WO2023084218A1 - A system for taking and analysing air samples - Google Patents

Abstract

A system (200) for taking and analysing air samples is presented. The system includes an air sampling unit (210), a sample processing unit (220), an imaging unit (230), and a transport mechanism (240). The air sampling unit receives an air filter cartridge having an air filter and a slide, and to passes a volume of air through the air filter. The sample processing unit (220) applies a solvent onto the air filter to obtain a sample matrix on the slide. The imaging unit (230) obtains an image of the sample matrix. The transport mechanism (240) takes the air filter cartridge and moves it from one unit to another within the system. A processor processes the sample matrix image and calculates a concentration of contaminant particles. An air filter cartridge for use with the system is also presented.

Description

A SYSTEM FOR TAKING AND ANALYSING AIR SAMPLES

Field of the Disclosure

The present disclosure relates to a system for taking and analysing air samples, and in particular to a system for analysing air samples and identify the presence of airborne contaminant particles such as asbestos fibres.

Background

The conventional technique for measuring airborne fibre levels, for instance a level of asbestos fibres, requires a trained laboratory analyst to obtain the air sample, prepare it, and follow a specific procedure to identify and count asbestos fibres using phase-contrast microscopy images of the sample. This manual technique is inherently slow and relatively expensive. While current detection limit is sufficient to establish a transient level of cleanliness following asbestos removal works, the manual method is still unsuitable for environmental monitoring for instance on demolition sites or building indoor air quality, in which fibre levels are typically 400-500 times lower. Electron microscopy techniques can achieve this level of sensitivity, but cannot be carried out on-site.

W02018071958 describes an air quality monitor for counting respirable fibres such as asbestos fibres. The system includes a robotic microscope platform, a sample imaging apparatus, and a computing apparatus. However, the system requires the air to be manually sampled, and the sample to be processed by hand before introduction to the microscopy component.

It is an object of the disclosure to address one or more of the above mentioned limitations. Summary of the disclosure

According to a first aspect of the disclosure there is provided a system for taking and analysing air samples, the system comprising an air sampling unit adapted to receive an air filter cartridge having an air filter and a slide, and to pass a volume of air through the air filter, a sample processing unit adapted to apply a solvent onto the air filter to obtain a sample matrix on the slide; an imaging unit adapted to obtain an image of the sample matrix; a transport mechanism adapted to take the air filter cartridge and move it from one unit to another within the system; and a processor configured to process the sample matrix image and to calculate a concentration of contaminant particles.

For instance, the air sampling unit may be adapted to pass a measured or pre-defined volume of air.

For instance, the sample processing unit may be adapted to spray the solvent onto the air filter to obtain the sample matrix on the slide.

Optionally, the air sampling unit comprises an air flow channel adapted to receive the air filter cartridge, and an air pump coupled to the air flow channel, the air pump being adapted to control the volume of air flowing through the air filter.

Optionally, the air flow channel comprises a first channel portion, a second channel portion, and an actuator adapted to move the first and second channel portions between an open state and a closed state, wherein in the close state, the first channel portion and the second channel portion form a seal around the air filter. Optionally, the air sampling unit comprises a container for storing a plurality of air filter cartridges and a cartridge dispenser adapted to insert the air filter cartridge between the first channel portion and the second channel portion.

Optionally, the sample processing unit comprises a pump connectable to a solvent reservoir, a solvent heating chamber, and a spray enclosure.

Optionally, the spray enclosure is provided with a slit aperture for receiving the air filter cartridge.

Optionally, the sample processing unit comprises the solvent reservoir. For instance the solvent reservoir may be a reservoir for storing acetone.

Optionally, the imaging unit comprises a light source, an optical apparatus, and a camera. For instance the light source may have an electromagnetic spectrum covering visible and near infrared wavelengths.

Optionally, the imaging unit is adapted to perform phase contrast microscopy.

Optionally, the imaging unit is adapted to capture images with a magnification factor equal or greater than 100 fold. For example between 200 and 600 folds. For instance with a magnification of about 500 folds.

Optionally, the transport mechanism comprises a translation stage coupled to a motion device.

Optionally, the motion device comprises a gripper for gripping the air filter cartridge. Optionally, the system comprises a storage unit for storing the sample matrix. For instance the storage unit may be a container for storing air filter cartridges that have been processed and analysed.

Optionally, the contaminant particles comprise fibres. For instance the fibres may be asbestos fibres or plastic microfibres or mineral fibres including refractory ceramic fibres.

Optionally, the contaminant particles comprise at least one of exhaust emission particulates, nanoparticles, and air spores.

Optionally, the processor is configured to execute a machine learning algorithm to identify countable contaminant particles and non-countable contaminant particles and to provide a number of countable contaminant particles.

Optionally, the machine learning algorithm is configured to perform of a pixel-level image classification method in which every pixel belongs to a specific class.

Optionally, the image classification method is a deep learning image classification method.

Optionally, the machine learning algorithm is trained using a set of training data comprising images of fibres manually annotated for differentiating between asbestos and non-asbestos fibres.

Optionally, the processor is configured to derive a toxicity index based on fibre characteristics comprising at least one of a fibre size and a fibre aspect ratio distribution. Optionally, the processor may be configured to generate one or more of a digital output signal, an encrypted certificate, and a text alert.

Optionally, the processor is configured to identify false positives among the countable contaminant particles.

According to a second aspect of the disclosure there is provided an air filter cartridge for use with a system for taking and analysing air samples according to the first aspect, the air filter cartridge comprising an air filter attached to a casing; and a slide movable between a first configuration in which the filter and the slide are non-overlapping and a second configuration in which the filter and the slide are overlapping.

Optionally, the air filter cartridge comprises a mechanical arrangement operable between a first state to hold the slide within in a first portion of the casing and a second state to release the slide to a second portion of the casing.

For instance the mechanical arrangement may comprise a first member and a second member.

Detailed Description

The disclosure is described in further detail below by way of example and with reference to the accompanying drawings, in which: figure 1 is a diagram of a system for analysing air samples according to the disclosure; figure 2 is a perspective view of an exemplary embodiment of the system of figure 1; figure 3A is a perspective front view of a housing for the system of figure 2; figure 3B is a perspective rear view of the housing for the system of figure 2; figure 4A is a top view, front view and side view of a cartridge for use in the system of figure 2; figure 4B is a front view the cartridge of figure 4A; figure 4C is a side view the cartridge of figure 4A; figure 4D is a cross sectional view of the cartridge of figure 4A; figure 4E is perspective view of a cartridge held by a gripper; figure 4F is a perspective view of the cartridge interacting with a system for moving the slide inside the cartridge; figure 5 is a flow chart illustrating the operation of the system of figure 2; figures 6A-6D are a series of diagrams illustrating the operation of a cartridge dispenser; figure 7 A is a perspective view of an air sampling assembly; figure 7B is a perspective view of the air sampling assembly integrated with the cartridge dispenser of figure 6; figure 8A is a cross sectional view of the air sampling assembly of figure 7 in a closed configuration; figure 8B is a cross sectional view of the air sampling assembly of figure 7 in an open configuration; figure 9 is a perspective view of a chemical processing unit; figure 10 is a flow chart of a method for analysing an air sample;

Figure 11A shows the pixel-level segmentation of asbestos fibres;

Figure 11B is a confusion matrix for the pixel-level segmentation of asbestos fibres;

Figure 12 shows the selection of classification threshold for pixel-level segmentation of asbestos fibres;

Figure 13 shows examples of fibre-level segmentation of asbestos fibres;

Figure 14 shows the Fibre length, width and aspect ratio (right) distributions obtain for a set of samples; Figure 15 shows the selected learning rate profile and the corresponding training curve during step 1 and step 2 of the training;

Figure 16 shows the confusion matrix for the predictions of the trained model on the validation set for a qualitative algorithm;

Figure 17 shows example predictions and confidence scores from the validation set.

Figure 1 illustrates a system for analysing air samples according to the disclosure. The system 100 includes an air sampling unit 110, a sample processing unit 120, an imaging unit 130, a transport mechanism 140 and a processor 150. The air sampling unit 110, also referred to as air sampler is adapted to receive an air filter cartridge having an air filter and a slide, and to pass a measured volume of air through the air filter. The volume of air may be a pre-defined volume of air. For instance the air sampling unit may be adapted to measure the volume of air to pass a measured pre-defined volume through the air filter. The sample processing unit 120 also referred to as sample processor, is adapted to apply a solvent onto the air filter to obtain a sample matrix on the slide, for instance by spraying the solvent. The imaging unit, also referred to as imaging system 130, is adapted to obtain an image of the sample matrix. The transport mechanism 140 is adapted to take the air filter cartridge and move it from one unit to another within the system. The processor 150 is configured to process the sample matrix image and to calculate a concentration of contaminant particles. A controller 160 is provided to control the air sampling unit 110, the sample processing unit, the imaging unit and the processor.

The system 100 is fully automated hence enabling airborne contaminant monitoring, such as fibres monitoring, to be conducted without the intervention of a trained specialist and improve the accuracy, speed, sensitivity and reliability of the results provided. Figure 2 is a perspective view of an exemplary implementation of the system of figure 1. The system 200 includes an air sampling unit 210, a sample processing unit 220, an imaging unit 230, a transport mechanism 240 and a processor 250.

The air sampling unit 210 includes a container or magazine 211 for storing a plurality of air filter cartridges, an air pump 212 coupled to a sampling stage 213 for receiving a cartridge. The sampling stage 213 is located upstream of the air pump 212. A cartridge dispenser (not shown) is coupled to the magazine 211 to position a cartridge onto the sampling stage 213. The cartridge dispenser may be configured to recognise an identifier such as a chip provided on each cartridge.

The sample processing unit 220 includes a solvent reservoir 221, a pump for pumping the solvent into a heating chamber, and a spray enclosure 222 in which the solvent is sprayed onto the air filter. The solvent reservoir 221 is user-replaceable and may have a volume of about lOOmL for storing acetone or another solvent.

The imaging unit 230 includes a phase-contrast microscope 231, a light source and a camera (not shown). The light source may have a electromagnetic spectrum extending beyond visible radiation into the near infrared. In this case the camera is adapted to detect visible as well as near- 1R spectrum. Extending the light source and camera into the near-lR spectrum may be used to improve the quantitative and qualitative analysis of the contaminant particles, for instance fibre characterisation. The imaging unit 230 is configured to perform automatic focussing and scanning of the sample slides. The microscope may be adapted to provide relatively high magnification for instance a magnification of x 100 or more, for example a 500 fold magnification. The imaging unit 230 may also be adapted to perform a quality check of the sample matrix. For example the imaging unit 230 may be configured to perform machine vision to check for filter clearance and integrity of sample for example by identifying the presence of bubbles, particulate coverage, moisture misting, or damages. Then a suitable sampling area can be identified for counting the contaminant particles. Otherwise the sample can be rejected outright as being unable to be analysed.

The transport mechanism 240 is used for moving a cartridge to different units within the system. The transport mechanism 240 may include a translation stage 241 provided with a gripper 242 for holding the cartridge. In operation the translation stage 241 moves the gripper 242 along a translation axis to a desired location. The gripper 242 is actuated to grab and release the cartridge to and from the desired unit. The translation stage 241 may be implemented as a gantry. The gripper 242 has a pair of actuatable jaws for clamping or releasing the cartridge. Additional mechanisms may be provided for instance to rotate the gripper to flip the cartridge.

A controller 260 is provided to control the operation of the air sampling unit, 210, the sample processing unit 220, the imaging unit 230, the transport mechanism 240 and the processor 250.

The system 200 may also be provided with a geo-location tracker to minimise a risk of mis-use or fraudulent reporting. The tracker would be used to show the position of a cartridge of interest at the time of sampling.

The system 200 may also be adapted to permit the introduction and analysis of samples obtained externally. External samples may be obtained from either wearable sampling devices used for worker inhalation exposure evaluation, or from several ambient air sample locations for multi-point environmental monitoring. The external sampling devices may include a custom-designed sampling head and sampling pump. In this scenario secure transfer of air sampling data may be provided to the parent analysis system 200 to ensure measurement integrity.

The sampling pump of the wearable sampling device records the sample duration and flow rate and retain this data within the individual sample cartridge. The cartridge is then manually removed from the sample pump and inserted into the analytical device system 200, which then process, analyse and, using the sample data, reports to contaminant (for example fibre) concentration.

Traditionally exposure monitoring has been carried out using a wearable pump connected by tubing to a sample filter holder (cowl) attached to the user’s collar. This is to measure the contaminant level around the breathing zone. The test is rarely done because of the cost and time involved. It doesn’t measure the actual exposure level inside of the respirator, so assumptions must be made as to the protection factor afforded to work this out to allow comparison against regulatory limits. By providing filter cassette that fits onto the visor of the user (or other respiratory protection devise) the system of the disclosure provides the option of sampling from the air within the respirator.

Figures 3A and 3B show the perspective front view and rear view of a housing for the system of figure 2, respectively. The housing 300 includes a casing for enclosing and protecting the various elements of the system. In this example the casing is made of a front portion and a rear portion attached to a top and bottom plate. The top plate is provided with an air sampling inlet and a hatch for replacing the magazine. The front casing has an opening for displaying various process status indicators. A user device including a touch screen and a processor is detachably mounted on the front casing. The rear casing has a hatch for replacing the solvent using by the system, for instance acetone. Various openings are provided to provide power and communication ports. The housing is airtight with air sampling line from inlet to pump and then exhaust isolated from internals of the device. For ease of use the whole system can be mounted on an adjustable stand.

Figure 4A-C show the top view, front view and side view of a cartridge also referred to as cassette for use in the system of figure 2. The cartridge 400 includes a filter 410 and a transparent slide 420, for instance a glass slide, provided in a casing or housing 430. The cartridge 400 has a front portion 400a that includes the filter 410 and a rear portion 400b for holding and releasing the transparent slide 420. The casing 430 has an opening located in the front portion 400a to expose an area of the filter 410. The filter 410 may be provided by a membrane or a mesh, for instance the filter may be a paper filter. The casing 430 is also provided with a pair of slits 440a and 440b extending along a length of the rear portion 400b. The casing 430 may be made of plastic. The cartridge 400 also includes a mechanical arrangement adapted to move the slide 420 between a first configuration in which the slide does not overlap with the filter 410, and a second configuration in which the slide overlaps with the filter 410. The mechanical arrangement may be actuated by the gripper after the air sampling step to cover the exposed surface of the filter prior to chemical processing.

Each cartridge or filter is provided with a unique identifier, hence allowing traceability of the sample. The cartridge facilitates manipulation of the air filter within the device.

Figure 4D shows a cross sectional view of the cartridge of figure 4A. The rear portion 400b has a mechanical arrangement comprising an upper longitudinal member 442 and a lower longitudinal member 444. Each member 442, 444 has a recess forming a chamber for holding the transparent slide 420. The longitudinal member 442 is flexible so that it can move up vertically and act as an upper jaw for releasing the slide 420, hence allowing motion of the slide from the rear portion 400b to the front portion 400a. Figure 4D shows the slide 420 after release. The slide 420 is under the air filter 410.

Figure 4E shows the cartridge of figure 4A held by a gripper 242. Figure 4F shows the cartridge interacting with a system for moving the slide 420 inside the cartridge. The mechanical system includes a pair of prongs or hooks 470a, 470b that can be moved vertically using a motor 460. The cartridge 400 is positioned by the gripper 242 so that when lowered the prongs 470a, 470b engage the slits 440a and 440b. The gripper 242 is then pulled backward so that the slide 420 moves from the rear portion 400b to the front portion 400a of the cartridge. Once the slide 420 is in place the member 442 provides an edge that holds the slide 420 in place in the front portion 400a and prevents it front moving. The cartridge 400 can then be inserted by the gripper into the spray enclosure 222.

Figure 5 is a flow chart illustrating the operation of the system of figure 2. In operation a user places the system in a location of interest where air monitoring should be performed. Once powered the user may be presented with a user interface for starting the air monitoring procedure. The user interface may be designed to restrict the settings of the device to a minimal. For instance the user may only control where the device is physically positioned and when it is turned on.

In operation the controller 260 controls the various components of the system to implement the following steps: air sampling (step 510), air filter processing (step 520), imaging (step 530), image analysis (step 540), reporting (step 550) and storing the sample withing the device (step 560).

Starting at step 510, a cartridge is automatically selected from the magazine 211 and positioned by micro-automation horizontally on the sampling stage 213. The pump 212 is activated to pass ambient air through the air filter. The air passes through the pump and exhausts from the device via a valve to minimise risk of ingress. The air flow may be controlled to use a relatively small volume of air hence reducing the probability of catching large particles. The flow rate can be significantly increased over current typical flows for battery-powered devices to further speed up analysis reporting, or to improve measurement sensitivity.

Various parameters may be monitored including air pressure, flow rate, temperature, moisture and dust levels. A pressure sensor may be used to monitor a desired pressure differential across the air filter and identify any filter damage such as breach or puncture. The air sampling flow rate may be measured using various types of sensors including internal hot-wire or ultrasonic sensors. Air temperature and atmospheric pressure may also be monitored and taken into consideration for calculating the concentration of contaminant particles in the air sample.

Dust and moisture levels may impact on filter clearance and countability. Moisture and small particle matter, for instance PM10 particles having a diameter of 10 micrometres or less may be monitored using a PM10 particle sensor.

After air sampling, the transport mechanism 240 removes the cartridge from the sampling stage 213 and inserts the cartridge into the sample processing unit 220 for chemical treatment.

During step 520, a sprung glass slide incorporated within the cartridge is released and slides over the filter surface. The cartridge is then rotated by 180 degrees and positioned into a spray enclosure or chamber and enclosed. A small volume of solvent, for instance about 50 pL of acetone, is then automatically extracted from the solvent reservoir and passed through a heated nozzle and sprayed as vapour onto the rear (under)-side of the filter onto the glass surface behind. This serves to render the filter transparent and also seals the filter material onto the glass slide, hence retaining the sampled fibres within as a fixative mechanism. The transparent filter on the glass slide is referred to as sample matrix. The glass slide (or other transparent substrate) is provided on top of the air filter and no additional coverslip is required, hence reducing bubble formation. This sample matrix is then ready for phase contrast microscopy measurement. Optionally, if desired a coverslip could be incorporated to sandwich the sample matrix between the glass slide and the coverslip.

The transport mechanism 240 removes the cartridge from the sample processing unit 220 and inserts the cartridge into the imaging unit 230.

At the start of step 530, a pre-imaging check may be used, for instance using machine vision imaging of the filter to identify any imperfections that may have been introduced during the air filter processing step 520 and could lead to fiber counting errors. The phase-contrast microscope is controlled to perform phase contrast. Asbestos and other materials have partial spectroscopic fingerprints in the near-lR and analytical performance may be enhanced by extension of the spectrum beyond the visible region of the spectrum.

The gripper 242 positions the cartridge onto the microscope stage. The microscope stage then moves the cartridge to enable an identifier (for instance a 2D barcode or other identifier label) to be identified and the correct sampling plane starting depth to be focussed on.

The camera then captures several field images of the sample matrix to ensure that a representative proportion of the sample is obtained. This is performed by scanning the sample matrix and recording phase contrast images. Image capture may be performed at relatively high magnification for instance greater than 100 fold magnification, for example 600 fold magnification. At each field, a depth-of-field set of images is obtained to ensure that any objects are captured at optimum focus in a stacked image. The intensity of the light source may be monitored and adjusted to ensure that it remains within a pre-defined range.

The imaging unit 230 may be automatically checked for calibration at regular intervals or even constantly. For instance the alignment of the light source- condenser-phase rings-camera may be checked to be within required tolerance by machine vision. This may be achieved at pre-defined times. The performance of the optics and image capture may be calibrated automatically for resolution and magnification at regular intervals, for instance before and after each test, against a reference sample slide. A vibration sensor may be utilised along with vibration dampening mechanism to ensure to prevent operation above a pre-determined vibration thresholds for which image capture performance may be compromised.

At step 540 the images are processed using a machine learning algorithm to identify objects or bodies meeting the criteria for “fibrous” shapes or other types of contaminants. These objects are then filtered based on size rules (length, diameter and aspect ratio) before being confirmed as “countable” or “non-countable”. The total number of fibres or other contaminant in each sample is then aggregated and reported (step 550) against the sample corrected air volume as a number of contaminant (for instance fibres) per unit volume of air (eg f/mL). A qualitative discriminatory analysis output may be reported for the countable fibres identified, as a level of confidence (X%) that Y% of the countable fibres are asbestos, or further resolved into different asbestos fibre types. Non-asbestos fibres may also be characterised.

The test result can be reported in the form of digital, attributable, transparent, and incorruptible certification, backed up by location data and sampling photographic evidence of non-interference. The certificate could be provided with additional panoramic digital images of the device location to assure the certificate reader as to the location of the device. Images could be obtained of the sample inlet triggered by close movement around the inlet in order to flag interference when unattended by the user (eg deliberate contamination or blocking of the inlet). The presence of this functionality would provide a strong deterrent against such actions.

It will be appreciated that the system 200 may include various means of communication to transmit and or receive data. For instance, the system 200 may be adapted to communicate with an external server. In some embodiments the external server may be configured to perform various functionalities of the processor 150.

Optionally an additional storage step may be implemented. In this scenario the system is provided with a storage unit to store the imaged samples. Once probed each sample is stored securely within the device beyond user access. Each sampled is available for archiving and subsequent review as required, for instance upon request by a client, a duty-holder, a regulator or other stake-holder.

Therefore, in contrast with manual paper based systems, the system of the disclosure permits to issue secure, attributable, incorruptible digitized results certification.

The system of the disclosure provides several advantages over the conventional manual counting method.

In the existing manual method the human analyst can only use a small field size of 100 pm diameter, as it is too difficult to count fibres in a larger field size without getting mixed-up and potentially double-counting. This field size is typically 10% of the actual field under the same magnification conditions. The system of the disclosure can view a field that is at least 10 times larger with comparable pixel resolution. Only 20 fields (image captures) are required to get to the same analytical sensitivity as the human method.

The human analyst is only permitted to view 2400 field per 24-hour period. Essentially this amounts to 12 samples (with each sample typically needing 200 fields analysed). On any one sample the number of fields can be increased but this becomes very time-consuming and reduces the number of individual samples that can be analysed per day. Doubling the numbers of fields will halve the detection limit but can only be done for 6 samples per day. The system of the disclosure can count any number of fields without fatigue and with only limited impact on speed of reporting.

The standard method assumes a sample volume of 480 litres and 200 fields counted on the filter. The system of the disclosure can sample at higher flow rates. This permits to reach the 480 litre threshold much quicker if required, or to extend the sample volume to drive down the detection limit.

The current manual technique does not permit the human analyst to run field blanks and subtract the count obtained from the sample result. In contrast, the system of the disclosure may also be set-up to automatically run field blanks at a pre-defined frequency. The frequency may be fixed and beyond the user’s control. The field blank counts may then be subtracted from the sample counts hence further reducing the analytical detection limit. The combination of low detection limit together with qualitative analysis of the contaminant (for instance fibres) permits to reliably discount false positives. This permits the use of the system for the monitoring of indoor air (i.e. for evaluating building occupant exposure risk) or external ambient air (for instance downwind of hazardous waste sites, demolition sites etc).

Figures 6A-6D illustrate the operation of a cartridge dispenser. The cartridge dispenser 600 includes two parts: a first guide or channel 610 on which a magazine 605 can be translated along a longitudinal axis, and a receiving stage 650 provided underneath the first guide 610. The first guide 610 is provided with a slot 612 for receiving a cartridge 630/400. The receiving stage 650 has a flipper mechanism 615 and a second guide 620 substantially perpendicular to the first guide 610. The magazine 605 is pre-loaded with several cartridges, in this example 25 cartridges. In operation, a stepper motor pushes the magazine 605 in a forward direction. The first cartridge present in the magazine drops through the slot 612 into the flipper mechanism 615. The flipper mechanism 615 then releases a panel to place the cartridge onto the second guide 620. An actuator including a mechanical arm 660 is then used to push the cartridge 630/400 along the second guide onto the sampling stage. The second guide 620 is provided with an aperture 622. When pushed the filter area of the cartridge is substantially aligned with the aperture 622. The cartridge dispenser 600 can be replenished by the user without cross-contaminating or otherwise compromising sample integrity.

Figure 7A is a perspective view of an air sampling assembly 700. The air sampling assembly includes an air pump 710 coupled to an air flow channel comprising a first channel portion (lower cowl portion 720), a second channel portion (upper cowl portion 730). An actuator is provided to move the first and second channel portions 720/730 between an open state and a closed state. In the close state, the first channel portion 720 and the second channel portion 730 form a seal around the air filter. In this example the actuator includes two pairs of solenoids to move the lower cowl portion and the upper cowl portion with respect to each other. The first pair of solenoids (732a, 732b) is provided on a lower mounting plate 740. Similarly, the second pair of solenoids (734a, 734b) is provided on an upper mounting plate 750. The lower cowl portion 720 is connected to the pump 710. The profile of the lower cowl portion may be designed to provide a laminar air flow. An air flow sensor may be provided within the first channel portion 720 or a duct connecting the pump 710 to the first channel portion. Figure 7B is a perspective view of the air sampling assembly 700 integrated with the cartridge dispenser 600 of figure 6. To minimize electrostatic build-up the cartridge dispenser 600 may be connected to a ground port (earthed). This permits to reduce the potential loss of fibers on the cowl.

Figure 8 shows cross sectional views across the lower cowl portion and the upper cowl portion. Figure 8A illustrates the cowl pipe assembly in a closed position, and figure 8B illustrates the cowl pipe assembly in an open position. The same reference numerals have been used to represent the features already described in figures 6 and 7. In the closed position the cartridge 630/400 is clamped between the lower cowl pipe 720 and the upper cowl pipe 730. The filter portion of the cartridge 630 is substantially aligned with the aperture 622 provided in the guide 620. An air flow from the pump can take place to capture ambient air within the air filter. In the open position the cartridge 630/400 can be removed and another cartridge can be inserted for performing another air sampling.

Figure 9 is a perspective view of a sample processing unit. The sample processing unit 900 includes a solvent reservoir 910, a pump 920, a heating chamber 930, a spray enclosure 940 and optionally an extractor pump 950. Several channel portions 961-964 are provided to couple the various components of the processing units. The solvent reservoir 910 is user- replaceable and may have a volume of about lOOmL for storing acetone or another solvent. The spray enclosure 940 is provided with slit aperture 942 for receiving the cartridge.

In operation the cartridge is inserted in the slit 942 so that the filter is within the enclosure 940. The pump 920 is actuated to pump a pre-determined volume of solvent from the reservoir 910 into the heating block 930 via channel portions 961 and 962. The heating block 930 then heats the acetone which evaporates and is transported to the enclosure 940 via the channel 963. The solvent vapours then condense onto the filter. As a result the filter sticks onto the slide to form a sample matrix. The sample matrix is transparent hence allowing subsequent imaging to be performed. The extractor pump 950 is then used to remove any solvent excess from the enclosure. This may be achieved using a charcoal filter to absorb the extracted solvent.

Figure 10 is a flow chart of a method for analysing an air sample. At step 1010, the imaging unit performs a quality check to identify one or more areas that meet a quality criteria to perform phase contrast imaging. At step 1020, the imaging unit performs phase contrast imaging on the area previously identified. At step 1030, the processor executes a machine learning algorithm to identify countable and non-countable contaminant particles. This is achieved based on a set of trained criteria allowing to perform pattern recognition of specific contaminant particles. The set of criteria is specific for the type of contaminant particles including fibres such as asbestos fibres, plastic fibres and refractory ceramic fibres, or other types of particles including diesel engine exhaust emission particulates, fungal spores and nanoparticles. In the case of asbestos, machine learning classifiers are trained using a large set of real-world fibre images (thousands of images) that have been manually annotated by trained microscopist to identify countable asbestos fibres. At step 1040, the processor calculates a concentration of contaminant particles. The contaminant particles meeting the trained criteria are counted then computed to generate a concentration value. At step 1050, the processor reports the concentration of contaminant particles. For instance the processor may provide one or more of a digital output signal, an encrypted certificate, and a text alert.

The machine learning algorithm may be designed to recognise specific types of fibres. For instance, the machine learning algorithm may be configured to recognise fibres meeting certain morphological criteria including diameter, length and aspect ratio. The machine learning algorithm may also be adapted to discriminate qualitatively between asbestos and non-asbestos fibres allowing the proportion of confirmed asbestos fibres to be reported. In addition different types of asbestos fibres may also be differentiated.

In a specific implementation the machine learning algorithm uses a pixellevel image classification method which outputs a high resolution image in which every pixel belongs to a specific class. A deep Learning model is used performs semantic segmentation of the image in which a label is assigned to every pixel of the image. Then a qualitative identification of the type of fibre is performed.

The characteristics of the fibres, for instance the particle size and aspect ratio distributions, may be used to derive a toxicity index that considers the increased risk carried by longer and straighter fibres. Fibre toxicity is closely related to fibre length (directly) and thickness (inversely). The proposed system of the disclosure enables to establish the fibre morphological distribution for the sample as a whole thereby allowing a toxicity/carcinogenicity rating criteria to be applied to the sample - based on fibre length and diameter - in addition to countable fibre numbers. For instance, the characterisation of fibre morphologies can be used to generate a histogram of the fibre characteristics enabling carcinogenicity risk categorisation.

Figures 11 to 17 describe the machine learning algorithm used to analyse the phase contrast images.

Fibre segmentation and counting

This section describes the Deep Learning algorithm developed to automatically identify and count asbestos fibres from microscope images. The goal is to replicate and improve the performance of trained human analysts at discriminating and counting asbestos fibres. The current method for detection of asbestos fibres in air samples follows a standardised procedure defined on "HSG248 Asbestos: The analysts' guide for sampling, analysis and clearance procedures" (Health and Safety Executive, "Asbestos: The analysts' guide for sampling, analysis and clearance procedures," 2021). Once the sample has been mounted on the microscope, the analyst proceeds to count the fibres according to a set of specific rules: i) “a countable fibre is defined as any object which is longer than 5 pm, with an average width less than 3 pm and having an aspect (length /width) ratio greater than 3: 1 (fibres attached to particles are assessed as if the particle does not exist and are counted if the visible part of the fibre meets the above definition).” ii) “a countable fibre with both ends within the graticule area is recorded as one fibre.” iii) “a countable fibre with only one end in the graticule area is recorded as half a fibre.” iv) “a countable fibre passing through the graticule area, and having no ends within that area, is not counted.” v) “a split fibre is taken to be one countable fibre if it meets the definition above, otherwise it should be ignored; a split fibre is defined as an agglomerate of fibres which at one or more points on its length appears to be solid and undivided, but at other points appears to divide into separate strands; the width is measured across the undivided part, not across the split part.” vi) “loose agglomerates of fibres are counted individually if they can be distinguished sufficiently to determine that they meet the definition above.” vii) “fibres in a bundle and tight agglomerates of fibres, where no individual fibres meeting the definition of a countable fibre can be distinguished, are taken to be one countable fibre if the bundle or agglomerate as a whole meets the definition above.” viii ) “if the width of the fibre varies along its length, a representative average width should be considered.”

The number of asbestos fibres N counted through this method can be used to calculate the concentration of fibres C in the air sample, in fibres/ml, as:

Where D is the diameter of the exposed filter area, in mm, V is the volume of air sampled, in litres, n is the number of graticulate areas examined and d is the diameter of the Walton-Beckett graticule, in pm. In the UK, the control limit for asbestos exposure is 0.1 fibres per millilitre of air, averaged over any continuous period of 4 hours.

An important aspect to consider is the accuracy of the manual fibre counting method. This is particularly relevant for evaluation of the performance of image analysis algorithms that aim to replicate this method. Several factors including the subjectivity of a human-based measurement but also incorrect microscope adjustment, poor lighting or fatigue of the analyst can influence the fibre counting process.

Algorithm for semantic segmentation

The computer vision algorithm developed to automate the task of counting asbestos fibres is based on a Deep Learning model that performs semantic segmentation of the images (A. Garcia-Garcia, S. Orts-Escolano, S. Oprea, V. Villena-Martinez and J. Garcia-Rodriguez, "A Review on Deep Learning Techniques Applied to Semantic Segmentation," arXiv:l 704.06857 [cs.CV], 22 Apr 2017). The goal of this technique is to assign a label to every pixel of the image (i.e. fibre or background in our case). It is therefore a pixel-level image classification method which output is a high resolution image where every pixel belongs to a specific class. The selected model is a modified version of the U-NetDeep Learning architecture (O. Ronneberger, P. Fischer and T. Brox, "U-Net: Convolutional Networks for Biomedical Image Segmentation," arXiv:1505.04597 [cs.CV], 18 May 2015).

The U-Net model is a fully convolutional network formed by two sections: an encoder that extracts the main features from the images and a symmetric decoder that reconstructs the image and classifies each pixel. Additionally, the network incorporates skip-connections at every block, which concatenate feature maps from the decoder with previous feature maps from the corresponding level of the encoder section. This is essential for object localisation since it enables the network to retain relevant features lost by the encoder path.

The model was modified to introduction ResNet blocks in encoder and decoder instead of the original U-Net blocks (K. He, X. Zhang, S. Ren and J. Sun, "Deep Residual Learning for Image Recognition," arXiv:1512.03385vl [cs.CV], 10 Dec 2015). More specifically a ResNet-34 backbone was introduced to enhance feature extraction. The input image size is 480x640 pixels, which is reduced to 1024 15x20 feature maps after 34 ResNet operations.

Training such a neural network from scratch is challenging due to the large number of parameters involved. The U-Net model used in this work is formed of 41,221,168 parameters. The task is further complicated when the available dataset is small as it is in this case. Transfer learning is a very useful technique in these occasions since it uses a model that has already been trained on a different but larger dataset as a starting point and then adapts this model to fit the characteristics of our dataset. The rationale behind this concept is that in computer vision most of the features extracted from images using convolutional neural networks are common for different datasets. This algorithm was implemented using the vision module of the fastai vl library (J. Howard, S. Gugger, S. Bekman, F. Ingham, F. Monroe, A. Shaw and R. Thomas, "fastai," 2019), which seats on top of PyTorch 1.0. The data was shuffled and randomly split into training set (70% - 424 images) and validation set (30% - 181 images). Additionally, the images were normalised using the statistics of the ImageNet dataset that was employed to generate the pre-trained model used for transfer learning. Furthermore, the images fed to the neural network were augmented to balance out the small size of the annotated dataset. Image augmentation consists of applying slight transformations to the input images to increase data variability. This avoids overfitting and results in a more generalised model that will perform better when presented with previously unseen data. In our case, after testing several augmentation options (i.e. flip, zoom, rotation, lighting, warp), the best results were achieved by applying a combination of random horizontal flips, rotations of up to 10°, zooms of magnification up to l.lx, lighting and contrast alterations of up to 20% and symmetric warp of up to 20%.

The training set was used to fit the model using the Adam optimiser (D. P. Kingma and J. Ba, "Adam: A Method for Stochastic Optimization," arXiv:1412.6980v9 [cs.LG], 30 Jan 2017. ) by minimising the pixel-level crossentropy loss function, (R. Rubinstein, "Optimization of Computer Simulation Models with Rare Events," European Journal of Operations Research, vol. 99, no. 1, pp. 89-112, 1997. R. Rubinstein, "The Cross-Entropy Method for Combinatorial and Continuous Optimization," Methodology And Computing In Applied Probability, vol. 1, no. 2, pp. 127-190, 1999) commonly used for semantic segmentation tasks. Initially, the network was trained using input images of half the original resolution (i.e. 240x320). The rationale behind this procedure is that the network will be able to learn most of the important features in the dataset even if it is presented with images of lower resolution. The advantage is that the memory and computational times required to train the network are significantly reduced. Once the network has been pre- trained on low resolution images, the high resolution images are fed to the model to fine-tune its weights and enhance its performance.

The Icycle policy (L. N. Smith, "A disciplined approach to neural network hyper-parameters: Part 1 -- learning rate, batch size, momentum, and weight decay," arXiv:1803.09820v2 [cs.LG], 24 Apr 2018.) was employed to further speed up the training process. This method uses a two-phase procedure consisting of increasing the learning rate in the initial phase of training to prevent the model from landing in steep areas of the loss function, to then proceed to a second phase where the learning rate is gradually decreased when a flatter region is found. This learning rate profile acts as a regularization method and avoids overfitting. Once the model has been trained, the performance evaluation is conducted purely on the validation set, which is formed by images that had not previously been seen by the model.

Results

Pixel-level classification

As mentioned above, the image segmentation algorithm performs per-pixel classification to discriminate between pixels belonging to a fibre and pixels forming part of the image background.

Figure 11 shows an example of the predictions obtained on the validation dataset. Most pixels are classified correctly as belonging to a fibre (1110) or background. However, a number of false negatives (1130) and false positives (1120) are also observed in the predictions. False negatives correspond to pixels that were annotated by the analysts as belonging to a fibre but were not predicted as so by the algorithm. In contrast, false positives represent pixels annotated as background that the algorithm considered to be part of a fibre. In both cases, slight deviations when the analyst depicts the outline of the particle during the annotation process can have an important contribution, especially for large fibres. This is an issue that is difficult to avoid without recurring to an extremely long and tedious annotation process and has to be taken into account when evaluating pixel-level segmentation results.

Figure 11B shows the confusion matrix summarizing the pixel-level classification results for the entire validation dataset. The confusion matrix clearly reflects the nature of the dataset where asbestos fibres only occupy a minor proportion of the image. Therefore, it is not surprising, that most pixels belong to the background. However, for this application, the performance should be evaluated on the ability of the algorithm to detect fibres. 11B shows how more pixels are classified correctly than incorrectly, either as false positives or false negatives.

For systems with large imbalance between object and background pixels, a more adequate measure of the performance of the model is achieved through metrics such as precision P, recall R and F_± score-.

TP P = -

TP + FP

TP R ~ TP + FN

P - R

F. score = 2 — — -

¹ P + R

The precision is a measure of the fraction of pixels correctly classified as a fibre among all the pixels classified as a fibre by the model. Recall or sensitivity represents the fraction of pixels correctly classified as a fibre with respect to the pixels belonging to a fibre according to the ground truth. The F_± score is the harmonic mean of precision and recall and provides a way to combine these two metrics into a single performance evaluation value when precision and recall are equally significant. Evaluating the model using these performance indicators gives a better sense of how the algorithm performs at classifying pixels belonging to asbestos fibres.

Furthermore, for every pixel and every class (i.e. fibre and background), the image segmentation algorithm provides a score between 0 and 1 indicating the probability of that specific pixel to belong to a particular class. A classification threshold has to be set in order to assign a class to the pixel. For imbalanced classification tasks, the best performance is obtained when this threshold is selected so it provides simultaneously the largest precision and recall or, in other words, so it maximises the Fi score.

Figure 12A shows the Precision-Recall curve constructed through calculating pairs of precision and recall values at different thresholds. Figure 12B shows the corresponding evolution of the Fi score for every threshold. The optimum threshold corresponds to the maximum Fi score but also provides the best balance of precision and recall (i.e. the one that provides the closest point to the top-right corner on Figure 12A. In our case, the optimum classification threshold for asbestos fibres is 0.837, and provides a precision of 72.0%, a recall of 73.1% and an Fi score of 72.5% on the pixel classification task for the full validation set. This threshold was used to generate the results shown previously in Figures 11A and 11B.

Fibre-level classification

The image analysis algorithm is trained to classify individual pixels as belonging to fibre or background. The goal is to count the fibres present in a particular image and this does not require classifying every single pixel in every fibre correctly. So long as enough pixels are properly detected to make it a countable fibre, it will be equivalent for fibre counting purposes.

The regionprops library from the Python scikit-image module was used to draw bounding boxes around groups of adjacent pixels. Figure 13 shows the results of this procedure when applied to the example presented in 11A. The boxes on the right hand side (for instance box 1310 in Figure A) represent ground truth fibres provided by analysts’ annotations. These boxes are also shown in figure B alongside additional bounding boxes that correspond to algorithm predictions (for instance box 1320 in Figure B). Furthermore, through this library it was also possible to extract information regarding the size and shape of the detected objects. Based on this, the counting rules described above were implemented to discard objects that do not meet the size and shape criteria of countable fibres (i.e. length > 5 pm and aspect ratio > 3:1). The length and width of the detected objects are calculated as the dimensions of the major and minor axes of the rotated ellipse with the same normalised second central moments as the region. The aspect ratio is the ratio of length to width. This approach allows to take the orientation of the fibres into account and will provide representative sizes for straight fibres, which constitute a large proportion of the dataset. However, it can lead to the rejection of curly fibres where these dimensions are less defined. Therefore, an exemption is introduced to only discard objects which aspect ratio < 3 when their solidity is larger than 0.8. The solidity is the ratio of the area of the object to the area of its convex hull and is closer to 1 for straight fibres than for curly fibres. For consistency, this filter was applied to both ground truth annotations and predictions.

At the fibre level, the performance of the algorithm is measured in terms of its ability to predict both the number and the location of fibres annotated by the analysts. In order to systematically decide whether a fibre was predicted correctly, we make use of the intersection over union (loU) metric. The loU is simply the ratio between the area of overlap of ground truth and prediction bounding boxes with respect to the union of these two bounding boxes. It gives an indication of the degree of overlap between annotations and predictions. Due to the nature of object detection algorithms and the variability of the manual annotation process, it would be unrealistic to aim for full match between expected and predicted bounding boxes. Therefore, objects with loU greater than 25% and which meet the size and shape criteria of countable fibres are considered as true positives.

Fibre size and shape distributions

One of the add-ons of implementing an automatic image analysis algorithm is the ability to generate fibre size and shape distributions. This is a feature that is not available through the current method of analysis of asbestos fibres and provides significant insights into the evaluation of the toxicity of the sample.

Figure 14 shows the length, width and aspect ratio distributions for the validation set of the sample analysed in this work. Most fibres have a length of approximately 10 pm, a width of 3 pm and an aspect ratio close to 4. However, the presence of fibres with lengths up to 181 pm can also be observed. It may be useful to be able to detect and quantify the amount of these longer fibres since they carry a higher toxicity.

Qualitative analysis - fibre type classification

Quantifying the concentration and characteristics of asbestos fibres in air samples is essential to assess the risk of exposure. The automatic fibrecounting algorithm will contribute to speed up the analysis and improve detection limits. Nevertheless, simply detecting the presence of asbestos fibres in air samples has not been a problem historically where the focus has been on occupational exposure. In this context, samples are generally characterised by relatively high levels of countable fibres and, in most cases (e.g. demolition, building refurbishment), they are almost certainly known to contain asbestos anyway. However, there is rising concern regarding non-occupational exposure due to growing numbers of people dying of asbestos diseases that have no obvious source of exposure. For this reason there is now increasing interest in monitoring the health impacts of urban exposure (24/7) of the general population to background low-levels of asbestos (predominantly from buildings and traffic). Research and regulation in this area has always been handicapped by the fact that non-asbestos fibres - usually irrelevant in the occupational environment - can be expected to be proportionately more significant in ambient air quality samples. As such, it can be hard (without using costly techniques such as SEM or TEM) to characterise background fibre levels, and harder still to exclude the possible influence of non-asbestos fibres in the measured levels.

A measurement technique that can combine improved sensitivity with the ability to discriminate and confirm the presence of asbestos fibres opens up the potential for research and the needs for regulation in a new exposure vector. In this section, we summarise the findings of a preliminary study on the ability of Deep Learning algorithms to discriminate automatically asbestos fibres from other types of fibres or particles in images of air samples.

Dataset

This work aims to evaluate the performance of state-of-the-art image classification algorithms to differentiate between three categories of microparticles: i) Amosite amphibole asbestos fibres (25 images): distinctive needlelike very straight fibres. ii) Chrysotile serpentine asbestos fibres (26 images): dark, fine and morphologically varying fibres with characteristic curly shape. iii) Non-asbestos fibres (99 images): a range of fibrous non-asbestos building materials. These are the most likely fibre types to come across in any sort of quantity that may be countable under the counting rules. These are man-made vitreous/glass-type fibres which come in a narrow range of chemical compositions and a broader range of physical parameters. Many of them are unlikely to be countable due to large diameter but a proportion may come within the counting rule criteria.

The main differences between non-asbestos and asbestos fibres obvious to the human eye are the greater typical thickness, the parallelism of fibre sides and the high degree of transparency/light transmittance of the internal mass of the fibre which contrasts with a distinctive "edge" around the fibre.

As regards qualitative fibre discrimination, providing a degree of confidence as to the likely presence of asbestos fibres within the countable fibre loading of a sample would be a significant advance with respect to the current method. Furthermore, the potential to discriminate within asbestos types (i.e. amosite vs chrysotile) or non-asbestos types is an important add-on since different types carry different toxicity categories.

The size of the dataset (150 images) provides a good indication as for the likelihood of success of the classification method and enables an informed decision on the estimated size of the dataset required to obtain statistically conclusive results.

Classification algorithm

The Deep Learning image classification algorithm used in this work is formed by a Convolutional Neural Network (CNN) backbone that extracts features from the images and which final convolutional layer is connected to a fully connected head that transforms these features into predictions for each class. Transfer learning was also applied in this case, making use of a pretrained ResNet-34 (K. He, X. Zhang, S. Ren and J. Sun, "Deep Residual Learning for Image Recognition," arXiv:1512.03385vl [cs.CV], 10 Dec 2015) backbone and modifying the model heads to adapt it to the three output classes expected in this study. The heads include: i) a concatenation of adaptive average and adaptive pooling layers ii) a flatten layer to reduce the output to a single dimension iiij two fully connected blocks with batch normalisation, dropout, linear transformation and ReLU activation (only on the first block).

This leads to a model with 21,814,083 parameters which was also implemented using the vision module of the fastai vl library (J. Howard, S. Gugger, S. Bekman, F. Ingham, F. Monroe, A. Shaw and R. Thomas, "fastai," 2019).

In image classification tasks, it is generally important to use a balanced dataset with similar representation of the different classes. For this reason, we decided to use only 25 images per class for the analysis. The data was shuffled and randomly split into training set (80% - 60 images) and validation set (20% - 15 images). Additionally, the images were normalised using the statistics of the ImageNet dataset that was employed to generate the pre-trained model used for transfer learning (J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li and L. Fei-Fei, "ImageNet: A Large-Scale Hierarchical Image Database," in CVPR, Miami, 2009).

The model was trained on the training set using the Adam optimiser (D. P. Kingma and J. Ba, "Adam: A Method for Stochastic Optimization," arXiv:1412.6980v9 [cs.LG], 30 Jan 2017. ) to minimise the cross-entropy loss function (R. Rubinstein, "Optimization of Computer Simulation Models with Rare Events," European Journal of Operations Research, vol. 99, no. 1, pp. 89- 112, 1997.), commonly used for classification problems in Deep Learning. The Icycle policy was also employed in this case with a batch size of 4 images and without image augmentation (L. N. Smith, "A disciplined approach to neural network hyper-parameters: Part 1 -- learning rate, batch size, momentum, and weight decay," arXiv:1803.09820v2 [cs.LG], 24 Apr 2018. ). Training was carried out in two steps:

Figure 15 shows the selected learning rate profile and the corresponding training curve during step 1 and step 2 of the training, respectively. Step 1: The pre-trained weights for the ResNet-34 architecture trained on the ImageNet dataset are loaded and the weights for the model head are initialised randomly. Then, the backbone weights are kept frozen while the model head is trained for 4 epochs to generate predictions from the features extracted by the CNN backbone. Step 2 Once an initial model has been obtained with a pre-trained architecture, we unfreeze the backbone and retrain the entire network for 8 epochs to fine-tune the model and adapt it to the characteristics of our own dataset.

Preliminary results

Figure 16 shows the confusion matrix for the predictions of the trained model on the validation set. All the 15 images forming this set were classified into the correct category. As mentioned previously, the validation set is independent from the training set and it is used to evaluate the performance of the model on images that it has not seen before. For every input image run through the model, the output is a set of probabilities for each of the three possible classes. The final prediction is the one with the highest probability.

Figure 17 summarises the predictions and confidence scores for some image forming the validation set. While in most cases the selected class was clearly defined, some images had a stronger contribution from alternative categories. In order to train the model using a balanced dataset, it was necessary to discard 74 images of the ‘non-asbestos’ class. When the image classification algorithm is applied to this additional set of images, 13 ‘nonasbestos’ images are misclassified as containing chrysotile asbestos fibres. This shows how a larger dataset is required to obtain a more accurate model with further extrapolation capabilities.

The system for analysing air samples as described in the present disclosure can be used by a non-expert and near real-time quantification of particle contaminants in near real time. It also provides enhanced sensitivity over existing manual methods.

The system may be used attended by the user for clearance testing or unattended for area monitoring. The system can also be used in combination with several separate sampling units, hence providing flexibility of use for different types of applications.

The system can be used to perform air monitoring as a reactive measure for instance during higher-risk remedial asbestos work, but also as a precautionary risk management tool for instance for duty-holders with responsibility for buildings with asbestos.

It can also be evaluated against quality control performance schemes for human analysts such as the UK HSE Regular Interlaboratory Counting Exchanges (RICE) scheme.

A skilled person will appreciate that variations of the disclosed arrangements are possible without departing from the disclosure. For instance although the training of the machine learning algorithm has been described using a specific example it will be appreciated that the architecture, software library, train/test split, augmentation among other parameters may be changed. Accordingly, the above description of the specific embodiments of the system of the disclosure is made by way of example only and not for the purposes of limitation. It will be clear to the skilled person that minor modifications may be made without significant changes to the operation described.

Claims

36

1. A system for taking and analysing air samples, the system comprising an air sampling unit adapted to receive an air filter cartridge having an air filter and a slide, and to pass a volume of air through the air filter, a sample processing unit adapted to apply a solvent onto the air filter to obtain a sample matrix on the slide; an imaging unit adapted to obtain an image of the sample matrix; a transport mechanism adapted to take the air filter cartridge and move it from one unit to another within the system; and a processor configured to process the sample matrix image and to calculate a concentration of contaminant particles.

2. The system as claimed in claim 1, wherein the air sampling unit comprises an air flow channel adapted to receive the air filter cartridge, and an air pump coupled to the air flow channel, the air pump being adapted to control the volume of air flowing through the air filter.

3. The system as claimed in claim 2, wherein the air flow channel comprises a first channel portion, a second channel portion, and an actuator adapted to move the first and second channel portions between an open state and a closed state, wherein in the close state, the first channel portion and the second channel portion form a seal around the air filter.

4. The system as claimed in claim 3, wherein the air sampling unit comprises a container for storing a plurality of air filter cartridges 37 and a cartridge dispenser adapted to insert the air filter cartridge between the first channel portion and the second channel portion. The system as claimed in any preceding claims, wherein the sample processing unit comprises a pump connectable to a solvent reservoir, a solvent heating chamber, and a spray enclosure. The system as claimed in claim 5, wherein the spray enclosure is provided with a slit aperture for receiving the air filter cartridge. The system as claimed in claim 5 or 6, wherein the sample processing unit comprises the solvent reservoir. The system as claimed in any of the preceding claims, wherein the imaging unit comprises a light source, an optical apparatus, and a camera. The system as claimed in claim 7 or 8, wherein the imaging unit is adapted to perform phase contrast microscopy. The system as claimed in any of the claims 7 to 9, wherein the imaging unit is adapted to capture images with a magnification factor equal or greater than 100 fold. The system as claimed in any of the preceding claims, wherein the transport mechanism comprises a translation stage coupled to a motion device. The system as claimed in claim 11, wherein the motion device comprises a gripper for gripping the air filter cartridge.

13. The system as claimed in any of the preceding claims comprising a storage unit for storing the sample matrix.

14. The system as claimed in any preceding claims wherein the contaminant particles comprise fibres.

15. The system as claimed in any preceding claims wherein the contaminant particles comprise at least one of exhaust emission particulates, nanoparticles and air spores.

16. The system as claimed in any of the preceding claims, wherein the processor is configured to execute a machine learning algorithm to identify countable contaminant particles and non-countable contaminant particles and to provide a number of countable contaminant particles.

17. The system as claimed in claim 16, wherein the machine learning algorithm is configured to perform of a pixel-level image classification method in which every pixel belongs to a specific class.

18. The system as claimed in claim 17, wherein the image classification method is a deep learning image classification method.

19. The system as claimed in any of the claims claim 16 to 18, wherein the machine learning algorithm is trained using a set of training data comprising images of fibres manually annotated for differentiating between asbestos and non-asbestos fibres.

20. The system as claimed in any of the claims 16 to 19, wherein the processor is configured to derive a toxicity index based on fibre characteristics comprising at least one of a fibre size and a fibre aspect ratio distribution. The system as claimed in any of the claims claim 16 to 20, wherein the processor is configured to identify false positives among the countable contaminant particles. An air filter cartridge for use with a system for taking and analysing air samples as claimed in any of the preceding claims, the air filter cartridge comprising an air filter attached to a casing; and a slide movable between a first configuration in which the filter and the slide are non-overlapping and a second configuration in which the filter and the slide are overlapping. The air filter cartridge as claimed in claim 22, wherein the air filter cartridge comprises a mechanical arrangement operable between a first state to hold the slide within in a first portion of the casing and a second state to release the slide to a second portion of the casing.