Method of estimating filter lifetime

Abstract

A method of estimating S130 a lifetime of a particulate filter (120, fig. 2) installed in an apparatus (100, fig. 2) using retrieved datasets including: a first dataset S110 indicative of performance

Description

Field of the Invention The present invention relates to a method of estimating filter lifetime for use with an apparatus with a particulate filter. Background Particulate filters are known to be used, for example, in applications of air purification. An air purifier may draw in ambient air, channel said air through a suitable particulate filter to remove particulates carried by said air, and release filtered air from the air purifier. Particulate filters have a finite design lifetime following which filter loading may adversely affect performance. Filter loading refers to a build-up of particulates retained in the particulate filter, i.e. particulates filtered out by the particulate filter, and may eventually reduce performance. In such a case, performance may be restored by replacing the particulate filter, e.g. with a cleaned particulate filter or a new replacement filter. A priori estimation of when the particulate filter will reach the end of its design lifetime may be difficult, however, because the design lifetime is dependent on the conditions under which a given particulate filter is used. These conditions may differ significantly, especially for domestic applications, due to a wide range of air quality and use cases. This may make blanket predictions of the filter lifetime in terms of hours of operating time unsatisfactorily inaccurate. Yet an accurate estimation of filter lifetime would ensure that product operation is optimised and that users can be confident that the air purifier is cleaning ambient air in a cost efficient and resource efficient way. The present invention has been devised in light of the above considerations. Summary of the Invention Increased accuracy for estimating filter lifetime may be achievable using a pressure sensor configured to monitor a pressure drop across a particulate filter. However, addition of a pressure sensor adds cost and complexity. This may be undesirable, since it would mean an increase in upfront cost of a filter apparatus and hence offset, at least in part, any reduction in operating cost of the filter apparatus. Moreover, additional maintenance may be necessitated. Another alternative for estimating filter lifetime may utilise monitoring of the performance of a motor generating airflow through the particulate filter. For example, current draw of the motor may be monitored as part of the control process and will decrease overtime fora given setting as the particulate filter clogs. However, current draw is affected also by ambient atmospheric conditions, such as temperature, humidity, barometric pressure. The changes in atmospheric conditions may affect motor performance to a similar level as filter loading. Another alternative for estimating filter lifetime may utilise a particle sensor as this will give measurement of the air quality that the filter is operating in. However, the interrelation of particle counts and filter life is complex and dependent on a range of additional factors, such as temperature, humidity, aerosol material (oil, solid), charge state, size distribution (instantaneous and overtime), meaning the data from only a particle counter may only provide an estimation of the particulate matter removed by the filter. The present inventors devised a method of combing datasets from two sensors, monitoring motor performance and particle count respectively, in an air purification apparatus to more accurately infer the current status of the filter. Through the combination of two datasets a prediction of the filter lifetime / state may be generated that has a greater accuracy. Moreover, this improved method may be implementable at no additional hardware cost, since some examples of known air purification apparatuses may already be provided with sensors for monitoring motor performance and for monitoring particle number. According to a first aspect of the invention, there is provided a method of estimating a lifetime of a particulate filter installed in an apparatus, the method comprising: retrieving a first dataset indicative of a motor performance of a motor for generating an airflow through the particulate filter over a period of time since an installation of the particulate filter in the apparatus; retrieving a second dataset indicative of an amount of particulates in the airflow over the period of time since the installation of the particulate filter in the apparatus; using the first dataset and the second dataset to generate an estimate of a lifetime of the particulate filter. By utilising the first dataset and the second dataset for generating the estimate of the lifetime of the particulate filter, the accuracy of the estimate may be improved without relying on a pressure sensor. The method may further include determining that the estimate of the lifetime of the filter has reached a design lifetime; and providing an indication to a user that the lifetime has been reached. The period of time since the installation of the particulate filter in the apparatus may include the entire time since the installation of the particulate filter in the apparatus, but the first dataset and the second dataset may contain data corresponding to one or more lengths of time since the installation of the particulate filter. For example, no data may be recorded for periods of time in which the apparatus is switched off. Similarly, recording of the first dataset and the second dataset may not commence immediately with the installation of the particulate filter since in some examples the apparatus is switched off for replacing the particulate filter. However, since in this case no airflow is generated through the newly installed particulate filter, this time period from the installation to first operation following the installation may not have an effect on the lifetime of the particulate filter. The method may further include obtaining a subset of the first dataset indicative of the motor performance at the time of installation of the particulate filter in the apparatus; and using the subset indicative of the motor performance at the time of installation of the particulate filter when generating the estimate of the lifetime of the particulate filter. By recording the subset of the first dataset which is indicative of the motor performance at the time of installation, a baseline value of motor performance may be established. Comparing against said baseline value of motor performance may further increase accuracy of the estimate of filter lifetime. For example, thus the impact of manufacturing tolerances may be reduced. The method may further include recording a data indicative of the motor performance at the time of installation of a replacement filter in the apparatus. By recording data indicative of motor performance at the time of installation of the replacement filter, the impact of machine wear on the estimate of filter lifetime may be reduced since the effect of machine wear on the baseline value and on subsequent motor performance throughout the lifetime of a given filter may be comparable. The method may further include calculating an estimate of the airflow through the particulate filter over the time interval based on at least a subset of the first dataset corresponding to the time interval; and calculating an estimate of the number of particulates ingested by the apparatus over the time interval based on the estimate of the airflow and at least a subset of the second data set corresponding to the time interval; and using the estimate of the number of particles ingested by the apparatus when generating the estimate of the lifetime of the particulate filter. By using the estimate of the number of particles ingested when generating the estimate of the lifetime of the particulate filter, the accuracy of the estimate of the lifetime of the particulate filter may be improved. The particle number estimate over the time interval may be calculated based on at least a subset of the second dataset corresponding to the time interval, for example as a time average or integration. The method may further include applying a low-pass filter to the first dataset. The first dataset may include data indicative of the motor performance at a relatively high frequency, e.g. at least once per minute in operation, resulting in the first dataset capturing short-term variations in the motor performance. Variations in the motor performance as a result of atmospheric conditions may also occur on a relatively short timescale, e.g. days or weeks, and therefore be captured in the first dataset. By contrast, filter loading may occur on a relatively long timescale, e.g. months to years. Application of a low-pass filter to the first dataset may therefore reduce the effect of short-term trends and help identify longterm trends considered more likely to be associated with filter loading. The method may further include using the first dataset and the second dataset to generate the estimate of the lifetime of the particulate filter by: determining the estimate of the lifetime using a look-up table. The look-up table may be produced using previously acquired experimental data, e.g. experimental data which shows how actual lifetime of the particulate filter varies with experimentally obtained first and second datasets. The method may further include using the first dataset and the second dataset to generate the estimate of the lifetime of the particulate filter by: determining the estimate of the lifetime using a trained machine-learned model that has been trained with labelled training data. The labelled training data may include labelled pairs of first and second datasets (e.g. as discussed in more detail below). The trained machine-learned model may be a Gaussian Processes Regression model. The method may further include retrieving a temperature dataset indicative of an ambient temperature of the apparatus since the installation of the particulate filter in the apparatus; and using the temperature dataset when generating the estimate of the lifetime of the particulate filter. The method may further include retrieving a barometric pressure dataset indicative of an ambient barometric pressure of the apparatus since the installation of the particulate filter in the apparatus; and using the barometric pressure dataset when generating the estimate of the lifetime of the particulate filter. The method may further include retrieving a humidity dataset indicative of an ambient humidity of the apparatus since the installation of the particulate filter in the apparatus; and using the humidity dataset when generating the estimate of the lifetime of the particulate filter. By tracking the changes in ambient conditions, the effect of ambient conditions on motor performance may be considered when generating the estimate of filter lifetime. Such ambient conditions may include temperature, barometric pressure and humidity. The method may further include generating the second dataset using a sensor configured to detect particulates of 4.5 microns or smaller in diameter. The sensor to detect particulates may be configured to detect particulates of 4.5 microns or smaller; may be configured to detect particulates of 3.5 microns or smaller; or may be configured to detect particulates of 2.5 microns or smaller. By utilising a sensor to detect particulates of relatively small size, e.g. those having a diameter of 4.5 microns or smaller, accuracy of the estimate of filter lifetime may be improved. By contrast, a particle sensor configured to detect relatively large particulates, e.g. 10 microns or larger, may provide a relatively inaccurate estimate of particulates suspended in the ambient air and therefore lead to reduce the accuracy of an estimate based thereon. This may be considered surprising, since a particulate sensor detecting particulates of a diameter of, say, 10 microns or smaller could be viewed as providing a better sample than a particulate sensor detecting particulates of a diameter of, say, 2.5 microns or smaller. However, the inventors found that from smaller particulate sizes a better estimate of particulate numbers may be derivable. There may be provided an air purification system including an air purification apparatus comprising the particulate filter, the motor, a first sensor system configured to acquire the first dataset and a second sensor system configured to acquire the second dataset; and the air purification system configured to carry out the method as described above. The motor may be configured to drive a rotor at a fixed speed and the first dataset indicative of the motor performance corresponds to power draw of the motor. The air purification apparatus may be configured to carry out the method as described above. Alternatively, the air purification system may comprise a remote server configured to carry out the method as described above. According to another aspect of the invention, there is provided a method of training a machine learning model for use in the method as described above; including: obtaining multiple pairs of first and second datasets, wherein each pair of first and second datasets includes: a first dataset indicative of a motor performance of a motor for generating an airflow through the particulate filter over a period of time since an installation of the particulate filter in an apparatus; a second dataset indicative of an amount of particulates in the airflow over the period of time since the installation of the particulate filter in the apparatus; producing labelled training data by labelling each pair of first and second datasets to associate those first and second datasets with an actual lifetime of the particulate filter determined in relation to the particulate filter used to obtain the pair of first and second datasets; using the labelled training data to train a machine learning model for calculating an estimated lifetime of a particulate filter based on a pair of first and second datasets. The method of training the machine learning model may include obtaining further labelled datasets to be used fortraining of the machine-learned model. The further labelled datasets may correspond to datasets indicative of ambient temperature, humidity or barometric pressure. The invention includes the combination of the aspects and preferred features described except where such a combination is clearly impermissible or expressly avoided. Summary of the Figures Embodiments and experiments illustrating the principles of the invention will now be discussed with reference to the accompanying figures in which: Figure 1 is an illustration of an air purification apparatus. Figure 2 is another illustration of the air purification apparatus, also showing certain internal components. Figure 3 is an illustration of an air purification system including the air purification apparatus and a remote server. Figure 4 illustrates a method of estimating a lifetime of a particulate filter of the air purification apparatus. Figure 5 shows a schematic representation of generating an estimate of filter lifetime based on motor performance. Figure 6 shows a schematic representation of generating an estimate of filter lifetime based on the number of particulates. Figure 7 shows a schematic representation of an ensemble estimate combining the estimates of Figures 5 and 6. Figure 8 is a schematic representation of generating an estimate of the lifetime of the particulate filter based on a single model. Figure 9 illustrates a method of training a machine learning model for estimating the lifetime of the particulate filter. Detailed Description of the Invention Aspects and embodiments of the present invention will now be discussed with reference to the accompanying figures. Further aspects and embodiments will be apparent to those skilled in the art. Figures 1 and 2 illustrate an exemplary air purification apparatus 100 configured to carry out a method according to the invention. In use, the apparatus 100 draws in ambient air and expels purified air by generating an airflow 110 through a particulate filter 120. The particulate filter 120 is configured to retain particulates 1000, e.g. dust and other contaminants, carried in the airflow 110. Suitably, the apparatus 100 comprises a motor 130 and a rotor 140 driven by the motor 130 to generate the airflow 110. In combination, the motor 130 and the rotor 140 may also be referred to as a compressor. In this example, the particulate filter 120 is located upstream of the motor 130 and the rotor 140. The apparatus 100 comprises a first sensor system 150, a second sensor system 160, and a third sensor system 170. The first sensor system 150 is configured to generate a first dataset indicative of a motor performance of the motor 130 when generating the airflow 110 through the particulate filter 120. The first dataset is a measure of efficiency of generating the airflow 110 over a period of time since an installation of the particulate filter 120 in the apparatus 100. The first sensor system 150 includes at least one sensor. In this example, the first sensor system 150 is a single sensor arranged to monitor electrical power consumption by the motor 130, current draw. The second sensor system 160 is configured to generate a second dataset indicative of a number of particulates 1000 in the airflow 110 over the period of time since the installation of the particulate filter 120 in the apparatus 100. In this example, the second sensor system 160 is a single sensor arranged to sample the particulates 1000 in the airflow 110. More particularly, the second sensor system 160 is a PM2.5 sensor configured to detect particulates having a diameter of 2.5 microns or less. The third sensor system 170 is configured to generate a third dataset. The third dataset includes a temperature dataset indicative of an ambient temperature of the apparatus 100 over the period of time since the installation of the particulate filter 120 in the apparatus 100. Suitably, the third sensor system 170 includes a temperature sensor. Further, the third dataset includes a humidity dataset indicative of an ambient humidity of the apparatus 100 over the period of time since the installation ofthe particulate filter 120 in the apparatus 100. Suitably, the third sensor system 170 includes a humidity sensor. Figure 3 illustrates an air purification system 10. The air purification system 10 includes the apparatus 100 and a remote server 200. In this example, the method of estimating a lifetime ofthe particulate filter 120 is performed by the remote server 200 based on the first dataset and the second dataset generated by the apparatus 100 and transmitted to the remote server 200. The remote server 200 also retrieves a barometric dataset indicative of an ambient barometric pressure ofthe apparatus 100 since installation ofthe particulate filter 120. In this example, the barometric dataset is obtained not by an onboard sensor ofthe apparatus 100, but is instead obtained using location information ofthe apparatus 100 and retrieved by the remote server 200 from a suitable source, such as an online repository of weather data. In this example, the location information ofthe apparatus 100 is provided directly to the remote server 200 through a smart device of a user ofthe apparatus 100. Figure 4 illustrates the method of estimating the lifetime ofthe particulate filter 120 installed in the apparatus. The method comprises a first step S110 of retrieving the first dataset indicative of the motor performance of the motor 130 when generating the airflow 110 through the particulate filter 120 over the period of time since the installation of the particulate filter 120 in the apparatus 100. The method comprises a second step S120 of retrieving the second dataset indicative of the amount of particulates in the airflow 110 overthe period of time since the installation of the particulate filter 120 in the apparatus 100. It is noted that the first step S110 and the second step S120 are not be carried out in any particular order, i.e. the first step S110 may be performed before the second step S120, or the second step S120 may be performed before the first step S110, or they may be performed simultaneously. In this example, the first dataset and the second dataset are gathered by the apparatus 100 and transmitted to the remote server 200. The remote server 200 stores these datasets, e.g. temporarily, and retrieves the stored datasets for purposes of generating the estimate of lifetime of the particulate filter 120 as part of a third step S130 of the method. The third step S130 of the method is using the first dataset and the second dataset to generate the estimate of the lifetime of the particulate filter 120. The estimate of the lifetime of the particulate filter 120 may be generated in various ways, of which examples are discussed below, but before doing so the underlying considerations providing relevant context are discussed. Filter lifetime is dependent upon the amount of particulate matter that is removed from the air by the apparatus 100. Overtime, particulates clog the particulate filter 120 and affect performance of the particulate filter 120. This is manifested as a rise in pressure drop across the particulate filter 120, which in turn is reflected by a change in motor performance when generating the airflow 110. In this example, the motor 130 is configured to drive the rotor 140 at a fixed running speed for each setting (e.g. “silent” at a low speed, “normal” at a medium speed, “turbo” at a high speed). As filter loading occurs, a gradual reduction in the power drawn by the motor 130 at a given running speed is observed. However, the instantaneous current draw also depends on ambient conditions, e.g. local temperature and humidity. The instantaneous current draw also depends on barometric pressure. While the temperature and humidity of the air are measured locally by the third sensor system 170, the apparatus 100 includes no barometric sensor. The local air density can vary, e.g. due to natural fluctuations in barometric pressure or due to differences in the local elevation with respect to sea level of the apparatus 100. It may therefore not be possible to determine whether the instantaneous measure of the current drawn by the motor 130 is influenced primarily by filter loading, or local changes in the atmospheric pressure. This makes it difficult to use the current draw alone to accurately determine the instantaneous filter state. While filter loading occurs on the time scale of months to years, local variations in atmospheric conditions may occur on the scale of days or weeks. The first dataset therefore may contain fluctuations not representative of filter loading. By contrast, from the second dataset indicative of particulates 1000 in the airflow 110 a gradually changing profile may be derived, for example by estimating a total number of particulates 1000 overthe period of time based on individual sampling at discrete times. Thus, estimating based on a combination of the first dataset and the second dataset may provide valuable synergy. For example, the first dataset and the second dataset can be cross-referenced to identify any rapid fluctuations in the filter load estimate based on the first dataset. Likewise, a filter loading estimate based on the second dataset is easily biased by assumptions made about the aerosol composition. Using the first dataset, filter loading as a function of time based on the second dataset can be compared to the change in power consumption as a function of time based on the first dataset. If these two metrics do not align, adjustments can be made to the filter loading estimate based on the second dataset to obtain a more accurate estimate of the filter status. The first sensor system 150 monitors power consumption of the motor 130 as part of the motor control process. The second sensor system 160 samples particulates 1000 to feedback room air quality to the user and for automated control of the motor setting (e.g. silent / normal / turbo). As such, these sensor systems and the data gathered by the sensor systems are unrelated to each other. Each of these could be used to infer a change in filter state: the current draw will decrease over time for a given setting as the filter clogs; the total number of particles cleaned from a room can be estimated from machine on time, setting and measured particle counts. However, neither of these datasets alone is considered to provide a sufficiently accurate indicator of filter loading / status. As noted above, current draw measurements are affected by ambient conditions (temperature, humidity, barometric pressure) and machine wear. Equally, the particulate matter counts offer low resolution measurement and are only an estimation of the particulate matter removed by the particulate filter 120. Moreover, the relationship between filter pressure rise and particle size / amount is complex and influenced by ambient conditions. Using the PM2.5 measurements also allows an estimation of the amount of material that the filter has removed from the air. This is implemented by integrating the product of airflow 110 through the apparatus 100 by the instantaneous particulate reading. In this way a quantity for the total amount of pollution that the particulate filter 120 has been exposed to over the operating time of the particulate filter 120 can be estimated. This approach is limited due to the resolution of the PM2.5 sensor of the second sensor system 160, and the particular assumptions made relating the PM2.5 concentration to the global concentration of all pollutants in the local environment of the apparatus 100. Some household aerosols can be generated that would either not be identified by the PM2.5 sensor, but which would nevertheless result in filter loading (e.g., cooking smoke or burnt toast) or would skew the total distribution of particulates. Moreover, changes in the local atmospheric pressure also affects the rate of filter loading and the mass flow through the machine. Filter loading is sensitive to the face velocity of the aerosols. As the local density changes due to fluctuations in the atmospheric pressure, the velocity of the air through the apparatus 100 changes, adjusting the face velocity of aerosols impinging on the particulate filter 120. The total volume of air ingested by the apparatus 100 also changes with respect to local changes in the barometric pressure and corresponding air density. This adds further uncertainty to the filter loading prediction based only on PM2.5 readings and the operating point of the apparatus 100. Hence, the output of the second sensor system 160 alone is considered insufficient to accurately quantify the actual particulate matter challenging the particulate filter 120. There is a valuable synergy in combining these two filter life estimation techniques. The integration approach used in the PM2.5 estimation provides a naturally smooth and gradually changing profile. This can be cross-referenced against any rapid fluctuations in the filter load estimate provided by the current loading method. In this way fluctuations due to changing barometric pressure can be easily identified. Using appropriate statistical methods, it becomes possible to determine the confidence in the instantaneous readings meaning the lag or delay time associated with a low pass filter can be significantly reduced. If the location of the device changes - owner moved to a new location at different elevation - it will be possible to compare the different filter loading metrics and identify inconsistency. As a change in elevation will result in a sustained change in the local operating conditions, it is possible to identify by the respective changes in the two filter loading estimates. In this example, the method further includes applying a low pass filter to the first dataset representative of the instantaneous current draw to separate the influence of local changes in barometric pressure from the long-term trend introduced by the particulates 1000 loading the particulate filter 120. It is noted that application of the low-pass filter results in a lag time (or ‘delay’) for obtaining a reliable estimate of filter loading based on the first dataset following application of the low-pass filter. Thus, the estimate of the filter lifetime based on the first dataset following said application may not be accurately inferred for a period of days or even months after the particulate filter 120 has become fully loaded. Using appropriate statistical methods, it is possible to determine the confidence in the instantaneous readings meaning the lag time associated with a low-pass filter can be significantly reduced. In this example, the method of estimating the lifetime of the particulate filter 120 further includes obtaining a subset of the first dataset indicative of the motor performance at the time of installation of the particulate filter in the apparatus. That is to say, the first dataset includes a subset relating to when the particulate filter 120 was newly installed and had experienced no or little filter loading. This subset of the first dataset therefore may be utilised as a baseline value or reference value when generating the estimate of the lifetime of the particulate filter. In this example, the method further includes using a look-up table (e.g. trained using experimental data) to calculate an estimate of the airflow 110 through the particulate filter 120 over the time period based on at least a subset of the first dataset corresponding to the time period and calculating an estimate of the number of particulates ingested by the apparatus over the time interval based on the estimate of the airflow and at least a subset of the second data set corresponding to the time period. The estimate of the number of particles ingested by the apparatus 100 is used when generating the estimate of the lifetime of the particulate filter 120. Figures 5 and 6 show schematic representations of generating independent lifetime estimates based on the first dataset of the motor performance and the second dataset of the amount of particulates, respectively. The apparatus 100 has no differential pressure sensor from which the airflow 110 through the particulate filter 120 can be derived directly. However, it is possible to infer the amount of particulates 1000 that have been removed by looking at the historical records generated by the second sensor system 160. Additionally, the current drawn by the motor 130 can be used to infer the pressure drop across the particulate filter 120. By benchmarking the current draw on the apparatus 100 when the particulate filter 120 is installed, e.g. a 'clean filter state' is detected. Monitoring the current draw allows an algorithm to estimate the lifetime of the particulate filter 120. This may optionally also include altitude, temperature and humidity sensors. Combining this with the PM2.5 logs, will allow for a more accurate measurement of system performance and thus filter life, without directly measuring the pressure drop across the filter. Figure 7 shows a schematic representation of generating a lifetime estimate based on the outputs of Figures 5 and 6 in an ensemble model providing an ensemble filter status. The ensemble filter status according to Figure 7 makes use of the estimates of the lifetime provided by both of the filter status estimation techniques of Figures 5 and 6 to provide a more accurate inference of the filter lifetime. The ensemble filter status can be an instantaneous estimate, making use of only the most recent outputs from the two independent filter status models. Alternatively, the ensemble filter status can contain a memory block to record a window of historical filter state estimates. In this way temporal patterns and changes can be used to further enhance the capabilities of the filter state estimate. The choice of appropriate algorithm or machine learning model chosen for this purpose may be based on the required accuracy, computational cost and memory footprint. Candidate machine learning models include decision trees, random forests, XGBoost, Gaussian Process Regression, as well as neural network approaches. For neural architectures, historical information can be included explicitly through number of input channels, or by implementing a convolutional input layer to encode the historical input vectors. Alternatively recursive elements (RNN) can be used where an LSTM (Long Short-Term Memory) or GRU (Gated Recurrent Unit) can be used to enable historical data to influence instantaneous filter status estimates. Figure 8 is a schematic representation of generating an estimate of the lifetime of the particulate filter 120 based on a single model which takes as inputs the atmospheric conditions (e.g. temperature and relative humidity); the motor performance (e.g. current draw, voltage and coil temperature); the second dataset (e.g. the PM2.5 sampling data); a target running speed of the rotor 140 and a current running speed of the rotor 140; as well as any other measurements performed by onboard sensor systems of the apparatus; and optionally information provided from an external server. The single model of Figure 8 can provide a consolidated estimate of the filter status, establishing the relationships between each of the data inputs to provide an optimised functional model for filter status estimation. The choice of appropriate algorithm or machine learning model chosen may be based on the required accuracy, computational cost and memory footprint. Candidate machine learning models include decision trees, random forests, XGBoost, Gaussian Process Regression, as well as neural network approaches. For neural architectures, historical information can be included explicitly through number of input channels, or by implementing a convolutional input layer to encode the historical input vectors. Alternatively recursive elements (RNN) can be used where an LSTM or GRU can be used to enable historical data to influence instantaneous filter status estimates. In some examples machine learning may be utilised. In such examples where machine learning is utilised, using the first dataset and the second dataset to generate the estimate of the lifetime of the particulate filter may include determining the estimate of the lifetime using a trained machine-learned model that has been trained with labelled pairs of first datasets and labelled second datasets. In some examples, the trained machine-learned model may be a Gaussian Processes Regression model but any suitable machine-learned model may be utilised. Training a machine-learning model may use a dataset of labelled examples. This data can be generated via a range of methods. Particulate filters 120 can be loaded overtime in representative conditions with accurate pressure drop measurements performed at regular intervals to obtain accurate filter restriction values. Data can also be generated using accelerated testing where particulate filters 120 are challenged with abnormally high levels of filter loading and accurate measurements are made of the actual pressure rise across the particulate filter 120 as a function of filter loading and PM2.5 ingested by the apparatus 100. Another approach to collecting data can be based on artificially blocking particulate filters 120 to a known degree using masking of the filter area equivalent to a certain filter restriction level. Collecting experimental data can be expensive and limit the amount of data available in a short time frame. Synthetic data can be generated to augment the number of examples and expand the coverage of the data set to include extreme examples of atmospheric pressure variations for example. Synthetic data can be generated using computer simulation models to simulate the current draw of the motor 130 fora given machine tolerance profile, atmospheric conditions and filter restriction. The models are trained to minimise a given loss function. The loss function is chosen to ensure that when converged the algorithm performance is maximised in terms of accuracy of predicting the current filter status and remaining lifetime. In optimising the machine learning model different model hyperparameters can be tuned, as well as evaluating the importance of different input signals to the model. A method of training a machine learning model may include obtaining multiple pairs of the first datasets and the second datasets. The pairs of datasets are obtained under the same conditions, e.g. the same apparatus 100. More particularly, each pair of first and second datasets includes a first dataset indicative of a motor performance of a motor for generating an airflow through the particulate filter over a time period since an installation of the particulate filter in an apparatus, and a second dataset indicative of an amount of particulates in the airflow over the time period since the installation of the particulate filter in the apparatus. The method of training the machine learning model may further include producing labelled training data by labelling each pair of first and second datasets to associate those first and second datasets with an actual lifetime of the particulate filter determined in relation to the particulate filter used to obtain the pair of first and second datasets. The method of training the machine learning model further includes using the labelled training data to train a machine learning model for calculating an estimated lifetime of a particulate filter based on a pair of first and second datasets. The first sensor system 150 described above utilises a single sensor to monitor the motor performance. In other examples, multiple sensors may be utilised. These sensors may monitor the same physical quantity, such as power or current drawn by the motor 130. Alternatively, the sensors may monitor different physical quantities, each indicative of motor performance, such as a combination of power and current drawn by the motor 130, the rotational frequency of the rotor 140, or others. The second sensor system 160 described above utilises a single sensor to sample the airflow 110. In other examples, the sensor 160 is not located in the airflow 110 but in or at the apparatus 100 such that air indicative of particulates in the airflow 110 is sampled. Moreover, multiple sensors may be provided. Some of these sensors may be located within the apparatus 110, others may be arranged on the outside or even externally. The air purification system 10 described above utilises the remote server 200 for determining the lifetime of the particulate filter 120. According to other examples, this is done at the apparatus 100. The location information relating to the apparatus 100 is transmitted directly to the remote server 200 by a user’s smart device in the example described above. In other examples, the location information may be provided to the apparatus 100 and transmitted by the apparatus 100 to the remote server 200. Also, the smart device is merely an example of suitable means for obtaining location information, but alternatively other means may be utilised, such as an onboard location sensor of the apparatus 100 or location information may be inferred on the basis of details of an internet connection of apparatus 100. The apparatus 100 described above sees a decrease in power required to maintain mass airflow rates as filter loading occurs. In other examples, an increase in power consumption may occur as a result of a different rotor arrangement. In yet further examples, airflow may be generated according to different parameters, e.g. at fixed power. In this case, a drop in mass airflow occurs overtime as a result of filter loading and the rotational frequency of the rotor may change. Hence, motor performance may in this case be monitored through rotational frequency, e.g. rotations per minute, rather than power consumption. The method of estimating filter lifetime as described above allows operating a purifier the necessary control methods and algorithms to determine an improved estimate of the filter lifetime. For example, this may be achieved by calculating the filter pressure drop and stage through the lifetime of the filter. Once the filter has removed pollutants representative of its design lifetime, the apparatus can then highlight the need to exchange or maintain the filter to the user. The features disclosed in the foregoing description, or in the following claims, or in the accompanying drawings, expressed in their specific forms or in terms of a means for performing the disclosed function, or a method or process for obtaining the disclosed results, as appropriate, may, separately, or in any combination of such features, be utilised for realising the invention in diverse forms thereof. While the invention has been described in conjunction with the exemplary embodiments described above, many equivalent modifications and variations will be apparent to those skilled in the art when given this disclosure. Accordingly, the exemplary embodiments of the invention set forth above are considered to be illustrative and not limiting. Various changes to the described embodiments may be made without departing from the spirit and scope of the invention. For the avoidance of any doubt, any theoretical explanations provided herein are provided for the purposes of improving the understanding of a reader. The inventors do not wish to be bound by any of these theoretical explanations. Any section headings used herein are for organizational purposes only and are not to be construed as 5 limiting the subject matter described. Throughout this specification, including the claims which follow, unless the context requires otherwise, the word “comprise” and “include”, and variations such as “comprises”, “comprising”, and “including” will be understood to imply the inclusion of a stated integer or step or group of integers or steps but not the exclusion of any other integer or step or group of integers or steps. 10 It must be noted that, as used in the specification and the appended claims, the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise. Ranges may be expressed herein as from “about” one particular value, and / or to “about” another particular value. When such a range is expressed, another embodiment includes from the one particular value and / or to the other particular value. Similarly, when values are expressed as approximations, by the use of the antecedent “about,” it 15 will be understood that the particular value forms another embodiment. The term “about” in relation to a numerical value is optional and means for example + / -10%.

Claims

1. A method of estimating a lifetime of a particulate filter installed in an apparatus, the method comprising:retrieving a first dataset indicative of a motor performance of a motor for generating an airflow through the particulate filter over a time period since an installation of the particulate filter in the apparatus, the first dataset generated by a first sensor system comprising a sensor arranged to monitor electrical power consumption by the motor;retrieving a second dataset indicative of an amount of particulates in the airflow over the time period since the installation of the particulate filter in the apparatus, the second dataset generated by a second sensor system comprising a sensor arranged to detect particulates having a diameter of 2.5 microns or less;applying a low-pass filter to the first dataset, the low-pass filter arranged to reduce any effects of short-term trends on any long-term trends introduced by particulates loading the particulate filter; and,using the first dataset and the second dataset to generate an estimate of a lifetime of the particulate filter.

2. The method of claim 1, further comprising:determining that the estimate of the lifetime of the filter has reached a design lifetime; and providing an indication to a user that the lifetime has been reached.

3. The method according to any preceding claim, further comprising:obtaining a subset of the first dataset indicative of the motor performance at the time of installation of the particulate filter in the apparatus; andusing the subset indicative of the motor performance at the time of installation of the particulate filter when generating the estimate of the lifetime of the particulate filter.

4. The method of claim 3, further comprising:recording a data indicative of the motor performance at the time of installation of a replacement filter in the apparatus.

5. The method of according to any preceding claim, further comprising:calculating an estimate of the airflowthrough the particulate filter over the time period based on at least a subset of the first dataset corresponding to the time period; andcalculating an estimate of the number of particulates ingested by the apparatus over the time period based on the estimate of the airflow and at least a subset of the second data set corresponding to the time period; andusing the estimate of the number of particles ingested by the apparatus when generating the estimate of the lifetime of the particulate filter.

6. The method according to any preceding claim,wherein using the first dataset and the second dataset to generate the estimate of the lifetime of the particulate filter comprises:determining the estimate of the lifetime using a trained machine-learned model that has been trained with labelled pairs of first datasets and labelled second datasets.

7. The method according to claim 6,wherein the trained machine-learned model is a Gaussian Processes Regression model.

8. The method according to any preceding claim, further comprising:retrieving a temperature dataset indicative of an ambient temperature of the apparatus since the installation of the particulate filter in the apparatus; andusing the temperature dataset when generating the estimate of the lifetime of the particulate filter.

9. The method of claim 1, further comprising:retrieving a barometric pressure dataset indicative of an ambient barometric pressure of the apparatus since the installation of the particulate filter in the apparatus; andusing the barometric pressure dataset when generating the estimate of the lifetime of the particulate filter.

10. The method according to any preceding claim, further comprising:retrieving a humidity dataset indicative of an ambient humidity of the apparatus since the installation of the particulate filter in the apparatus; andusing the humidity dataset when generating the estimate of the lifetime of the particulate filter.

11. An air purification system including:an air purification apparatus comprising the particulate filter, the motor, a first sensor system configured to acquire the first dataset and a second sensor system configured to acquire the second dataset; andthe air purification system is configured to carry out the method of any previous claim.

12. The air purification system of claim 11, wherein the air purification apparatus is configured to carry out the method of any previous claim.

13. The air purification system of claim 11 or 12, wherein the motor is configured to drive a rotor at a fixed speed and the first dataset indicative of the motor performance corresponds to power draw of the motor.5 14. The air purification system of claim 11 further comprising a remote server, wherein theremote server is configured to carry out the method according to any one of claims 1 to 10.

15. A method of training a machine learning model for use in a method according to any one of claims 1 to 10, wherein the method includes:10 obtaining multiple pairs of the first datasets and the second datasets, wherein each pair offirst and second datasets includes:a first dataset indicative of a motor performance of a motor for generating an airflow through the particulate filter over a time period since an installation of the particulate filter in an apparatus;15 a second dataset indicative of an amount of particulates in the airflow over the timeperiod since the installation of the particulate filter in the apparatus;producing labelled training data by labelling each pair of first and second datasets to associate those first and second datasets with an actual lifetime of the particulate filter determined in relation to the particulate filter used to obtain the pair of first and second datasets;20 using the labelled training data to train a machine learning model for calculating anestimated lifetime of a particulate filter based on a pair of first and second datasets.

Images