Fire detection

Abstract

A method of detecting a fire comprises receiving at least one infrared image comprising at least one pixel corresponding to a respective temperature, determining a probability of the presence of a fir

Description

Field The present invention relates to fire detection. In particular, the present invention relates to methods of detecting a fire and to a fire suppression system. Background Fire sprinklers and other fire suppression systems are commonly used in residential and commercial properties as a means to improve the inherent safety of the property. Common fire suppression systems are configured to activate in response to the detection of signs of a fire, such as smoke and increased temperature. However, these detection techniques can increase the likelihood of false positives, for example by causing activation in response to a safe fire or heat source, which may lead to unnecessary water damage. There is also the risk of such fire suppression systems activating when a dangerous fire has already significantly developed. Thus, there is a need in the art for fire detection techniques which enable a fire suppression system to respond rapidly to a dangerous fire without increasing the likelihood of false positives. Summary According to a first aspect of the present invention, there is provided a method of detecting a fire. The method comprises receiving at least one infrared image, the at least one infrared image comprising at least one pixel corresponding to a respective temperature, determining a probability of the presence of a fire based on the at least one infrared image, and signalling a fire condition based on the determination. The fire condition may be a positive fire condition or a negative fire condition. A positive fire condition indicates the presence of a dangerous fire, rather than a safe or controlled fire, such as a hob fire or an open fire in a grate. A negative fire condition indicates the lack of a dangerous fire and / or that a pre-set condition or threshold has not been satisfied or reached during the determination step. The method may comprise, wherein the fire condition is a positive fire condition, activating a spray head to deploy fire-suppressant material. Wherein the at least one infrared image includes a plurality of pixels each corresponding to a respective temperature, the probability of the presence of the fire may be determined using at least one artificial neural network. Wherein the at least one infrared image is a stream of infrared images including an nth infrared image and n is a non-zero positive integer, determining the probability may comprise: wherein n <a threshold value, inputting the nth infrared image into a first artificial neural network, the first artificial neural network comprising an output layer and a plurality of hidden layers, and the output layer outputting the probability of the presence of the fire. Determining the probability may further comprise: wherein n >the threshold value, a hidden layer of the plurality of hidden layers of the first neural network outputting an intermediate output, extracting at least one feature from a plurality of infrared images preceding the nth infrared image, inputting the at least one feature and the intermediate output into a second artificial neural network, the second artificial neural network comprising an output layer and a plurality of hidden layers, and an output layer of the second artificial neural network outputting the probability of the presence of the fire. The first artificial neural network may be a convolutional neural network. The second artificial neural network may be a convolutional neural network. The second neural network may be a three-dimensional neural network. The at least one feature may be one or more of: i) the maximum temperature value of each image of the plurality of infrared images; ii) the 98th temperature percentile value of each image of the plurality of infrared images; iii) the 100th temperature percentile value of each image of the plurality of infrared images; iv) slope of the best fit line for a plot of the 98th temperature percentile for each image of the plurality of infrared images; v) slope of the best fit line for a plot of the 100th temperature percentile for each image of the plurality of infrared images; vi) the result of the subtraction of the 98th temperature percentile of the 1st image in the plurality of images from the 98th temperature percentile of the last image in the plurality of images; and vil) the result of the subtraction of the 100th temperature percentile of the 1st image in the plurality of images from the 100th temperature percentile of the last image in the plurality of images. Optionally, n >6. The method may further comprise determining a number of hot objects, m, captured by the nth infrared image, wherein m is a non-zero positive integer, and determining a location of the fire. The determination comprises: wherein m >1, the hidden layer of the plurality of hidden layers of the first artificial neural network outputting the intermediate output, inputting the intermediate output into a third artificial neural network, and determining the location of the fire based on outputs of the third artificial neural network. The determination comprises: wherein m = 1, determining the location of the fire based on a location of a peak in temperature in the nth infrared image. The location of the fire may be expressed as a column index of the multi-pixel infrared image. The third artificial neural network may be a convolutional neural network. The at least one infrared image may be processed prior to being input into the first artificial neural network. Processing the at least one infrared image may involve normalising the at least one infrared image. Wherein a spray head is deploying fire suppressant material at an original angular position corresponding to an original fire location in response to a positive fire condition, the method may comprise: capturing a stream of infrared images at the original angular position, determining whether one of the images in the stream satisfies a condition, if the condition is satisfied, deactivating the spray head, capturing at least one first infrared image at a first adjusted angular position, determining whether at least one hot object is present in the at least one first infrared image, and if present, determining an adjusted location of the fire based on the hot object(s). The method may further comprise: after deactivating the spray head, capturing at least one second infrared image at a second adjusted angular position, and determining whether at least one hot object is present in the at least one second infrared image. The fire may be a travelling fire. The condition may be one of the infrared images in the stream having an absolute 98th temperature percentile of at least 40°C, wherein the average temperature rise between each preceding image is at least 0.1°C over the last 20 seconds. In other examples, the absolute 100th temperature percentile may be considered instead. The first adjusted angular position may be at least ±7° from the original location of the fire. The second adjusted angular position may be between the first adjusted angular position and the original location of the fire. The method may further comprise activating the spray head to deploy fire suppressant material towards the adjusted location of the fire. The method may further comprise: wherein for a given infrared image in the stream of infrared images captured at the original angular position, performing at least one division of the given infrared image which intersects the central point of the given infrared image, wherein for each division: dividing the plurality of pixels into a first group and a second group, subtracting a pixel of the first group from a mirroring pixel of the second group, and determining a number of pixels that differ in temperature value based on the subtraction. The method may further comprise: calculating a symmetry score based on the determination. Wherein the symmetry score >a threshold symmetry score, consider the given infrared image in the determination of whether the image satisfies the condition. Wherein the symmetry score <a threshold symmetry score, discard the given infrared image. The threshold symmetry score may be 300. The division may be along the vertical or the horizontal axis. Two divisions may be performed, for example along the vertical axis and along the horizontal axis. The given infrared image may be processed before calculating the symmetry score. This processing may involve adjusting the temperature readings at the periphery of the infrared image and applying a median filter to the infrared image. The method may further comprise focussing scanning at a given angular position in response to a given infrared image retuning a negative fire condition but exceeding a pre-set probability of the presence of the fire. Wherein a negative fire condition is signalled in response to receiving a given infrared image and the probability of the presence of the fire for the given infrared image is greater than a first pre-set probability, the method may further comprise: if the probability for the given infrared image is less than a second pre-set probability, capturing a first stream of infrared images at the angular position corresponding to the given infrared image, or if the probability for the given infrared image meets or exceeds the second pre-set probability and is less than a third-pre-set probability, capturing a second stream of infrared images at the angular position corresponding to the given infrared image. The first stream of infrared images may be captured until one of the following conditions is satisfied: i) a pre-set period elapses; or II) the probability of the presence of the fire for an Infrared image in the first stream meets or exceeds the second pre-set probability. The second stream of infrared images may be captured until the probability of the presence of the fire for an infrared image in the second stream: i) exceeds the third pre-set probability; or ii) falls below the second pre-set probability. The first pre-set probability may be between 10% and 20%, for example 15%, and the second pre-set probability may be between 50% and 70%, for example 60%. The third pre-set probability may be between 80% and 90%, for example 86.1%. The pre-set period may be between 8 second and 12 seconds, for example 10 seconds. The plurality of infrared images may be captured at a rate of 1 to 2 infrared images per second. If the probability exceeds the third pre-set probability, a positive fire condition may be signalled. Wherein the at least one infrared image includes a given infrared image capturing a hot object, the method may further comprise: determining the location of the hot object in the given infrared image, and, in response to a determination that the hot object is not at the centre or substantially at the centre of the given infrared image, adjusting an angular position of an infrared sensor which captured the given image such that the hot object is at the centre of the infrared sensor's field of view. Wherein the at least one infrared image consists of a single pixel corresponding to a respective temperature, the method may further comprise: in a first scan, capturing scan data defining a baseline, wherein the scan data consists of a plurality of Infrared images captured at different azimuthal angles of uniform increment, performing a first fire detection method based on the baseline and determining the probability of the presence of the fire based on the first fire detection method. The method may further comprise: in a subsequent scan, capturing subsequent scan data, performing the first fire detection method and / or one or more of a second, third, and fourth fire detection method based on the subsequent scan data, and determining the probability of the presence of the fire for each performed fire detection method. The scan data may be smoothed prior to performing a fire detection method. The first, second, third, and fourth fire detection methods may be for detecting fires of different growth rates. The first fire detection method may be for detecting ultra-fast developing fires and the second, third, and fourth fire detection methods may be for detecting slower-growing fires. The second fire detection method may be for detecting slowly developing fires. The third fire detection method may be for detecting fast developing fires. The fourth fire detection method may be for detecting medium speed developing fires. The first fire detection method may comprise: receiving the scan data of either the first or a subsequent scan, discarding the infrared image(s) In the scan data for which their temperature is less than a threshold value, and identifying the highest temperature out of any remining infrared images. A positive fire condition may be signalled based on the infrared image having the highest temperature. The second fire detection method may comprise: receiving the scan data of a given scan, and, for each infrared image in the given scan: calculating a difference between the temperature of the infrared Image and the corresponding temperature of the infrared image in the immediately preceding scan to produce a gradient, and calculating a final gradient score, wherein the score is based on the gradient and a peak in temperature within the given scan. The second fire detection method may further comprise: discarding any infrared images for which their final gradient score does not exceed a threshold value and identifying the highest final gradient score out of any remaining infrared images. A positive fire condition may be signalled based on the infrared image having the highest final gradient score. The final gradient score may be Indicative of the consecutive change in temperature across scans at the corresponding azimuthal angle. The third fire detection method may comprise: receiving the scan data of a given scan, discarding the infrared image(s) in the scan data for which their temperature is below a threshold value, for each of any remaining infrared images, calculating the difference between its temperature and the corresponding baseline temperature, discarding any infrared images for which their difference is less than a threshold value, and identifying the highest temperature out of any remaining infrared images. A positive fire condition may be signalled based on the infrared image having the highest temperature. The fourth fire detection method may comprise: receiving the scan data of a given scan, discarding the infrared image(s) in the scan data for which their temperature is below a threshold value, for each of any remaining infrared images, calculating the difference between its temperature and the corresponding baseline temperature, discarding any infrared images for which their difference is less than a threshold value, calculating the final gradient score for each of any remaining infrared images, discarding any infrared images for which their final gradient score is below a threshold value, and identifying the highest final gradient score out of any remaining infrared images. A positive fire condition may be signalled based on the Infrared image having the highest final gradient score. The method of detecting a fire may comprise initiating the method in response to at least one triggering event. The triggering event may be detection of smoke by a smoke detector. The triggering event may be detection of flames by a camera. The triggering event may be detection of an input by a user into a user input device. The at least one triggering event may be a single triggering event, such as detection of smoke by the smoke detector. The method of detecting a fire may be initiated in response to a combination of triggering events, such as detection of flames and smoke. According to a second aspect of the present invention, there Is provided a method of training an artificial neural network. The method comprises: feeding a plurality of training Infrared images into an untrained artificial neural network, wherein each training infrared image is assigned a sample weight which is set according to the timestamp of the training infrared image and the untrained artificial neural network has a set of neural network weights. The method further comprises, for each training infrared image, adjusting the set of neural network weights according to its sample weight. The training Infrared images may be processed prior to being fed into the artificial neural network. Wherein for a set of training infrared images captured over a period in which a fire ignites, the sample weight of each training infrared image may be set according to the stage of development of the fire which its timestamp corresponds to. Wherein an earliest activation time is the timestamp for which its corresponding training Infrared image is the first Infrared image in the set to clearly capture the fire: the training infrared images having a timestamp preceding the earliest activation time may be assigned a maximum sample weight, the training infrared image having a timestamp at the earliest activation time may be assigned a minimum sample weight, and the training infrared images having a timestamp immediately following the earliest activation time may be assigned sample weights between the minimum and maximum sample weight. An infrared image is considered to have clearly captured a fire when the fire can be distinguished from the ambient temperature background of the infrared image. The earliest activation time may be the time at which the fire ignites. The maximum sample weight may be 1. The minimum sample weight may be 0.1. A latest activation time is the timestamp following the earliest activation time after which the sample weights return to the maximum sample weight. Timestamps from the earliest activation time to the latest activation time fall within a buffer zone. The artificial neural network may be trained such that it is more likely to activate within the buffer zone, preferably at a fire point time. Sample weights within the buffer zone may increase according to functions based on a natural logarithm. According to a third aspect of the present invention, there is provided a computer program which, when executed by at least one or more processors, causes the processor(s) to perform the method according to the first or second aspects. According to a fourth aspect of the present invention, there is provided a computer readable medium, optionally a non-transitory computer readable medium, which stores or carries the computer program according to the third aspect. According to a fifth aspect of the present invention, there is provided a hardware processor configured to perform the method according to the first or second aspects. According to a sixth aspect of the present invention, there is provided a system comprising a controller configured to perform the method according to the first or second aspects, at least one spray head unit in communication with the controller, the at least one spray head unit comprising a spray nozzle configured to deploy fire suppressant material, and at least one infrared sensor. The at least one spray head unit may comprise the at least one infrared sensor. The system may further comprise a smoke detector, a camera, and / or a user input device. The user input device may be a touch panel or a button. The user input device may be comprised in the spray head unit. The at least one infrared sensor may be configured to capture a plurality of infrared images at respective angular positions. The at least one infrared sensor may be configured to rotate through 130° and capture an infrared image at each 5° increment. The at least one infrared sensor may be configured to capture an infrared image at angular positions 35°, 90°, and 145°. The angular positions may be defined relative to a wall to which the at least one infrared sensor is mounted. The controller may be configured to terminate scanning of the at least one infrared sensor once a termination condition is satisfied. The termination condition may be an absence of smoke detected by the smoke detector and / or flames detected by the camera. In response to an absence of smoke detected by the smoke detector, the controller may be configured to terminate scanning after a pre-set time-out period. The time-out period may be between 35 and 55 seconds, for example 45 seconds. The system may comprise a plurality of spray head units. Each spray head unit may be in communication with the controller, or each spray head unit may be in communication with a respective controller configured to perform the method of any preceding aspect. 5 Each spray head unit may be in the same room, or each spray head unit may be in a separate room. Brief Description of Drawings Certain embodiments of the present invention will now be described, by way of example, with reference to the accompanying drawings in which: Figure 1 schematically illustrates a fire suppression system; Figure 2 schematically illustrates a fire suppression system in a room; Figure 3 schematically illustrates a field of view of an infrared sensor; Figure 4 schematically illustrates a fire suppression system in a room; Figure 5 shows a process flow diagram of a method of operating a fire suppression system; Figure 6 shows a process flow diagram of a fire detection method; Figure 7 shows a process flow diagram of a fire detection method; Figure 8 shows a process flow diagram of a fire detection method; Figure 9 shows a process flow diagram of a fire detection method; Figure 10a is a data plot of temperature against scan number; Figure 10b is a data plot of temperature against scan number; Figure 10c is a data plot of probability of the presence of a fire against scan number; Figure 11 schematically illustrates a field of view of an infrared sensor; Figure 12 shows a process flow diagram of a method of operating a fire suppression system; Figure 13 shows a process flow diagram of a method of determining the probability of the presence of a fire; Figure 14 shows an example architecture of a neural network; Figure 15 shows an example architecture of a neural network; Figure 16 shows a process flow diagram of a method of determining the location of a fire; Figure 17a shows an infrared image capturing a hot object and corresponding data plots; Figure 17b shows an infrared image capturing a hot object and corresponding data plots; Figure 18 shows an example architecture of a neural network; Figure 19 shows a process flow diagram of a method of determining the location of a fire; Figure 20a shows an infrared image capturing two hot objects and corresponding data plots; Figure 20b shows an infrared image capturing two hot objects and corresponding data plots; Figure 21 shows a plurality of data plots corresponding to a focus mode of a fire suppression system; Figure 22a shows an infrared image and corresponding data plots; Figure 22b shows an infrared image and corresponding data plots; Figure 22c shows an infrared image and corresponding data plots; Figure 22d shows an infrared image and corresponding data plots; Figure 23 shows a process flow diagram of a method of fire chasing; Figure 24a shows an infrared image and corresponding data plots; Figure 24b shows an infrared image and corresponding data plots; Figure 24c shows an infrared image and corresponding data plots; Figure 24d shows an infrared image and corresponding data plots; Figure 25a shows an infrared image captured using a wet lens; Figure 25b shows a processed infrared image captured using a wet lens; Figure 25c shows a processed infrared image captured using a wet lens; Figure 26a shows a plot of symmetry score for a plurality of infrared images; Figure 26b shows a plot of symmetry score for a plurality of infrared images; Figure 27 shows a plurality of data plots relating to neural network training; Figure 28a shows a plurality of data plots relating to neural network training; Figure 28b shows a plurality of data plots relating to neural network training; Figure 28c shows a plurality of data plots relating to neural network training; and Figure 28d shows a plurality of data plots relating to neural network training. Detailed Description of Certain Embodiments In the following, like parts are denoted by like references. The present application is concerned with methods of detecting a fire. The methods of fire detection according to the present application have reduced activation times and instances of false positives compared with conventional fire detection methods. First fire suppression system 1, li Referring to Figure 1, a first fire suppression system 1, li (herein "first system") is shown. The first system li comprises one or more wall-mountable spray head units 2 (herein "spray head unit"), a controller 3, a pump 4 for supplying fire suppressing material 5, for example water, from a source 6 via piping 7 to the spray head unit(s) 2. The source 6 may be a mains water supply. The controller 3 and the spray head unit(s) 2 are connected by communication line 8. Each spray head unit 2 comprises a rotatable spray head assembly 9 (herein "spray head"). A spray nozzle 10 is comprised in the spray head 9. The spray nozzle 10 is set into a surface of the spray head 9 such that the spray nozzle 10 rotates with the spray head 9. The pump 4 is configured to deliver the fire supressing material 5 at high pressure, for example 80 MPa, to the spray nozzle 10 via the piping 7. Thus, the spray nozzle 10 is configured to deploy the fire suppressing material 5 from the spray head 9. The spray head 9 is configured to turn on one vertical axis (not shown), which allows the spray nozzle 10 to be faced in multiple directions. Specifically, the nozzle 10 is configured to deploy the fire supressing material 5 in a specific azimuthal direction, i.e., at an angle 0°, dependent on the rotary position of the spray head assembly 8. This angle of the nozzle 10 is defined relative to a wall (not shown) or other surface / object into which the spray head unit 2 is installed. The angle of the nozzle 10 parallel to the wall (or other surface / object) is 0°. The spray nozzle 10 may be orientated to deliver a mist of the fire-suppressant material 5 radially in a plane defined by the first axis and a second axis which is perpendicular to the first axis. Referring also to Figure 2, the one or more spray head units 2 can be installed in the same room 11 and preferably mounted to separate walls 12 of the room 11. The vertical axis is substantially parallel to the wall 12. The spray head 9 is usually kept in a first position ("parked position") which does not expose the spray nozzle 10 to the room 11 (as shown in Figure 2). When the first system li is activated, the spray head 9 is rotated so that the spray nozzle 10 is directed at a fire in the room 11. The controller 3 is configured to activate the one or more spray head units 2, as will be hereinafter explained. Each spray head unit 2 further comprises an infrared sensor 13 (Figure 1). The infrared sensor 13 is configured to capture images, each image consisting of a single pixel corresponding to a value of temperature. The infrared sensor 13 may be a Melexis (TM) infrared temperature sensor MLX90614KSF-ACC, although other single pixel infrared sensors could be used. The infrared sensor 13 is configured to capture the single pixel images at multiple azimuthal directions. The infrared sensor 13 may be provided on a rotatable part of the spray head unit 2, such as the spray head 9 (such that the infrared sensor 13 rotates with spray nozzle 10 and may be aligned with the nozzle 10). In these examples, the infrared sensor 13 is provided on the spray head 8 so as to rotate to different azimuthal directions. Alternatively, the infrared sensor 13 may be provided on a static part of the spray head unit 2, such as the faceplate 14 (Figure 2). In these examples, the infrared sensor 13 may scan the room 11 even when the spray head 9 is in the parked position. The infrared sensor 13 is configured to rotate through a pre-set angular range, e.g. 130°, and capture images at unform increments across that range (herein "scan data"). In a preferred embodiment, the infrared sensor 13 is configured to capture a first image at an azimuthal angle of 25° and a proceeding image at each 5° increment, with the last image captured at an azimuthal angle of 155°. In this example, the infrared image 13 captures 27 images during a complete sweep of the pre-set angular range. The azimuthal direction of the infrared sensor 13 (also referred to as the azimuthal angle) is defined relative to the wall 12 (or other surface / object) into which the spray head unit 2 is installed. The azimuthal angle parallel to the wall (or other surface / object) is 0°. An example of a full sweep of the Infrared sensor 13 is shown in Figure 3, with the field of view of the sensor 13 represented as a set of overlapping triangles. The first system li may further comprise other sensors (not shown), either incorporated into or separate from the spray head unit(s) 2. The other sensors may include at least one thermal sensor, camera, smoke detector, and / or microphone. Thus, the system li may detect non-thermal signs of a fire. The controller may be connected to an external communications network, such as a mobile network {e.g. GSM, 4G, 5G), wireless local area network (WLAN), Long Range Wide Area Network (LoRaWAN), or low-power wide-area (LPWA) network. In a preferred example of the first system li, multiple spray head units 2 address a single room 11 at different positions. The spray head units 2 are positioned such that each respective infrared sensor 13 has a different field of view (which may overlap). In this way, all or a majority of the room is being monitored by the system li at once. In the example shown in Figure 4, the system li includes three spray head units 3 only. A fire 15 is captured by each infrared sensor 13 at different azimuthal angles: p°, r°, s°. As will be hereinafter explained, the controller 3 Is configured to select which spray head unit 3 to activate to extinguish the fire 15. Method of operating the first system 1, li The first system li is for detecting a fire in a room 11 and supressing the fire by activating one of the spray head units 2. The decision to activate one of the spray head units 2 is made by the controller 3. This decision is made following processing of the single pixel images captured by the infrared sensor(s) 13, as will now be explained. Referring to Figure 5, a method of operating the first system li (herein "first operation method") is shown. The first operation method is for detecting and supressing fires. The method may be initiated in response to a triggering event. For example, the first system li may activate to perform the method in response to the smoke detector detecting smoke in the room 11. Merely for the sake of illustration, the first operation method is herein explained using an example in which each infrared sensor 13 captures 27 images in a single sweep starting at an azimuthal angle of 25° and ending at an azimuthal angle of 155°, with an image captured at each 5° increment. In this example, the system li comprises multiple spray head units 2. When the system li is activated, the infrared sensors 13 perform a sweep of the room 11 and captures a plurality of single pixel images ("scan data") at uniform increments - as hereinbefore described. The scan data consists of sets of scan data, each set corresponding to a respective spray head unit 2. As each image consists of a single pixel, each set of scan data consists of a plurality of temperature values (herein "readings"). Each temperature value corresponds to a respective azimuthal angle at which the respective image was captured. In step Sl.l, the scan data is received or gathered by the controller 3 from the spray head units 2. In step SI.2, the scan data is smoothed. For each set of scan data, each reading is averaged with its adjacent readings. Thus, the 1st reading is averaged with the 2nd reading, the 3rd is averaged with the 2nd and 4th reading, and so on. Performing step SI.2 may help to remove or reduce the affect of the flickering of the fire and thus help to identify the actual position of the fire more accurately. In step SI.3, the controller 3 determines if the received scan data is gathered from a first sweep / scan of the room 11. In response to a positive determination in step SI.3, the controller 3 saves the scan data as a baseline and, if applicable, resets Gradient Scores (described hereinafter). Following step SI.4, the controller 3 performs the Highest Temperature Fire Detection method (shown in Figure 6; described hereinafter). By executing the highest temperature fire detection method, the controller 3 determines if a fire is present (step SI.6). If the received scan data is from a subsequent sweep (negative determination in step SI.3), one or more fire detection methods (described hereinafter) are performed (step SI.9). This may include performing the Highest Temperature Fire Detection method. Preferably, the controller 3 performs every fire detection method hereinafter described concurrently. In response to a positive determination in step SI.6, the controller 3 selects one of the spray head units 3 to activate and supress the fire. The spay nozzle 10 of the selected spray head unit 3 deploys the fire supressing material 5 as hereinbefore described (step SI.8). In an example in which the first system li includes a single spray head unit 2, step SI.7 is not performed. The selection of the spray head unit 2 is made according to which fire detection method returns a positive determination of the presence of a dangerous fire. This will be explained in more detail hereinafter. In response to a negative determination in step SI.6, the first system li performs a further sweep of the room 11. Thus, the first operation method may involve performing multiple sweeps of the room 11. The method terminates once a fire is detected and supressed or when the triggering event no longer applies, for example when smoke is no longer detected in the room 11. Referring now to Figures 6 to 9, various methods of fire detection involving processing single pixel images will now be described. Referring specifically to Figure 6, the Highest Temperature Fire Detection method (herein "first fire detection method") will now be described. The Highest Temperature Fire Detection method is particularly suited to detecting ultra-fast developing fires. An ultra-fast developing fire may be taken as a fire with an extremely high temperature very soon after smoke detection (or after some other triggering event) by the first system li. For example, an ultra-fast developing fire may be taken as a fire with a recorded temperature {i.e., temperature as recorded by the infrared sensor 13) of above 85°C about 9 to 12 seconds after the triggering event. In step S2.1, the scan data is received by the controller 3. In step S2.2, the controller 3 determines whether any readings are at or above 85°C. If there are, any readings below 85°C are discarded (step S2.3). Following step S2.3, the controller 3 identifies the highest temperature reading (step S2.4) and identifies the azimuthal angle corresponding to that reading (step S2.5). This is the position of the fire to the closest increment of the sweep performed by the infrared sensor 13. If two or more readings have the same highest temperature reading, the reading having the lower index will be selected in step S2.4. If the two or more readings have the same index (because they have been captured by different spray head units 2), the reading captured by the spray head unit 2 that was commissioned first will be selected. As used herein, the spray head unit 2 that was commissioned first is taken as the first spray head unit 2 of the relevant spray head units 2 to be paired or connected to the first system li during the initial installation / set-up of the first system li. Likewise, the order in which subsequent spray head units 2 are paired determines their order of precedence. The controller 3 then outputs a positive fire condition (step S2.6): that a fire has been detected. In no readings meet at least the threshold temperature, the controller 3 returns a negative fire condition (step S2.6): that no fire has been detected using the Highest Temperature Fire Detection method. Referring now to Figure 7, a Gradient Fire Detection method (herein "second fire detection method") will now be described. The Gradient Fire Detection method is particularly suited to detecting slowly developing fires. A slowly developing fire may be taken as a fire recorded as growing at about 0.15°C to 1°C per 10 seconds, with its absolute temperature (actual temperature) being less than 39°C. Examples of a slowly developing fire include fires from electric cables and some small plastic electrical appliances. In step S3.1, the scan data is received by the controller 3 (herein "received scan data"). In step S3.2, the controller 3 calculates a Gradient for each reading. For each set of scan data, this calculation consists of: subtracting the reading of the immediately preceding scan data from the respective reading of the received scan data. For example, the 3rd reading of the immediately preceding scan data is subtracted from the 3rd reading of the received scan data for the same spray head unit 2. For the present example, step S3.2 returns 27 Gradients for each spray head unit 2 (the infrared sensors 13 each capture 27 images in one sweep). Thus, the Gradient corresponding to each reading is the temperature value difference between the received scan data and the immediately preceding scan data for a given azimuthal angle for a given spray head unit 2. The Gradient will be positive if there is an increase in temperature from the previous sweep to the present sweep; the Gradient will be negative if there is a decrease. Following the calculation of the Gradients, the controller 3 calculates the Consecutive Rise for each reading in each set of scan data (step S3.3). Each reading has a value of Consecutive Rise prior to step S3.3 (herein "Current Consecutive Rise"). If the received scan data is from the second sweep, then every reading will have a Current Consecutive Rise of 0. To calculate the Consecutive Rise for a given reading: If Gradient >0.15°C, then Consecutive Rise = Current Consecutive Rise + 1 (1) If Gradient <0.15°C, then set Consecutive Rise to 0. Next, in step S3.4, the controller calculates the Modified Gradient Score for each reading. Each reading has a value of gradient score prior to step S3.4 (herein "Current Gradient Score"). If the received scan data is from the second sweep, then every reading will have a Current Gradient Score of 0. This is because the Gradient Scores are reset in the first sweep (see step SI.4). If Gradient >0.45°C, then Modified Gradient Score = Current Gradient Score + 14.1 (2) If Gradient <-0.25°C, then Modified Gradient Score = Current Gradient Score - 32 (3) If -0.25°C <Gradient <0.45°C, the gradient score will remain unchanged i.e., Modified Gradient Score = Current Gradient Score. For any Modified Gradient Score <-30, set the Modified Gradient Score to -30. This is to prevent any Modified Gradient Score <- 30 after completion of step S3.4. If Consecutive Rise (calculated In step S3.3) >2, then Modified Gradient Score = Current Gradient Score + (0.7 x Gradient x sqrt(Consecutive Rise)) (4) In an example where a given reading satisfies the conditions for execution of equations (2) and (4), then the modified gradient score after completing step S3.4 ("Final Modified Gradient Score") will be: Final Modified Gradient Score = Current Gradient Score + 14.1 + (0.7 x Gradient x sqrt(Consecutive Rise)) (5) On completion of step S3.4, a Modified Gradient Score has been calculated for each reading. In step S3.5, the Peakedness is calculated for each reading. New Peakedness values are calculated for each sweep and are not saved on completion of a sweep. Peakedness is an indicator of the presence and extent of peaking in temperature across the sweep for a given spray head unit 2. To calculate a value of Peakedness for each reading: First, define a Peakedness Threshold: Peakedness Threshold = (Max. temp. - average temp.) / 1.5 (6) "Max. temp." is the maximum temperature (reading) within the set of scan data of the relevant reading. As hereinbefore explained, the set of scan data is the scan data gathered from the same spray head unit 2 for the present sweep. "Average temp." is the average temperature (reading) of the set of scan data. Then, for the set of scan data, find the left and right sides of the temperature peak using the Peakedness Threshold. In this case, "left" refers to the reading of lower index and "right" refers to the reading of higher index. To find the left and right sides of the peak: Compare each reading of the relevant set of scan data to the Peakedness Threshold, from lowest index (1st reading) to highest index (in this example, 27th reading). The left side of the peak is taken as the first reading which is at or exceeds the Peakedness Threshold. The right side of the peak is taken as the next reading which is below the Peakedness Threshold. If the right side of the peak is not reached before the final reading in the set, then the final reading (in this example, 27th reading) is taken as the right side. Next, the left and right sides are used to determine the Peak Width: Peak Width = Index of right side - Index of left side (7) For example, if the left side of the peak is taken as the 5th reading and the right side of the peak is taken as the 11th reading, the Peak Width is 6 according to equation (7). The peak in the set of scan data is identified as a Corner Fire if: Peak Width >13 and one side of the peak is taken as an edge reading ( / .e. first or last reading within the set). The Peakedness for a given reading is dependent on Peakedness Weight, which is calculated as follows: Peakedness Weight = Peakedness Threshold / (Peak Width x 60) (8) If Peakedness Weight >26, then Peakedness Weight = 26 This is to prevent any Peakedness Weight exceeding 26. If the peak has been identified as a Corner Fire, the Peakedness Weight as calculated in equation (8) is then adjusted: Adjusted Peakedness Weight = Peakedness Weight - 145 (9) Then, the Peakedness for each reading (herein "Peakedness[index]") with the relevant set of scan data is calculated as follows: If Peakedness Weight >0, then: For readings between and including the right and left side peak readings: Peakedness[index] = Peakedness Weight I abs(index - max. index) (10) For the present example, the maximum index is 27. For all remaining readings in the set: Peakedness[index] = 0 If Peakedness Weight <0, then: For readings between and including the right and left side peak readings: Peakednessfindex] = Peakedness Weight x (abs(index - max. index) / Peak Width) (11) For all remaining readings in the set: Peakedness[index] = Peakedness Weight In the case of a Corner Fire, the maximum index is taken as the index of the left or right edge of the set of scan data (either the 1st or last reading), depending on which side the peak is located. Once Peakedness for each reading has been calculated, the controller 3 calculates a Combined Gradient Score for each reading (step S3.6). This Combined Gradient Score is calculated by summating the Modified Gradient Score calculated In step S3.4 and the Gradient Score from the immediately preceding sweep. This Combined Gradient Score is taken as the Current Gradient Score in the next sweep for step S3.4. In step S3.7, the controller 3 calculates the Final Gradient Score. The Final Gradient Score for each reading is calculated as follows: Final Gradient Score = Combined Gradient Score + Peakedness (12) At the completion of step S3.7, Final Gradient Scores have been calculated for each reading for each set of received scan data. Next, in step S3.8, the controller determines whether any Final Gradient Scores exceed 100. The controller then removes all readings which have a Final Gradient Score at or below 100 (step S3.10). In response to a negative determination in step S3.8 (no scores at or exceeding 100), then the controller 3 returns a negative fire condition (step S3.12): that no fire has been detected using the Gradient Fire Detection method. Once the relevant readings have been removed in step S3.9, the controller 3 identifies the reading having the highest Final Gradient Score (step S3.10) and identifies the azimuthal angle corresponding to that reading (step S3.11). If two or more readings have the same highest Final Gradient Score, the reading having the lower index will be selected in step S3.10. If the two or more readings have the same index (because they have been captured by different spray head units 2), the reading captured by the spray head unit 2 that was commissioned first will be selected. The controller 3 then outputs a positive fire condition (step S3.12): that a fire has been detected. Referring now to Figure 8, the Baseline Increase Fire Detection method (herein "third fire detection method") will now be described. The Baseline Increase Fire Detection method is particularly suited to detecting fast developing fires. A fast developing fire may be taken as a fire with a recorded temperature of more than 50°C which is recorded as growing between 2°C and 3°C per 10 seconds. In step S4.1, the scan data is received by the controller 3. In step S4.2, the controller 3 determines whether any readings are at or above 50°C. If there are, any readings below the threshold temperature are discarded (step S4.3). If there are no readings at or above 50°C, the controller returns a negative fire condition (step S4.9): no fire detected using the Baseline Increase Fire Detection method. Once the relevant readings have been discarded, the controller 3 calculates, for each reading, the difference between the reading and the corresponding baseline reading (saved in step SI.4). This is done by subtracting the baseline reading from the reading from the received scan data (step S4.4). In step S4.5, the controller 3 determines whether any readings have a difference (calculated in step S4.4) at or above 8°C. All readings below this temperature threshold are discarded (step S4.6). If there are no readings with a difference at or above 8 °C, the controller 3 returns a negative fire condition (step S4.9). In step S4.7, the controller 3 identifies the reading having the highest temperature (step S4.7) and identifies the azimuthal angle corresponding to that reading (step S4.8). If two or more readings have the same highest temperature, the reading having the lower index will be selected in step S4.7. If the two or more readings have the same index (because they have been captured by different spray head units 2), the reading captured by the spray head unit 2 that was commissioned first will be selected. The controller 3 then outputs a positive fire condition (step S4.9). Referring now to Figure 9, the Hybrid Fire Detection method (herein "fourth fire detection method") will now be described. The Hybrid Fire Detection method utilises aspects of the Gradient Fire Detection method and the Baseline Increase Fire Detection method. The Hybrid Fire Detection method is particularly suited to medium speed developing fires. A medium speed developing fire may be taken as a fire with a recorded temperature of more than 39°C which is recorded as growing about 1.5°C per 10 seconds. In step S5.1, the scan data is received by the controller 3. In step S5.2, the controller 3 determines whether any readings are at or above 39°C. If there are, any readings below the threshold temperature are discarded (step S5.3). If there are no readings at or above 39°C, the controller returns a negative fire condition (step S5.12): no fire detected using the Baseline Increase Fire Detection method. Once the relevant readings have been discarded, the controller 3 calculates, for each reading, the difference between the reading and the corresponding baseline reading (saved In step SI.4). This is done by subtracting the baseline reading from the reading from the received scan data (step S5.4). In step S5.5, the controller 3 determines whether any readings have a difference (calculated in step S5.4) at or above 8°C. All readings below this temperature threshold are discarded (step S5.6). If there are no readings with a difference at or above 8 °C, the controller 3 returns a negative fire condition (step S5.12). Once the relevant readings have been discarded, the controller calculates the Final Gradient Score for each remaining reading (step S5.7). This is done according to step S3.7 (Figure 7) hereinbefore described. If the Gradient Fire Detection method is being performed concurrently to the Hybrid Fire Detection method, the controller 3 simply uses the Final Gradient Scores already calculated for the Gradient Fire Detection method. In step S5.8, the controller 3 determines whether any Final Gradient Scores are at or exceed 46. Any readings having a Final Gradient Score below this threshold are discarded (S5.9). If there are no readings meeting the criterion in step S5.8, the controller 3 returns a negative fire condition (step S5.12). Once the relevant readings have been discarded, the controller 3 determines which remaining reading has the greatest Final Gradient Score (step S5.10). In step S5.ll, the controller 3 identifies the azimuthal angle corresponding to that reading (step S5.ll) and then outputs a positive fire condition (step S5.12). If two or more readings have the same highest Final Gradient Score, the reading having the lower index will be selected in step S5.10. If the two or more readings have the same index (because they have been captured by different spray head units 2), the reading captured by the spray head unit 2 that was commissioned first will be selected. It should be appreciated that the description of the first to fourth fire detection methods hereinbefore provided are specific implementations of these methods. Thus, different threshold values (e.g., for temperature, Gradient, Modified Gradient Score, Consecutive Rise, Peak Width, and / or Peak Weight) may be selected as appropriate. For example, the temperature threshold for step S2.2 may be between 70°C and 90°C, for example 80°C. For the same reason, different values may be used to adjust the modified gradient score in equations 2, 3, 4, and 5 as appropriate. Referring again to Figures 5 to 9, the operation method of the first system li will be further described. As hereinbefore explained, the controller may perform one, some or all of the fire detection methods as part of step SI.9. In examples in which some or all of the fire detection methods are performed, there is the possibility of more than one of the fire detection methods outputting a positive fire condition. In which case, the controller 3 will select which spray head unit 2 to activate (step SI.7) according to which fire detection method has output a positive fire condition and has the highest precedence. The fire detection methods are ranked in the following order of precedence: 1. Gradient Fire Detection Method 2. Baseline Increase Fire Detection Method 3. Hybrid Fire Detection Method 4. Highest Temperature Fire Detection Method As an example, if both the Gradient Fire Detection method and the Hybrid Fire Detection method output the positive fire condition, the controller 3 will consider the result of the Gradient Fire Detection method for extinguishing the fire. Specifically, the controller 3 selects the spray head unit 2 corresponding to the reading having the highest Final Gradient Score (step S3.10). The spray nozzle 10 of the selected spray head unit 2 deploys the material 5 at the azimuthal angle corresponding to that reading. In other words, the spray head unit 3 which is activated (and the deployment angle of its spray nozzle 10) is set by the reading satisfying all criteria of the fire detection method of highest precedence. The technical advantages the operation method will now be explained with reference to Figure 10a to 10c. Figure 10a shows a plot 16, 16i of temperature rise against scan number as the operation method is performed. The plot 16: illustrates how the temperature increases at an azimuthal angle of 25° as the infrared sensor 13 performs multiple sweeps / scans of a room 11. For example, the plot 16i indicates that the temperature increases by approximately 2°C between the 11th and 12th scans of the infrared sensor 13. Thus, the plot 16i indicates the consecutive growth of the fire at angle 25°. The absolute temperature measured by the infrared sensor 13 at each scan number is shown in Figure 10b. This plot 16, I62 shows that at the 30th scan the temperature is about 35.7°C. The Highest Temperature Fire Detection method considers the absolute temperature of the scan data to determine the presence of a dangerous fire. However, the other fire detection methods additionally consider at least one of the following variables: the rise in temperature over more than one scan, the consecutiveness of the temperature rise, and the extent of temperature peaking across the scan data. In other words, the operation method takes into account the distribution of the temperature across the room and the trends in temperature change over time. A fire behaves different according to its rate of development (e.g. slowly developing fire vs. fast developing fire). The variables hereinbefore mentioned take into account the behaviour of the fire, allowing for the operation method to identify the presence of a fire more accurately and at an earlier stage of development. In this way, the operation method may reduce fire suppression activation times and instances of false positives. This is illustrated in Figure 10c, which shows that the probability of the presence of a fire exceeds 100% (triggering activation) at the 17th scan when the operation method takes into account "Last Growth" (herein referred to as "Gradient), "Consecutiveness" (herein "Consecutive Rise"), and Peakedness. If the operation method only considered the variable Last Growth, the probability would only exceed 100% at the 19th scan. Second fire suppression system 1, 12 As hereinbefore described, the first fire suppression system li is for detecting and supressing a fire. The fire is detected by processing single pixel images captured by the infrared sensor(s) 13. This processing is performed by the controller 3. Alternatively, a fire may be detected by processing multi pixel images. For such examples, the first fire suppression system li is replaced by a second fire suppression system (not shown). The second fire suppression system (herein "second system") is the same as the first system, aside from the differences herein now described. The infrared sensor 13 of the second system is configured to capture images, wherein each image consists of multiple pixels. Each pixel corresponds to a value of temperature. Referring now to Figure 11, the infrared sensor 13 of the second system (herein "multi pixel sensor" 13, 13?) will now be described. Figure 11 shows an example of a full sweep of the multi pixel sensor 13z represented as a set of overlapping sectors. The multi pixel sensor 13? is configured to rotate through a pre-set angular range, which may be greater than 180°, and to capture infrared Images at pre-set azimuthal angles across that range. In a single sweep, more than one image may be captured at a given azimuthal angle. In the preferred example shown in Figure 11, the multi pixel sensor 132 is configured to capture one or more images at three pre-set azimuthal angles in the following sequence: 35°, 90°, and 145°. These pre-set azimuthal angles define the centre of the image(s) captured at each angle. In a preferred example, 2 images are captured at each of the azimuthal angles 35°, 90°, and 145°. Thus, in this example sweep, 6 images are captured, and 2 images may be taken per second. The number of images taken at a given angle during a given sweep may differ for each sweep. In a preferred example, in a first sweep, 2 images are taken at each azimuthal angle of 35°, 90°, and 145°, and in subsequent sweeps, 4 images are taken at each azimuthal angle of 35°, 90°, and 145°. Thus, In this preferred example, 18 images are captured in total over the first and second sweeps. Taking fewer images during the first sweep may provide the advantage of the second system reacting quickly if the fire Is already well developed at the time of activation of the second system. Each image has a field of view which allows a region of the room to be captured in the image. Each pixel in that image corresponds to a subregion of the captured region. In a preferred example shown in Figure 11, each image has an azimuthal angular field of view of 110°. The multi pixel sensor 13? may be configured to have no azimuthal blind spots. This is represented In Figure 11 by the field of view of the first and third images (left and right sectors) extending beyond the wall 12. The multi-pixel sensor 13; may be a Melexis (TM) infrared temperature sensor MLX90640ESF-BAA-000, although other multi pixel sensors may be used. When sensor MLX90640ESF-BAA-000 Is used, each image consists of 798 (32x24) pixels and the sensor captures an image In a field of 110°x75°. Referring now to Figure 12, a method of operating the second system (herein "second operation method") will now be described. As with the first operation method, the decision to activate one of the spray head units 2 is made by the controller 3 in the second operation method. The second operation method involves using one or more trained neural networks to identify the presence of a (dangerous) fire captured in a multi pixel image. The second operation method also involves outputting a location of the fire. Prior to performing the second operation method, the one or more neural networks are trained; this will be described hereinafter. The second operation method is for detecting and supressing fires. The method may be initiated in response to a triggering event, as with the first operation method. For example, the second system 12 may activate to perform the method in response to the smoke detector detecting smoke in the room 11. Merely for the sake of illustration, the second operation method is herein explained using an example in which a single spray head unit 2 is used and the multi pixel sensor 13z is configured to capture 2 images at each of the following three pre-set azimuthal angles in the following sequence: 35°, 90°, and 145°. The second operation method begins, in step S6.1, with the controller 3 receiving or gathering from the spray head unit 2 an infrared image comprising multiple pixels (herein "received image"). This received image is the first image captured at the azimuthal angle of 35° during the first sweep. In step S6.2, the received image may be processed. This processing consists of: Firstly, clipping the temperature values of the pixels to be between 0°C and 600°C inclusive. This means that any temperature values below 0°C are set to 0°C and any temperature values above 600°C are set to 600°C. Then, the ambient room temperature is determined. In a preferred example, the median value of the temperature values after clipping is taken as the ambient room temperature. The ambient room temperature may be determined some other way. For example, the ambient room temperature may be measured by a temperature sensor comprised in the second system 12, for example a temperature sensor installed into the ceiling. The smoke detector, if present, may be configured to measure the ambient room temperature. In the event that the determined ambient room temperature is below or exceeds a pre-set threshold, the ambient room temperature is taken as the threshold value. For example, in a preferred embodiment, the ambient room temperature can only take values from 0°C to 50°C. In this example, if an ambient room temperature value of 65°C is determined, then the ambient room temperature is set to 50°C. Finally, the temperature values are normalized by subtracting the ambient room temperature value from each temperature value and then dividing the resulting temperature values by 600°C. Thus, the processed received image consists of dimensionless values from 0 to 1, each value corresponding to a temperature value of a pixel in the image. It should be appreciated that other temperature threshold values may be used other than 0°C and 600°C, and thus a different, suitable temperature value may be used In the division step. The processing in step S6.2 allows the second system I2 to operate in various ambient temperatures (e.g., an ambient temperature of 45°C, ambient temperatures of rooms that may be naturally hot, such as kitchens, etc.), without high ambient temperatures causing the second system I2 to erroneously conclude that there is a fire. As hereinbefore mentioned, step S6.2 has been described in relation to the case in which the received image is the first image captured at the azimuthal angle of 35° during the first sweep. In the same sweep, a further image is captured at 35° and then 2 images at each of the following azimuthal angles: 90° and 145°. For subsequently received images, the ambient room temperature is either determined as hereinbefore described or taken as the same ambient room temperature as a preceding received image. In a preferred example, the ambient room temperature for the received image first captured at 35° is taken as the ambient room temperature for all received images of that sweep. In the next sweep, a new ambient room temperature is determined for the first received image. Thus, the ambient room temperature may be updated every e.g., 5 minutes while scanning. In some examples, the received image may not be processed (in other words, step S6.2 is omitted). In such examples, the neural network(s) are trained to operate with data that is not processed / normalized. Merely for the sake of illustration, the remaining steps of the second operation method are described with respect to an example in which the received image has been processed. Next, in step S6.3, the controller 3 determines the probability that a fire is present in the room 11 using the received image. A fire condition is output by the controller 3 on completion of step S6.3. This step will be described in more detail hereinafter with reference to Figure 13. In the present example, the received image is the first image received by the spray head unit 2 (e.g. in response to the triggering event) as part of the second operation method. Before, after, or concurrently to step 6.3 (but after step S6.2, if performed), the controller 3 determines the number and position of hot objects in the received image. A hot object is a region of the received image where spiking in the temperature occurs. A hot object may represent a dangerous fire or some other source of heat in the room 11. After determining the number and location of the hot objects, the location of the fire is determined by the controller 3 (step S6.4). Step S6.4 will be described in more detail hereinafter with reference to Figures 16 and 18. Using the fire condition output in step S6.3, the controller 3 determines whether a fire is present (step S6.5). In response to a positive determination: in the present example, the controller 3 simply Instructs the single spray head unit 2 to activate and deploy the fire supressing material 5 (step S6.7). The azimuthal angle at which the nozzle 10 deploys the fire supressing material 5 is dictated by the output of step S6.4. In other examples in which a plurality of spray head units 2 are used, the controller 2 selects which spray head unit 2 to deploy in step S6.6. The controller 3 selects the spray head unit 2 which corresponds to the received image which first results in a positive determination in step S6.5. If a positive determination occurs for two or more spray head units 2 simultaneously (or within a pre-set time window, such as 1 to 3 seconds), then the spray head unit 2 which was commissioned first is selected. In response to a negative determination in step S6.5, the second system 12 may continue the sweep of the room 11. In examples in which the received image is the last image captured in the sweep, the second system lz may begin a new sweep of the room 11. Thus, the second operation method may involve performing multiple sweeps of the room 11. In other examples in which a negative determination is returned, the controller 3 may activate a focus mode in which a subsection of the room is more precisely scanned instead of the sweep being continued / a new sweep being commenced. The focus mode will be described hereinafter in more detail. The method terminates once a fire is detected and suppressed or when the triggering event no longer applies, for example when smoke is no longer detected in the room 11. Referring now to Figure 13, the method of determining the probability of the presence of a dangerous fire (step S6.3; Figure 12) as part of the second operation method will now be described. Firstly, in step S7.1, the controller 3 determines the index, n, of the received image. The index of the received image is set according to the number of preceding images received from the same spray head unit 2 and captured at the same azimuthal angle since the second operation method was commenced. In the present example, the received image has index n = 1 because the image was taken in the first sweep of the spray head unit 2. If a second sweep was performed, the first received image captured at an azimuthal angle of 35° for that sweep would have index n= 3. If n is below a threshold index value, then the received image is fed or input into a first neural network (step S7.2) (herein "first neural network input"). The threshold index value is preferably 6. Thus, the first neural network input takes the form of a set of values, each value corresponding to a temperature value recorded for a respective pixel in the received image. If the received image has been processed in step S6.2, the set of values consists of dimensionless values from 0 to 1. In the present example, the received image consists of 768 (32x24) pixels. Thus, the set of values of the first neural network input consists of 768 values. The first neural network is a convolutional neural network. Referring to Figure 14, an example architecture of the first neural network is shown. The first neural network according to the example architecture comprises an input layer (labelled "Input_IR_Frame: InputLayer"), an output layer (labelled "Probability ^OLFire: Dense"), and several hidden layers. In the example in Figure 14, the received image consists of 768 (32x24) pixels, which are fed into the input layer. The hidden layers consist of several zero padding layers (labelled "zero_padding2d: ZeroPadding2D" and similar), two-dimensional convolutional layers (labelled "conv2d: Conv2D"), batch normalization layers (labelled "batch normalization: BatchNormalization" and similar), activation layers (labelled "activation: Activation" and similar), and max pooling layers (labelled "max_pooling2d: MaxPooling2D"). The penultimate layer of the example architecture is a dropout layer (labelled "dropout: Dropout") and the immediately preceding layer is a global average pooling layer (labelled "global_average_poollng2d: GlobalAveragePooling2D"). As will be hereinafter described, the first neural network has been trained to recognise the presence of a dangerous fire within a received image. The output of the first neural network is a percentage probability that such a fire is captured in the received image. As shown in Figure 14, the output layer is configured to produce this single output. Referring again to Figure 13, after the received image is input into the first neural network, the output layer of the network outputs a probability that the received image captures a fire (step S7.3). Then, the controller 3 outputs or signals a fire condition (step S7.4). If the percentage output in step S7.3 exceeds a threshold percentage, the controller 3 outputs a positive fire condition: that there is a fire captured in the received image. If the percentage output in step S7.3 does not exceed the threshold percentage, the controller 3 outputs a negative fire condition: that no fire has been captured in the received image or that there is insufficient certainty that a fire has been captured in the received image. As will be hereinafter described, outputting a negative fire condition, in some cases, causes the controller 3 to active the focus mode. In some examples, the received image (herein "present received image") is preceded by a set of received images from previous sweeps of the spray head unit 2 (herein "history of images") which have been captured at the same azimuthal angle. In which case, information regarding trends or changes across these images may be used as part of determining the probability that the present received image captures a fire. This can improve the accuracy of the second operation method. The history of images may consist of between 15 and 25 images, for example 18 images. The history of images may consist of the number of images taken over a pre-set period of time, for example 9 seconds. In the focus mode of the second operation method, which will be hereinafter described, the history of images may consist of the number of images taken over a pre-set period of time, for example 9 seconds. In the focus mode, the history of images may be set to have an upper limit of 18 images. Referring again to step S7.2, if the controller 3 determines that the index, n, of the present received image meets or exceeds the threshold Index value, the present received image is input into the first neural network (step S7.5). In response to this input, the first neural network outputs an intermediate output from a hidden layer of the network (step S7.6), rather than a probability output from the output layer. The intermediate output takes the form of a set of float values resulting from the partial processing of the first neural network input by one or more of the hidden layers. The set of float values is fewer than the number of values of the first neural network input (in other words, less than 768 values for the present example). For example, In the present case, the set of float numbers may consist of 64 values. Intermediate outputs are generally easier to subsequently process because processing fewer values saves computing power. Either before, concurrently, or after step S7.6, the controller 3 extracts one or more features from the history of images (step S7.7). Examples of features extracted from the history of images Include: the maximum temperature value in each previously received image, and the temperature value corresponding to the 98th percentile (or some other percentile) in each previously received Image. Other examples include the slope or gradient of the best fit line for a plot of the 100th percentile for each image in the history of images (herein "slope"). In this case, the 100th percentile data is plotted against the timestamp of the images. Further examples of extracted features include the slope or gradient of the best fit line for a plot of the 98th percentile for each image in the history of images (herein "slope"). Other examples of extracted features include the momentum of the 98th percentile for the history of images, which is the result of the subtraction of the 98th percentile of the 1st Image In the history of images from the 98th percentile of the last image In the history of images. Likewise, another example is the momentum of the 100th percentile for the history of images, which is calculated in the same way as the momentum for the 98th percentile. In a preferred example, 6 features are extracted from the history of images. Next, in step S7.8, the extracted feature(s) and the intermediate output are input into a second neural network. The second neural network is a convolutional neural network. Training of the second neural network will be hereinafter described. Referring now to Figure 15, an example architecture of the second neural network is shown. Similar to the first neural network, the second neural network according to the example architecture comprises an input layer (labelled "Input_Feature_Vector: InputLayer"), an output layer (labelled "Probability_Of_Fire: Dense"), and several hidden layers. As shown in Figure 15, the input layer is configured to receive fewer inputs in comparison to the input layer of the first neural network (Figure 14). As with the first neural network, the output layer of the second neural network is configured to produce a single output. The hidden layers consist of two batch normalization layers (labelled "batch_normalization: BatchNormalization" and similar) and two activation layers (labelled "activation: Activation" and similar), as well as several dense layers (labelled "dense: Dense" and similar), one of which is the penultimate layer. The layer immediately preceding the last dense layer is a dropout layer (labelled "dropout: Dropout"). Referring again to Figure 13: In response to the inputs in step S7.8, the output layer of the second neural network outputs a probability that the received image captures a fire (step S7.9). Then, the controller 3 outputs or signals a fire condition based on the probability (step S7.10) as previously described in reference to step S7.4. Referring now to Figure 16, the method of determining the location of the fire (step S6.4; Figure 12) as part of the second operation method will now be described. Firstly, in step S8.1, the controller 3 determines the number and position of hot objects present in the received image. The process of determining the number and location of hot objects is herein described with reference to the present example in which the received image has 32 x 24 pixels ( / .e. 32 columns, each column consisting of 24 pixels). Firstly, for each column, the controller 3 calculates a Temperature Variation value (herein "TV") using equation (13): TV[x] = Maximum temperature value of column[x] - median temperature value of image (13) "x" denotes the index of each column, wherein the left-most column has the lowest index (1) and the right-most column has the highest index (32). Then, equation (14) is used to calculate a Modified Temperature Variation value ("herein TV_m") for each column: TV_m[x] = TV[x] - median temperature value of column x (14) If TVm[x] <0, then set TV m[x] to 0. Then, a threshold value (herein "Column Threshold") is defined using equation (15): Column Threshold = maximum TV_m[x] value I 2.5 (15) This Column Threshold value is used to identify the left and right sides of a hot body. The left side of the hot body is defined as column having index "a" and the right side of the hot body is defined as column having index "b". To identify indices a and b: First, identify the maximum TV_m[x] value and compare each TV_m[x] value left of the maximum TV_m[x] value, starting with the TV_m[x] value directly adjacent to the maximum TV_m[x] value on the left. The index of the first TV_m[x] value (left of the maximum TV_m[x] value) exceeding the Column Threshold value is set as "a". Next, compare each TV_m[x] value right of the maximum TV_m[x] value, starting with the W m[x] value directly adjacent to the maximum TVm[x] value on the right. The index of the first TV_m[x] value (right of the maximum TV m[x] value) exceeding the Column Threshold value is set as "b". Next, a Location Index is calculated using equation (16): Location Index = (a + b) / 2 (16) Next, a Peak Height is calculated. The Peak Height is used to determine the presence of a hot object between indices a and b. Peak Height = maximum TV_m[x] value - minimum TV_m[x] value from a to b (17) The expression "minimum TV m[x] value from a to b" in equation (17) refers to the minimum value of 7V m[x] out of the values of 7V m[x] calculated for columns a through to b. If Peak Height <4, no hot object is present. If Peak Height >4, a hot object Is present from indices a to b. It should be appreciated that a Peak Height threshold value other than 4 may be used. The location of this hot object is defined by the Location Index hereinbefore calculated in equation (16). The Location Index is expressed in terms of column indices and corresponds to an azimuthal angle to which the spray nozzle 10 can be turned. There may be more than one hot object captured by the received image. Therefore, to identify additional hot objects, all values of TV_m[x] calculated for columns a through to b are set to -1. In some examples, one more values of TV_m[x] for columns adjacent to a and b are also set to -1 (to allow for gaps between hot objects). Then, the method of determining the number and location of a hot object herein described is repeated (second iteration). The second iteration begins by identifying the new minimum TV .m[x] value and identifying the new indices a to b as herein described. Multiple iterations may be performed. A new Column Threshold value Is calculated for each iteration. When no hot object is found in an Iteration (Peak Height <4), the method of determining the number and location of the hot object(s) is terminated and no subsequent iterations are performed. Referring also to Figures 17a and 17b, two infrared images are shown. Each infrared image captures a region identified as a single hot object according to the method in step S8.1. Each infrared image consists of pixels corresponding to different temperatures. Some of the pixels have been labelled to indicate their temperature (specifically, the range In which the temperature falls): "Br", "R", "O", "Y", "G", "T", "Bl", "N". Each infrared image has a corresponding graph aligned with the column indices of the infrared image. Each graph includes a plot of Modified Temperature Variation value, TV_m against column index. Indices a and b are defined by the column indices at which the Column Threshold value intersects the TV_m plot. Referring specifically to Figure 17a, the Location Index (labelled "LI" in the figure) Is identified as column index 17. Referring now to Figure 17b, the Location Index Is identified as column index 16. Each graph also includes a plot of fire location probability, Pnns, for each column index (calculated In step S8.5 hereinafter described). This probability Pnns for each column index is calculated by a third neural network (which is also hereinafter described). The plot is inverted; the top of the graph corresponds to a probability of 0% and the base of the graph corresponds to a probability of 100%. Once the method of determining the number and location of the hot object(s) is complete, the controller 3 determines whether more than one hot object has been found (steps S8.2). If only one hot object has been found, the controller 3 outputs a location of the fire (step S8.3) based on the Location Index hereinbefore calculated in equation (16). As explained, the Location Index corresponds to an azimuthal angle to which the spray nozzle 10 can turn to suppress the fire. Thus, in the example shown in Figure 17a, the location of the fire is identified at column Index 17. In the example shown in Figure 17b, the location of the fire is identified at column index 16. If more than one hot object is found, the third neural network is used to determine the location of the fire in step S8.4. Specifically, the intermediate output of the first neural network hereinbefore described is input into the third neural network. The third neural network is a convolutional neural network. The third neural network is trained to output a probability of a fire corresponding to each column of the received image. Thus, in the present example, the third neural network outputs 32 probabilities. Referring now to Figure 18, an example architecture of the third neural network is shown. Similar to the neural networks hereinbefore described, the third neural network according to the example architecture comprises an input layer (labelled "Input Feature^V InputLayer"), an output layer (labelled "FireLocation Probability: Dense"), and several hidden layers. As shown in Figure 18, the output layer is configured to produce a plurality of outputs: in the present example, 32 outputs. The hidden layers consist of two batch normalization layers (labelled "batch_normalization: BatchNormalization" and similar) and two activation layers (labelled "activation: Activation" and similar), as well as several dense layers (labelled "dense: Dense" and similar), one of which is the penultimate layer. The layer immediately preceding the last dense layer is a dropout layer (labelled "dropout: Dropout"). Referring again to Figure 14: In response to receiving the intermediate output of the first neural network, the third neural network outputs a probability corresponding to each column (steps S8.5). Based on the output of the third neural network, the location of the fire is determined (step S8.6). Referring also to Figure 19, the way in which the location of the fire is determined in step S8.6 will now be determined. First, an average of each column probability is calculated (step S9.1) from the present received image and previously received images. In a preferred example, for a given column index, x, the probability corresponding to that column for the present received image is summated with the probabilities for the preceding 3 images for that column. The total probability calculated in this step is divided by 4 to produce an average column probability, Pt_x. In other examples, a different number of previously received images may be used. If the present received image is the first received image for a given azimuthal angle, then the averaging in step S9.1 merely involves taking the probability corresponding to a given column x for the present received image as the average column probability, Pt_x. In the present example, 32 values of Pt x are calculated. Next, we consider the probability that a fire is present at the location in which a hot object has been identified (Figure 14; step S8.1). In step S9.2, a total hot object probability, Pho, is calculated for each identified hot object. For each hot object, the probability across the peak width (from indices a to b) is summated according to equation (18): p = yb p (is) In step S9.3, the total hot object probability, Pho, is then normalised for each identified hot object by dividing by the hot object width. This is done according to equation (19): P^o=^ (19) The greatest value of Pnho is determined for the received image (step S9.4). Then, it is determined whether this value is the same as or exceeds 0.1 (step S9.5). In response to a positive determination in step S9.5, the location of the fire for step S8.6 is taken to be the location of the hot object corresponding to the greatest value of Pnho (step S9.6). In response to a negative determination in step S9.5, the results / output of the third neural network are disregarded. Instead, the location of the fire for step S8.6 is taken to be the hot object having the greatest Peak Height (see equation (17)). As hereinbefore described, if the controller determines that a fire is present in step S6.5, the nozzle 10 is controlled to deploy fire supresslng material 5. The direction at which the nozzle 10 deploys the material 5 is towards the location of the fire, as determined in either step S9.6 or step S9.8. Referring to Figures 20a and 20b, two infrared images are shown. Each image captures two hot objects (as identified according to the method in step S8.1 hereinbefore described). In each Image, one of the hot objects is a heater and the other hot object is a dangerous fire. As with Figures 17a and 17b, each infrared image has a corresponding graph aligned with the column indices of the infrared image. Each graph includes a plot of Modified Temperature Variation value, TV_m against column index and a plot of fire location probability, Pnns, for each column index - as hereinbefore described. Referring specifically to Figure 20a, the two hot objects have been identified with Location Indices of column index 13 and column index 19. Referring now to Figure 20b, the two hot objects have been identified with Location Indices of column index 16 and column Index 25. By performing the method according to step S8.6 hereinbefore described, the location of the dangerous fire is identified at column index 13 in Figure 20a and column Index 16 in Figure 20b. In both Figures 20a and 20b, the peak in the probability Pnn3 occurs at the same column index as the location of the fire. As hereinbefore explained, if only one hot object has been found (such as in Figures 17a and 17b), the third neural network is not used to determine the location of the fire - and Instead the Location Index is used. Thus, in Figure 17a, although peaks in the probability Pnns occur at column indices 15 and 25, the fire is identified at column index 17. Likewise, in Figure 17b, peaks in the probability Pnn3 occur at column indices 4, 12, and 30, but the fire is identified at column index 16. Focus mode In some examples of the second operation method, the controller 3 is configured to activate a focus mode in response to retuning a negative fire condition in step S6.3 (Figure 12), such as in step S7.4 (Figure 13) or step S7.10 (Figure 13), for a given infrared image (herein referred to as "activating infrared image"). The focus mode allows for more precise monitoring of a room compared to when a standard sweep of the room is performed. This is because this operational mode involves focussing fire monitoring to a subsection of the room, where a probability reading suggests that there is a small but developing fire in that subsection. Thus, the focus mode helps to identify the presence of a dangerous fire more quickly. As hereinbefore described, a negative fire condition occurs when the probability of a fire output in step S7.3 or step S7.9 (herein referred to as "fire probability") is at or below a threshold percentage. The threshold percentage may be 86.1%, or greater. The focus mode may not be activated if the fire probability of the infrared image returning a negative fire condition is no more than 15%. In the focus mode: If the fire probability of the activating infrared image is between a first pre-set probability and a second pre-set probability, the controller 3 controls the multi-pixel sensor 132 to take a plurality of infrared images in succession at the azimuthal angle of the activating infrared image for a pre-set period. This operation sequence is herein referred to as the "timed focus operation". The timed focus operation terminates when either i) the pre-set period elapses, or ii) a fire probability of one of the images of the plurality of infrared images meets or exceeds the second pre-set probability. This will be described in more detail hereinafter. If the pre-set period elapses without condition ii) being satisfied, the second system returns to standard operation. In other words, the second system returns to performing the sweep of the room as hereinbefore described. The controller 3 considers the infrared images sequentially. This means that as soon as either condition i) or ii) herein described is satisfied, the timed focus operation terminates - regardless of the fire probability of later infrared images of the plurality of infrared images. The pre-set period may be between 8 second and 12 seconds, for example 10 seconds. The first pre-set probability may be between 10% and 20%, for example 15%, and the second pre-set probability may be between 50% and 70%, for example 60%. If the fire probability of the activating infrared image is at least the second pre-set probability but no more than a third pre-set probability, the controller 3 controls the multi-pixel sensor 13? to take a plurality of infrared images in succession at the azimuthal angle of the activating infrared image. This operation sequence occurs until either i) a fire probability of one of the images of the plurality of infrared images exceeds the third pre-set probability, or ii) a fire probability of one of the images of the plurality of infrared images falls below the second pre-set probability. This operation sequence is herein referred to as "advanced focus operation". In the case where a fire probability of one of the images of the plurality of infrared images exceeds the third pre-set probability, a positive fire condition is signalled. As with the timed focus operation, the controller 3 considers the infrared images sequentially during the advanced focus operation. This means that as soon as either condition i) or ii) herein described is satisfied, the timed focus operation terminates -regardless of the fire probability of later infrared images. The second pre-set probability may be between 50% and 70%, for example 60%. The third pre-set probability may be between 80% and 90%, for example 86.1%. As hereinbefore explained, the timed focus operation may terminate when a fire probability of one of the infrared images meets or exceeds the second pre-set probability. If the fire probability of this infrared image is at least the second pre-set probability but no more than the third pre-set probability, the advanced focus operation Is then performed. If the fire probability of this Infrared image exceeds the third pre-set probability, then a positive fire condition is signalled. For both the timed focus operation and the advanced focus operation, the fire probability for each image of the plurality of infrared images is determined according to steps S7.1 to S7.3 or steps S7.1 to S7.9 hereinbefore described. The method used to determine the fire probability may depend on the history of images gathered, as explained in more detail below. For both the timed focus operation and the advanced focus operation, the plurality of infrared images may be captured at a rate of 1 to 2 infrared images per second, preferably 2 Images per second. If the focus mode results in a positive fire condition being signalled, the location of the fire is determined as hereinbefore described. The location of the fire is determined based on the infrared image, captured in the focus mode, which caused the positive fire condition. Referring to Figure 21, a plot of fire probability for a succession of infrared images taken during the focus mode hereinbefore described is shown, labelled plot "a" ("smoothed 6 p final"). Each fire probability corresponds to an infrared image taken between times 08:54:30 and 08:57:00 at the same azimuthal angle. The fire probabilities of plot "a" were calculated according to step S7.9 (Figure 13); in other words, plot "a" is a plot of second neural network outputs. Figure 21 also shows another plot of fire probability for the succession of infrared images, labelled plot "b" ("smoothed_6_prob_nn_base"). The fire probabilities of plot "b" were calculated according to step S7.3 (Figure 13); in other words, plot "b" is a plot of first neural network outputs. The 100th temperature percentile for each infrared image is also plotted in Figure 21, labelled plot "c" ("single_frame_100.0_Percentile"). Likewise, the 98th temperature percentile for each infrared image is plotted and labelled plot "d" ("single_frame_98.0_Percentile"). Throughout the focus mode, the 100th temperature percentage is constant at 65°C and the 98th temperature percentage is constant at 40°C. Both the 100th and 98th temperature percentiles, and their derivatives of slope and momentum, are examples of features input into the second neural network during step S7.8 (Figure 13). In the present example, each of these four features are input into the second neural network. The fire probability used to compare with the first / second / third pre-set probability may depend on the index of the image in the succession of images taken during the timed focus operation / advanced focus operation. For example, if the threshold Index value is 6, and an image captured during the timed focus operation has an index of 5, the probability used to compare with the first / second pre-set probability is that output by the first neural network (plot "b"). Whereas, for a later image captured during the same timed focus operation having an index of 7, the probability used to compare with the first / second pre-set probability is that output by the second neural network (plot "a"). In a preferred example, the threshold index value is 6. At time ti in Figure 21, the threshold index value has not been exceeded. At this time, the "snioothed6^p fire probability (plot "b") exceeds 0.15 and the timed focus operation is activated. A first succession of infrared images is taken at the relevant azimuthal angle until the pre-set period elapses at time t2. The timed focus operation also occurs from ts to U, from ts to te, and from t? to ts. In each case, the timed focus operation terminates when the pre-set period elapses. By time tg, a sufficient history of images has been collected and the threshold value has been exceeded, and so the "smoothed6_p probability (plot "a") is considered instead. Just after this time, the "smoothed_6_prob_nn_final" probability is at least 0.6 and so the advanced focus operation is activated. There is no time-out condition for the advanced focus operation. Thus, the succession of infrared images continues to be taken beyond time 08:57:00. Figure 21 further includes plots relating to the location of the fire (plots "e" and "f"). Plot "e" ("Location Combined Predictions") is a plot of the fire location, as calculated in step S9.6 (Figure 19), for each infrared image. Plot "f" ("Hottest Object Location") is a plot of the location of the hottest object for each infrared image. For the example in Figure 21, a single hot object was identified and taken as the location of the fire, which was at column index 20 for each infrared image. Thus, as only a single hot object was identified, plots "e" and "f" are overlaid in Figure 21. Precise Angle Detection In some examples of the second operation method, the controller 3 is configured to perform a precise angle detection method, which will now be explained with reference to Figures 22a to 22d. The precise angle detection method may be performed when a single hot object has been identified in the received image during step S8.1 (Figure 14). If the hot object is not at the centre or substantially at the centre of the received image, the azimuthal angle of the multi-pixel sensor 132 is adjusted (to an "adjusted azimuthal angle") such that the hot object is at the centre of the sensor's field of view (and, thus, at the centre of an infrared image captured at the adjusted azimuthal angle). At the adjusted azimuthal angle, the hot object has an adjusted Location Index, which corresponds to the adjusted azimuthal angle. If this hot object yields a positive fire condition, the spray nozzle 10 can then be turned to the adjusted azimuthal angle and deploy the fire supressing material 5. Referring to the example shown in Figure 22a, the received image is captured at an azimuthal angle of 145° and comprises a hot object at Location Index 21. The multiple-pixel sensor 13? is then adjusted to an angle of 131° such that the hot object is at Location Index 15 ( / .e. at the centre of the received image); the infrared image corresponding to the adjusted azimuthal angle is shown in Figure 22b. Figures 22c and 22d capture the deployment of the fire supressing material 5 by the spray nozzle 10 towards the area of the room corresponding to Location Index 15. The mist first obscures, then extinguishes the hot object. Fire chasing In some examples of the second operation method, the controller 3 is configured to perform a fire chasing method. The fire chasing method is particularly suited for tackling travelling fires. This is because the method involves readjusting the angle at which the multi-pixel sensor 132 scans the room. The fire chasing method will now be explained with reference to Figure 23. As hereinbefore described, fire supressing material 5 may be deployed as part of the second operation method (Figure 12). The direction at which the spray nozzle 10 deploys the material 5 is towards the location of the fire (herein "original fire location"), as determined in step S8.3 (Figure 16), step S9.6, or step S9.8 (Figure 19). The fire chasing method only occurs when the second system 12 has already been activated to extinguish a fire (step S10.1) at the original fire location. As the material 5 is being deployed, the multi-pixel sensor 13? captures a stream of infrared images at the azimuthal angle corresponding to the original fire location (step S10.2). The controller 3 then determines whether one of the images of the stream of infrared images satisfies a condition (step S10.3), triggering activation of the fire chasing method. In a preferred example, the condition is one of the infrared images having an absolute 98th temperature percentile of at least 40°C, wherein the average temperature rise between each preceding image is at least 0.1°C over the last 20 seconds. In other examples, the absolute 100th temperature percentile may be considered instead. In other examples, other temperature thresholds and / or other average temperature rises may be considered. If the condition is not satisfied, the multi-pixel sensor 132 continues to scan at the original fire location whilst the spray nozzle 10 continues to deploy the material 5. If the condition is satisfied, the pump 4 is deactivated (step S10.4). Once the pump 4 is deactivated, the controller 3 adjusts the azimuthal angle of the multi-pixel sensor 132 to a first adjusted angular position (step S10.5). Immediately after the pump 4 is deactivated, the pressure in the second system I2 may be too high for the multi-pixel sensor 132 to be turned. Therefore, in a preferred example, the multi-pixel sensor 132 is adjusted between 3 and 5 seconds after the pump 4 is deactivated. In some examples, the spray nozzle 10 may still be deploying material 5 when the sensor 132 is moving to the first adjusted angular position and, optionally, the second angular position (described hereinbelow). Next, the multi-pixel sensor 132 captures an infrared image (herein "first adjusted infrared image") at the first adjusted angular position (step S10.6). During deployment of the material 5, the mist produced by the spray nozzle 10 obstructs a region of the room from accurate temperature sensing, typically ±7° from the location of the fire. Therefore, preferably the first adjusted angular position is at least ±7° from the original fire location. If the original fire location, X, is less than 90°, the first adjusted angular position may be X + 37°. If the original fire location, X, is 90° or greater, the first adjusted angular position may be X - 37°. In other examples, the first adjusted angular position may be at other angles. In some examples of the fire chasing method, the controller 3 adjusts the azimuthal angle of the multi-pixel sensor 132 to a second adjusted angular position (step S10.7) after capturing the first adjusted infrared image. The multi-pixel sensor 13z captures an infrared image (herein "second adjusted infrared image") at the second adjusted angular position (step S10.8). The second adjusted angular position may be between the first adjusted angular position and the original fire location. If the original fire location, X, is less than 90°, the first adjusted angular position may be X + 22°. If the original fire location, X, is 90° or greater, the first adjusted angular position may be X - 22°. In other examples, the second adjusted angular position may be at other angles, such as angles beyond the first adjusted angular position. The controller 3 determines whether a hot object is present (step S10.9) in the first adjusted infrared image and, if captured, the second adjusted infrared image. This is done according to the method in step S8.1 (Figure 16). Determining whether a hot object is present in the first infrared image may be done before, during, or after the second adjusted infrared image is captured. If no hot object is identified in step S10.9, the fire chasing method is terminated and the controller 3 simply controls the spray nozzle 10 to deploy the material 5 towards the original fire location. If at least one hot object is identified in step S10.9, the controller 3 then determines an adjusted location of the fire (step S10.10) and controls the spray nozzle 10 to deploy the material 5 towards this adjusted fire location (step S10.ll). Thus, the second system 12 can "chase" a fire as it travels to extinguish it. To determines the adjusted location of the fire, the controller 3 first determines the location of the hot object(s) in the first adjusted infrared image and, if captured, the second adjusted infrared image. The location of each hot object is determined according to step S8.1. Typically, the first adjusted infrared Image (and, optionally, the second adjust infrared image) will capture a single hot object. However, if more than one hot object Is captured in one of the adjusted infrared images, the hot object having the greatest Peak Height (see equation (17)) Is taken as the relevant hot object used in step S10.10. If only one of the adjusted infrared images contains a hot object(s), the adjusted location of the fire is taken as the location of the relevant hot object of that image. If a hot object(s) is found in both the first and second adjusted infrared images, there are two relevant hot objects to consider - one from each image. In which case, the adjusted location of the fire is taken as the average of the locations of the relevant hot objects. In alternative examples of the fire chasing method, more than one first adjusted infrared image may be taken during step S10.6 and, optionally, more than one second adjusted Infrared image may be taken during step S10.8. This may be done if a hot object is not captured / cannot be detected in the first image taken during step S10.6 and / or step S10.8. In other alternative examples of the fire chasing method, the controller 3 may determine the probability of a fire being present in the first adjusted infrared image(s) (and second adjusted infrared image(s), if captured) during step S10.9 when a hot object has been identified in the Image(s). In which case, the image having the highest probability may be used in the determination of the adjusted fire location in step S10.10. Preferably, the probability of a fire being present in the first / second adjusted infrared image(s) is determined according to step S7.9 (Figure 13). This may involve using alternative versions of the first and second neural networks which have been trained on infrared images which capture the material 5 being deployed by the spray nozzle 10. The fire chasing method terminates when either the material 5 is deployed at the adjusted fire location or the spray nozzle 10 resumes deploying the material 5 at the original fire location. In some examples, the controller 3 is configured to "lock" the fire chasing method for a pre-set period, preferably 2 minutes, when the fire chasing method terminates. The second system 12 does not perform the fire chasing method when locked, regardless of whether the condition in step S10.3 is satisfied. This is to avoid effective fire suppression being prevented by unnecessary interruptions from the sensor 132 scanning at adjusted positions. A sequence of infrared images is shown in Figures 24a to 24d; these images were captured during an example performance of the fire chasing method. In Figure 24a, the spay nozzle 10 is deploying material 5 towards the original fire location and the fire is obscured by the mist. The azimuthal angle of the multi-pixel sensor 132 is at 135°. In Figure 24b, the multi-pixel sensor 132 has been adjusted to the first angular position of 98°, revealing a hot object. The hot object is identified as a fire at Location Index 3. In Figure 24c, the multi-pixel sensor 132 has been adjusted to the second angular position of 113° and the hot object is now identified as a fire at Location Index 7. In Figure 24d, the spay nozzle 10 is deploying material 5 towards the adjusted location of the fire and the fire is obscured by the mist. The azimuthal angle of the multi-pixel sensor 132 is now at 136°. This new azimuthal angle of the multi-pixel sensor 132 is set by the adjusted fire location determined in step S10.10. Wet sensor During deployment of the fire supressing material 5, the lens of the multi-pixel sensor 132 is likely to get wet. The infrared image in Figure 24a shows an example infrared imaged for which the lens is dry. The mist created by the deploying material 5 is shown in this image by the colder columnar region between approximately column index 12 and 19. When the lens gets wet, the captured image is distorted. Figure 25a shows an example infrared image in which the lens is wet. The heat map now forms a substantially concentric circular pattern, and the columnar region is absent. This distortion by the wet lens reduces the accuracy of the infrared image and may prevent or delay activation of the fire chasing method. Therefore, the controller 3 can be configured to disregard an infrared image (captured in step S10.2; Figure 23) which has become too distorted by the wet lens. The extent of distortion is measured by a symmetry score, which will now be described with reference to Figures 25a to 25c. Figure 25a shows an example received infrared image which has been distorted by the wet lens. The symmetry score is calculated by processing such a received infrared image. Firstly, the median temperature value (50th temperature percentile) is calculated for the received infrared image. The corners of the received infrared image are then replaced with the median temperature value, as shown in Figure 25b. This step is done because readings at the edge or periphery of an infrared image are typically less accurate than readings towards the centre. In a preferred example, the 21 pixels at each corner are replaced by the median temperature value. Secondly, a median filter (3x3 window) is applied to the received infrared image to reduce noise. In the example shown in Figure 25c, the infrared image of Figure 25b has been applied with the median filter. Once the median filter has been applied, the received infrared image (herein "filtered image") can be used to calculate the symmetry score. To calculate the symmetry score of the filtered image, the image is divided into a first group of pixels and a second group of pixels. The divide may be made along the vertical of the filtered image, intersecting its central point (in other words, the image is divided symmetrically). This vertical divide is shown In Figure 25c as the dashed line "v". Next, each pixel of the first group is subtracted from a respective, mirroring pixel of the second group. Herein, "mirroring pixel" refers to the pixel having a mirrored location, relative to the divide, to the relevant pixel of the other group. For each division, if the result is a non-zero value, then the first group pixel and the mirroring pixel of the second group differ in temperature value. Before, subsequently, or concurrently to the subtraction step hereinbefore described, the filtered image may be divided along the horizontal, intersecting its central point. This horizontal divide is shown in Figure 25c as the dashed line "h". The horizontal divide defines a further two groups of pixels: a third group and a fourth group. Each pixel of the third group is subtracted from a respective, mirroring pixel of the fourth group. For each division, if the result is a non-zero value, then the third group pixel and the mirroring pixel of the fourth group differ in temperature value. The symmetry score of the filtered image is the number of mirroring pixels which differ in temperature value (in other words, the number of subtractions resulting in a non-zero value). The lower the symmetry score, the more distorted the image is due to the wet lens. Figure 26a shows a plot of symmetry score for a plurality of infrared images captured sequentially. The lens of the multi-pixel sensor 13? becomes wet approximately when the 140th infrared image is captured. After the lens becomes wet, the symmetry score plateaus to around 100. Figure 26b, by contrast, shows a plot of symmetry score for a plurality of infrared images captured sequentially in which the lens remains dry. Only a single symmetry score falls below 300 and none are recorded at 100. If a symmetry score of a given received infrared image is the same or more than a threshold symmetry score, the given received image is considered in step S10.3 (Figure 23). If the symmetry score of the given received infrared image is less than the threshold symmetry score, the given received infrared image is discarded. The threshold symmetry score may be 300. In some examples, the filtered image may only be divided along the vertical as part of the calculation of the symmetry score. In other examples, the filtered image may only be divided along the horizontal. Alternatively or additionally, the filtered image may be divided via the centre point in some other way (e.g., diagonally). Training As hereinbefore explained, the first neural network is a convolutional neural network which has been trained to recognise the presence of a dangerous fire within a received image. The untrained first neural network is fed with a plurality of infrared images (herein "training images"), each training image either capturing a dangerous fire or not capturing a dangerous fire. The training images may include images which capture the same background / environment, but at different times and / or in different conditions (e.g., at different ambient temperatures). The training images may include images which capture different backgrounds / environments. Training images which capture a dangerous fire are labelled "positive" (100% fire probability) and training images which do not capture a dangerous fire are labelled "negative" (0% fire probability). Both negative and positive images may capture other hot objects, which are not dangerous fires, such as an oven, a hob, or an electric heater. The labelling of the training images will be described in more detail hereinafter. A training image is considered to have captured a dangerous fire when the fire has become visible to the sensor 13?, as will be hereinafter explained. The training images are taken from one or more sets of images. A set of images ("image set") consists of a plurality of infrared images captured sequentially over a period of time. Some, or all, of the image sets consist of a plurality of infrared images captured sequentially over a period in which there is initially no dangerous fire, a dangerous fire ignites and grows, and is then supressed. Such image sets are referred to as "active image sets". Each active image set includes negative images (captured during preignition of the fire) and positive images (captured during ignition and post-ignition of the fire). Images which are captured once the dangerous fire is being supressed ( / .e., when the mist from the spray nozzle 10 is extinguishing the fire) are not used as training images. All images in the active image set correspond to the same azimuthal position - in other words, they capture the same region of the room in which the dangerous fire occurs. Some of the image sets may consist of a plurality of infrared images captured sequentially over a period in which a dangerous fire never ignites. Such image sets are referred to as "passive image sets". Passive image sets only consist of negative images. The infrared images in a given passive image set correspond to the same azimuthal position. Both active and passive image sets may include images which capture other hot objects. These other hot objects may increase and / or decrease in temperature over the period of time of the set. Before the first neural network is trained, the training images are labelled either "positive" or "negative", as hereinbefore mentioned. Additionally, each training image is assigned a "sample weight". The sample weight is a value from 0 to 1. The labelling of each training image, and the value of the assigned sample weight, depends on the timestamp of the training image. The timestamp of the training image is the time at which the Image is captured as part of the image set. For active image sets, one of the timestamps is assigned the "earliest activation time". This is the timestamp for which its corresponding training image is considered to have clearly captured the dangerous fire. The training image is considered to have captured the dangerous fire clearly when the dangerous fire can be distinguished from the ambient temperature background (and so is "visible" to the sensor 13?). In cases where the dangerous fire is suddenly large and grows rapidly (for example, an explosion), the "earliest activation time" will be the time at which the fire ignites. In other cases, the "earliest activation time" will be after the fire ignites. For example, in cases where the dangerous fire begins small and is far away from the sensor, a training image taken at or soon after ignition captures the dangerous fire in only a small fraction of one pixel; this makes it impossible to adequately distinguish the dangerous fire signature in the Image, and therefore the dangerous fire must develop further before it becomes "visible" to the sensor 132. Some infrared images of a given active image set have a timestamp preceding the earliest activation time (in other words, earlier-captured images); therefore, these infrared images are labelled "negative". Some infrared images of the given active image set have a timestamp at or following the earliest activation time (in other words, later-captured images), which are therefore labelled "positive". In this way, the timestamp of the training image dictates whether it is a positive or negative image. For active image sets, the time at which the spray nozzle 10 begins to deploy the fire supressing material 5 to extinguish the fire is herein referred to as "suppression". Infrared images having timestamps at or following suppression are not labelled and are discarded for training purposes. Although these images capture the dangerous fire, each image is obscured by the mist from the spray nozzle 10, which make them unsuitable for training. For passive image sets, there is no earliest activation time as a dangerous fire never ignites. Therefore, all infrared images of a given passive image set are labelled "negative". For active image sets, the sample weight value associated with each timestamp depends on that timestamp's proximity to the earliest activation time, the fire point (hereinafter described), and / or the latest activation time (also hereinafter described). In other words, the sample weight value for a given timestamp is set according to the stage of development of the dangerous fire which that timestamp corresponds to. This will now be explained with reference to Figure 27. Figure 27 shows a plurality of data plots corresponding to a period from time 16:21:30 to time 16:25:30. This is the period over which an example active image set is captured. The plurality of data plots includes a plot of sample weight varying over this period (labelled plot "c"; "sample_weight"). Figure 27 also Indicates the earliest activation time, occurring at time 16:22:55, and suppression, occurring at time 16:24:50. As shown in plot c, the timestamps preceding the earliest activation time are assigned the maximum sample weight value of 1. As hereinbefore explained, images captured before the earliest activation time are negative. Thus, negative images have a maximum sample weight value of 1. It is advantageous for the second system 12 to identify the dangerous fire soon after ignition. Most preferably, the second system I2 should identify the fire at the time "fire point", indicated in Figure 27. However, it is still acceptable for the second system I2 to identify the fire slightly before or after the fire point, namely during a period referred to as the "buffer zone". This buffer zone is also indicated in Figure 27. As will be hereinafter explained, the first neural network is trained so it is more likely to activate within the buffer zone. In the example shown in Figure 27, the buffer zone is from time 16:22:50 to time 16:23:35. Timestamps falling within this period are assigned sample weight values less than 1. The latest time in the buffer zone is the "latest activation time", indicated in Figure 27. It is preferred for the second system I2 to activate before the fire has significantly developed. Therefore, the timestamps immediately following the latest activation time are assigned the maximum weight sample value of 1. As hereinbefore explained, the images captured from the earliest activation time to immediately before suppression are positive images. The positive and negative images are indicated in Figure 27 using dashed arrows, along with the images not used for the training ("discarded images"). The timestamp at the earliest activation time is assigned a minimum sample weight value of 0.1. The timestamps following the earliest activation time are assigned sample weights of gradually increasing value from 0.1, such that the earlier timestamps following the earliest activation time have a lower sample weight than the later timestamps. From the earliest activation time to the fire point, the sample weights increase according to a function based on a natural logarithm. From the fire point to the latest activation time, the sample weights increase according to a different function based on the same natural logarithm. The functions are determined according to the values of sample weight set for the timestamps of the earliest activation time, the fire point, and the latest activation time. In the example shown in Figure 27, the timestamp at the fire point is assigned a sample weight value of 0.5. In other examples, sample weight at the fire point may take a value other than 0.5. Furthermore, in other examples, the minimum sample weight may take a different value, for example 0. In other examples, the maximum sample weight may take a different value, for example 0.9. As shown in the example in Figure 27, all negative images are assigned the maximum sample weight, whereas different positive images are assigned different sample weights. The sample weights are set according to the timestamps, as hereinbefore explained. For passive image sets, all images are assigned the maximum sample weight as none of the images capture a dangerous fire. The training of the first neural network will now be described in greater detail. The training of the first neural network is agnostic to the sequence in which the training images are fed into the network. In other words, the training images do not need to be fed according to image set and / or in chronological order. Furthermore, there is no requirement to feed a complete image set into the first neural network. The first neural network may be fed a random shuffle of positive and negative images for training. The training images may be processed according to step S6.2 (Figure 12) prior to being fed into the untrained first neural network. This is done when it is intended that the second operation method is performed using received images which have also been processed according to step S6.2. As the training images are input into the untrained first neural network, the neural network weights are adjusted according to the accuracy of the output of the first neural network. In other words, the neural network weights are adjusted according to whether the outputted percentage probability matches the labelling of the training image causing the positive fire condition (herein "activating training image"). For example, if a negative image fed into the first neural network results in a positive fire condition, the neural network weights are penalised / adjusted. If a positive image results in a positive fire condition, the neural network weights are not adjusted / minimally adjusted. The neural network weights are also adjusted according to the sample weight assigned to each training image hereinbefore described. The greater the sample weight of the activating training image, the greater the extent of penalty / adjustment applied to the neural network weights. In an example in which the activating training image has the maximum sample weight, the neural network weights receive the maximum penalty / adjustment. This adjustment reduces the likelihood of the first neural network outputting a positive fire condition based on an infrared image which does not capture a dangerous fire. In an example in which the activating training image has a minimum or less than maximum sample weight, the neural network weights receive no penalty / adjustment or a reduced penalty / adjustment. This means that the first neural network is more likely to output a positive fire condition based on an infrared image which captures a dangerous fire in the buffer region - in other words, at an earlier stage of growth. As previously explained, the sample weights in the buffer zone reduce with increased proximity to the earliest activation time. Therefore, adjusting the neural network weights according to these sample weights increases the likelihood of the first neural network outputting a positive fire condition based on an infrared image which captures a dangerous fire at or around the fire point, rather than later in the buffer region. If the activating training image is at the fire point, only a moderate penalty / adjustment is applied to the neural network weights, defined by the sample weight of 0.5. As hereinbefore explained, the sample weights are set according to the stage of development of the dangerous fire. Thus, use of the sample weights in the training according to the present application gives rise to more precise adjustment of the neural network weights. In turn, the trained neural network is more likely to activate at the early stages of a dangerous fire ( / '.e., in the buffer zone), thereby reducing damage and threat to life. The first neural network is taken to be adequately trained once the network consistently outputs a fire probability corresponding to a positive fire condition at the fire point. Figure T7 further shows a plot of the 100th temperate percentile ( / '.e., maximum temperature reading within the image) for each infrared image captured over the period from time 16:21:30 to time 16:25:30, labelled plot "a". Plot "a" only begins to peak near the suppression point. Figure 27 further shows a plot of the 98th temperature percentile for each infrared image, labelled plot "b". As with plot "a", plot "b" only begins to peak near the suppression point. Thus, using the 100th temperature percentile / 98th temperature percentile to determine the presence of a dangerous fire results in a slower activation time compared to the trained first neural network. Plot "d" in Figure 27 indicates the annotated location of the fire by way of column index (approximately column index 14). During the set period, the location of the fire captured by the sequence of training images remains constant. The training of the second neural network will now be described. As hereinbefore explained, the second neural network is configured to output the probability of the presence of a dangerous fire based on an intermediate output of the first neural network and one or more extracted features of received infrared images. A set of intermediate outputs and extracted features are used to train the second neural network. The trained first neural network is fed with the training images hereinbefore described and outputs an intermediate output for each training image. In a preferred example, the intermediate output is the output of the ”global_average_pooling2d" layer (Figure 14). In a preferred example, for a given training image, the follow extracted features are determined: 1. 98th temperature percentile of the training image; 2. 100th temperature percentile of the training image; 3. Slope of 98th percentile best fit line for 18 infrared images immediately preceding the given training image; 4. Result of subtraction of 98th percentile of 1st infrared image from 98th percentile of 18th (last) infrared image 5. Slope of 100th percentile best fit line for 18 infrared images immediately preceding the given training image; 6. Result of subtraction of 100th percentile of 1st infrared image from 100th percentile of 18th (last) Infrared image. The best fit line for extracted feature (3) is the best fit line of the 98th percentile data plot for the preceding 18 infrared images, wherein the 98th percentiles are plotted against the timestamps of their infrared images. The slope or gradient of the best fit line is determined using linear regression. Extracted feature (5) is calculated in the same way as extracted feature (3). Unlike the training of the first neural network, training the second neural network involves feeding the first neural network training images in sequence. In other words, the first neural network is not fed a random shuffle of positive and negative images for training the second neural network. This is because the extracted features require a sufficient history of images for calculation. Preferably, extracted features (3) to (6) use a history of 18 images, but may use a history of 6 to 18 images. In other examples, the history of images is defined by a pre-set period of time, rather than a number of images. Once generated, the intermediate outputs and extracted features are fed into the untrained second neural network. A given intermediate output and its corresponding extracted features are input consecutively before other intermediate outputs and their extracted features are input. The untrained second neural network can be fed a random shuffle of sets of intermediate outputs and their corresponding extracted features; in other words, the sets of intermediate outputs and their corresponding extracted features do not have to be fed into the network in the same order as the training images the sets correspond to. As the intermediate outputs and extracted features are input into the untrained second neural network, its neural network weights are adjusted according to the accuracy of the output of the second neural network (in the same way as the first neural network). The neural network weights of the second neural network are also adjusted according to the sample weights hereinbefore described. Although training images are not directly input into the second neural network, a given intermediate output and the associated extracted features are Input into the second neural network; these Inputs correspond to a given training Image. The given training image has an assigned sample weight, as hereinbefore described, based on which the neural network weights are adjusted. The training of the third neural network will now be described. As hereinbefore explained, the third neural network is configured to output a probability of a fire for each column of the received infrared image based on an intermediate output of the first neural network. In a preferred example, the intermediate output is the output of the "global_average_pooling2d" layer (Figure 14). The probabilities corresponding to the columns are then used to determine the location of the dangerous fire. The trained first neural network is fed with training images and outputs an intermediate output for each training Image. These intermediate outputs are then used to train the third neural network. The training images used to generate the intermediate outputs capture a dangerous fire (positive images) and other non-dangerous hot objects. Each training image is labelled with the location of the dangerous fire, expressed as a column index. Training of the third neural network does not require the first neural network to be fed training Images in sequence; the first neural network can be fed a random shuffle of positive images. During training, the neural network weights of the third neural network are adjusted according to the output of the network. A given training image is input into the first neural network to produce an intermediate output. This intermediate output is then fed into the third neural network, which outputs a probability corresponding to each column of the given training image. The location of the dangerous fire captured in the given training image is determined using the probabilities corresponding to the columns, as hereinbefore explained in reference to Figure 19. The neural network weights are adjusted according to whether the outputted probability accurately indicates the fire location. For example, if the fire location determined based on the output of the third neural network (herein "NN3 fire location") does not match with the labelled location of the dangerous fire in the given training Image, a penalty is applied to the neural network weights. In some examples of the third neural network training, sample weights are additionally used. This may be done when the fire is spread over more than one column. In such examples, sample weights are assigned to each column of each training image. The columns which contain the dangerous fire are assigned the minimum or a lower sample weight value and the remaining columns are assigned the maximum or a higher sample weight value. If the NN3 fire location does not match with the labelled location of the dangerous fire, a penalty is applied to the neural network weights according to the sample weight of the column of the NN3 fire location. If the NN3 fire location does match with the labelled location of the dangerous fire, either no penalty is applied to the neural network weights or a minimum / reduced penalty is applied according to the sample weight of the column of the NN3 fire location. Referring now to Figure 28a, a plurality of data plots corresponding to a period from time 16:22:00 to time 16:25:30 is shown. This is the period over which an example active image set is captured. The plurality of data plots includes a plot of the probability percentage output by the trained first neural network ("smoothed 6 p for a sequence of infrared Images. As shown, the probability percentage output by the first neural network is 0 before the buffer zone, increases significantly at the fire point, and reaches 1 within the buffer zone. Thus, the first neural network trained according to the present application is configured to output a probability corresponding to a positive fire condition soon after ignition of the dangerous fire and before it significantly develops - in other words, within the buffer zone. Figure 28a further shows plot "f" ("Location Combined Predictions"), which is a plot of the fire location, as calculated in step S9.6 (Figure 19), for each infrared image. As indicated by this plot, the location of the fire remains constant during and after the buffer zone. Figure 28a further shows the plot of sample weights (plot c) hereinbefore described. Figure 28b is similar to Figure 28a. However, instead of plot e, Figure 28b shows a plot of the probability percentage output by the trained second neural network ("smoothed_6_prob_nn_final") for the sequence of infrared Images. As shown, the probability percentage output by the second neural network is 0 before the buffer zone, increases significantly at the fire point, and reaches 1 within the buffer zone. Thus, the second neural network trained according to the present application is configured to output a probability corresponding to a positive fire condition within the buffer zone. The plot "smoothed_6_prob_nn_final" is labelled plot "g" in Figure 28b. Referring to Figures 28c and 28d, the trained neural networks were fed different active image sets. Both figures show the trained first and second neural networks outputting a probability corresponding to a positive fire condition within the buffer zone. The figures show that the trained first and second neural networks activate promptly after ignition even when fed different sequences of infrared images. Figures 28c and 28d further show the plots of 100th temperature percentile (plot a) and 98th temperature percentile (plot b) for each image in the respective sequences of infrared images. Both Figures 28c and 28d show plots a and b beginning to peak soon before suppression, rather than within the buffer zone. Modifications It will be appreciated that various modifications may be made to the embodiments hereinbefore described. Such modifications may involve equivalent and other features which are already known. Features of one embodiment may be replaced or supplemented by features of another embodiment. Although claims have been formulated in this application to particular combinations of features, it should be understood that the scope of the disclosure of the present invention also includes any novel features or any novel combination of features disclosed herein either explicitly or implicitly or any generalization thereof, whether or not it relates to the same invention as presently claimed in any claim and whether or not it mitigates any or all of the same technical problems as does the present invention. The applicant hereby gives notice that new claims may be formulated to such features and / or combinations of such features during the prosecution of the present application or of any further application derived therefrom.

Claims

1. A method of detecting a fire, the method comprising:receiving at least one infrared image, the at least one infrared image comprising at least one pixel corresponding to a respective temperature;determining a probability of the presence of a fire based on the at least one infrared image; andsignalling a fire condition based on the determination.

2. The method of claim 1, the method comprising:wherein the fire condition is a positive fire condition, activating a spray head to deploy fire-suppressant material.

3. The method of claims 1 or 2, wherein:the at least one infrared image includes a plurality of pixels each corresponding to a respective temperature; andthe probability of the presence of the fire is determined using at least one artificial neural network.

4. The method of claim 3, wherein:the at least one infrared image is a stream of infrared images including an nth infrared Image and n is a non-zero positive integer; anddetermining the probability comprises:wherein n <a threshold value:inputting the nth Infrared image into a first artificial neural network, the first artificial neural network comprising an output layer and a plurality of hidden layers;the output layer outputting the probability of the presence of the fire; wherein n >the threshold value:a hidden layer of the plurality of hidden layers of the first neural network outputting an intermediate output;extracting at least one feature from a plurality of infrared images preceding the nth infrared Image;inputting the at least one feature and the intermediate output into a second artificial neural network, the second artificial neural network comprising an output layer and a plurality of hidden layers;an output layer of the second artificial neural network outputting the probability of the presence of the fire.

5. The method of claim 4, wherein n >6.

6. The method of any one of claims 4 to 5, the method further comprising:determining a number of hot objects, m, captured by the nth infrared image, wherein m is a non-zero positive integer; anddetermining a location of the fire, the determination comprising:wherein m >1:the hidden layer of the plurality of hidden layers of the first artificial neural network outputting the intermediate output;inputting the intermediate output into a third artificial neural network; determining the location of the fire based on outputs of the thirdartificial neural network;wherein m = 1:determining the location of the fire based on a location of a peak in temperature in the nth infrared image.

7. The method of any one of claims 4 to 6, wherein the at least one infrared image is processed prior to being input into the first artificial neural network.

8. The method of any one of claims 3 to 6, wherein a spray head is deploying fire suppressant material at an original angular position corresponding to an original fire location in response to a positive fire condition, the method comprising:capturing a stream of infrared images at the original angular position; determining whether one of the images in the stream satisfies a condition;if the condition is satisfied, deactivating the spray head;capturing at least one first Infrared image at a first adjusted angular position; determining whether at least one hot object is present in the at least one first infrared image;if present, determining an adjusted location of the fire based on the hot object(s).

9. The method of claim 8, the method further comprising:after deactivating the spray head, capturing at least one second infrared Image at a second adjusted angular position; anddetermining whether at least one hot object is present in the at least one second infrared image.

10. The method of claims 8 or 9, wherein for a given infrared image in the stream of infrared images captured at the original angular position:performing at least one division of the given infrared image which intersects the central point of the given infrared image, wherein for each division:dividing the plurality of pixels into a first group and a second group;subtracting a pixel of the first group from a mirroring pixel of the second group;determining a number of pixels that differ in temperature value basedon the subtraction;calculating a symmetry score based on the determination;wherein the symmetry score >a threshold symmetry score:consider the given infrared image in the determination of whether theimage satisfies the condition;wherein the symmetry score <a threshold symmetry score: discard the given infrared image.

11. The method of any one of claims 3 to 10, wherein:a negative fire condition is signalled in response to receiving a given infrared image; andthe probability of the presence of the fire for the given infrared image is greater than a first pre-set probability;the method comprising:if the probability for the given infrared image is less than a second pre-set probability, capturing a first stream of infrared images at the angular position corresponding to the given infrared image;if the probability for the given infrared image meets or exceeds the second preset probability and is less than a third-pre-set probability, capturing a second stream of infrared images at the angular position corresponding to the given infrared image.

12. The method of claim 11, wherein the first stream of infrared images is captured until one of the following conditions is satisfied:i) a pre-set period elapses; orii) the probability of the presence of the fire for an infrared image in the first stream meets or exceeds the second pre-set probability.

13. The method of claims 11 or 12, wherein the second stream of infrared images is captured until the probability of the presence of the fire for an infrared image in the second stream:i) exceeds the third pre-set probability; or II) falls bellow the second pre-set probability.

14. The method of any one of claims 3 to 13, wherein the at least one infrared image includes a given infrared image capturing a hot object, the method comprising: determining the location of the hot object in the given infrared image; and in response to a determination that the hot object is not at the centre or substantially at the centre of the given infrared image, adjusting an angular position of an infrared sensor which captured the given image such that the hot object is at the centre of the infrared sensor's field of view.

15. The method of claims 1 or 2, wherein the at least one infrared image consists of a single pixel corresponding to a respective temperature, the method comprising: in a first scan:capturing scan data defining a baseline, wherein the scan data consists of a plurality of infrared images captured at different azimuthal angles of uniform increment;performing a first fire detection method based on the baseline; and determining the probability of the presence of the fire based on the first fire detection method;in a subsequent scan:capturing subsequent scan data;performing the first fire detection method and / or one or more of a second, third, and fourth fire detection method based on the subsequent scan data;determining the probability of the presence of the fire for each performed fire detection method.

16. The method of claim 15, wherein the first, second, third, and fourth fire detection methods are for detecting fires of different growth rates.

17. The method of claim 16, wherein the first fire detection method is for detecting ultra-fast developing fires and the second, third, and fourth fire detection methods are for detecting slower-growing fires.

18. The method of any preceding claim, the method comprising:Initiating the method of any preceding claim in response to at least one triggering event.

19. A method of training an artificial neural network, the method comprising: feeding a plurality of training infrared images into an untrained artificial neural network, wherein:each training infrared image is assigned a sample weight which is set according to the timestamp of the training infrared image; andthe untrained artificial neural network has a set of neural network weights;for each training infrared image, adjusting the set of neural network weights according to its sample weight.

20. The method of claim 19, wherein for a set of training infrared images captured over a period in which a fire ignites:the sample weight of each training infrared image is set according to the stage of development of the fire which its timestamp corresponds to.

21. The method of claim 20, wherein:the training infrared images having a timestamp preceding an earliest activation time are assigned a maximum sample weight;the training infrared image having a timestamp at the earliest activation time is assigned a minimum sample weight; andthe training infrared images having a timestamp Immediately following the earliest activation time are assigned sample weights between the minimum and maximum sample weight;wherein the earliest activation time is the timestamp for which its corresponding training infrared Image is the first infrared image in the set to clearly capture the fire.

22. A computer program which, when executed by at least one or more processors, causes the processor(s) to perform the method of any one of claims 1 to 21.

23. A computer readable medium, optionally a non-transitory computer readable medium, which stores or carries the computer program of claim 22.

24. A hardware processor configured to perform the method of any one of claims 1 to 21.

25. A system comprising:a controller configured to perform the method of any one of claims 1 to 21;at least one spray head unit in communication with the controller, the at least one spray head unit comprising a spray nozzle configured to deploy fire suppressant material; and5 at least one infrared sensor.

Images