Sensing method and devices

Abstract

A computer implemented method for determining a status of a subject. The method comprises obtaining measurements of a plurality of parameters from a plurality of sensor; using one or more parameter-specific machine learning models to each process the measurements of a respective parameter of the plurality of parameters to derive a respective parameter-specific indication of the status of the subject; using a multi-parameter model to process the one or more parameter-specific indications and measurements of any of the plurality of parameters from which a parameter-specific indication was not derived to derive a multi-parameter indication of the status of the subject; and outputting a signal based on the multi-parameter indication of the status of the subject.

Description

[0001] SENSING METHOD AND DEVICES

[0002] FIELD OF THE INVENTION

[0003] Embodiments described herein relate to methods and systems for determining the status of a subject.

[0004] BACKGROUND

[0005] Standalone sensor devices, such as wireless sensor nodes, or Internet of Things (loT) sensor devices, are typically configured to collect data about their surroundings and to output sensed data, or a signal based on processing sensed data. Such devices may be used in a variety of industries and fields, such as environmental monitoring, industrial automation, or healthcare.

[0006] Such standalone sensor devices may be powered by an on board or local power source, such as a battery or energy harvesting device, enabling them to be arranged to monitor environments where no mains power connection is available. However, sensor devices may require power to perform measurements with their sensors, to transmit data to other devices, particularly data-intensive raw sensor data, and to locally process sensor data, particularly when using machine learning models.

[0007] Therefore, using a local power source limits the extent to which wireless sensor devices can capture, process and transmit data, or the duration for which they may be deployed. Locally powered sensor devices that utilise machine learning models to process sensor data are particularly limited as updating such machine learning models may require wirelessly receiving and implementing large updates which can consume a particularly large amount of energy. Existing locally powered sensor devices typically limit the number of sensors that they use and / or the extent to which they process captured data in order to reduce their energy consumption to acceptable levels.

[0008] SUMMARY OF THE INVENTION

[0009] According to a first aspect of the invention, there is provided a computer implemented method for determining a status of a subject, the method comprising: obtaining measurements of a plurality of parameters from a plurality of sensors; using one or more parameter-specific machine learning models to each process the measurements of a respective parameter of the plurality of parameters to derive a respective parameter-specific indication of the status of the subject; using a multi-parameter model to process the one or more parameter-specific indications and measurements of any of the plurality of parameters from which a parameter-specific indication was not derived to derive a multi-parameter indication of the status of the subject; and outputting a signal based on the multi-parameter indication of the status of the subject.

[0010] Using a separate multi-parameter model and one or more parameter-specific machine learning models advantageously enables individual models to be updated separately, allowing smaller updates to be received and implemented, reducing energy consumption when performing updates.

[0011] In some embodiments the one or more parameter-specific models and the multiparameter model are configured to be updated separately.

[0012] In some embodiments, the method further comprises: monitoring measurements of a first parameter from a subset of the plurality of sensors, while the remainder of the plurality of sensors and / or a processor implementing the multi-parameter model are in an inactive state; and in response to the monitored measurements of the first parameter meeting a first criteria, activating the remainder of the plurality of sensors and / or the processor, and obtaining the measurements of the remainder of the plurality of parameters from the plurality of sensors.

[0013] The remainder of the plurality of sensors and / or a processor being in the inactive state until monitored measurements meet a first criteria may advantageously reduce energy consumption before such a criteria is met.

[0014] In some embodiments, the measurements of the first parameter from the subset of the plurality of sensors may be monitored using a parameter-specific machine learning model.

[0015] In some embodiments, the method is performed by a standalone constrained computing device comprising a local power source, which may be a battery, capacitor, or energy harvesting device.

[0016] In some embodiments, the multi-parameter model is a machine learning model.

[0017] In some embodiments, the parameter-specific indications comprise confidence values indicating a certainty that the subject is in a particular state. In some embodiments, the subject is a pipe and the multi-parameter indication of the status of the pipe is an indication of a rate of fluid flow within the pipe.

[0018] In some embodiments, the subject is a pipe and the multi-parameter indication of the status of the pipe is an indication of whether the pipe is leaking.

[0019] In some embodiments, the subject is a pipe and the multi-parameter indication of the status of the pipe is an indication of whether water within the pipe is at risk of developing legionella bacteria.

[0020] According to a second aspect of the invention, there is provided a device for determining a status of a subject, the device comprising: one or more processors configured to: obtain measurements of a plurality of parameters from a plurality of sensors: use one or more parameter-specific machine learning models to each process the measurements of a respective parameter of the plurality of parameters to derive a respective parameter-specific indication of the status of the subject; and use a multiparameter model to process the one or more parameter-specific indications and measurements of any parameters of the plurality of parameters from which a parameter— specific indication was not derived to derive a multi-parameter indication of the status of the subject. The device further comprising an output interface configured to output a signal based on the multi-parameter indication of the status of the subject; and a local power source.

[0021] In some embodiments, the processor is a low-power neural decision processor optimised for performing machine learning models.

[0022] In some embodiments, the output interface is a wireless transceiver further configured to receive updates to the one or more sensor-specific models and / or the multiparameter model.

[0023] In some embodiments, the device comprises the plurality of sensors.

[0024] In some embodiments, the plurality of sensors are selected from: a subject temperature sensor, an ambient temperature sensor, a wideband audio sensor, and a narrowband audio sensor.

[0025] The device may be configured to perform any of the steps of methods as described herein.

[0026] According to a third aspect of the invention, there is provided a method of manufacturing a device for determining a status of a subject, the method comprising: providing one or more processors, a computer readable storage medium, an output interface, and a local power source; configuring the one or more processors to receive measurements from a plurality of sensors and to provide an output to the output interface according to the computer instructions; and storing computer instructions on the memory, the computer instructions when executed causing the one or more processors to perform a method as described above.

[0027] According to a fourth aspect of the invention, there is provided a method of training a machine-learning system for determining a status of a subject, the method comprising : constructing one or more parameter-specific training data sets each comprising a plurality of training examples, each training example comprising a ground truth measurement of a status of the subject and corresponding measurements of a respective parameter of a plurality of parameters; and using the one or more parameter-specific training data sets to each train a corresponding parameter-specific machine learning model to infer a respective parameter-specific indication of the status of the subject from measurements of the respective parameter of the plurality of parameters.

[0028] In some embodiments, the method of training may further comprise: constructing a multi-parameter training data set comprising a plurality of training examples, each training example comprising a ground truth measurement of the subject and corresponding parameter-specific indications of the status of the subject derived by the one or more parameter-specific machine learning models; and using the multiparameter training data set to train a multi-parameter machine learning model to infer a multi-parameter indication of the status of the subject from at least the one or more parameter-specific indications of the status of the subject inferred by the one or more parameter-specific models.

[0029] In some embodiments, the multi-parameter training data set further comprises measurements of one or more parameters from which a parameter-specific indication is not derived by a respective parameter-specific model; and the multi-parameter machine learning model is trained to infer the multi-parameter indication of the status of the subject from the at least the one or more parameter-indications of the status of the subject inferred by the one or more parameter-specific models and measurements of the one or more parameters from which a parameter-specific indication is not derived by a respective parameter-specific model.

[0030] In some embodiments, the step of using the one or more parameter-specific training data sets to train comprises determining a time interval over which the ground truth measurements are sampled.

[0031] According to a fifth aspect of the invention, there is provided a system comprising a one or more devices according to the second aspect, and a central controller configured to receive the signals output by the one or more devices.

[0032] In some embodiments, the central controller may be configured to post process the signals received from the one or more devices to identify and / or rectify errors therein.

[0033] Particular embodiments of the invention will be understood with reference to the following figures.

[0034] BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Examples of the present invention will now be described in detail with reference to the accompanying drawings, in which:

[0036] Fig. 1 is a diagram of a locally powered sensor device according to an embodiment of the invention;

[0037] Fig. 2a shows a first model structure for deriving an indication of a status of a subject from measurements by a plurality of sensors;

[0038] Fig, 2b shows a second model structure for deriving an indication of a status of a subject from measurements by a plurality of sensors;

[0039] Fig. 3 shows a method of operating a sensor device starting in an inactive state;

[0040] Fig. 4 shows an example of an embodiment of a locally powered device in an inactive state;

[0041] Fig. 5 shows training a multi-part model system using two data acquisition devices (DAQs) to derive an indication of a status of a subject based on parameters measured by a plurality of sensors;

[0042] Fig. 6 is a flowchart showing a method of training a multi-model arrangement to infer an indication of a status of a subject from parameter measurements;

[0043] Fig. 7 shows a training dataset for training a parameter-specific machine learning model; and,

[0044] Fig. 8 shows a third model structure for deriving an indication of a status of a subject from measurements by a plurality of sensors.

[0045] DETAILED DESCRIPTION

[0046] Embodiments described herein relate to methods and devices for monitoring statuses of subjects in which measurements of a plurality of parameters are obtained from a plurality of sensors, one or more parameter-specific machine learning models derive a parameter-specific subject status indication from measurements of a respective parameter, a multi-parameter model derives a multi-parameter subject status indication from the one or more parameter-specific subject status indication and measurements of any of the parameters that were not used to derive a parameterspecific subject status indication; and a signal is output based on the derived multiparameter subject status indication.

[0047] In such embodiments, sensor data is advantageously processed on a sensor device using a multi-stage model configuration. In a first stage, measurements of some or all of the measured parameters obtained from the plurality of sensors are processed into parameter-specific indications of the status of a monitored subject, and in a second stage these parameter-specific indications, as well as measurements of any of the parameters which were not processed by the first stage, are processed by a multiparameter model to produce a multi-parameter indication, based upon which a signal may be output. Such a model configuration advantageously enables individual models within the configuration to be updated separately, allowing smaller updates to be received and implemented, reducing the demands on a devices' lower power source during updates. Additionally, such configurations may enable only parts of the device to be active at certain times, limiting the power consumption of the device during these periods.

[0048] Fig. 1 is a diagram of an example of a sensor device 100 according to an embodiment of the invention, which may be used to perform methods according to other embodiments of the invention.

[0049] The illustrated device 100 comprises a plurality of sensors 110 configured to capture sensor data on the device's environment including a subject, a data processor 120 for processing the captured sensor data to derive an indication of a status of the subject, an output interface 130 for outputting a signal based on the derived indication, and a local power source 140 for powering the device 100. The subject may be a device or system, an environment, a living organism, such as a patient, or a body part thereof, or any other suitable object whose status may change and which may be monitored by suitable sensors.

[0050] It will be appreciated that variation may be made to the device 100 in different embodiments, for example, the number of sensors 110 may be varied and / or one, some, or all of the sensors 110 may not be comprised by the device and instead the device 100 may be configured to receive measurements from said sensors 110, for example, via associated wired connections. Alternatively, or additionally, the device may comprise additional components, such as a microcontroller unit (MCU), a digital signal processor (DSP), one or more additional processors, computer memory, one or more additional local power sources, a housing, one or more user controls, and / or one or more connectors or supports for mounting or installing the device 100 within an environment.

[0051] The sensors 110 may comprise a variety of different types of sensors, such as temperature, humidity, pressure, light, motion or audio sensors. The plurality of sensors 110 may comprise multiple different sensors of the same type, which may be configured to measure the same property of different aspects of the device's environment. For example, a first sensor may measure a property of a subject, and a second sensor may measure the same property of the subject's surroundings. For example, in an embodiment of a sensor device 100 for monitoring fluid flow within a pipe, the plurality of sensors may comprise an ambient temperature sensor, a pipe surface temperature sensor, a wideband audio sensor and a narrowband audio sensor.

[0052] The device is configured to provide measurements of plurality of sensed parameters obtained from the sensors 110 to the processor 120. One, some, or all of the sensed parameters may be measurements made by an individual sensor (such as temperatures measured by an individual temperature sensor). In such cases, parameter-specific models and indications may be sensor-specific models and indications.

[0053] Alternatively or additionally, one, some, or all of the measured parameters may be derived from measurements by a plurality of the sensors (such as a temperature difference between two temperature sensors measuring different locations). Such parameters may be derived from measurements from a plurality of the sensors by a digital signal processor (DSP) which may pre-process measurements from some or all of the plurality of sensors. Such a DSP may be an additional processor comprised by the device 100, or may be integrated with a processor 120 which derives indications of the status of the monitored subject.

[0054] The sensor device 100 and sensors 110 thereof may be non-invasive devices which may be installed to monitor a subject system or device without modifying said subject system or device, for example by being positioned, mounted or installed on or adjacent to said subject system or device. For example, an embodiment of a sensor device 100 for monitoring fluid flow within a pipe may be configured to be clipped onto - or otherwise mounted on or adjacent to - an exterior of the pipe, without requiring modification to the pipe to enable one or more of the sensors access to the pipe's interior.

[0055] The data processor 120 is configured to obtain measurements of each of a plurality of parameters obtained from the plurality of sensors 110 and to process the measurements using to derive an indication of a monitored subject. The data processor 120 is configured to implement at least one parameter-specific machine learning model 120 configured to derive a parameter-specific indication of the status of a monitored subject from measurements of a respective parameter, and a multi-parameter model configured to derive a multi-parameter indication of the status of the subject from the one or more parameter-specific indications and measurements of any parameter which were not used to derive corresponding parameter-specific indications. Various implementations of such models are discussed below with reference to figs. 2a and 2b. It will be appreciated that in alternative devices, such models may be implemented by a plurality of data processors 120.

[0056] The data processors 120 may be a processor specialised for performing machine learning operations, such as a neural network accelerator. The data processor 120 may additionally be specialised for performing machine learning operations in low- power and / or loT devices. For example, the data processor 120 may be a specialised Neural Decision Processor™ (www.syntiant.com / neural-decision-processors) as developed by Syntiant Corp, for use in low-power, always-on, high performance deep neural network edge Al applications. Using a specialised data processor 120 may enable the one or more parameter-specific machine learning models to be performed more efficiently, thereby reducing the power requirements for the device 100. In some embodiments the multi-parameter model may also be a machine learning model which may also benefit from being implemented using the specialised data processor 120. Alternatively, the data processor 120 may be a lower power microcontroller unit (MCU), a field programmable gate array (FPGA), or other suitable processor. In some embodiments, the data processor 120 comprises a digital signal processor (DSP) and a neural network processing unit, such as a deep neural network unit. The DSP may pre-process received sensor measurements, and / or may calculate measurements of one or more of the parameters from measurements of multiple sensors.

[0057] In alternative embodiments, a device 100 may comprise a plurality of data processors 120 as described above which may together implement the one or more parameterspecific machine learning models and the multi-parameter model. In such embodiments, processing tasks, such as different models may be divided between different processors. Each of the data processors 120 may comprise any of the optional features described herein.

[0058] Alternatively, or additionally, the device may comprise one or more additional processors, such as a separate digital signal processor which may pre-process sensor measurements as described above.

[0059] The output interface 130 may be or may comprise a wireless transmitter or transceiver, such as a radio transmitter or receiver, which may be configured to transmit a wireless signal based on the determined multi-parameter indication, for example, using a wireless technology such as Wi-Fi. Bluetooth, Zigbee, a cellular network or another suitable technology. Alternatively, or additionally, the output interface 130 may comprise other forms of transmitter or transceiver, such as a line-of-sight laser transmitter or transceiver, which may be configured to transmit such a signal. The wireless transceiver, or another receiver comprised by the device 100 may receive updates to the device, particularly updates to models implemented using the data processor 120 to derive indications from sensor measurements. Alternatively, or additionally, the output interface 130 may comprise a wired data output or data output port, which may be configured to output a signal based on the derived multi-parameter indication.

[0060] Alternatively, or additionally, the output interface 130 may comprise an output device configured to output a signal based on the derived multi-parameter indication to a user in the presence or vicinity of the device 100. For example, the output interface 130 may comprise one or more lights, displays, loudspeakers, printers, haptic feedback devices, and / or graphical user interfaces (GUIs), which may present a visible, audible, or haptic signal which may be received by a user based on the determined indication of the system. For example, if the indication indicates that the system is in a particular state, a light, loudspeaker, or other element of the output interface may be activated.

[0061] The output interface 130 is configured to output a signal based on the indication of the status of the subject derived by the data processor 120. In some embodiments, the output interface may also be configured to transmit other information, for example to a central controller of a system comprising one or more such devices 100. Such other information may include performance data, or outputs of one or more models implemented by the data processor which may be used to analyse performance of the models and the device 100.

[0062] In some embodiments the output interface 130 may be an input / output interface (or the device 100 may further comprise a separate input interface) which may be configured to receive information, such as configuration settings, instructions or updates for models implemented by the data processor.

[0063] The local power source 140 may comprise one or more batteries, such as rechargeable batteries, (super) capacitors, and / or may comprise one or more energy harvesting devices, such as solar cells, thermoelectric generators and / or piezoelectric generators.

[0064] In some embodiments, a system for monitoring statuses of one or more subjects may comprise a plurality of devices 100 as described herein, which may be arranged to monitor different subjects and / or different locations on one or more given subjects.

[0065] Alternatively, or additionally, such a system may further comprise one or more central or backend controllers, such as servers, which may be configured to communicate with the one or more devices 100 of the system. Such a central controller may be configured to receive the signals based on indications of the subject statuses form the plurality of devices 100 and optionally to receive other information from said devices. Alternatively, or additionally, such central controllers may be configured to transmit signals to the plurality of devices 100, such as configuration settings, instructions or updates for models implemented by their data processors.

[0066] In some embodiments, such a central controller may be configured to perform postprocessing of status indications received from individual devices 100, for example to identify and / or rectify potential errors or short-lived misdetections in status indications received from one or more devices 100. In some embodiments, the central controller may be configured to perform post processing to amalgamate a series of shorter duration status detections (such as a plurality of brief leak detections in a monitored pipe) which have been incorrectly separated, into a longer duration status detection (such as a flush detection in the monitored pipe). This may enhance the detection of longer duration status which may be categorised or treated differently than shorter duration status detections, thereby providing more reliable and accurate monitoring.

[0067] Alternatively, such post processing may be performed by an individual device 100 or a processor 120 thereof. However, in some embodiments the post processing may be more preferably performed by a central controller which may be less power constrained than a sensor device 100 (for example, due to being mains powered instead of battery powered).

[0068] Figs. 2a and 2b are diagrams showing the structure of two multi-part model configurations 200, 205 for use in deriving indications of the status of a monitored subject according to embodiments of the invention, for example using a device 100 as described above with reference to Fig. 1, in which the model configurations 200, 205 may be implemented by the one or more data processors 120.

[0069] Each model configuration 200, 205 is configured to receive measurements of each of a plurality of parameters 210 as inputs. The measurements are obtained from a plurality of sensors 110, which may be all or a subset of the sensors 110 comprised by or providing measurements to the device 100. Each of the illustrated example configurations receives measurements 210 of four different parameters. However, it will be appreciated that measurements of other numbers of parameter may received in different devices 100 and / or in different configurations of a given device 100.

[0070] The measurements 210 of at least one of the plurality of parameters are processed by a corresponding parameter-specific machine learning model 220 to derive a corresponding parameter-specific indication 230 of the status of a monitored subject.

[0071] Each parameter-specific machine learning model 220 may be an artificial neural network (ANN), such as a deep neural network.

[0072] Each parameter-specific machine learning model 220 may use a series of measurements of its corresponding parameter as inputs. For example, the inputs to a model may comprise a series of measurements of the parameter taken at predetermined intervals over a window of time. The intervals between parameter measurements and the duration of windows over which they are taken may depend upon the parameter-specific model and may vary between the different parameters. The measurements 210 of a parameter received as inputs by the configurations 200, 205 may be raw sensor data received directly from the sensors, or may be pre- processed, for example by a digital signal processor (DSP).

[0073] It will be appreciated that the measurements 210 of one, some, or all of the plurality of parameters are processed by corresponding parameter-specific machine learning models 220. The example model configuration shown in Fig, 2a comprises a single parameter-specific machine learning model 220, which processes measurements 210 of one of the parameters, while the remaining parameters' measurements 210 are used as inputs by the multi-parameter model 240, along with the parameter-specific indication 230 derived by the parameter-specific machine learning model 220. In the configuration shown in Fig, 2b, four parameter-specific machine learning models 220 are used and each of the parameters' measurements 210 are processed into a corresponding parameter-specific status indication 230.

[0074] The parameter-specific indications 230 derived by the one or more parameter-specific machine learning models 220 may comprise a binary indication of whether or not the monitored subject is in a given state (for example, whether or not a monitored pipe is leaking), a probability that the monitored subject is in a given state (for example, a probability that a monitored pipe is leaking), an estimated parameter of the monitored subject (for example, an estimation of a flow rate through a monitored pipe), an indication of which of a plurality of states the monitored subject is in (for example, whether a pipe is in a no-flow, normal-flow and leak-flow state), and / or classification probabilities that the monitored subject is in each of a plurality of states (for example, probabilities that a pipe is in each of a no-flow, normal-flow and leak-flow states). The parameter-specific indications 230 may be in the form of output activations of their corresponding machine learning models 220.

[0075] In embodiments comprising multiple parameter-specific machine learning models, the different parameter-specific models may all derive respective values of the same types of indication (for example, they may each derive a respective probability that a monitored pipe is leaking). Alternatively, some or all of the different parameterspecific models may derive different types of indication. In some embodiments, the different inputs of the multi-parameter model may be weighted. In some embodiments, some or all of the parameter-specific machine learning models may output a plurality of indications of the status of the subject. Each parameter-specific machine learning model 220 may be trained on a training data set comprising a plurality training examples, each training example comprising measurements of its respective parameter and a corresponding ground truth measurement of the indicated status of the subject.

[0076] The one or more parameter-specific indications 230 and measurements 210 of any parameters which were not used to derive a parameter-specific indication 230 are received as inputs by the multi-parameter model 240, which processes them to derive a multi-parameter indication 250 of a status of the subject. In some embodiments, the multi-parameter model 240 may also receive copies of some or all of the measured parameters 210 from which parameter-specific indications 230 were derived as inputs.

[0077] The multi-parameter indication 250 of the status of the subject may be a binary indication or probability of whether the subject is in a given state, an estimated parameter of the subject, and / or an indication or probability of which of a plurality of states the subject is in, as described above with reference to the parameter-specific indications. The multi-parameter indication 250 may be the same type of indication as is derived by one, some, or all of the parameter-specific machine learning models 220. In some embodiments, the multi-parameter model may output a plurality of different multi-parameter indications 250 of the status of the subject. The multiparameter indication 250 may be in the form of output activations of the multiparameter model 220.

[0078] Alternatively, the multi-parameter model 240 may derive a different type of indication to the indications derived by the one or more parameter-specific models 220. For example, the multi-parameter model 240 may output an indication 350 of whether or not - or the probability that - a monitored pipe is in a condition risking the development of Legionella bacteria, based on a parameter-specific indication 230 of the flow rate in the pipe and measurements 210 (or parameter-specific indications 230) of a temperature in the pipe.

[0079] The status of the monitored subject that is output by the device 100 using the output interface thereof 130 may be or may comprise the multi-parameter indication 250 derived by the multi-parameter model 240, or may be derived therefrom.

[0080] In some embodiments, the multi-parameter model 240 may be a machine learning model, such that the model configuration 200, 205 is a machine learning model configuration. For example, the multi-parameter model 240 may be an artificial neural network (ANN) such as a deep neural network. In alternative embodiments, the multiparameter model may a traditional or classical non-machine learning model, for example, using a simple algorithm, decision tree, and / or analytically derived mathematical rules to derive the multi-parameter indication from the one or more parameter-specific indications 230 and measurements 210 of any parameters from which no parameter-specific indications were derived.

[0081] In embodiments in which the multi-parameter model 240 is a machine learning model, the multi-parameter machine learning model 240 may be trained on a training data set comprising a plurality training examples, each training example comprising one or more parameter-specific indications (which may be derived by trained parameterspecific models from measurements of respective parameters), measurements of any parameters from which parameter-specific indications are not derived, and a corresponding ground truth measurement of the indicated status of the subject.

[0082] The inputs of the multi-parameter model may further comprise one or more additional variables, such as the outputs of other machine learning models, user set variables, or variables defining a monitored subject (for example, variables specifying the diameter, material, or other properties of a pipe monitored by the subject).

[0083] Model configurations 200, 205 as described above may be implemented using a single data processor 120 of a device 100 as described above in relation to Fig. 1. However, in other embodiments, the model configurations 200, 205 may be implemented across a plurality of different data processors. For example, one, some, or all of the parameter-specific machine learning models may be implemented on separate data processors to a processor implementing the multi-parameter model.

[0084] A multi-part model configuration comprising a separate multi-parameter model and one or more machine learning models as described above can advantageously be retrained and / or updated part-by-part. For example, an individual parameter-specific machine learning model may be updated on a device 100 - for example, after an instance of the model has been further trained by a central system or controller - without updating any other parameter-specific machine learning models, or the multiparameter model. When an individual model is updated, that model, or weights thereof, may be received by the device on which the model arrangement 200, 205 is implemented, for example, using a wireless receiver or transceiver, without transmitting an update to one, some, or all of the other models on that device. This minimises the amount of data that must be received and implemented by the device 100, thereby minimising the amount of energy that must be used by the device 100 in performing the update. Such configurations of models are therefore particularly advantageous on locally powered wireless sensor devices 100, such as devices as described above with reference to Fig. 1.

[0085] The model configurations 200, 205 described above comprise two stages, a first stage of one or more parameter specific models, and a second state consisting of the multiparameter model. However, it will be appreciated that in other embodiments the model configuration may comprise three or more stages, which may include a first stage of one or more intermediate models, a second stage consisting of one or more multiparameter models, and one or more additional stages comprising one or more additional models. Fig. 8 shows an example of such a model configuration 800, consisting of a first stage with four parameter-specific machine learning models 820, an intermediate stage with two multi-parameter models 840, and a third stage with the final model 860. The parameter-specific models 820 each process measurements of a single parameter 810 to derive parameter-specific indications 830. The multiparameter models 840 each process parameter specific indications 830 and parameter measurements 810 to derive multi-parameter indications 850. The final model 860 processes the multi-parameter indications 850 and parameter measurements 810 to derive a final indication 870. An output interface may output a signal based on the final indication (and by extension upon the multi-parameter indications output by the multi-parameter models).

[0086] Model configurations for use in identifying a state of a monitored subject have been described above with reference to Figs. 2a and 2b. In some embodiments, such model configurations may only be used during an active state of a wireless sensor device, in which measurements of multiple parameters are obtained from a plurality of sensors 110 and are processed to derive a multi-parameter indication of the status of a monitored subject. Such methods or devices may further utilise an inactive state in which some or all of the models described above are not implemented and used to derive indications of the status of the subject. Such an inactive state may be used to conserve power until it is "interrupted" resulting in the method or device transitioning into the active state and deriving an indication of the status of the monitored subject as described above.

[0087] In some embodiments, in addition to outputting an indication of the status of a monitored subject given by or derived from the multi-parameter indication 250, the device 100 may be configured or configurable to output some or all of the parameter specific indications 230 derived by its parameter-specific machine learning models 220. For example, to transmit said parameter specific indications 230 to a central controller. The parameter specific indications 230 may be output periodically, or in response to a certain status of the subject being determined from the multi-parameter indication 250 (for example, when the multi-parameter indication 250 makes a positive detection of a certain status, such as a detection that a monitored pipe is leaking). The outputting of the parameter specific indications 230 may be optimised to be power efficient, for example by compressing the parameter specific indications 230, such as from Sbit to 2bit.

[0088] Transmitting individual parameter-specific models' parameter specific indications 230 to a central controller may enable an analysis of the individual parameter-specific models 220 to be performed, for example to determine if they are behaving as intended or unexpectedly in unfamiliar environments or other situations, such as by outputting a indication that conflicts or contradicts with indications output by the other models 220, 230. For example, if an audio-parameter-specific model determines that a monitored pipe is leaking while other models determine that it is not, transmitting its indications may enable its sub-optimal performance to be identified and potentially corrected, for example by providing an update or replacement for that specific model. Such functionality may therefore aid in making informed decisions regarding the provision of updates to parameter-specific models 220, enhancing the overall adaptability and effectiveness of the system in varying conditions.

[0089] Additionally, providing the parameter specific indications 230 to a central controller may allow for an evaluation of the multi-parameter model's reasoning.

[0090] In some embodiments, the subject monitored by a method, device 100 and / or model configuration 200, 205 outlined above may be a pipe. In some embodiments, the multi-parameter indication output by the multi-parameter model may be comprise indication of a flow rate through the pipe. Such an indication may be derived based on measured parameters of a temperature of the pipe, a temperature difference between the pipe and its surroundings, sounds in the vicinity of the pipe, and / or vibrations of the pipe.

[0091] Alternatively, or additionally, the multi-parameter indication may comprise an indication of whether the pipe is leaking or of a category of such a leak. The indication that the pipe is leaking may be (or may be based on) an indication that the pipe has a non-zero flow rate, a flow rate above some threshold flow rate (such as 0.5 litres per minute) or a flow rate within some range of flow rates or may be an indication that the pipe has had such a flow rate for more than a threshold period of time. When a pipe is leaking it will have a continual flow rate as its contents moves to exit the pipe. The magnitude of this continual flow rate depends upon the size of the leak, with small or 'pinhole' leaks resulting in continual flow rates that are below those of the pipe in use. The indication may differentiate between such small leaks and larger gushing leaks, for example, based on a determined flow rate.

[0092] In some embodiments, the multi-parameter indication may comprise an indication of whether the pipe is in a flush state. A flush is a flow-event of a certain duration, which may be based on user requirements, for example 3 minutes or 7 minutes. Such flushes may be performed periodically, such as once per week, and embodiments may be configured to identify such events as one continuous flow and not as a leak.

[0093] Alternatively, or additionally, the multi-parameter indication may comprise an indication of whether the monitored is in a condition risking the development of Legionella bacteria in water therein. Such an indication may be (or may be based on) an indication that the flow rate through the pipe has been zero, or below some threshold flow rate, for more than a threshold period of time. Legionella bacteria can develop within water that is still or very slow flowing for an extended period of time, allowing bacteria to accumulate and create biofilms on the interior surface of pipes.

[0094] It will be appreciated that in alternative embodiments, methods, devices 100 and / or model configurations as described herein may monitor other subjects, such as other infrastructure elements, pieces of machinery, or biological organisms. The sensors and indications used may vary depending upon the subject.

[0095] Fig. 3 is a flowchart 300 of an embodiment of a method of operating a sensor device, such as a device 100 as described above, including an inactive state. The device begins 310 in an inactive state, in which measurements of one or more monitored parameters are obtained 320 from a monitored subset of a plurality of sensors and in which one or more other sensors and / or processors are in a low-powered, non-powered state to conserve power. Whether the one or more monitored parameters measured by the monitored sensors satisfy interruption requirements is determined 330, and if they do not, the device remains inactive and the subset of sensors continue to be monitored 310.

[0096] If the interruption requirements are satisfied by the one or more monitored parameters, the device transitions 340 to the active state, and the low- or non-powered sensors and / or processors are powered. Within the active state, measurements of a plurality of parameters are obtained 350 from the plurality of low- or non-powered sensors, one or more parameter-specific status indications are derived 360 using parameter-specific machine learning models, a multi-parameter indication is derived 370 using the multi-parameter model and a signal is output 380 based on the determined multi-parameter indication.

[0097] In the inactive state, some of a device's sensors may be in a low- or non-powered 'sleeping' state, for example, all of the sensors except the one or more monitored sensors may be in such a state. In some embodiments, all but one of the plurality of sensors may be low- or non-powered and the single other sensor may be monitored. Using such low- or non-powered states may advantageously minimise energy use by the sensors in the inactive state.

[0098] The remaining sensors, which may be or may include the monitored sensors may be in a powered and / or sampling state. At least one of the powered and / or sampling sensors is a monitored sensor. The one or more monitored sensors are powered and take measurements, these measurements provide the one or more monitored parameters based on which whether to transition to the active state is determined. Some or all of the powered, sampling, and / or monitored sensors may be in a low- powered state, such as a state in which they are not passed a clock signal by a processor or other device.

[0099] The one or more monitored sensors may be sensors that are more power efficient in a their powered and / or sampling state - and / or which have smaller power savings in their low- or non-powered state - compared to the unmonitored sensors which are in a low- or non-powered state. The one or more monitored sensors may be more sensitive to valid interruption criteria and / or robust against unintentional interruption detections than the unmonitored sensors which are in a low- or non-powered state.

[0100] Alternatively, or additionally, one or more processors of the device performing the method may be low-powered or unpowered in the inactive state, which may also advantageously minimise energy use. These processors may be or may include some or all of the processors that implement the multi-parameter and parameter-specific models in the active state. For example, the one or more unpowered processors may include the processor which implements the multi-parameter model in use and / or the one or more processors that implement the parameter-specific models of parameters measured by any unpowered sensors.

[0101] In some embodiments, at least one processor remains in a powered and / or sampling state and is configured to evaluate whether the measurements of the one or more monitored parameters obtained from the one or more powered monitored sensors satisfy one or more activation requirements to trigger an "interrupt" and transition the device or method to the active state. The powered / and / or sampling state of this at least one processor may be a low-power state. In some embodiments, this processor may be a processor which implements the multi-parameter model and / or one or more of the parameter-specific models in use. Alternatively, it may be a different processor, such as a digital signal processor (DSP) which may remain powered while the main model-implementing processor is unpowered. This processor may act as a gatekeeper device. Alternatively, no processors may remain in a sampling or powered state. In such embodiments, whether the one or more monitored parameters satisfy the activation requirements may be evaluated by the one or more monitored sensors themselves.

[0102] The measurements of the one or more monitored parameters are processed to determine whether they satisfy some activation criteria and if they do, the device or method transitions to the active state. In some embodiments, the measurements taken by a powered sensor may be processed by comparing a value of individual measurements, or of a plurality of measurements, to one or more thresholds or other criteria. For example, the device or method may transition to the active stage if a monitored audio sensor detects more than a threshold amount of energy in a certain frequency range.

[0103] Alternatively, or additionally, the measurements taken by a powered sensor may be processed using a machine learning model, which may derive a parameter-specific indication of the status of the subject therefrom. Such a parameter-specific machine learning model may be one of the parameter-specific machine learning models used in the active state, or a different model, thereby utilising the multi-part nature of the active model arrangement to further minimise power consumption of the device. The derived parameter-specific indication of the status of the model may be compared to one or more thresholds or other criteria.

[0104] The one or more monitored parameters may be one or some of the parameters of which measurements are obtained and processed in the active state, such parameters may indicate a possibility that the system is in a given status being monitored for in the actives state and the analysis performed in the inactive state may indicate that there is some minimal threshold probability that the monitored subject is in that state, such that it is worthwhile to perform the full processing of the active state.

[0105] In embodiments in which multiple parameters are monitored in the inactive state, the method or device may transition to the active state if any one of the monitored parameters satisfies a respective requirement or may only transition to the active state if all the monitored parameters satisfy respective requirements.

[0106] Transitioning to the active state may comprise powering the unpowered sensors or processors, and / or loading the one or more sensor-specific machine learning models and the multi-parameter model onto one or more processors. Transitioning to the active state may consume a relatively large amount of energy and it may therefore be desirable not to switch back and forth between the active and inactive states too frequently.

[0107] The method or device may transition from the active state to the inactive state after a predetermined period of time has elapsed since transitioning from the inactive state and / or after a predetermined period of time has elapsed in which the determined indications of the status of the system indicate a that the subject has been in a certain state (such as not being in the state indicated by the interruption). Alternatively, or additionally, within the active state, one or more additional "deactivation" models and / or criteria may be applied to one or more measured parameters and may be configured to trigger a transition from the active state to the inactive state.

[0108] Methods as described above utilise an inactive state in conjunction with an active state in which a multi-parameter indication is determined from measurements of multiple sensors using a multi-part model configuration with one or more parameter specific machine learning models and a separate multi-parameter model. In alternative aspects of the invention an inactive state as described herein may be utilised in conjunction with an active state employing a monolithic model, such as a monolithic machine learning model.

[0109] Fig. 4 shows an example of an embodiment of a device 400 operating in an inactive state. The device 400 comprises four sensors 410, a digital signal processor (DSP) 415, a neural network accelerator processor 420, a microcontroller unit (MCU) 425, and a radio transceiver 430. In the inactive state, three of the sensors 410 and the neural network accelerator processor 420 are in "sleeping" low-powered states. One of the sensors 410 is in a "sampling" state in which it takes measurements of a parameter and the digital signal processor (DSP) 415 receives the measurements and compares them to an activation threshold. When the DSP 415 determines that the activation threshold has been met, the device 400 transitions to the active state and the sensors 410 and neural accelerator processor 420 all become powered.

[0110] In the active state, the digital signal processor (DSP) 415 receives measurements made by all four of the sensors 410 and pre-processes them and provides measurements of a plurality of parameters measured by the sensor to the neural accelerator data processor 420 which implements a multi-part model configuration as described above with reference to figs. 2a and 2b to derive an indication of the status of a monitored subject. The indication of the status of the monitored subject is transmitted by the wireless transceiver 430 which defines an output interface of the device 400. The radio transceiver 430 may also receive updates to the models implemented by the device.

[0111] The microcontroller unit (MCU) 425 may control the overall operation of the device, may control the operation of the radio transceiver, may signal the state of the device, and / or may control the distribution of computer instructions implementing the machine learning models from a computer memory of the device to the neural network accelerator processor 420 when the device transitions to the active state.

[0112] Fig. 5 shows an example of training a system including multi-part model and an inactive state as described above with reference to figs. 2a, 2b, and 3, which utilises machine learning models to derive parameter-specific indications of the status of a subject and to trigger a transition to the active state. A plurality of sensors are arranged to measure parameters of an example subject and / or its surroundings and a reference sensor device is arranged to measure a ground truth status of the example subject.

[0113] When training models for monitoring a pipe, an invasive in-line flow sensor may be provided within the pipe in order to provide a ground truth measurement of a flow rate within the pipe (or of whether or not the pipe has a leak, in which case there will be a low flow rate, or not, in which case there will be no flow rate) and a plurality of non- invasive sensors may be arranged to monitor parameters of the exterior and / or surroundings of the pipe. For example, a pair of temperature sensors may be arranged to measure a temperature of the pipe and of the surrounding environment respectively, together providing a temperature difference parameter, a wideband audio sensor may provide measurements of a sounds in the vicinity of the pipe and a narrowband audio sensor or accelerometer may provide measurements of vibrations of the pipe itself.

[0114] When fluid flows within a pipe, the temperature of that pipe may change relative to its surroundings, thereby allowing a flow of liquid within a pipe to be monitored using temperature sensors as described above. In normal operations, after water ceases to flow within the pipe, the temperature of the pipe may gradually return to the level of its surroundings. However, if a leak is present within the pipe, the temperature may return more slowly, and may not reach the temperature of its surroundings.

[0115] The parameter measurements produced by the sensors could be provided by a slave data acquisition device (DAQ) and a master data acquisition device (DAQ) as shown in Fig. 5, or a combined data acquisition device (DAQ). Such a slave DAQ may collect parameter measurements by sensors in 'sleeping' or low-power states for use in training an interruption model for use in an inactive state. The master DAQ may collect parameter measurements by 'active' or sampling sensors for use in training one or more models to derive indications of the status of the system in an active state. A combined DAQ may collect parameter measurements from both 'active' and 'sleeping' sensors for training both types of model.

[0116] Fig. 6 is a flowchart showing a method of training a multi-model arrangement to infer an indication of the status of a system. The method comprising constructing 610 at least one parameter-specific training data sets, and training 620 at least one parameter specific machine learning model to infer a parameter-specific indication of the status of the system from measurements of a respective parameter using said training data set. Fig. 7 shows an example of such a training data set including a plurality of training examples, each including a ground truth status of the system and one or more associated parameter measurements. The parameter measurements in a training example may be a series of measurements captured over time windows which may be specific to each of the machine learning models.

[0117] After training 630 the at least one parameter-specific machine learning models, the method further comprises configuring a multi-parameter model to infer a multiparameter indication of the status of the subject from a parameter-specific probability for a status from each trained parameter-specific machine learning model (and optionally from parameters from which a parameter-specific indication is not obtained). In some embodiments, the multi-parameter model may be a machine learning model, which may be configured in an additional training step and over its own time window.

[0118] According to a further embodiment, there is provided a method of manufacturing a device as described herein. Such a method may comprise providing one or more processors, a computer readable storage medium, an output interface, and a local power source and any other components of the device. The one or more processors may be configured to receive measurements from a plurality of sensors and to provide an output to the output interface according to the computer instructions. Computer instructions may be stored on the memory, or within the one or more processors, the computer instructions when executed causing the one or more processors to perform a method as described herein.

[0119] While certain embodiments have been described above, these embodiments have been presented by way of example, and it will be appreciated that various omissions, substitutions, changes or additions may be made without departing from the scope of the invention as defined by the claims.

Claims

CLAIMS1. A computer implemented method for determining a status of a subject, the method comprising: obtaining measurements of a plurality of parameters from a plurality of sensors; using one or more parameter-specific machine learning models to each process the measurements of a respective parameter of the plurality of parameters to derive a respective parameter-specific indication of the status of the subject; using a multi-parameter model to process the one or more parameter-specific indications and measurements of any of the plurality of parameters from which a parameter-specific indication was not derived to derive a multi-parameter indication of the status of the subject; and outputting a signal based on the multi-parameter indication of the status of the subject.

2. A method according to claim 1 wherein the one or more parameter-specific models and the multi-parameter model are configured to be updated separately.

3. A method according to claim 1 or claim 2, further comprising: monitoring measurements of a first parameter from a subset of the plurality of sensors, while the remainder of the plurality of sensors and / or a processor implementing the multi-parameter model are in an inactive state; and in response to the monitored measurements of the first parameter meeting a first criteria, activating the remainder of the plurality of sensors and / or the processor, and obtaining the measurements of the remainder of the plurality of parameters from the plurality of sensors.

4. A method according to claim 3, wherein the measurements of the first parameter from the subset of the plurality of sensors are monitored using a parameterspecific machine learning model.

5. A method according to any preceding claim, performed by a standalone constrained computing device comprising a local power source.

6. A method according to any preceding claim, wherein the multi-parameter model is a machine learning model.

7. A method according to any preceding claim, wherein the parameter-specificindications comprise confidence values indicating a certainty that the subject is in a particular state.

8. A method according to any preceding claim, wherein the subject is a pipe and the multi-parameter indication of the status of the pipe is an indication of a rate of fluid flow within the pipe.

9. A method according to any of claims 1 to 7, wherein the subject is a pipe and the multi-parameter indication of the status of the pipe is an indication of whether the pipe is leaking.

10. A method according to any of claims 1 to 7, wherein the subject is a pipe and the multi-parameter indication of the status of the pipe is an indication of whether water within the pipe is at risk of developing legionella bacteria.

11. A device for determining a status of a subject, the device comprising: one or more processors configured to: obtain measurements of a plurality of parameters from a plurality of sensors; use one or more parameter-specific machine learning models to each process the measurements of a respective parameter of the plurality of parameters to derive a respective parameter-specific indication of the status of the subject; and, use a multi-parameter model to process the one or more parameterspecific indications and measurements of any parameters of the plurality of parameters from which a parameter-specific indication was not derived to derive a multi-parameter indication of the status of the subject; an output interface configured to output a signal based on the multi-parameter indication of the status of the subject; and, a local power source.

12. A device according to claim 11, wherein the processor is a low-power neural decision processor optimised for performing machine learning models.

13. A device according to claim 11 or claim 12, wherein the output interface is a wireless transceiver further configured to receive updates to the one or more sensorspecific models and / or the multi-parameter model.

14. A device according to any of claims 11 to 13, comprising the plurality of sensors.

15. A device according to claim 14, wherein the plurality of sensors are selected from: a subject temperature sensor, an ambient temperature sensor, a wideband audio sensor, and a narrowband audio sensor.

16. A non-transitory computer readable storage medium comprising computer instructions that when executed by one or more processors, cause the one or more processors to perform a method according to any of claims 1 to 10.

17. A method of manufacturing a device for determining a status of a subject, the method comprising: providing one or more processors, a computer readable storage medium, an output interface, and a local power source; configuring the one or more processors to receive measurements from a plurality of sensor inputs and to provide an output to the output interface according to the computer instructions; and, storing computer instructions on the memory, the computer instructions when executed causing the one or more processors to perform a method according to any of claims 1 to 10.

18. A method of training a machine-learning system for determining a status of a subject, the method comprising: constructing one or more parameter-specific training data sets each comprising a plurality of training examples, each training example comprising a ground truth measurement of a status of the subject and corresponding measurements of a respective parameter of a plurality of parameters; and, using the one or more parameter-specific training data sets to each train a corresponding parameter-specific machine learning model to infer a respective parameter-specific indication of the status of the subject from measurements of the respective parameter of the plurality of parameters.

19. A method according to claim 18, further comprising: constructing a multi-parameter training data set comprising a plurality of training examples, each training example comprising a ground truth measurement of the subject and corresponding parameter-specific indications of the status of the subject derived by the one or more parameter-specific machine learning models; and, using the multi-parameter training data set to train a multi-parameter machinelearning model to infer a multi-parameter indication of the status of the subject from at least the one or more parameter-specific indications of the status of the subject inferred by the one or more parameter-specific models.

20. A method according to claim 19, wherein the multi-parameter training data set further comprises measurements of one or more parameters from which a parameterspecific indication is not derived by a respective parameter-specific model, and wherein the multi-parameter machine learning model is trained to infer the multi-parameter indication of the status of the subject from the at least the one or more parameter- indications of the status of the subject inferred by the one or more parameter-specific models and measurements of the one or more parameters from which a parameterspecific indication is not derived by a respective parameter-specific model.

21. The method of any of claims 18 to 20, wherein the step of using the one or more parameter-specific training data sets to train comprises determining a time interval over which the ground truth measurements are sampled.

Images