heated water supplied into the water circulation system; and the temperature of a conduit (e.g. pipe) carrying the heated water. The return flow temperature may be one of: the temperature of water returning from the water circulation system to the heating component; and the temperature of a conduit carrying the returning water. Conduit temperatures may be temperatures of outer surfaces of those conduits (e.g. pipes). Thus, the method may comprise receiving the temperature measurements from respective sensors mounted on supply and/or return flow conduits (and/or a hot water supply conduit), the sensors preferably arranged to measure a temperature of an exterior surface of a respective conduit.

The features of the training and/or test samples preferably comprise one or more of: one or more of the measured temperatures; rate of change of one or more of the measured temperatures; a minimum or maximum of a measured temperature over a time window; a temperature increase or temperature decrease of a -5 -measured temperature over a time window; an average value or standard deviation of two or more (or each) of the measured temperatures. For example, any of the above features may be obtained based on any (or all) of the supply flow, return flow and hot water supply temperatures.

The operational data and/or features may comprise a measure of electricity consumption of a temperature control appliance (e.g. boiler) of the temperature control system.

Preferably, the anomaly indications are anomaly classifications, and the prediction model is a classification model. More specifically, the prediction model may be a binary classification model classifying a sample as being either a normal sample indicative of normal operation or an abnormal sample indicative of an anomalous operating state (however, more than two classes could alternatively be supported e.g. to allow for classification of distinct fault types). The associating step preferably associates an anomaly classification with each training sample based on comparison of the anomaly score output by the anomaly detection model to a classification threshold (e.g. assigning a classification indicative of a fault being present to samples that exceed the threshold). However, in other embodiments the anomaly indications could be other values derived from the scores or even the scores themselves, with the prediction model directly predicting those values or scores.

The prediction model preferably comprises a decision tree (preferably a binary decision tree outputting either "normal/no fault" or "abnormal/fault" classifications, though additional classifications could be supported). The method preferably comprises identifying an anomalous operating state in response to the prediction model outputting a predetermined anomaly classification associated with anomalous operation (e.g. the "fault" classification) for a test sample.

The method may comprise training a plurality of prediction models, each trained on a subset of the training features (or more specifically each trained on a respective different subset), wherein the step of applying the prediction model comprises applying multiple ones or each of the trained prediction models to the test samples, and preferably further comprising identifying an anomalous operating state in -6 -response to any of the prediction models outputting an anomaly indication indicative of anomalous operation. Each prediction model may be associated with a respective fault type, the method comprising, in response to a given one of the prediction models outputting an anomaly indication indicative of anomalous operation, diagnosing a fault of the corresponding fault type and outputting a notification of the diagnosed fault type.

The anomaly detection model preferably comprises one or more partitioning trees, each defining a hierarchy of partitioning operations that recursively partition the training sample set into subsets, with each training sample assigned to a leaf node of the partitioning tree representing a partition of the training sample set. The partitioning operations in each partitioning tree are preferably randomly determined, preferably by random selection of a partitioning feature and/or a split point for the partitioning feature for each partitioning operation. The anomaly detection model preferably assigns an anomaly score to each training sample based on the depths in the partitioning trees of the leaf nodes to which the training sample is assigned.

The anomaly detection model preferably comprises (or is based on) an isolation forest comprising a plurality of isolation trees generated by random partitioning of the training sample set.

The method is preferably performed at a computing device (e.g. a central monitoring server), the method comprising receiving the operational data from sensors and/or a control system associated with the temperature control system being monitored at the computing device, processing the operational data to identify presence or absence of the anomalous operating state (using the method set out above), and transmitting a message from the computing device to another device in dependence on detection of the anomalous operating state. Training of the anomaly detection model and prediction model may be performed at the same computing device as the monitoring or at a different computing device (for example training may occur at one device, with the prediction model provided to a second device which then processes data for the monitored system to identify the anomalous operating state). -7 -

The method preferably comprises, in response to identifying presence of an anomalous operating state, outputting a fault alert, notification or control command. This may comprise performing one or more of sending a notification to a user device associated with a user of the monitored temperature control system, the notification optionally comprising an electronic message transmitted via a public messaging system (e.g. email, instant messaging, SMS) or a notification displayed in a monitoring application provided or accessed at the user device; sending a control command to the monitored temperature control system to deactivate, or modify an operating configuration of, the temperature control system; and initiating scheduling of a visit by a maintenance engineer to perform maintenance and/or repair of the temperature control system.

In a further aspect of the invention, there is provided a computer-implemented method for detecting an anomalous operating state of a heating system, 15 comprising: receiving operational data indicating operational characteristics of a population of heating systems, the operational data including temperature measurements from one or more temperature sensors of the heating systems and preferably further including heating demand signals; generating a set of training samples based on the operational data, each training sample comprising values for a plurality of training features derived from operational data associated with a given heating system; processing training samples of the set of training samples using an isolation forest anomaly detection model, the anomaly detection model arranged to output for each training sample an anomaly score that is indicative of whether the training sample represents an anomalous outlier in the training sample set; associating with each processed training sample an anomaly classification based on the anomaly score assigned to the training sample by the anomaly detection model; training a decision tree classifier on the processed training samples and the associated anomaly classifications to obtain a trained decision tree classifier that is arranged to predict a corresponding anomaly classification for a test sample of feature values; -8 -receiving further operational data including temperature measurements (and preferably a demand signal) from a heating system being monitored and deriving test samples of feature values from the further sensor data; applying the trained decision tree classifier to the test samples to obtain anomaly classifications for the test samples; and identifying an anomalous operating state of the monitored heating system based on the anomaly classifications.

The method in this aspect may include any of the further steps or features of the first aspect of the invention as set out above.

The invention also provides a system or computing device having means, optionally comprising one or more processors and associated memory, for performing any method as set herein.

The invention further provides a computer program, computer program product or tangible computer readable medium comprising software code adapted, when executed by a data processing system, to perform any method as set out herein.

Any feature in one aspect of the invention may be applied to other aspects of the invention, in any appropriate combination. In particular, method features may be applied to apparatus and computer program aspects, and vice versa.

Furthermore, features implemented in hardware may generally be implemented in software, and vice versa. Any reference to software and hardware features herein should be construed accordingly.

Preferred features of the present invention will now be described, purely by way of example, with reference to the accompanying drawings, in which:-Figure 1 illustrates a monitoring system for a boiler; Figure 2 shows various operating characteristics of a boiler, such as a demand signal and temperature measurements, during normal operation of the boiler; -9 -Figure 3 shows operating characteristics of a boiler during faulty operation; Figure 4 illustrates the effect of hot water suppression on boiler operating characteristics; Figures 5 and 6 illustrate further fault scenarios; Figure 7 illustrates a process for training a diagnostic model for diagnosing heating system faults; Figure 8 illustrates a diagnostic method employing the trained diagnostic model; Figure 9 illustrates a process for refining an anomaly detection model; Figure 10 illustrates outliers identified by the anomaly detection model; Figure 11 illustrates a detected anomaly; Figure 12 illustrates a relationship between a contamination parameter used for constructing the isolation forest model, and a number of anomaly alerts produced by the system; Figure 13 illustrated detection of anomalies when the contamination parameter is set too high; Figures 14 and 15 illustrate examples of decision tree classifiers trained based on the output of the isolation forest anomaly detection model; Figure 16A-16C illustrate a user interface for fault notification and resolution; and Figure 17 illustrates a processing system for implementing described diagnostic techniques.

Overview Presently described embodiments seek to detect abnormal operating states of a boiler that might be indicative of existing or developing faults, based on sensor data from various sensors monitoring the boiler.

A central heating boiler provides space heating on demand by supplying heated water to radiators placed in areas of interest in a building. Control of the system may be via a thermostat and control interface allowing a user to configure a heating schedule specifying when the heating should be on and a target temperature for the environment during any activation periods. The system then automatically controls activation of the boiler based on the schedule and current environmental temperature measured by the thermostat, though the user may also override the schedule to selectively activate/deactivate space heating on an ad hoc basis.

When space heating is required, the control system notifies the boiler via a control signal, which is referred to as the demand flag. This signal is either ON, indicating the boiler should be activated to perform heating, or OFF, to indicate that no heating is to be performed. While in the activated state (demand flag ON), the boiler heats water and pumps the heated water toward the radiators via a supply flow pipe. After passing through the radiators the water flows back to the boiler via a return pipe. Water continues circulating in this way until the desired environmental temperature is reached, as detected by the thermostat, which then causes the demand flag to be turned off.

The boiler may additionally supply domestic hot water to water outlets (e.g. to taps and showers), either on demand (e.g. in the form of a combi boiler), or by heating water for storage in a hot water tank (e.g. on a schedule).

It is often a symptom of a faulty boiler when the temperatures on the pipes that carry water do not rise when the user demands space heating. However, on-demand or scheduled heating of domestic hot water for supply to hot water outlets or a tank can interrupt space heating, since the boiler typically only heats water for one of the subsystems (space heating or domestic hot water) at any one time. Thus, space heating may be interrupted or delayed when an occupant turns on a hot water tap or shower. This is also referred to herein as hot water suppression.

The effects of hot water suppression mean that a lack of expected temperature rise on heating pipes after heating is requested does not always indicate a fault. This makes reliable diagnostics of heating faults more challenging and a system that simply monitors pipe temperatures could thus easily generate spurious alerts.

Described embodiments provide the ability to detect fault symptoms remotely based on monitoring operational data obtained from the boiler, enabling an energy provider or maintenance service provider to inform their customer that the boiler is not working as expected and ask the customer to perform some basic recovery procedures (boiler repressurization is one of the most common examples). This avoids an engineer visit in the cases that can be solved directly by the user, reducing cost to the consumer and service provider. The system can also enable an engineer to be sent to repair the boiler in cases where the basic corrective procedures fail Fault detection techniques are described herein mainly in relation to the detection of space heating failures, but the techniques can be extended to the detection of faults in the supply of domestic hot water.

Figure 1 shows a monitoring system for a boiler 104. The boiler 104 receives water from a domestic cold water supply 106 and produces a domestic hot water (DHVV) output 105. The boiler is also connected to a central heating system through supply and return pipes 108, 109 which lead to a series of radiators (not shown). The pipe network and radiators form a closed-loop water circulation system. The skilled person would understand that variations in the boiler design and operation are possible, including further features and boilers which do not feed radiators directly. For instance, different boiler types may have a hot water output 105 which feeds directly to domestic supply outlets or may feed into a hot water storage tank from which outlets are then supplied.

The monitoring system of Figure 1 includes sensors attached to various system components, including heating supply flow temperature sensor 111 and heating return flow temperature sensor 112. Heating supply flow temperature sensor 111 obtains measurements indicative of the temperature of heated water being supplied into the circulation system via supply pipe 108 after heating by the boiler, referred to herein as the "supply flow temperature". Heating return temperature sensor 112 obtains measurements indicative (during normal operation) of the temperature of water returning to the boiler for heating via return flow pipe 109 after passing through heating pipes and radiators, referred to herein as the "return flow temperature".

The supply and return flow temperature sensors may measure the temperature of the water itself or may provide indirect measurements. In a preferred embodiment, to avoid the need for invasive sensors, the sensors are installed externally, being placed against the pipes so as to measure the temperatures of the exterior pipe surfaces, which are used as a representative temperatures, since the pipe surface temperatures are expected to track the temperature of the water flowing within. Thus, the temperature sensors can be used to identify a change in temperature of the pipe indicating the presence of a flow. In this approach, sensor 111 may thus be affixed to an exterior of the supply flow pipe 108 and sensor 112 may be attached to the exterior of the return flow pipe 109.

Similarly, hot water sensor 110 associated or attached to hot water output pipe 105 may be a temperature sensor indicating flow of domestic hot water from the boiler 104 e.g. to taps, showers and/or a hot water tank. The sensors 111, 112, 110 may alternatively be other sensors capable of providing indications of temperature and/or flow by measuring flow or related phenomena. Additionally, an electrical energy sensor 114 may be attached or associated with a power connection of the boiler 104 and is configured to detect the amount of electrical energy being used by the boiler. The electrical energy sensor 114 may be a current sensor or voltage sensor, or other suitable means for measuring electrical energy.

The sensors communicate with a controller 102. The controller 102 may be a microcontroller or other control device. The controller is connected to the sensors by a connection means 113 which may be wired (e.g. a wire) or wireless (e.g. BluetoothTm or ZigbeeTM) for each sensor. In typical embodiments a wired connection is used to allow the controller to power the sensors. The controller 102 is preferably linked via wired or wireless connection to communicate with one or more servers 101 (either directly or through one or more intermediate devices and/or networks e.g. the Internet). The server may also be in communication with a user device 103, e.g. via the Internet, although it is also possible for a user device to connect on a local area network (including to the controller or an intermediary device such as a hub). The user device 103 (for instance a personal electronic device such as a mobile telephone or smartphone, or other user interactive device) may allow a user to interact with the controller or server and/or to send instructions to the controller or server and/or to receive notifications from the controller or server, for instance regarding a diagnostic problem or boiler fault. The notification may be by text message or notification in an application on the user device 103. The notification may prompt the user to take an action to control the boiler or may inform the user of an action already taken by a boiler controller. Alternatively, the server may control, or send instructions to, the boiler directly based on a fault determination.

Preferably one or more of the sensors (or all of them) are attachable to the boiler, for example by clipping onto or grasping the wire/pipe they measure. For example, temperature sensors can be affixed to the exterior of the supply flow pipe and return flow pipe connecting the boiler to the pipe network and radiators to measure the pipe temperatures. In this way the system may easily be retrofitted to a boiler. The controller and/or sensors may be provided as a kit for installation by an engineer or user in an existing heating installation. In alternative implementations some or all of the sensors may be integrated into the boiler.

The exact positions of the sensors will typically differ in different installations depending on the type of boiler and type of sensor kit used, as well as the way in which installation was carried out.

The controller collects measurements from the various sensors and sends them to the server, where the data is processed by a diagnostic algorithm. The controller may also send other operational data to the server, such as the demand signal used to control the boiler (or processed/summarised data describing the demand signal, e.g. specifying ON/OFF transitions). When the server detects a fault based on the supplied data, the user is notified, e.g. via an application on the user device.

Figure 2 shows a graph of operational data including various sensor measurements for a heating appliance (e.g. boiler) and heating system working correctly over a period of time (the horizontal axis indicates time and the vertical axis is labelled to show temperature values in degrees Celsius). The curve 202 shows the heating demand signal (call-for-heat), which is a binary ON/OFF signal. The controller transitions this signal to the ON state when space heating is required and to the OFF state when space heating is no longer needed e.g. because a target temperature has been attained in the environment. Curve 204 shows the supply flow temperature and curve 206 shows the return flow temperature to/from the heating appliance. It can be seen that when the demand flag transitions to the ON state, e.g. in region 208, the temperatures on the supply flow and return flow pipes rise after a few minutes. This indicates that hot water is correctly circulating through the system and the radiators are warming the air. Also shown are the DHW temperature 210 which is not significantly affected by the activation, and the current draw 212 of the heating appliance, which tracks the demand signal to some extent.

Figure 3 shows an example of a boiler not able to provide space heating. Here we notice that when the demand flag goes up (region 300), the pipe temperatures remain stable for hours around 20°C. This indicates that heating is not being provided, and that the user is waiting for the indoor temperature to be raised.

Embodiments of the invention aim to detect if a boiler is not providing heating when requested by the user (either manually or via a preconfigured heating schedule). Preferably, the failure is detected as soon as possible to enable the user to be notified promptly, possibly even before they notice themselves that there is a problem. However, detecting when a boiler is failing before the user realizes that there is an issue is challenging because the response time is different for each boiler and depends on the outdoor temperature. More importantly, a combi boiler will not provide heating when domestic hot water is requested at the same time (hot water suppression).

Figure 4 illustrates an example of a boiler that is requested to provide space heating while also providing domestic hot water. As illustrated in region 400, the temperature of the domestic hot water (210) rises above 40°C before the demand flag 202 goes up and the temperatures on supply flow (204) and return flow (206) rise only when the temperature 210 on the domestic hot water pipe starts decreasing. Space heating is thus provided with some delay due to hot water suppression (the vertical bar at 402 shows when the boiler starts providing space heating). This is normal behaviour of a boiler, and no fault should be identified or alert issued in this case.

While Figure 3 shows a failure of the heating system that is present from the start of a heating cycle (i.e. where the boiler never responds to activation of the demand flag), failures may also occur during a heating cycle, as illustrated in Figures 5 and 6. In particular, Figure 5 shows a situation where the boiler initially responds correctly to the demand for heat but then stops providing heating after less than one hour, and then subsequently recovers (see region 500). Figure 6 illustrates a boiler -b -failing at the very end of the heating cycle (see region 600), where it can be seen that supply flow and return flow temperatures start to drop before the demand flag turns off.

Embodiments of the invention provide a diagnostics system that is able to detect a variety of failure scenarios, including failure at the start of a cycle, during a cycle or at the end of a cycle, as shown in Figures 3, 5 and 6, whilst avoiding false detections, e.g., due to DHW suppression as shown in Figure 4. Note while reference is made herein to faults or failures, this shall be taken to include any anomalous operating conditions. In addition to situations where the boiler has failed completely (temporarily or permanently) and so is at that time not able to provide any heating, this may also encompass situations where the boiler is working but not achieving expected performance levels.

Fault detection method The proposed fault detection method is based on identifying cases in which the behaviour of the boiler is anomalous and considering these cases as indicative of faults.

The detection method is based on a decision tree classifier that is trained on the output of an anomaly detection model applied to training data from a population of installed boilers. The training stage involves three main steps: feature extraction, creation of an anomaly model, and training of the decision tree classifier. The process is illustrated in overview in Figure 7.

In step 702, sensor measurements and any other operational data (such as the demand signal) are received from a population of boilers, over a time period to be analysed. These are boilers installed in heating systems in building environments, e.g. houses/apartments, which are provided with sensors and data collection systems as shown in Figure 1. Thus, in an embodiment, the operational data includes data from the following sensors/sources: * Supply flow temperature 111 * Return flow temperature 112 * Domestic hot water (DHVV) supply temperature 110 * Boiler current draw 114 (or other power consumption measure) * Demand flag Note the process may collect live data until a sufficient volume of data has been obtained, or the data may be obtained from a database of historical data. For each boiler, a data set consists of separate time series of sensor values for each sensor or other data source (e.g. as illustrated in Figures 2-6). Each sensor reading has a timestamp indicating the measurement time of that reading (similarly for the demand flag).

In step 704, values are computed for a set of features for each boiler, at each of a sequence of time increments extending over the time period to be analysed (e.g. as determined by the timestamps of the sensor data). Note that where the time resolution of different sensors and sources differs, this may involve interpolating data values as needed. The features are values derived from the raw sensor data that may be used by the detection model to identify faults. Features may include derived / computed values obtained from the sensor data but may also include the sensor data values themselves. In an embodiment, the following features are obtained/computed at each timestamp in the sensor data for all the boilers available: * Amount of time the demand flag has been ON since the last request for heating (e.g. duration since last transition of the demand flag from OFF to ON) * Temperature on the supply flow pipe * Temperature on the return flow pipe * Temperature on the domestic hot water pipe * Temperature slope on the supply flow pipe * Temperature slope on the return flow pipe * Temperature slope on the domestic hot water pipe * Maximum of the temperature on the domestic hot water pipe over a preceding period, e.g. over the last 15 minutes (from the current time) * The temperature increase on the domestic hot water pipe over a preceding period, e.g. in the last 15 minutes (e.g. as a temperature delta, i.e. difference between current temperature and earlier temperature) * Average of the temperatures on flow, return and domestic hot water pipe (computed as the average of the three temperature values at the present time instant) * Standard deviation of the three temperature values of the flow, return and domestic hot water pipes at the current time instant * Current drawn by the boiler Note the above are examples of features that may be used. Embodiments may use a subset of those features and/or may use alternative or additional features.

Furthermore, the final fault detection model derived from the obtained features may not use all of the derived features to perform fault classification. The set of feature values of the chosen features derived at a given time instant (timestamp value) from the data for a particular boiler forms a training sample.

Step 704 results in a time series of training samples for each boiler in the population being used for training.

The training samples are then used to derive a diagnostic model for identifying faults (and more generally abnormal operating states). The process is divided into two stages: firstly, an outlier detection model is used to identifier outliers in the training samples, which are labelled as being indicative of faults (the terms outlier detection model and anomaly detection model are used interchangeably herein). Secondly, a prediction model is trained which predicts the fault classifications directly from the training samples.

Specifically, in step 706 an outlier detection model is created and used to generate anomaly scores for each training sample. The model assigns a score to each sample indicating a degree of isolation of the training sample from other samples, with isolated samples considered as outliers and thus potentially indicative of abnormal operation. Since boilers generally have predictable behaviour, any feature sets identified as anomalous are considered to be indicative of a failure.

In step 708 fault classifications are assigned to training samples based on the anomaly scores. For example, any training samples with an anomaly score exceeding some predetermined threshold may be assigned a "fault" classification, while training samples whose scores do not exceed the threshold are assigned a "no fault" classification.

The second stage involves, in step 710, training the prediction model, using a suitable machine learning algorithm. In preferred embodiments, the prediction model is a decision tree. The decision tree (or other prediction model) is learnt to predict the fault classifications obtained in the first stage using the outlier detection model directly from the feature data of the training samples. The resulting prediction model can then be used for direct classification of future boiler data.

Application of the prediction model obtained by the Figure 7 process is illustrated in the monitoring / diagnostics process of Figure 8. In step 802, further operational data (e.g. sensor measurements and the demand signal) are received from one or more boilers being monitored for faults. The monitored boilers may be the same as those used for creating the diagnostic model or may be different. Sets of feature values are derived from the raw sensor measurements and other data in step 804 in the same way as described above, to produce a time series of test samples, where each test sample includes a set of feature values for a given time instant derived from the boiler data. The features are the same as used during model creation (though in some cases if the final fault classifier does not use all features then a reduced set of features could be produced in this step).

In step 806, the previously trained prediction model is applied to each test sample, resulting in the prediction model outputting a "fault" or "no fault" classification for the test sample. If the sample was classified as representing a "fault" (test 810), then the process continues to generate a fault notification (step 812). The process may then continue to step 802 to continue classifying samples. If the sample is classified as "no fault" then the process returns immediately to step 802 to continue classifying samples.

The classification loop of Figure 8 may in practice be implemented in various ways. For example, the classification may be performed in real-time / online, individually for each sample as the sensor data is received from a specific boiler and formed into the test sample of derived feature values. Alternatively, sensor data and/or test samples may be buffered and processed in a batch. For example, the classification could be run periodically to process data for a given boiler generated over a given collection window, e.g. the preceding 24 hours. The algorithm may process data from multiple boilers sequentially or in parallel.

Outlier detection model In preferred embodiments, the outlier detection model used to obtain the initial fault classifications that are used during decision tree learning is an isolation forest. The isolation forest algorithm generates a representation of a data set using random partitioning that allows outliers to be identified on the assumption that outliers will require fewer random partitioning steps to become isolated in a partition (i.e. fewer partitioning steps are required to end up with an outlier on its own in a partition compared to non-outliers).

Creation of the isolation forest (as performed in step 706 of Figure 7) is illustrated in more detail in Figure 9.

In step 902, the training samples (generated as part of the Figure 7 process) are received as input to the isolation forest algorithm.

In step 904, the training parameters for the isolation forest algorithm are initialised (or updated on subsequent iterations). The training parameters are discussed in more detail below.

In step 906, the isolation forest algorithm is then run on the training samples in accordance with the configured training parameters to derive an isolation forest model in step 904. The isolation forest is trained using an algorithm as described in Liu, Fei Tony, Kai Ming Ting, and Zhi-Hua Zhou. "Isolation forest." 2008 Eighth IEEE International Conference on Data Mining, IEEE, 2008.

Briefly, the process involves building a set of isolation trees from the input feature data (i.e. the feature data values derived from the source boiler data as discussed above). A tree is built by randomly selecting, at each node starting with a root node, a feature from the feature set, and then randomly selecting a split point (partitioning value) between the minimum and maximum values of that feature in the data set.

-20 -Samples with a value for the selected feature below the split point are assigned to a left subtree of the node and samples with a value for the selected feature greater than or equal to the selected split point are assigned to a right subtree of the node. The left and right subtrees are then built recursively in the same manner from the assigned samples, until subtrees include a single sample (forming leaf nodes of the tree) or a maximum tree depth has been reached. An anomaly score is then assigned to each input sample. This involves computing the path length from the root to the sample in each tree, and deriving the score from the average path lengths across all trees. Corrections are applied for the case where building of the tree was terminated early due to the tree depth constraint, resulting in a leaf node containing more than one input sample.

The number of isolation trees produced by the algorithm may be varied as needed. In an embodiment, running the isolation forest with 100 isolation trees has been found to be effective. The maximum depth of the trees is set by the isolation forest algorithm (but could in some implementations also be a tuneable parameter). A further parameter of the isolation forest algorithm is the contamination coefficient, which is discussed in more detail below. These parameters are set in step 904.

Once the isolation forest has been built, anomaly scores are computed for each training sample in step 908, and a classification is assigned to each training sample based on a threshold for the anomaly score.

The isolation forest model highlights all the time instants at which the boiler exhibits anomalous behaviour (e.g. as in the Figure 5 example), since the isolation forest will give high path lengths and hence a high anomaly score to unusual configurations of the features, identifying them as outliers.

In step 910, the model is evaluated against expert assessment of the data. In an embodiment, this involves a human expert reviewing the relevant feature data and/or source data for all the situations where samples were classified as faults by the algorithm to determine whether each fault was correctly identified by the algorithm or not, and assigning a score to the performance of the model. A standard precision metric may be used as the performance score, e.g. based on the ratio of -21 -correct fault classifications (true positives) to total number of faults classified (see also https://en.wikipedia.org/wiki/Precision_and_recall for details).

If the score is not acceptable (test 912), the training parameters for the training of the isolation forest model are modified (step 904) and the training is repeated (906) with the modified parameters until an acceptable score is reached. Parameters may be tuned manually; alternatively this can be automated by searching over predefined sets/ranges of values for each parameter being adjusted. In preferred embodiments, only the contamination coefficient (discussed below) is modified and the tree count and maximum tree depth remain fixed as the contamination parameter typically has greater impact on the final classification performance. However, in other implementations additional parameters could be tuned at this stage. Acceptability of the score may e.g. be determined based on a score threshold, with training repeated until the threshold is reached or exceeded.

When the accuracy of the isolation forest is considered sufficient (step 912) and so a suitable isolation forest meeting the required score threshold has been trained, the system can then proceed to the second stage of training a prediction model, as described above.

Note that the iterative refinement of the isolation forest may not always be needed. Unlike supervised machine learning algorithms, isolation forests are not based on training from previous labelled examples but rather evaluate the degree to which data points may be deemed to be outliers or normal with respect to the input data set without any prior knowledge of what constitutes an outlier / anomaly, purely based on the properties of the data set itself. Thus, in some cases, an isolation forest built from the input data using fixed training parameters may produce a sufficiently accurate result. Nevertheless, in some cases, the iterative refinement based on expert evaluation of the output as described above can allow classification performance of the system to be further improved.

Anomalies are classified based on a threshold on the anomaly score assigned to each training sample by the isolation forest algorithm -samples that exceed the score threshold are classified as anomalous and thus indicative of a potential fault, while samples that do not exceed the threshold are classified as indicative of -22 -normal boiler behaviour. The score threshold may be fixed (e.g. a value of 0.5 may be used -though it should be noted that how the threshold is set and appropriate values to use may depend on the specific isolation forest implementation being used). Alternatively, the score threshold may be adjusted to improve performance of the model as needed.

Note that full details of the algorithm for building the isolation forest and computing anomaly scores are provided in the above-referenced paper. Furthermore, variations of the standard isolation forest algorithm described in that paper and/or subsequent improvements to the basic isolation forest algorithm may also be used.

Figure 10 visually illustrates the result of outlier classification using the isolation forest method. Specifically, this shows a subset of features segmented by the isolation forest with a contamination coefficient of 0.001. The marked points are anomalies identified by the model. The following three feature dimensions are depicted: * "timeon" is the time since the demand flag transitioned to ON (e.g. measured in seconds or in number of samples received since the transition) * "Appliance_Flow_Temperature" is the supply flow temperature On degrees Celsius) * "Appliance_Return_Temperature" is the return flow temperature (in degrees Celsius) Each subchart shows a two-dimensional view of the data set for a given pairing of two of the features (the subcharts across the diagonal show histograms for the relevant parameters). It can be seen in this case that the model considers as anomalies (marked points) the cases in which the temperatures are low after the demand has been on for about 500 flmestamps (i.e. 500 received samples). Figure 11 shows an anomaly identified by the same model on one of the boilers in the field and confirmed by a human expert.

Contamination parameter The training of the Isolation Forest requires specification of the contamination parameter. This parameter reflects the proportion of outliers expected in the -23 -training dataset. Picking this parameter appropriately may improve effectiveness as setting the parameter too high can result in a flood of alerts and setting it too low can result in boiler faults not being detected.

Figure 12 shows the relationship between number of alerts per year for each boiler and the contamination parameter while Figure 13 shows what would happen with a contamination parameter that is too high in the case illustrated in Figure 11 (each vertical dashed line represents a detected anomaly). In this case it can be noted that the algorithm identifies as anomalies cases that are not frequent but represent normal behaviour of the boiler.

One approach that can be used to choose a suitable contamination parameter is to count the number of anomalies found for different values of the contamination parameter and select the value that gives a number of detected anomalies close to the number of failures expected in a year.

The contamination parameter value can be refined using the iterative approach illustrated in Figure 9, based on evaluation of the isolation forest output by a human expert, until the detection accuracy is considered sufficient.

Prediction model As previously described, once the isolation forest has been created, the anomaly scores it produces for each sample are converted into fault classifications using an anomaly score threshold. Those classifications are then provided with the training samples as input to a learning algorithm to derive a prediction model that predicts fault classifications directly from the training samples.

In a preferred embodiment, the prediction model is a decision tree. The input to the decision tree learning process consists of the set of training samples from which the isolation forest was built, together with the fault/no-fault classifications assigned to each training sample based on the isolation forest scores. The output of the decision tree learning process is a decision tree that predicts a fault/no-fault classification for any set of feature values.

-24 -Any suitable decision tree learning algorithm may be used. In one embodiment, a standard implementation provided by the scikit-learn library is used, which implements an optimized version of the CART algorithm.

An advantage of the described approach is that (unlike some machine learning approaches, e.g. neural networks), the decisions taken by a Decision Tree are easily explainable. Decision trees can be inspected and/or visualised to understand how the model is classifying anomalies. This may allow users to place greater trust in the classifications made by the model.

Maximum tree depth for the tree learning can be controlled based on the desired complexity of the resulting trees. For example, limiting tree depth to 2 or 3 can produce tree models that are easy to inspect, understand and explain. However, tree depth could be increased at the cost of increased complexity.

An example of a decision tree derived by the described techniques is shown in Figure 14, showing the decision nodes in the tree, terminating in leaf nodes indicating anomalies (labelled as "failure"). Note that only decision branches leading to further decision nodes or to a failure classification are shown for clarity; decision branches that lead to a classification as non-anomalous, i.e. normal, have been omitted (e.g. in Figure 14, at the root node, when the condition "Demand on >20 min" is false, this results in a normal, non-failure classification).

The model shown in Figure 14 has been obtained using the following features: * Amount of time the demand flag has been ON since the last request * Temperature on the supply flow pipe * Maximum of the temperature on the domestic hot water pipe in the last 15 minutes This model has been found to be effective at identifying failures in the middle of heating cycles because it only checks for failures in which the demand flag has been ON for a long time and uses high thresholds on the temperature features.

In an extension of the above described techniques, the methodology described above is used to build different decision tree models, each decision tree trained on -25 -a subset of the features mentioned above. This can result in different classifiers that are useful in specific situations.

Figure 15 illustrates a further classifier that was obtained using the following features: * Amount of time the demand flag has been ON since the last request * Temperature on the supply flow pipe * The temperature increase on the domestic hot water pipe over the last 15 minutes * Average of the temperatures on supply flow, return flow and domestic hot water pipes * Standard deviation of the temperatures on supply flow, return flow and domestic hot water pipes This model has been found effective for detecting failures that happen at the beginning of the heating cycle because it can be triggered after just 7 minutes of the demand flag being ON, and is particularly robust to hot water suppression as it has multiple checks to ensure that there was no significant temperature increase on the pipe carrying domestic hot water.

In one approach, the combinations of features to use for each specific classifier are picked by a human expert via trial and error. Alternatively, decision tree learning could be based on all features or on random feature subsets. In either case, the tree depth constraint will force the decision tree learning algorithm to select the features from the provided feature set that are most useful for producing an accurate classification result. Generated decision trees can be evaluated by computing precision and recall (and/or FI-scores or other metrics), and models (trees) with the best scores are selected for use.

Using this approach, a group of decision tree classifiers may be produced, which can detect different categories of faults. Each of the models can subsequently be applied to live boiler data, with a fault identified when any of the models detects an anomaly.

-26 -Fault notification The controller 102 (Figure 1) for each monitored boiler continually collects data from the sensors and transmits the data to the server 101 over a network, e.g. the Internet. After initial training (Figure 7), the server runs the diagnostic algorithm (Figure 8) continuously on the received data. Typically a population of multiple boilers is monitored in this way.

When sensor data is received from a boiler controller, the following steps are performed, as discussed in relation to Figure 8: * the features are extracted to form a test sample of feature values corresponding to a particular point in time * the test sample is input to each of the trained decision tree models * if any of the models predicts an anomaly/fault the user is alerted.

Additionally, where multiple trained models have been obtained, a fault code/type can be presented to the user identifying which model triggered the alert (or identifying the fault classification associated with that model).

Fault alerts or notifications provided to users could take the form of: * An electronic message (e.g. SMS, email etc.) sent to the user device via a communications network using a public messaging system * A message / notification sent to a specific application operating on the user device, e.g. a smart home control application, for display to the user within the application at the device * A notification displayed at the boiler / controller. For example, the server could send a control message to the controller 102 which then activates an indicator light or shows an error code or message on a display. The indicator light or display may be integrated into the boiler, controller, or another system component.

* A notification sent to a further system, e.g. a maintenance / customer management system, for example to trigger scheduling of an engineer visit In some embodiments, users (e.g. occupants of properties where a monitored boiler is installed) can be provided with a software application for monitoring their -27 -heating system on a personal computing / communications device such as a smartphone or tablet computer.

Example screenshots of a smartphone app are shown in Figures 16A -160. Figure 16A illustrates a monitoring interface for monitoring operation of a boiler, where a currently selected "Event log" tab lists alerts generating by the described diagnostic algorithms. Here, the top entry shows an alert for a currently detected issue. By selecting the alert, the user may be taken to a further interface providing information and instructions for attempting to resolve the issue. As illustrated in Figure 16B, this may include a series of information messages, queries and instructions, requesting the user to carry out diagnostic / corrective steps and/or provide information. In some cases, these steps and interactions may lead to resolution of the issue, in which case the process may terminate and the issue may be marked as "resolved".

If the issue cannot be resolved, the interface may prompt the user to perform additional steps, such as creating a booking for a maintenance engineer to visit the users property and perform further diagnostics / repair. An example is illustrated in Figure 160. For example, this interface may link to a further automated booking service provided as part of the application, or a separate booking application or web service, to allow an engineer visit to be booked. An engineer then attends the property to perform in situ diagnostics and repair on the boiler as needed.

While illustrated as a smartphone app, a similar interface could also be provided as a web application accessible via a web browser on any Internet-connected computer or device.

In some embodiments, instead of (or in addition to) notifying the user, a control signal may be transmitted to the boiler, controller or another system component in response to detection of a fault, for example to deactivate the heating system/boiler (e.g. to prevent damage due to a leak or low pressure), or to reconfigure the system / boiler or take some other control action. As another example, the control signal could be to change the target temperature to see if the system can recover itself by alleviating the workload. Where the system is able to distinguish between fault categories using different decision tree models, certain fault categories (e.g. those -28 -with risk of catastrophic failure or damage) may result in automatic control actions e.g. boiler shutdown (plus user notification), whilst others may only result in user notification as previously described.

The described diagnostic techniques thus allow customers of an energy provider or boiler maintenance provider to be notified when their heating systems are exhibiting anomalous operating characteristics and faults. The method can identify, for example: * Pump faults.

* Ignition problems.

* Low pressure inside the boiler.

System architecture Figure 17 illustrates a server device 101 in more detail. The server includes one or more processors 1702 together with volatile / random access memory 1704 for storing temporary data and software code being executed.

A network interface 1706 is provided for communication with other system components (e.g. controller 102 and user device 103) over one or more networks 1716 (e.g. Local and/or Wide Area Networks, including the Internet).

Persistent storage 1708 (e.g. in the form of hard disk storage, optical storage and the like) persistently stores software for performing the described functions. This includes the learning algorithms, in particular the outlier model (isolation forest) algorithm 1710 and the decision tree learning algorithm 1711, along with a live monitoring and diagnostics process 1712 for real-time monitoring of the boiler population using the learned decision tree models. Persistent storage also stores data used by these processes such as a database 1714 of sensor data and derived feature data and training samples. The persistent storage also includes other server software and data (not shown), such as a server operating system.

The server will include other conventional hardware and software components as known to those skilled in the ad, and the components are interconnected by a data -29 -bus (this may in practice consist of several distinct buses such as a memory bus and I/O bus).

User device 103, e.g. in the form of a smartphone or other personal computing / communications device, connects to the central server via the network and may run an application 1718 for boiler monitoring (as described in relation to Figures 16A to 160). User device 103 is representative of a population of such devices associated with different households / customers and controller 102 is representative of a population of such controllers associated with different households / heating system installations. Sensors 1720 correspond, for example, to the various sensors shown in Figure 1, for measurement of various operating characteristics of heating systems.

Sensor data and other operational data (e.g. demand signals) is received from controllers 102 of multiple households! heating system installations, stored in the sensor data database 1714 and processed to produce training samples for the learning algorithms, the samples including sets of feature values derived from the raw sensor data as discussed previously. The training samples are input to the outlier model 1710 and decision tree learning algorithm 1711 to obtain one or more trained decision trees. Further sensor data (which may be from the same population of boilers or from a different population of boilers) is then received and processed in the same way to generate test samples. The monitoring and diagnostics process 1712 applies the trained decision trees to the test samples to detect faults. Alerts / notifications are sent to user devices and/or other system components when a fault condition is detected.

The diagnostic process may operate essentially in real-time (to monitor sensor data as it is received from the controllers) or alternatively, diagnostics may be performed periodically on collected sensor data. For example, the fault detection may be run once a day on data collected for a particular boiler over the past 24 hours.

While a specific architecture is shown by way of example, any appropriate hardware/software architecture may be employed.

-30 -Furthermore, functional components indicated as separate may be combined and vice versa. For example, the database 1714 may be stored at the server 101 as shown or may be provided as a separate database server. Furthermore, the functions of server 101 may in practice be implemented by multiple separate server devices (e.g. to distribute monitoring for different boiler subpopulations over multiple servers in a server cluster).

Instead of running the diagnostics method at the central server, it could be run locally, e.g. on controller 102, or another local computing device connected to the controller. In that case, the server may push the appropriate decision tree models to the local device after centrally performing the learning processes on data sets received from a large population of boilers.

It will be understood that the present invention has been described above purely by way of example, and modification of detail can be made within the scope of the invention.

For example, while the described algorithm applies a threshold to convert anomaly scores of samples produced by the isolation forest to classifications, with a decision tree then trained to predict those classifications from the sample features, an alternative implementation could use a different prediction model (e.g. a regression model) to predict the isolation forest scores directly (or other values derived from the scores), and thresholding could then be applied to predicted scores to obtain a final fault classification (or the predicted score could be used to provide quantitative information e.g. on fault likelihood).

Furthermore, other outlier/anomaly detection algorithms or models could be used instead of isolation forests (for example based on fitting probability distributions e.g. Gaussian distributions to the data which provide probabilities for given data points in the data, which may correspond to the anomaly scores of the isolation forest, and setting a probability threshold to detect anomalies). Similarly, prediction/classification models other than decision trees could be used.

The above embodiments concern diagnostic techniques applied to a heating system, where the heating system comprises a heating component for heating a -31 -fluid that is circulated through a fluid circulation system. The heating component is typically a boiler and the fluid is typically water. In such systems the fluid circulation system is typically closed, e.g. comprising pipe networks and radiators in closed loops. However, the techniques may be applied to other heating systems (e.g. air heating systems) and/or other heating components (such as heat exchangers and furnaces). In such systems, heated air may be supplied directly into rooms via a network of ducts and thus the circulation system is open with the heated spaces forming part of the circulation system.

The described techniques may be implemented in domestic settings, e.g. for boilers and heating systems installed in homes, as well as in commercial / industrial settings, e.g. for heating systems installed in offices, retail and other commercial buildings and spaces, factories and the like.

The described techniques may also be extended to any temperature control system adapted to manage and control the temperature in an environment, including cooling systems, e.g. where a fluid is cooled and circulated to cool an environment (for example a refrigeration system or air conditioning system).

Claims (25)

1. -32 -CLAIMS1. A computer-implemented method for detecting an anomalous operating state of a temperature control system for controlling a temperature in an environment, the method comprising: receiving operational data indicating operational characteristics of one or more temperature control systems; generating a set of training samples based on the operational data, each training sample comprising values for a plurality of training features derived from operational data associated with a given temperature control system; processing training samples of the set of training samples using an anomaly detection model, the anomaly detection model arranged to output for each training sample an anomaly score that is indicative of whether the training sample represents an anomalous outlier in the set of training samples; associating with each processed training sample an anomaly indication based on the anomaly score assigned to the training sample by the anomaly detection model; training a prediction model on the processed training samples and the associated anomaly indications to obtain a trained prediction model that is arranged to predict a corresponding anomaly indication for a test sample of feature values; and applying the prediction model to identify an anomalous operating state of a temperature control system being monitored, based on further operational data received from the monitored temperature control system.
2. 2. A method according to claim 1, comprising: receiving the further operational data from the temperature control system being monitored and deriving one or more test samples of feature values from the further operational data; applying the trained prediction model to the test samples to obtain predicted anomaly indications for the test samples; and identifying an anomalous operating state of the monitored temperature control system based on the predicted anomaly indications.
3. -33 - 3. A method according to claim 1 or 2, wherein each training or test sample comprises values for a plurality of features derived from operational data associated with a given temperature control system corresponding to a particular measurement time.
4. 4. A method according to any of the preceding claims, wherein the operational data comprises one or more of: sensor data from one or more sensors measuring operating characteristics of a temperature control system; and control data associated with a temperature control system.
5. 5. A method according to any of the preceding claims, wherein the features of the training and test samples comprise one or both of: values of the received operational data or sensor data, and derived values computed from the received operational or sensor data.
6. 6. A method according to any of the preceding claims, wherein the operational data comprises a demand signal indicating a request for activation of a temperature control function, preferably a heating function, of the temperature control system.
7. 7. A method according to claim 6, wherein the features comprise a time duration since the demand signal transitioned to an ON state, the ON state indicating that activation of the temperature control function is being requested.
8. 8. A method according to any of the preceding claims, wherein the operational data comprises temperature measurements obtained using one or more temperature sensors associated with a given temperature control system.
9. 9. A method according to any of the preceding claims, wherein the temperature control system is a heating system comprising a heating appliance such as a boiler for heating water that is circulated through a water circulation system to provide space heating and/or for supplying heated water to hot water outlets, the operational data comprising one or more of: -34 -measurement data indicating a temperature at a supply flow pipe supplying heated water from the heating appliance into the circulation system; measurement data indicating a temperature at a return flow pipe returning water from the circulation system to the heating appliance; measurement data indicating a temperature at a hot water supply pipe for supplying heated water to hot water outlets.
10. 10. A method according to claim 9, wherein the features of the training and/or test samples comprise one or more of: one or more of the measured temperatures; rate of change of one or more of the measured temperatures; a minimum or maximum of a measured temperature over a time window; a temperature increase or temperature decrease of a measured temperature over a time window; an average value or standard deviation of two or more of the measured temperatures.
11. 11. A method according to any of the preceding claims, wherein the operational data and/or features comprises a measure of electricity consumption of a temperature control appliance of the temperature control system.
12. 12. A method according to any of the preceding claims, wherein the anomaly indications are anomaly classifications, and wherein the prediction model is a classification model, optionally wherein the prediction model is a binary classification model classifying a sample as being either a normal sample indicative of normal operation or an abnormal sample indicative of an anomalous operating state.
13. 13. A method according to any of the preceding claims, wherein the associating step associates an anomaly classification with each training sample based on comparison of the anomaly score output by the anomaly detection model to a classification threshold.
14. 14. A method according any of the preceding claims, wherein the prediction model comprises a decision tree.
15. -35 - 15. A method according to any of claims 12 to 14, comprising identifying an anomalous operating state in response to the prediction model outputting an anomaly classification associated with anomalous operation for a test sample.
16. 16. A method according to any of the preceding claims, comprising training a plurality of prediction models, each trained on a subset of the training features, wherein the step of applying the prediction model comprises applying multiple ones or each of the trained prediction models to the test samples, and preferably further comprising identifying an anomalous operating state in response to any of the prediction models outputting an anomaly indication indicative of anomalous operation.
17. 17. A method according to claim 16, wherein each prediction model is associated with a respective fault type, the method comprising, in response to a given one of the prediction models outputting an anomaly indication indicative of anomalous operation, diagnosing a fault of the corresponding fault type and outputting a notification of the diagnosed fault type.
18. 18. A method according to any of the preceding claims, wherein the anomaly detection model comprises one or more partitioning trees, each defining a hierarchy of partitioning operations that recursively partition the training sample set into subsets, with each training sample assigned to a leaf node of the partitioning tree representing a partition of the training sample set, wherein the partitioning operations in each partitioning tree are preferably randomly determined, preferably by random selection of a partitioning feature and/or a split point for the partitioning feature for each partitioning operation.
19. 19. A method according to claim 18, wherein the anomaly detection model assigns an anomaly score to each training sample based on the depths in the partitioning trees of the leaf nodes to which the training sample is assigned.
20. 20. A method according to any of the preceding claims, wherein the anomaly detection model comprises an isolation forest comprising a plurality of isolation trees generated by random partitioning of the training sample set.
21. -36 - 21. A method according to any of the preceding claims, performed at a computing device, the method comprising receiving the operational data from sensors and/or a control system associated with the temperature control system being monitored at the computing device, processing the operational data to identify presence or absence of the anomalous operating state, and transmitting a message from the computing device to another device in dependence on detection of the anomalous operating state.
22. 22. A method according to any of the preceding claims, comprising, in response to identifying presence of an anomalous operating state, outputting a fault alert, notification or control command, optionally comprising performing one or more of: sending a notification to a user device associated with a user of the monitored temperature control system, the notification optionally comprising an electronic message transmitted via a public messaging system or a notification displayed in a monitoring application provided at the user device; sending a control command to the monitored temperature control system to deactivate, or modify an operating configuration of, the temperature control system; and initiating scheduling of a visit by a maintenance engineer to perform maintenance and/or repair of the temperature control system.
23. 23. A computer-implemented method for detecting an anomalous operating state of a heating system, comprising: receiving operational data indicating operational characteristics of a population of heating systems, the operational data including temperature measurements from one or more temperature sensors of the heating systems; generating a set of training samples based on the operational data, each training sample comprising values for a plurality of training features derived from operational data associated with a given heating system; processing training samples of the set of training samples using an isolation forest anomaly detection model, the anomaly detection model arranged to output for each training sample an anomaly score that is indicative of whether the training sample represents an anomalous outlier in the training sample set; -37 -associating with each processed training sample an anomaly classification based on the anomaly score assigned to the training sample by the anomaly detection model; training a decision tree classifier on the processed training samples and the associated anomaly classifications to obtain a trained decision tree classifier that is arranged to predict a corresponding anomaly classification for a test sample of feature values; receiving further operational data including temperature measurements from a heating system being monitored and deriving test samples of feature values from the further sensor data; applying the trained decision tree classifier to the test samples to obtain anomaly classifications for the test samples; and identifying an anomalous operating state of the monitored heating system based on the anomaly classifications.
24. 24. A system having means, optionally comprising one or more processors and associated memory, for performing a method as set out in any of the preceding claims.
25. 25. A computer program, computer program product or tangible computer readable medium comprising software code adapted, when executed by a data processing system, to perform a method as set out in any of claims 1 to 23.