Pipe condition assessment

Abstract

Abstract Systems, methods, devices and assemblies which enable classification and / or localisation of a condition within a conduit. A soundwave is emitted from a device into the conduit and reflected when it meets a condition. The reflected wave may be received by a receiver. The received signal is processed and analysed by a model to predict the classification of a condition within the conduit and / or the distance of a condition from the end of the conduit. a condition within a conduit. A soundwave is emitted from a device into the conduit and condition within the conduit and / or the distance of a condition from the end of the conduit. 20 26 20 15 58 27 F eb 2 02 6 2 0 2 6 2 0 1 5 5 8 2 7 F e b 2 0 2 6 a c o n d i t i o n w i t h i n a c o n d u i t . A s o u n d w a v e i s e m i t t e d f r o m a d e v i c e i n t o t h e c o n d u i t a n d 2 0 2 6 2 0 1 5 5 8 2 7 F e b 2 0 2 6 a c o n d i t i o n w i t h i n a c o n d u i t . A s o u n d w a v e i s e m i t t e d f r o m a d e v i c e i n t o t h e c o n d u i t a n d 1 / 12 310 140 120 100 110 200 240 356 230 234 242 232 210 220 130 Figure 1 501 Raw Acoustic Signal 502 506 505 507 508 Data Preprocessing 510 CWT CNN Transformer Classification and Time-Freq Localization Peak Scalogram Detection Preprocessed Signal 503 504 Figure 2 20 26 20 15 58 27 F eb 2 02 6 1 / 1 2 2 0 2 6 2 0 1 5 5 8 2 7 F e b 2 0 2 6 3 1 0 1 4 0 1 2 0 1 0 0 1 1 0 2 0 0 2 4 0 3 5 6 2 3 0 2 3 4 2 4 2 2 3 2 2 1 0 2 2 0 1 3 0 5 0 1 R a w 5 0 2 5 0 6 5 0 5 5 0 7 5 0 8 D a t a 5 1 0 C W T C N N T r a n s f o r m e r a n d T i m e - F r e q P e a k P r e p r o c e s s e d 5 0 3 5 0 4 1 / 1 2 2 0 2 6 2 0 1 5 5 8 2 7 F e b 2 0 2 6 2 4 0 2 3 4 2 4 2 2 3 2 2 1 0 2 2 0 1 3 0 5 0 1 R a w 5 0 2 5 0 6 5 0 5 5 0 7 5 0 8 D a t a 5 1 0 C W T C N N T r a n s f o r m e r a n d T i m e - F r e q P e a k D e t e c t i o n P r e p r o c e s s e d 5 0 3 5 0 4

Description

Technical Field

[0001] The invention relates generally to devices, systems and methods for assessing the condition of a pipe, which may include classifying and / or localising the condition of the pipe. Specific embodiments relate to devices, systems and methods that utilise acoustic reflectometry in determining the condition of broadband pipes. Background

[0002] The deployment of fiber-optic networks is critical for supporting the increasing demand for high-speed internet. Governments and telecommunication providers worldwide are prioritising the transition from outdated copper networks to high speed fibre-optic broadband to meet growing connectivity demands. As this transition is made, technicians or operatives will attempt to deploy fibre-optics into the existing infrastructure via a rodding process. The technicians often run into difficulties during this process. There is little information provided to the technician about the conditions within a conduit during the rodding process. This may lead to wasted efforts or ordering of remediation when technicians run into problems such as when conduits appear to be blocked.

[0003] Underground pipes have traditionally been used to house copper wiring and are also being used to house modern fibre-optic cabling. The pipes are often decades old and highly susceptible to blockages and degradation. The pipes are often formed form a polyvinyl chloride (PVC) material, or may be from other materials in the case of older conduits, such as galvanised iron, earthenware / ceramic or asbestos. The pipes which run from a user premises to a pit in the street may have a small diameter which may be in the region of 10-35mm. Such conduits may typically run under / through a front garden of the user.

[0004] Currently, field technicians rely on traditional fiberglass rods and mandrels to assess pipe viability and install fibre-optic cables via the rodding process, a method that has significant limitations. This process involves a technician manually pushing the rods / mandrels into the pipe in question until it meets resistance. For example, a technician may be required to perform the rodding process repeatedly when resistance is met. Manual rodding is labourintensive, offers no digital integration, and cannot reliably characterise blockages or 2026201558   27 Feb 2026 determine the remaining pipe capacity. Having to perform manual rodding repeatedly places stress on the bodies of the technicians performing the process and can lead to injury.

[0005] Even when rodding has been performed, excavation is often necessary to investigate potential blockages. The excavation may be manually performed by the technician using hand tools and / or may need to be performed on a different day using power tools or equipment. This can result in increased cost and delays to the installation process. The same process of rodding and / or excavation may also be used to assess the positioning and direction of the conduits, including locating bends. The cause of resistance in the rodding process may be unclear which can also lead to excavation. Some common causes of resistance are a blockage, a bend in the conduit, the end of the rod catching on a coupling between pipe sections, or the rod becoming entangled with existing cabling.

[0006] Ground-penetrating radar (GPR) has been used in other industries to locate underground infrastructure. However, this technique remains inaccurate. The signals produced via GPR are noisy and make it challenging or impossible to assess non-metallic pipes and to distinguish, classify or accurately locate blockages or other conditions therein. GPR equipment is also expensive and bulky in size meaning it is difficult to transport to a job and carry, and unsuitable for everyday use by technicians.

[0007] These inefficiencies highlight the need for an alternative and / or more advanced or accurate approach to underground pipe condition assessment. There is currently no comprehensive non-invasive method for fast and accurate condition assessment of existing pipes to be used for fibre-optic cable deployment. It may be beneficial to provide a way to perform an assessment of the condition of the conduit, which may improve efficiency of the fibre-optic deployment process, reduce wasted efforts and / or may enable a technician to select the most appropriate approach for a deployment job. For example, it may be desirable to provide a system, method, device and / or apparatus which improves on conventional methods of pipe condition assessment or provides a useful alternative. Summary

[0008] The present disclosure provides devices, apparatus, methods and systems that may be used for assessing the state of a conduit and determining classification and location of conditions within the conduit. The conditions within a conduit may be classified and / or localised in near real-time. The present disclosures are particularly useful in determining 2026201558   27 Feb 2026 condition classifications and / or locations in conduits of small diameter, for example around 20-100mm in diameter, such as fibre optic, communication or broadband conduits. Systems according to the disclosure may be non-invasive compared to existing methods, such as rodding or excavation. Devices of the present disclosure may be handheld and easily portable. The device herein may require access to one open end of a conduit.

[0009] Devices according to the present disclosure comprise an emitter for emitting soundwaves, a receiver for receiving reflected soundwaves, and an interface configured to engage with an open end of a conduit.

[0010] According to one aspect, there is provided a device for use in assessing conditions in a conduit, the device comprising: an emitter configured to emit soundwaves; a receiver configured to receive soundwaves; and an interface comprising an inner volume acting as a wave guide for soundwaves from the emitter and soundwaves to the receiver, the interface configured to engage an open end of the conduit such that the inner volume is in communication with an interior of the conduit, wherein the device is configured such that, in use, the emitter emits a soundwave through the inner volume of the interface into the conduit, and the receiver receives a reflected soundwave from the conduit through the inner volume.

[0011] According to embodiments, the device comprises a body. The body comprises an interior volume. The emitter and / or the receiver are housed within the interior volume of the body. The emitter and / or receiver may be removably inserted into the body. The emitter and / or receiver may be removably contained within the interior volume of the body. The body may comprise a wall defining the interior volume. The body may comprise an area for housing the emitter. The device may comprise a backplate configured to be removably affixed to the body. The backplate may be configured to hold the emitter in place within the interior volume of the body.

[0012] According to embodiments, the interface is formed integrally with the body. The body may comprise a tapered section. A wall of the body may, in the tapered section, taper inwardly from an area adjacent to the emitter towards an end at which the interface is to engage with a conduit.

[0013] According to embodiments, the interface is in the form of an adapter. The adapter may be removably attachable to the body of the device. 2026201558   27 Feb 2026

[0014] According to embodiments, the device comprises a plurality of interfaces. Each interface may be configured to engage with an open end of a conduit of a different respective size than the other interfaces. The body may be configured to removably attach to one interface at a time. In use, a suitable adapter may be selected to correspond to a diameter of the conduit. The adapter may be attached to the body before engaging with the open end of the conduit. Additionally or alternatively, the adapter may engage with an open end of the conduit before the adapter is attached to the body.

[0015] According to embodiments, the interface may be removably attachable to the body. The body may comprise a neck. The interface may comprise a first opening at a proximal end, the proximal end being the end configured for attachment to the body. The neck of the body may connect to the interface at the first opening. The interface may comprise first engagement surfaces and the body comprises second engagement surfaces. The first engagement surfaces may be located adjacent to the first opening on the interface. The second engagement surfaces may be located on or adjacent to the neck of the body. The first engagement surfaces and second engagement surfaces may be configured to interact with one another. The interface and the body may be releasably lockable together via a relative rotation of the interface and the body when brought into contact with one another resulting in interactions between the first engagement surfaces and the second engagement surfaces. The interface and the body may each comprise respective alignment indicators. The alignment indicators may be aligned with one another prior to releasable attachment between the interface and the body.

[0016] According to embodiments, the or each interface comprises an extension. The extension may be an elongate section of the interface. The extension may be substantially cylindrical. The extension may have an opening at a distal end. The extension of the interface may comprise an outer wall defining a hollow interior. The distal end of the extension may be configured for insertion into a conduit. The distal end of the extension may have a second opening of the interface. The extension may be inserted into a conduit of internal diameter that is substantially the same as the external diameter of the extension. Where the device comprises a plurality of adapters, the extension of each adapter may have a different respective diameter. A respective interface may have an extension with a diameter of about 20mm, about 35mm, about 50mm, or about 100mm. 2026201558   27 Feb 2026

[0017] According to embodiments, the extension has an external diameter sized for insertion into a conduit of internal diameter of 20-100 mm. The extension may be of any suitable length. According to examples the extension may be about 1 to 30cm, about 2 to 20cm, about 5 to 15cm, about 5cm, or about 10cm in length.

[0018] The interface may be formed from any suitable material. For example, the interface may be formed from a plastic material, such as a polypropylene, polyethylene, or high density polyethylene, a metal, a wood, for example. The interface may be substantially rigid. The interface may be partially flexible.

[0019] According to embodiments, the device is configured to be held in position adjacent to the open end of the conduit. The device may be held in place via a friction engagement between an exterior surface of the extension and an internal surface of the conduit when the extension is inserted into the conduit.

[0020] According to embodiments, the interface may be funnel shaped. The interface may comprise a portion which is funnel shaped. The interface may comprise a funnel section. The funnel section may be shaped substantially like a truncated cone. The funnel section may be attached to the extension. The extension of the interface may extend from a distal end of the funnel section. A proximal end of the funnel section may be adjacent to the first opening of the interface, The proximal end of the funnel section may be configured for removable attachment to the neck of the body. The funnel section may taper inwardly from a first internal diameter at a proximal end to a second diameter at the distal end. The funnel section may be substantially circular in cross sectional shape.

[0021] According to embodiments, the interface comprises a cut out region. A cut out region may be a gap or space in an outer wall of the interface. The cut out region may be located at least in the extension of the interface. The cut out region may extend longitudinally along the interface from the distal end of the extension. The cut out region may extend along the entire length of the extension. The cut out region may extend partially from the extension into the funnel section. The cut out region may be configured such that the extension may be inserted into a conduit which contains at least one cable, wire or fibre. The at least one cable, wire or fibre in the conduit may be positioned within the cut out region when the extension is inserted into the conduit. 2026201558   27 Feb 2026

[0022] According to embodiments, the device comprises a handle. The device may be handheld. The device may be portable and configured to be handheld prior to and after the interface engages a conduit. The device may comprise at least one hollow portion configured to convey one or more wires to and from the device. The handle may comprise the hollow portion.

[0023] According to embodiments, the emitter is a speaker. The receiver may be a microphone. The emitter and / or receiver may be configured to communicate with a master device respectively via a wired or wireless connection. The master device is located remotely to one or both of the emitter and the receiver. The master device may comprise a display screen. The master device may comprise a processor. The master device may comprise a memory or storage. The master device may comprise communication means, for example means for communication by cellular network, Wi-Fi, wired communication connection, Bluetooth, near field communication (NFC), or any suitable communication means. The master device may control emission of soundwaves from the emitter. The master device may receive signals from the receiver representative of received soundwaves, for example the master device may receive a raw acoustic signal from the receiver.

[0024] According to embodiments, the receiver and emitter are part of a master device. The interface may be a cradle. The interface may be attachable to a cradle. The cradle may be configured to hold a master device.

[0025] Another aspect of the present disclosure provides a system for assessing conditions in a conduit, wherein the system comprises: a device as provided in any aspect, embodiment or example herein, the device comprising, at least, an emitter, a receiver and an interface; a trained machine learning model that receives data, the data comprising a raw signal from the receiver and / or the data comprising a preprocessed signal, processed signal, image or scalogram derived from the raw signal, wherein the trained machine learning model makes a classification prediction based on the data to classify at least one condition within the conduit.

[0026] According to embodiments, the system comprises a localisation module. The localisation module may be configured to estimate a distance of the at least one condition from the end of the conduit at which the interface is engaged. The localisation module may utilise time-of-flight information of a soundwave emitted and received by the device. The localisation module may estimate the distance using at least the equation: d = (v x n) / (2 x 2026201558   27 Feb 2026 Fsample), where d is a distance of the condition from the open end of the conduit, v is the speed of the wave, n is a sample index, and Fsample is the sampling rate of the receiver.

[0027] According to embodiments, the system comprises a processor. The processor may be configured to process the raw signal received by the receiver into a preprocessed signal. The processor may perform peak detection on a signal to locate one or more peaks. The preprocessed signal may be analysed to determine at least one window associated with a peak or peaks in the preprocessed signal. The at least one window may correspond to the at least one condition within the conduit. The window may be a section of the signal associated with a condition. The trained machine learning model may predict a condition for each window. The localisation module may estimate a distance of a condition for each window.

[0028] According to embodiments, the system comprises a display screen. The display screen may be configured to display the preprocessed signal. The display screen may be configured to display the at least one classified condition. The display screen may be configured to display the estimated distance(s). The display screen may be interactive and / or a touch screen.

[0029] According to embodiments, visual indicators may be used to easily distinguish between classified conditions of different types, each type of condition may include a respective visual indicator. The at least one classified condition may be colour coded to represent its respective classification. According to embodiments, different classifications are represented in different colours.

[0030] According to embodiments, the system comprises an input configured to receive user inputs. The input may be configured to receive a new classification of a condition from the user when the user disagrees with a predicted classification. The new classification of the condition may be used to further refine the trained machine learning model.

[0031] According to embodiments, the trained machine learning model comprises a convolutional neural network (CNN) and / or a Multi-Layer Perceptron (MLP) used as feature extractor(s) to extract features from the signal or preprocessed signal or an image or scalogram based on the preprocessed signal. According to embodiments, the preprocessed signal is converted into a scalogram and a CNN performs feature extraction on the scalogram. 2026201558   27 Feb 2026

[0032] According to embodiments, the trained machine learning model has been trained on known condition data to classify one or more of the following conditions within the conduit: a bend in the conduit; a shape and / or radius of the bend; an opening in a conduit; a second end of the conduit; a blockage in the conduit; a type of blockage; a size of the blockage and / or a percentage or proportion of the conduit diameter that is blocked; a damage to the conduit; a type of the damage to the conduit; a clear length of pipe free from blockage; and / or a liquid or water in the conduit.

[0033] An aspect of the present disclosure relates to a method of assessing conditions in a conduit. The method may comprise: engaging an interface with an open end of a conduit, the interface having an inner volume acting as a waveguide; emitting by an emitter a soundwave through the inner volume of the interface into the conduit; receiving by a receiver a reflected soundwave from the conduit through the inner volume of the interface; transferring a data to a trained machine learning model, the data comprising a raw signal representative of the reflected wave from the receiver and / or the data comprising a preprocessed signal, processed signal, image or scalogram derived from the raw signal, wherein the trained machine learning model makes a classification prediction based on the data to classify at least one condition within the conduit; and displaying details or a representation of the classification of the at least one condition in the conduit to a user.

[0034] According to embodiments, the method comprises emitting a plurality of soundwaves. Each soundwave may be a chirp having a respective frequency and duration. A plurality of reflected soundwaves representative of each chirp may be received by the receiver. A processor may process each reflected soundwave separately to form processed data, such as a preprocessed signal. The processed data may be combined and averaged, wherein a signal representative of the averaged processed data is displayed to the user.

[0035] According to embodiments, the frequency and duration of each chirp is selected based on at least one of a diameter of the conduit, and an estimated length of the conduit. According to embodiments, subsets of the chirps have frequency and duration selected in respect of a distance range. Each subset may have a different frequency and duration compared to each other subset. Each subset may correspond to a different distance range than each other subset.

[0036] The methods, systems or devices above or elsewhere herein may incorporate or be combined with any feature or element described herein. For example, the method or 2026201558   27 Feb 2026 system may employ any device or devices as provided herein. It is envisaged that a display device or master device may comprise a mobile device, such as a smartphone or tablet. The machine learning model may be stored on a device or may be stored remotely, for example at a server with which a user device may communicate. The following provides further aspects or embodiments which may stand alone or be combined with any of the above aspects or embodiments or any other example herein.

[0037] A further aspect of the present disclosure provides a system for classifying a condition of a conduit, the system comprising: an emitter configured to emit a wave into the conduit, wherein the wave is reflected when it meets a condition within the conduit; an interface enabling the emitter to emit the wave into an end of the conduit; a receiver configured to receive a signal representative of the reflected wave; and a processor configured to process the signal and determine the condition of the conduit.

[0038] According to embodiments, the interface comprises a waveguide. According to embodiments, the interface comprises a flexible source tube. According to embodiments, the source tube is elongate with a length of about 0.5m to 3m.

[0039] According to embodiments, the emitter and receiver are positioned adjacent to one another.

[0040] According to embodiments, the waveguide is sized and / or shaped to direct waves from the emitter to the end of the conduit, and the waveguide is sized and / or shaped to direct waves from the end of the conduit to the receiver.

[0041] According to embodiments, the wave is a soundwave. According to embodiments, the emitter is a speaker. According to embodiments, the receiver is a microphone. According to embodiments, the emitter and receiver are components of a device. According to embodiments, the system further comprises a cradle configured to hold the device. According to embodiments, the cradle comprises the interface. According to embodiments, the interface is positioned at an end of the cradle. According to embodiments, the cradle comprises an adjustor configured to adjust a size to enable the cradle to hold devices of varying size, length and / or width.

[0042] A further aspect of the present disclosure provides a device for classifying a condition of a conduit, the device comprising: an input configured to receive a user command; 2026201558   27 Feb 2026 an emitter configured to emit a wave into the conduit; a receiver configured to receive a signal representative of the reflected wave; a processor configured to process the signal as a processed signal; a display configured to display an image and / or text representative of the condition and / or representative of the signal; and a software application, wherein the software application comprises a trained model, and the software application is configured to: control the emitter to emit the wave in response to a user command; control the processor to process the signal as a processed signal; and use the trained model to, based on the processed signal, make inferences to classify the condition of the conduit.

[0043] According to embodiments, the trained model comprises a convolutional neural network (CNN) used as a feature extractor to extract features from the signal or processed signal. According to embodiments, the trained model comprises a transformer configured to receive the features extracted by the CNN to classify the condition of the conduit. According to embodiments, the CNN uses strided convolutions to refine the extracted features and flattens the extracted features into tokens, wherein the tokens are processed by the transformer.

[0044] According to embodiments, the trained model has been trained from signals received by the receiver in test conduits in a test bed, where at least one known condition is present within the test conduit.

[0045] According to embodiments, the processor performs preprocessing to format the signal for input to the trained model. According to embodiments, the preprocessing comprises at least one of: normalisation to reduce any bias from variations in amplitude of signals received; cross-correlation to align the received signal with the emitted wave; applying a bandpass filter to remove noise outside a predetermined frequency range; and segmentation based on a predefined distance of the conduit.

[0046] According to embodiments, a continuous wavelet transform (CWT) is used to generate time-frequency scalograms of the signals. According to embodiments, the scalograms are analysed by the trained model to classify the condition of the conduit.

[0047] According to embodiments, the emitted wave is a chirp signal.

[0048] According to embodiments, the device is configured to visualise the signal, the signal visualised being a raw unprocessed signal or a processed signal, the signal may be a 2026201558   27 Feb 2026 cross-correlated signal. According to embodiments, the device is configured to display a waveform graph that visualises the signal, where peaks on the graph indicate the presence of conditions within the conduit.

[0049] According to embodiments, the device is a mobile device, smart device or smartphone.

[0050] A further aspect of the present disclosure provides an assembly for use in classifying a condition of a conduit, comprising: the device as provided in any aspect, embodiment or example herein; and an interface configured to be position between the end of the conduit and the device.

[0051] According to embodiments, the interface comprises a waveguide. According to embodiments, the interface is configured to direct a wave from the emitter into the end of the pipe, and the interface is configured to direct a wave from the end of the pipe to the receiver.

[0052] According to embodiments, the assembly comprises a cradle configured to hold the device in a fixed position with respect to the end of the conduit. According to embodiments, the cradle comprises the interface.

[0053] According to embodiments of any system, device, assembly or method herein, the condition may comprise at least one of: presence of at least one bend in the conduit; location of each bend in the conduit; shape and / or radius of bend in the conduit; presence of a second end of the conduit; location of a second end of the conduit; a length of the conduit; presence of a blockage within the conduit; location of a blockage within the conduit; type of blockage within the conduit; size of a blockage within the conduit; presence of damage to the conduit; type of damage to the conduit; and / or location of damage to the conduit.

[0054] According to embodiments, the system or device may be further configured to localise the condition of the conduit.According to embodiments, localising the condition of the conduit comprises use of the equation: d=(v x n) / (2 x f_sample), where d is a distance of the condition from the end of the conduit, v is the speed of the wave, n is a sample index, and f_sample is the sampling rate of the receiver.

[0055] According to embodiments, the conduit is a narrow diameter pipe, and / or a broadband pipe. According to embodiments, the conduit has an internal diameter of about 20mm to about 30mm. 2026201558   27 Feb 2026

[0056] According to embodiments, the condition is assessed using acoustic reflectometry sensing.

[0057] A further aspect of the present disclosure provides a method for classifying a condition of a conduit, wherein the method comprises: emitting a wave into an end of the conduit, the wave reflecting from a condition within a conduit; receiving the reflected wave as a signal; processing the signal into a form analysable by a machine learning model; analysing the signal by the machine learning model to make inferences, the machine learning model being trained on signal data related to known conditions; classifying the condition of the conduit based on the inferences.

[0058] According to embodiments, processing the signal comprises preprocessing the signal, wherein preprocessing the signal comprises at least one of: normalisation to reduce any bias from variations in amplitude of signals received; cross-correlation to align the received signal with the emitted wave; applying a filter to remove noise outside a predetermined frequency range; and segmentation based on a predefined distance of the conduit.

[0059] According to embodiments, a continuous wavelet transform (CWT) is used to generate a time-frequency scalogram of the signal. According to embodiments, analysing the signal by the machine learning model comprises analysing the scalogram. According to embodiments, the machine learning model comprises a convolutional neural network (CNN).

[0060] According to embodiments, the machine learning model comprises a transformer.

[0061] According to embodiments, the method further comprises analysing the signal to localise the condition.

[0062] Another aspect of the present disclosure provides a method for training a machine learning model to classify a condition within a conduit; wherein the method may comprise: providing a plurality of signals representative of known classifications of condition within a conduit; for each signal of the plurality of signals: (i) preprocessing the signal; (ii) applying a continuous wave transform (CWT) to create a time- frequency scalogram representative of the signal; (iii) analysing the time-frequency scalogram by a convolutional neural network (CNN) to extract features; (iv) analysing the extracted features by a transformer to classify the extracted features into categories; (v) inputting the known classification of condition 2026201558   27 Feb 2026 associated with the signal; repeating the steps (i) to (v) for the plurality of signals such that the machine learning model correlates the categories of extracted features with classifications of condition within a conduit.

[0063] Further aspects, embodiments and examples of an invention according to the present disclosure will be apparent to a person skilled in the art with reference to the following description. The present invention is not limited to those features provided above. Brief Description of Drawings

[0064] Embodiments of the invention will now be described with reference to the accompanying drawings. It is to be understood that the embodiments are given by way of illustration only and the invention is not limited by this illustration. In the drawings:

[0065] Figure 1 shows a device and cradle assembly according to an embodiment of the present disclosure;

[0066] Figure 2 is a flow chart showing an embodiment of a model used for classifying and localising a condition of a conduit according to the present disclosure;

[0067] Figure 3 is an image of a test bed of the present disclosure;

[0068] Figure 4 shows four synthetic blockages utilised in training of a system according to the present disclosure;

[0069] Figure 5 shows a graph of a signal from a validation test that predicted a 25% blockage located at a distance of 2.51m;

[0070] Figure 6 shows a scalogram of the signal of Figure 5 generated by a system of the present disclosure;

[0071] Figure 7 shows a graph of a signal from a validation test that predicted a 100% blockage located at a distance of 1.55m and another 100% blockage at 3.09m;

[0072] Figure 8 shows a scalogram of the signal of Figure 7 generated by a system of the present disclosure; 2026201558   27 Feb 2026

[0073] Figure 9 shows a graph of a signal from a validation test that predicted a 25% blockage located at a distance of 1.37m, a 25% blockage at 2.10m, a 25% blockage at 3.14m, a 25% blockage at 3.56m, and a 25% blockage at 3.95m;

[0074] Figure 10 shows a scalogram of the signal of Figure 9 generated by a system of the present disclosure;

[0075] Figures 11 and 12 show a device configured to emit waves into and receive waves from a conduit;

[0076] Figure 13 shows adapters (a), (b), (c), (d) configured for use with the device of Figures 11 and 12;

[0077] Figure 14 shows the device of Figures 11 and 12 with the adapter removed from the body of the device;

[0078] Figure 15 shows a device similar to the device of Figures 11 and 12 with an adapter partially inserted into a conduit;

[0079] Figure 16 shows the same device of Figure 15 with the adapter fully inserted into the conduit;

[0080] Figures 17 shows a plot of a signal received from a conduit as displayed on a display screen;

[0081] Figure 18 shows another plot of a signal received from a conduit as displayed on a display screen with trained machine learning model inferences;

[0082] Figure 19 shows the plot of Figure 18 with a selection menu;

[0083] Figure 20 is a flow chart showing process steps associated with training a machine learning model to be used for classifying a condition in a conduit according to the present disclosure; and

[0084] Figure 21 is a flow chart showing process steps associated with using a trained machine learning model to predict a classification of a condition in a conduit and estimating a distance of the condition. 2026201558   27 Feb 2026 Detailed Description

[0085] As discussed in the background section of the present disclosure, there is a need for an improved method of assessing the condition of underground pipes and conduits, particularly for those used in the telecommunications industry to hold copper cables and / or fibre-optic cables. These conduits are generally formed from a plastic material, such as polyvinyl chloride (PVC), although some older conduits may be formed from other materials, including galvanised iron, earthenware, or asbestos. The conduits used for these purposes have a narrow diameter, generally about 10mm to 35mm, and often around 25mm. Assessing pipes of such small diameters may be difficult due to reduced accessibility and to their underground positioning. It is also desirable to provide a system and method which may also be used for assessing pipes of other diameters, such as 50mm, 100mm or above these diameters.

[0086] Initially, several sensing technologies were explored by the inventors. These included: ground-penetrating radar (GPR); acoustic reflectometry (AR); and acoustic transillumination (AT). GPR has been used in the past to locate underground assets, however, it produces noisy outputs and may not be sufficiently accurate for the present purposes. AT requires access to both ends of a pipe which may be less desirable as this can reduce its ease of use and practicality. AR in its conventional form is not suited to reduced diameter conduits such as for telecommunications cabling. A simultaneous combination of GPR, AR, and AT for telecommunication conduit inspection may be optimal, however such a combination would be prohibitively costly and / or operationally difficult to employ. The system or apparatus to assess the condition of the conduits will preferably address at least one of the issues described herein.

[0087] The present inventors overcame the issues mentioned above to develop and employ an apparatus, method and system that employs AR in a manner which may be well suited to assessment of conditions within telecommunication conduits. The research undertaken resulted in a new and inventive approach for the detection, classification, and localization of blockages or other features or defects in telecommunication conduits. It is envisaged that the same technology may be used for other assessment or classification purposes beyond that of small diameter underground communications conduits.

[0088] The present disclosure provides, among other things, a pipe condition assessment system which employs AR as a non-invasive object sensing method, and 2026201558   27 Feb 2026 leverages a machine learning and / or deep learning techniques for classification of blockage types and / or other features or defects in conduits. The system may also provide the ability to accurately locate the blockage or feature. This may provide a user with a clear indication of a pipe’s internal condition without requiring manual rodding methods or excavation.

[0089] The present disclosure includes but is not limited to: devices and / or apparatus for use in assessing a conduit; a system for comprehensive blockage, feature or condition detection, classification, and localization; a custom test bed; a mobile device program; and demonstrations of the performance of the system based on quantitative and qualitative evaluations and comparison with known values.

[0090] The present system may utilise acoustic reflectometry (AR) which involves the emission of waves which are received after reflection from a surface or object. The waves are preferably soundwaves. According to the present disclosure other types of reflectometry may be utilised which emit and receive other types of waves, although acoustic or sound waves are a preferred method. AR may be used to determine the presence or absence of an object and / or distance to an object. In the present context, AR is used by emission of waves into an end of a conduit, the waves are then received after reflection from a surface or object within the conduit. The waves may reflect from a surface of the conduit itself, an end of the conduit, an object within the conduit, a partial blockage, a complete blockage, damage to the conduit, a bend or a change of direction of the conduit, an irregularity in the conduit, or any other possible element or aspect of the pipe not listed here. Herein, the term signal or wave may be used herein interchangeably in the context of a signal or wave emitted and received for the purposes of reflectometry. It is noted that where an emitted wave may be reflected back and then received, the amplitude of the received wave may vary from that of the emitted wave but the wavelength should remain the same.

[0091] The waves may be emitted and received at the same location. The waves may be emitted and received by the same device. In embodiments, a mobile device, such as but not limited to a mobile telephone, a tablet or a purpose built device, emits and receives the waves. According to other embodiments, the wave may be emitted by one device and received by another device. For example, the wave may be emitted by a separate device connected via a wire or wires or wirelessly to a main device. The emitter may be a speaker. The wave may be received by a separate device connected via a wire or wires or wirelessly to the main device. The receiver may be a microphone. The emitter and / or receiver may be a 2026201558   27 Feb 2026 peripheral device to a main device. In the present disclosure, the term device may be used to reference a single device or a plurality of devices which are interconnected or combine to perform one or more tasks.

[0092] An emitter of waves utilised in the present disclosure may be a speaker, transducer, oscillator, transmitter, ultrasound transducer, or piezoelectric transducer, or any other suitable component. The emitter may be a speaker of a mobile device. The emitter may be a speaker which connects wirelessly or via a wired connection to a mobile device. A receiver of waves utilised in the present disclosure may be a microphone or transducer or any other suitable component. The receiver may be a microphone of a mobile device. The receiver may be a peripheral microphone device that connects wirelessly or via a wired connection to a mobile device.

[0093] The emitter and receiver may be provided together in a device. The device with the emitter and receiver may be separate to or remote from a main device. The device may comprise a body. The body may comprise at least one internal cavity which houses the emitter and / or the receiver. The body may comprise or connect to an interface that is configured to connect to a conduit. The body may be integral with the interface or the interface may be removably connectable to the body. The interface may comprise an adapter and / or a waveguide. The emitter may emit soundwaves that travel into the conduit and the receiver may receive soundwaves reflected back from within the conduit. The emitter and / or receiver may be peripheral devices configured to connect to a main device. Additionally or alternatively, the main device may comprise the body, emitter and / or receiver. The main device may be a computer, laptop, tablet, mobile device, smartphone or other device. The main device may provide commands to the emitter via a wired or wireless connection. The main device may receive signals from the receiver via a wired or wireless connection. The main device may display a representation of signals received from the receiver on a display screen to a user of the device. The main device may connect to a server. The main device and / or server may process the received signals prior to display. The device and / or server may comprise a machine learning model which is configured to analyse the received signal or signals and to provide inferences of location and / or type of condition or feature within a conduit.

[0094] An interface may be used to acoustically connect the emitter / receiver to an end of a conduit. The interface may be or comprise an adapter. The interface may have an end 2026201558   27 Feb 2026 sized to engage with the end of the conduit. The interface may be or include a waveguide to enable transmission of waves from the speaker into the conduit and / or transmission of waves from the conduit to the receiver. Where the device comprises the emitter and receiver in proximity to one another, the interface may be positionable adjacent to the emitter and receiver. A holder may comprise the interface or be connected to the interface. The holder may include a cradle. The holder may hold the device. The device may be held in a fixed position relative to the end of the conduit. The holder may include an adjuster which adjusts a size of the holder. The adjuster may enable the holder to hold devices of varying sizes, preferably in a fixed relationship thereto.

[0095] The interface may be a unitary piece or may be formed of a plurality of interrelated parts. The interface may be attached to or form part of the holder. The interface may be attached to or part of the cradle. The interface may include a source tube. The source tube may be able to be elongated. The source tube may be flexible. The source tube may be connectable to an end of the conduit. The waveguide may comprise the source tube. The source tube may enable the device to be distanced away from the end of the conduit. For example, in real world conditions the conduit may be at least partly underground and / or the end of the conduit may be beneath ground level. The source tube may enable the device to be positioned above ground level and to emit and receive waves / signals through the source tube to and from the end of the conduit. The source tube may assist in overcoming noise in early results due to sound transmission occurring and / or may reduce noise present in a received signal.

[0096] The wave(s) received by the device may be converted into a signal or may be received by the device as a signal. The signal may be a two dimensional waveform. The signal may undergo processing and / or preprocessing. The device may comprise or connect to a processor which processes the signal and / or performs preprocessing. A memory may be utilised to record the signals or save copies of the signals.

[0097] The device may include communications means for transmission and / or reception of data and / or instructions. The device may include a transmitter, a receiver or a transceiver. The device may include a connection to a wired or wireless communications network. The communications network may include a telephone network, a cellular network, a radio network, a Wi-Fi network, a local area network, a wide area network or any other suitable network. The device may include Bluetooth communication and / or near field communication 2026201558   27 Feb 2026 (NFC) capabilities. The device may include or connect to a modem. The device may include or connect to a server.

[0098] The device may include a display. The display may be an integrated display screen. The display may be an LED or other type of display. The device may connect to a display, such as a monitor or other display device. The device may include an input. The input may be any one or more of: switch(es), button(s), control panel, mouse, keyboard, sound input, voice input, gesture input, touchscreen. The device may include an output to output a signal or data. The output may include any one or more of: display screen, printer, communications device, cable, wireless device, electronic file output.

[0099] The system may analyse a signal to determine at least one condition and / or location of a condition. The system according to the present disclosure may include a machine learning model (model). The model may predict a condition, and / or location based on an input signal. The input signal may be obtained by a device according to the present disclosure. The system may take into consideration the time between emission of a signal / wave and when the correlated reflected signal is received. The device of the system may emit signals at different frequencies intermittently. The difference in frequency may aid in correlating the emitted signals to the received signals. This correlation may be used to determine the time taken to travel out and be reflected back and, hence, the distance to the condition. The signal(s) received by the model may be processed or preprocessed signals or may be raw wave signals. The signal(s) may be converted into an image for input into the model. The signal(s) may be converted into scalograms for input into the model. A continuous wavelet transform (CWT) may be used to convert a signal(s) into a scalogram(s).

[0100] The model may be trained using a plurality of input signals of known conduit conditions and / or known conduit condition locations. The model may include a neural network, such as a convoluted neural network (CNN). The model may include a transformer. During a training pipeline, the model may be trained on input signals of known conditions. The model may be trained to recognise, differentiate between and classify conditions within a conduit. The conditions may include, but are not limited to: presence of at least one bend in the conduit; location of each bend in the conduit; shape of bend in the conduit; radius of bend in the conduit; direction of bend in the conduit; angle of bend in the conduit; direction of conduit or change of direction of conduit; presence of a second end of the conduit; location of a second end of the conduit; a length of the conduit; presence of a coupling between two or 2026201558   27 Feb 2026 more sections of pipe forming the conduit; location of a coupling between sections of pipe; presence of a blockage within the conduit; location of a blockage within the conduit; type of blockage within the conduit; size of a blockage within the conduit; presence of damage to the conduit; type of damage to the conduit; and / or location of damage to the conduit. Once trained, the model may perform an inference pipeline where a signal, or image / scalogram representative of the signal, is input and the trained model predicts a condition. Signals relating to conduits of known condition may be input into the trained model and the output may be compared to the known condition to assess the accuracy of the trained model. A trained model may be further trained to improve accuracy and / or to train on different conditions on which it was not previously trained.

[0101] The model may be part of a software application. The software application may be run on the device and / or saved to a memory of the device. The model and / or software application may be saved to a storage. The storage may comprise a memory. The storage may be remote from the device. The storage may be a cloud storage. A server may comprise the storage. The storage may be remotely accessible by the device. The system may enable a user to perform conduit condition assessment utilising a mobile device. TEST BED

[0102] In order to train the machine learning model, a test bed may be utilised to simulate real world conditions of a conduit in the ground. The test bed may comprise a box, large container or at least one wall defining a volume. The volume may be partially or entirely filled with a debris. The debris may comprise one or more of soil, sand, clay, stone, rocks, dirt, brick, vegetation, and / or waste. Within the debris at least one conduit may be buried or submerged. The buried conduit may simulate the conditions of an underground pipe, such as a telecommunications pipe. The conduit may have any desired diameter. The diameter of the conduit may be the same as or similar to a telecommunications conduit. The conduit may have a diameter of about 10mm to about 100mm, about 15mm to about 50mm, about 15mm to about 40mm, about 20mm to about 30mm, about 21mm to about 26mm, about 21mm, about 22mm,, about 23mm, about 24mm, or about 25mm. The conduit may have an internal diameter of about 23.3mm. An end of the or each conduit is exposed or directly accessible. An end of the conduit may extend through a wall of the test bed. An end of the conduit may extend from an upper surface of the debris. 2026201558   27 Feb 2026

[0103] Within the test bed each conduit may comprise at least one straight section. The conduit may comprise at least one bend. At least one bend may be a 90o bend. The conduit may comprise a plurality of bends. The bend(s) in the conduit may each be 90o. Each bend may be an angle selected from one of about 10o, 20o, 30o, 40o, 50o, 60o, 70o, 80o, 90o, 100o, 110o, 120o, 130o, 140o, 150o, 160o, 170o, 5o, 15o, 25o, 35o, 45o, 55o, 65o, 75o, 85o, 95o, 105o, 115o, 125o, 135o, 145o, 155o, 165o, or 175o. . The bend(s) may have a desired radius. The radius may be about 50mm, about 100mm, about 150mm, 200mm, about 250m, about 300mm, 350mm, about 400mm, about 450mm, 500mm, about 600mm, about 700mm, about 800mm, about 900mm, or any other desired radii of bend to be utilised in training or testing the model.

[0104] Figure 3 shows an example of a test bed 300. The test bed may be any desired shape or size. The test bed 300 may be substantially the shape of a rectangular prism. The test bed 300 may be any desired length, width and / or height. The length of the test bed 300 may be about 1m to 10m, 1m to 5m, 2m to 5m, 2m to 4m, about 2m, about 2.5m, about 3m, about 3.5m or about 4m. The width of the test bed 300 may be about 0.5m to 8m, 0.5m to 5m, 0.5m to 4m, 0.5m to 3m, about 0.5m, about 1m, about 1.5m, about 2m or about 2.5m. The height of the test bed 300 may be about 0.2m to 3m, 0.2m to 2m, 0.5m to 2m, 0.5m to 1.5m, about 0.3m, about 0.5m, about 0.8m, about 1m or about 1.2m. The test bed 300 shown in Figure 3 comprises four sides 310. The sides 310 may be each formed from substantially solid material, such as wood. An internal portion of each side may be covered with a protective material 320 that may form a barrier between an inner volume of the test bed 300 and the sides 310. A test bed 300 may also be provided which does not include a protective material 320. The protective material 320 may comprise a thin flexible plastic material, such as, but not restricted to, low-density polyethylene (LDPE), linear low-density polyethylene (LLDPE) or high-density polyethylene (HDPE).

[0105] A debris 330 is placed into the test bed 300. In the embodiment shown, the debris 330 is a sandy loam mixture. The debris 330 is preferably a soil or earth mixture to simulate the conditions in which a telecommunications conduit will be buried. At least one conduit 350, 351, 352 may be positioned in the test bed 300. The conduit(s) 350, 351, 352 may be placed int the test bed 300 before it has been entirely filled with debris 330. The example shown in Figure 3 includes three conduits 350, 351, 352. The conduits may each represent different conduit conditions, sizes and / or lengths. The conduits 350, 351, 352 in Figure 3 are each about 3m to 4m in length. Conduits 350, 351, 352 comprise straight sections 353. The 2026201558   27 Feb 2026 conduits further comprise pre-fabricated 90o bends. First bends 354 have a radius of 105mm. Second bends 355 have a radius of 305mm. Once the conduit(s) are positioned within the test bed 300 they may be covered with debris 330. The conduit(s) 350, 351, 352 may be covered with debris 330 up to a desired depth. The depth may be selected to match the depth at which a telecommunications conduit may be buried in the ground in use. The depth at which the conduit(s) 350, 351, 352 is buried in the debris 3300 may be 100mm to 1000mm, 100mm to 500mm, 200mm to 400mm, or about 300mm. Figure 3 shows the test bed 300 before the conduits 350, 351, 352 have been covered with debris 330.

[0106] A first end 356 of the conduits 350, 351, 352 may extend to and / or into a wall 310 of the test bed 300. A first end 356 of the conduits 350, 351, 352 may extend through a wall 310 of the test bed 300. A conduit opening at the first end 356 may be accessible on an outer side of the wall 310, as shown in Figure 1. In other embodiments, the conduit opening 356 at the first end 356 may be accessible out of an upper opening of the test bed 300. A second end 357 of the conduits 350, 351, 352 may extend towards or from an upper opening of the test bed 300.

[0107] Different conditions of a conduit may be simulated in the test bed. Figure 4 shows examples of synthetic blockages which may be positioned within a respective conduit for training and / or testing purposes. The blockages include: a 25% blockage 401; a 50% blockage 402; a 75% blockage 403; and a 100% blockage 404. The percentage of blockage refers to the amount of internal cross-sectional area this is blocked. The blockages may be formed of any suitable material. For example, the blockages may be formed from a plastic material, such as polyethylene, polypropylene, polystyrene or derivatives thereof, or wood, paper or metal. The blockages may be printed by additive manufacturing. The blockages are sized to fit inside the conduit. The blockages may be configured such that they can be manoeuvred along the conduit internally to synthesize blockages at any desired location. A line of material may connect to the respective blockage and be used for manoeuvring the blockage within the conduit. The line may include, for example, a string, a thread a wire or a nylon line. The line may be accessible from an end of the conduit. In the embodiment shown in Figure 3, the line 360 extends from the second end 357 of conduit 352. Each conduit may be marked at set intervals at which the blockage or other synthesized condition may be positioned. For example, the conduit may be marked at 5cm, 10cm or 20cm intervals along its length. 2026201558   27 Feb 2026

[0108] Other types of conditions may be simulated within the conduit(s) in the test bed, including but not limited to: bends of varying angle and / or radius; ends of pipe at varying distances from the device, e.g. pipes of varying length; conduit extending at varying angles to the horizonal; conduit that deviates from a straight line; conduit that deviates, sags or peaks in one or more locations; conduit is crushed and / or internal diameter reduced at one or more locations, e.g. tree roots crushing a conduit; external matter protruding into a conduit, e.g. tree roots have damaged and protrude into a conduit; soil or other material has entered a conduit; blockages formed from different materials; blockages of different shapes; blockages extending at different lengths along the conduit; presence of lifeforms within conduit, dead or alive, e.g. animals such as mice or insects, worms or other organisms; damage to the conduit, including holes or cracks or sections of pipe that are separated or misaligned; presence of coupling between sections of pipe forming the conduit; deformed sections of conduit; or water and / or mud within the conduit. DEVICE(S) AND APPARATUS

[0109] Figure 1 shows an example of a mobile device 100 configured for use in the disclosed system. The device 100 may be a smart phone. As previously discussed, another device may be used or purpose built for this purpose. A cradle 200 is configured to hold the device 100. The cradle 200 is configured to connect to an end of a conduit. The device 100 and cradle 200 may be configured for use with the test bed 300.

[0110] The device 100 includes a display 110. The device may include at least one input 120. The device may include other forms of input, including a touch screen. The touch screen may coincide with the display screen 110. The device 100 includes a wave emitter. The wave emitter may be a speaker. The device 100 includes a wave receiver. The wave receiver may be a microphone. In the embodiment shown in Figure 1, the wave emitter is the smartphone's speaker and the wave receiver is the smartphone's microphone, located at the second end 140 of device 100. The speaker and microphone in this embodiment are located proximal to one another. The speaker and microphone may be located at the same end 140 of the device 100.

[0111] The cradle 200 is designed to hold the device 100 in a fixed position. The cradle 200 includes a body 210. The device 100 may rest on the body 210. The cradle 200 may include an end wall 220. The cradle 200 may include an adjustor 230. The adjustor 230 may be adjusted such that the cradle 200 can hold devices 100 of different sizes. The adjustor 230 2026201558   27 Feb 2026 may be configured to engage with the device 100 to hold the device in place on the cradle 200. The adjustor 230 may comprise a knob 234 that may be user manipulated. The adjustor 230 may include an adjusting mechanism which is driven by the knob 234 or by another means. The adjusting mechanism may be in the form of a thread 232, as shown in Figure 1. The thread 232 may pass through a correspondingly sized hole in the end wall 220 of the cradle 200. The hole in the end wall 220 may include a corresponding threaded surface that engages with the thread 232. The knob 234 may be turned to cause the thread 232 to move inwards or outwards of the hole in the end wall 220. In this manner the adjustor 230 may be used to engage with and hold the device 100 in place. The adjustor may engage a first end 130 of the mobile device 100.

[0112] An interface may enable the wave emitter to emit waves into the conduit. The interface may enable the wave receiver to receive waves from the conduit. The interface may be attached to or part of a cradle. The interface may be connected to or connectable to the device. The interface may be connected to or connectable to an end of the conduit. The interface may have a portion which is sized to engage with an end of the conduit.

[0113] The interface may be positioned adjacent to the wave emitter and wave receiver. The interface may comprise or be in the form of a waveguide. In Figure 1, the cradle includes a waveguide 240. According to other possible embodiments, the waveguide may not be connected to a cradle. The waveguide 240 may be substantially funnel shaped. The waveguide 240 may have a width that reduces or tapers away from the wave emitter and wave receiver. The waveguide 240 may include an outlet 242. The outlet 242 may be sized to fit inside an end 356 of a conduit. The outlet 242 may have an outer edge which engages with an inner edge of the end 356 of the conduit. A width of the waveguide 240 at the outlet 242 may substantially match an inner diameter of the conduit. Figure 1 shows the smartphone device 100 held within the cradle 200 by the adjustor 230, with the waveguide 240 outlet 242 engaging with an end 356 of a conduit of test bed 300.

[0114] The interface may comprise an adapter which is sized and shaped to engage with an opening / end of the conduit. The adapter may be part of or connected (directly or indirectly) to the waveguide. The adapter may comprise the outlet of the waveguide. An adapter may be provided which can engage with conduits of varying diameter. Additionally or alternatively, a plurality of adapters may be provided which are respectively adapted to engage with conduits 2026201558   27 Feb 2026 of different diameters. The adapter and / or waveguide may be formed by additive manufacturing / 3D printing.

[0115] According to other possible examples, the interface may include a source tube. The source tube may be a flexible length of tube. The source tube may be able to be elongated. The source tube may form part of or be connected to the waveguide 240. The source tube may be connectable to an end of the conduit. The source tube may connect to the adapter. The source tube may enable the device 100 to be located distally of the end of the conduit. According to examples, the source tube may be 50cm to 3 metres in length, other desired lengths of source tube are also possible. The source tube may be extendible and / or retractable.

[0116] In use, the device 100 will emit waves into the conduit. The emitted waves may be soundwaves. The emitted waves may be a chirp signal output by the device's speaker. The waves travel into the conduit until they meet a physical feature. Waves will rebound back from the physical feature along the conduit and be received by the device 100. In the embodiment shown, the device includes a processor. The device may include a memory. The device may include a read only memory (ROM), a random access memory (RAM), a cache, a memory card and / or any other suitable type of memory. The received wave signals will be processed by the device 100. The device 100 may include a software which is configured for controlling the emission of waves and processing received signals. The software may be a mobile device application (app).

[0117] The device 100 may be connected to a communications network. The network may be a wireless network. The device 100 may include a modem. The device 100 may be configured to transmit data. The device 100 may communicate with other devices and / or with a server. The device may communicate with cloud based server.

[0118] According to embodiments of the present disclosure, the device(s) which emit and / or receive sound wave signals are remote from or separate from a main device. The main device may be a mobile device, such as a laptop, tablet or smartphone, or any other suitable device. The device(s) which emit and / or receive soundwave signals may be one or more peripheral devices. The peripheral device(s) may connect to the main device via one or more wires and / or wirelessly. The peripheral device(s) may include a speaker and / or a microphone. The main device may control emission of soundwaves from the speaker. The main device may receive signals and / or data from the microphone. The main device may 2026201558   27 Feb 2026 have a display screen to display received and / or processed signals. The main device may also be called a master device. The peripheral device(s) may be held within a body that can connect directly or indirectly via an adapter to a conduit. The speaker may emit soundwaves into the conduit and the microphone may receive soundwaves reflected back from within the conduit.

[0119] Figures 11, 12 and 14 show an embodiment of a device 1000 according to the present disclosure which is configured to emit and receive soundwave signals. The device 1000 may be a portable device and may be of a size and weight to be easily carried by a user. The compact size of the device 1000 may be of assistance when used in a location with congestion around the end of a conduit. In this example, the device 1000 comprises a speaker 1200 and a microphone 1300. The speaker 1200 and / or microphone 1300 can connect to a master device (not shown) via a wired and / or wireless connection. The speaker 1200 may be a peripheral speaker and the microphone 1300 may be a peripheral microphone. The device 1000 comprises a body 1100. The body 1100 may include an outer wall that defines an interior volume 1160. The outer wall may have any desired thickness, such as 1-10 mm. The outer wall may be substantially uniform in thickness or may have varying thickness in different parts of the body 1100. The interior volume 1160 houses the speaker 1200 and the microphone 1300. In other possible embodiments, only one of the speaker 1200 and microphone 1300 may be housed within the interior volume 1160 of the body 1100. In other embodiments, the speaker 1200 and microphone 1300 may be incorporated into a single unit that is housed within the interior volume 1160. The speaker 1200 and microphone 1300 may be removably held within the body 1100. An adapter 1500 is configured to connect to the body 1100.

[0120] In the embodiment shown, the speaker 1200 is housed in a rear section of the device 1000. The device 1000 includes a backplate 1110, which may be affixed to the rear of the body 1100. The backplate 1110 may be affixed by screws 1115 or other fasteners. The backplate 1110 may connect to the body 1100 via a releasable lock or locking mechanism. Affixing the backplate 1110 to the body 1100 maintains the speaker 1200 in a desired orientation relative to the interior volume 1160 of the body 1100.

[0121] The body 1100 includes a tapered section 1140 adjacent to the region that houses the speaker 1200. The tapered section 1140 relates to a tapering of the interior volume 1160 of the body 1100. The tapered section 1140 tapers inwards from the region 2026201558   27 Feb 2026 housing the speaker 1200 to a neck 1150. In the embodiment shown, the neck 1150 is a hollow cylindrical section at an opposed end of the body 1100 to the speaker housing region. The body 1100 of the device 1000 is shaped and sized such that the interior volume 1160 directs soundwaves emitted by the speaker 1200 towards the neck 1150 . The speaker 1200 may be any suitable size and shape. In the embodiment shown, the speaker 1200 is a substantially rectangular prism. The speaker housing area and the neck may be any desired shape and / or size. The neck 1150 may be shaped and sized to connect to the adapter 1500. In the embodiment shown, the speaker housing area is substantially rectangular in crosssectional shape to match the shape of the speaker 1200. The neck 1150 is substantially circular in cross-sectional shape. Thus, the tapered section changes in cross section area as it tapers inwards to the neck from a substantially rectangular to a substantially circular crosssectional shape. The body 1100 is sized and shaped to house the speaker 1200. The speaker 1200 may be an off-the-shelf product or may be a purpose-built component. Where the speaker 1200 is an off-the-shelf product, it may be received within and / or inserted into the body 1100. Where the speaker 1200 is a purpose-built component, it may be integral with the device 1000 or may be received within and / or inserted into the body 1100.

[0122] The speaker 1200 may include Bluetooth communication means to enable it to receive commands from a master device. In other embodiments, the speaker 1200 may receive commands via other means, such as near field communication, via Wi-Fi, and / or via a wired connection. In other embodiments the speaker 1200 may be integral with the master device. The speaker 1200 may include control actuators 1250. The control actuators 1250 may be switches or buttons. In the example shown in Figure 11, the control actuators 1250 include power (on / off), Bluetooth connection, volume up, volume down and 'play'. The body 1100 includes a speaker aperture 1120. The speaker aperture 1120 may provide access to the control actuators 1250 of the speaker 1200. The speaker 1200 and / or microphone 1300 may be waterproof or water resistant. In use, the device 1000 could be used in outdoor conditions or environments that could lead to it getting wet, thus the speaker 1200 and / or microphone 1300 being resistant to water may be beneficial in preventing unwanted damage.

[0123] In the present example, the body 1100 includes a handle 1400. The handle 1400 includes two supports 1410, 1420. The handle 1400 includes a handgrip 1450. The handgrip 1450 is connected to the body 1100 via the supports 1410, 1420. The handgrip 1450 is hollow in this example. A rear of the handgrip 1450 includes a first opening 1460. An underside of the handgrip 1450 includes a second opening 1470. A passage may be provided 2026201558   27 Feb 2026 through the handgrip between the first and second openings 1460, 1470. In the embodiment shown, the second opening 1470 is adjacent to one of the supports 1420. In other examples, the handgrip 1450 may not be hollow. The handle 1400 may allow a technician to carry, hold and manoeuvre the device 1000 prior to inspection of a conduit.

[0124] As shown in Figure 11, the body 1100 includes a microphone aperture 1130. The microphone 1300 may be inserted through the microphone aperture 1130 into the interior volume of the body 1100. Additionally or alternatively, a wire 1350 may be inserted to the microphone 1300 through the microphone aperture 1130. The microphone 1300 is configured to receive soundwave signals that have been emitted from the speaker 1200 and that have entered and been reflected back from within a conduit. The microphone 1300 is configured to send signals and / or data to the master device. In this example, the microphone 1300 is a wired device and connects to the master device via a wire 1350. In other examples, the microphone 1300 may communicate wirelessly, such as via Bluetooth, NFC or Wi-Fi. The passage within the handgrip 1450 may be used to convey the wire 1350 of the microphone 1300. Passage of the wire 1350 through a hollow passages, such as the hollow handgrip 1450, may prevent the wire 1350 from becoming tangled or caught on part of the device 1000 or other elements when in use, and / or may prevent interference that could be caused to the wire 1350 and / or microphone 1300 when a user holds the handgrip 1450. In the example shown, the wire 1350 from the microphone 1300 passes from the microphone 1300 through the second opening 1470 and out of the first opening 1460 of the handgrip 1450. In other examples, the handgrip 1450 does not include a passage. The microphone 1300 and / or wire 1350 of the microphone 1300 may be held in place relative to the body 1100. For example, the microphone 1300 may be held in place via a friction fit with the microphone aperture 1130. In the example of Figure 11, the wire 1350 of the microphone 1300 is held in place adjacent to the handle support 1420 via a strap 1320 that attaches around the support 1420 and the wire 1350. The strap 1320 in this example includes hook-and-loop (Velcro) connectors. Other means for connecting the strap 1320 in position are possible, including, but not limited to, adhesive or magnets. The microphone 1300 may be an off-the-shelf product or may be a purpose-built component. Where the microphone 1300 is an off-the-shelf product, it may be received within and / or inserted into the body 1100. Where the microphone 1300 is a purpose-built component, it may be integral with the device 1000 or may be received within and / or inserted into the body 1100. 2026201558   27 Feb 2026

[0125] The device of this embodiment comprises at least one adapter 1500. The adapter 1500 serves as a connector between the body 1100 and the end of a conduit 1600 to be inspected. In this example, the adapter 1500 is formed separately from the body 1100. The adapter 1500 is removably connected to the body 1100. According to other examples, the body 1100 may comprise an integral adapter 1500. In other words, the body 1100 could itself be configured to connect to the end of a conduit 1600. A separate adapter 1500 may be beneficial as it allows the adapter 1500 to be replaced by another adapter 1500 of different size. In this way, an appropriate adapter may be chosen to suit the diameter of a conduit to be inspected. Another potential benefit of having a separate adapter 1500 may be that, when there is limited space adjacent to the conduit end 1600, before the adapter 1500 is connected to the body 1100, the adapter 1500 may be inserted into the conduit end 1600 to be appropriately positioned before connecting the body 1100. The adapter 1500 includes an inner volume 1560. The adapter 1500 may be configured as a waveguide. As a waveguide, the adapter 1500 may be configured to guide soundwaves emitted from the speaker into the conduit and to guide soundwaves out of the conduit to be received by the microphone 1300. The adapter 1500 may be funnel shaped.

[0126] Figure 13 shows four examples of differently sized adapters 1500a, 1500b, 1500c and 1500d. Each adapter 1500 shown is configured to connect to the end of a conduit 1600 having a diameter of a chosen size. For example: adapter 1500a is configured to connect to the end of a conduit having a diameter of about 20mm; adapter 1500b is configured to connect to the end of a conduit having a diameter of about 35mm; adapter 1500c is configured to connect to the end of a conduit having a diameter of about 50mm; and adapter 1500d is configured connect to the end of a conduit having a diameter of about 100mm. Adapters 1500a, 1500b and 1500c are substantially funnel shaped. The funnel shaped adapters 1500 have a funnel section 1520. The funnel section 1520 reflects a tapering of the diameter of the inner volume 1560. The funnel section 1520 may act in a similar manner to the tapered section 1140 of the body 1100. The adapter 1500 may direct soundwaves from the body 1100 into a conduit , in use. The adapter 1500 may direct reflected soundwaves from the conduit back to the body 1100, in use. Each adapter 1500 includes an extension 1530. The extension 1530 is configured for insertion into a conduit end 1600 of a predetermined size (e.g. about 20mm, 35mm, 50mm or 100mm in diameter in the examples shown in Figure 13). Adapter 1500d of Figure 13 does not include a funnel section 1520 and has a substantially consistent outer and / or inner diameter along its length, with a majority of 2026201558   27 Feb 2026 its length being the extension 1530. The outer circumference of the extension 1530 of each adapter may be configured to substantially match the inner diameter of a conduit into which it is to be inserted. Figures 15 and 16 show how the extension 1530 of a device 1000 may be inserted into a conduit end 1600, in use. The device 1000 is configured to be held in place after insertion into a conduit through friction between the outside surface of the extension 1530 and the inner surface of the conduit into which it is inserted. In this manner, the extension may be a conduit engaging extension. In other examples, the extension 1530 may not be configured to hold the device 1000 in place when inserted into a conduit.

[0127] Each adapter 1500 is configured to removably and / or releasably connect to the neck 1150 of the body 1100. The adapter 1500 and neck 1150 may connect to one another via a relative rotational interaction. Figure 14 shows an example of a suitable mechanism for connection between an adapter 1500 and body 1100. In this example, the adapter 1500 includes a pair of engagement surfaces 1590 and the neck 1150 of the body 1100 includes a corresponding pair of engagement surfaces 1190. Each engagement surface 1590 of the adapter 1500 is configured to mate with a respective engagement surface 1190 of the body 1100. Each of the body 1100 and adapter 1500 includes a respective alignment indicator 1180, 1580. To connect the adapter 1500 to the body 1100, the adapter 1500 is brought into contact with the body 1100 with the alignment indicator 1580 of the adapter 1500 in alignment with the alignment indicator 1180 of the body 1100. The adapter 1500 may be twisted relative to the body 1100 (or vice versa) such that the engagement surfaces 1190, 1590 on the adapter 1500 and body 1100 are brought into engagement with one another. The engagement surfaces 1190, 1590 may be configured such that a relative twisting motion between the body 1100 and adapter 1500 may reach a point where no further twisting is possible and the adapter 1500 is connected to the body 1100. The adapter 1500 may lock onto the body 1100. Relative twisting between the adapter 1500 and the body 1100 in the opposite direction may release the adapter 1500 from the body 1100. The end of the adapter 1500 that connects to the neck 1150 of the body 1100 is at an opposed end of the adapter 1500 to a distal end of the extension 1530.

[0128] Depending on the type of conduit, in many cases, there may be cables, wires or other material congesting an inner volume of the end of the conduit 1600 into which the extension 1530 of the adapter 1500 is to be inserted. To account for this congestion, the extension 1530 of the adapter 1500 includes a cut-out section 1540. The cut-out section 1540 may be configured to be positioned adjacent to a lower part of the inner surface of the 2026201558   27 Feb 2026 conduit, in use. The cut-out section 1540 may allow elements, such as cabling or wiring, to be maintained in position within the conduit and / or extending out of the conduit, while at the same time allowing the extension 1530 of the adapter 1500 to be inserted into an end 1600 of a conduit.

[0129] In some cases, the end of the conduit 1600 into which the adapter 1500 is to be inserted will be so congested that an adapter 1500 of corresponding extension diameter cannot be inserted. In such cases it is envisaged that an adapter having an extension 1530 of a smaller diameter may instead be inserted into the conduit.

[0130] As discussed above, in some embodiments the body 1100 of the device 1000 may comprise the adapter 1500. In other words, the adapter 1500 including the extension 1530 may be integral with the body 1100 and not removably connected. It is noted that any of the discussion herein of the adapter 1500 and / or extension 1530 may be applicable to an adapter 1500 that is integral to the body 1100 of a device 1000. In some examples, the device 1000 may be shaped such that it does not include a neck 1150. In such cases, the shape and size of the body 1100 may taper to the diameter of the extension 1530. The tapering of the body 1100 and / or tapering of a funnel section 1520 of the adapter 1500 may be a substantially linear taper as shown in the accompanying figures, or alternatively may be a non-linear taper.

[0131] The extension 1530 of the adapter 1500 is configured for insertion into the conduit, as discussed, and may engage with the inner surface of the conduit to substantially keep the device 1000 in position. Other examples are possible which do not include an extension 1530. Such examples may have an adapter that contacts or only slightly enters the conduit and / or which is not configured to hold the device 1000 in place relative to the conduit. Other means of holding the device 1000 in place may be used or the device 1000 may be held in place by a user in alternative embodiments.

[0132] According to embodiments, the interior volume 1160 of the body 1100 may be insulated with an acoustic dampening material. The acoustic dampening material may comprise a foam. The acoustic dampening material may comprise any suitable material, for example the acoustic dampening material may comprise a foam formed from any of polyurethane, melamine, polyester fibre or polyethylene. The foam may be open cell or closed cell. The acoustic dampening material may be a high density material, which may be beneficial to absorb lower frequencies, or a lower density material, which may be beneficial to 2026201558   27 Feb 2026 absorb higher frequencies. According to one example, the acoustic dampening material is an open cell polyurethane foam. The acoustic dampening material may absorb soundwaves and / or reduce echoes within the body 1100. The acoustic dampening material may be configured to absorb external noise to reduce interference in signals received by the microphone 1300. The acoustic dampening material may be positioned and / or configured so as not to interfere with the transmission of soundwaves from the speaker 1200 to a conduit, and reception of soundwaves from the conduit to the microphone 1300. The acoustic dampening material may be 1-100mm in thickness, 5-100mm in thickness, 10-50mm in thickness, or about 25mm in thickness. The acoustic dampening material may be provided within the interior volume 1160 corresponding to the tapered section 1140.According to other embodiments, a similar acoustic dampening material may be provided within the interface and / or adapter 1500.

[0133] Important aspects of the device 1000 or example shown in Figure 1 include: a soundwave emitter (for example, a speaker in the example shown), a soundwave receiver (for example a microphone in the example), and an adapter configured to direct waves from the soundwave emitter into a conduit and from the conduit to the soundwave receiver. Many other possible embodiments of the device are envisaged beyond those shown in Figures 1 and 11-16 and the present disclosure incorporates such embodiments.

[0134] As described, the device 1000 may connect to a master device. The master device may be a computer, laptop, mobile device, tablet, smartphone or any suitable device. Optionally, the master device comprises a display screen. The master device may comprise communication means configured to transmit and / or receive data. The master device may command the speaker 1200 to emit soundwaves of a desired frequency and / or volume. The master device may receive signals representative of soundwaves received from the microphone 1300.

[0135] Although the device 1000 shown in Figures 11, 12 and 14-16 is configured to connect to a master device (not shown), it is understood that in other embodiments of the present disclosure, the device 1000 is a standalone device that comprises all of the required features of the peripheral device and master device. For example, the device 1000 may in other embodiments include a display to display wave signals (processed or unprocessed) received from the microphone. The device 1000 may include an internal processor to process the received signals and / or a memory to store received signals. The device 1000 may include 2026201558   27 Feb 2026 a communication means that enables it to communicate with one or more other devices and / or one or more servers.

[0136] According to possible embodiments, the device 1000 may include or house the mobile device 100. Alternatively, the cradle 200 of Figure 1 could be configured to connect to different sized conduits - which may be achieved via the cradle 200 connecting to one of a number of replaceable adapters, in a similar manner to body 1100 connecting to adapters 1500. Alternatively, different cradles 200 may be provided which are configured to receive mobile device 100 and each of which have differently sized waveguides that enable acoustic communication between the device 100 and conduits of different diameters. Any combination of features shown in Figures 1 and / or 11-16 are possible within the scope of the present disclosure. SOFTWARE

[0137] A software was developed to utilise a device, such as mobile device 100 or device 1000, for the purposes of AR conduit condition assessment. The software used in the example shown in Figure 1 was a mobile application. The software may enable the device to transmit a chirp signal (or other suitable wave signal) into a conduit and visualise acoustic reflections. The chirp signal may be an exponential chirp. The software may enable the device to visualise the acoustic reflections in real time.

[0138] Mobile applications are available which enable basic AR principles on a smartphone device, in general these are limited to basic tasks such as measuring distances from the smartphone to a wall in a room. However, the current mobile applications are for general use and do not enable condition classification and localisation in a conduit or provide sufficient control to a user, for example they may not allow customisation, data extraction or feature characterisation.

[0139] The software according to the present disclosure may provide fine-tuned parameter control. The software may be tailored for use with underground telecommunications conduits / pipes. A software application was developed using Flutter version 3.22.3 and deployed on a Samsung S24 smartphone running Android version 14. Other software applications are possible and may use any other operating system, such as Apple iOS. A user may interface with the software application on the device using at least one input on the device. The input may include at least one button and / or at least one 2026201558   27 Feb 2026 touchscreen. According to alternative embodiments, the device may receive voice input(s) from a user. When instructed via a user input, the software application causes the mobile device to transmit a chirp signal from a speaker. The chirp signal is a form of soundwave. As discussed above, other waves may be emitted by the device. The chirp signal propagates through the pipe as a sinusoidal waveform with a predefined duration and frequency sweep within a frequency range. The frequency range may be within the audible range (20 Hz to 20 kHz) or any other desired frequency range. The device's microphone or other signal receiver records the reflected signals. The microphone may be adjacent to the speaker on the device. The reflected signals may be formatted by the device and / or software application. According to this example, the reflected signals (waves) are captured by the device in 16-bit pulse-code modulation (PCM) format. The formatted signal is used for further analysis.

[0140] In the present example, when extracting information from the received reflected signal(s), a cross-correlation function was employed. The cross-correlation function is shown in Equation 1. z-T r T                            r -I r               t l—          i ■ CXy[n] = Lm^x[m] -y[m + n] - Equation 1

[0141] For a given signal index m, this function calculates the similarity between the transmitted signal x and the received signal y at a given time delay n. Further insight into the relationship between these signals is obtained through covariance analysis. This is shown in Equation 2: Covxy[n] = ^m -(^-^(^- ---yj - Equation 2

[0142] Here, x and y represent the mean values of the respective signals, and N is the total number of samples. A positive covariance indicates that the signals vary together, while a negative value suggests inverse variation. A covariance of zero signifies no correlation.

[0143] To further enhance signal interpretation, a normalised cross-correlation is used, which provides a standardised measure of similarity with values ranging from -1 to 1. The equation used in this example to provide normalised cross-correlation is shown in Equation 3. Cnorm[n] = Em=i(xm-x)-(ym-n-:y) ^X^=i(xm-x)2E^=i(ym-n = - Equation 3 ■ y)2 2026201558   27 Feb 2026

[0144] An optimised computational form of the normalised cross-correlation was implemented. This may provide enhanced real-time processing on the device. This is shown in Equation 4. ovxv(n) cnorm(n) = ^,0^,0)^0^.,,0)  Equation 4

[0145] The software application may enable a display screen of the device to visualise received reflected signal(s). The visualisation of the signal(s) may occur in real-time. The signals displayed may be a raw cross-correlated signal or a normalised cross-correlated signal. The visualisation may be in the form of a graph. Peaks in a displayed graph may indicate the presence a condition or feature of a conduit. The condition / feature may include any of a blockage, damage, bend, an end of a conduit or other structural feature inside the conduit. A software application in this manner may provide an intuitive diagnostic tool for conduit condition assessment. MACHINE LEARNING ALGORITHM AND SYSTEM FRAMEWORK

[0146] Reflected wave signals received by the device may be inherently noisy. The presently disclosed process may involve data pre-processing and / or data enhancement and may be followed by feature extraction and / or classification methods. Time-frequency analysis techniques may be effective in enhancing signal representation and mitigating noise. An approach of applying a continuous wavelet transform to generate time-frequency scalogram images which are processed using machine / deep learning classification may be beneficial for improving signal representation and noise suppression. The scalograms may be a twodimensional representation of a time-frequency distribution of a signal, where the x-axis is representative of time and the y-axis is representative of the frequency. The scalogram image may be colour or greyscale with the amplitude of the signal is depicted by a respective colour or brightness of the image at a given point. This approach may be utilised in underground conduits, including, but not limited to, PVC pipes. Time-frequency scalogram analysis may enhance the clarity of reflections, and may improve the distinction between signal components and noise. The present, apparatus systems and methods may utilise a convolutional neural network (CNN). The CNN may be utilised for local feature extraction. The present, apparatus systems and methods may utilise a transformer. The transformer may by used for global feature learning. Combining the use of a CNN with a transformer may 2026201558   27 Feb 2026 improve classification accuracy, including in acoustic signal analysis. It is noted that other types of machine learning or neural network architectures may be utilised. The present disclosure incorporates any such alternative methods.

[0147] According to the present example, a CNN-transformer hybrid model is trained to detect and classify objects of interest within the time-series signal data. The model may be formed or coded on any suitable platform. The present example was implemented using the application processing interfaces (APIs) TensorFlow and Keras. CNN may act as a feature extractor. The CNN may be used to capture spatial and / or frequency-based representations from scalogram images. The transformer may enhance the model’s ability to recognize long-range dependencies and / or complex patterns within features extracted by the CNN.

[0148] The CNN may comprise a plurality of convolutional layers. A convolutional layer may employs kernels or filters. The filters / kernels may be used to analyse an input data. The input data may be an image, which in the present context may be a scalogram or spectrogram that represents the time-frequency domain of a reflected signal. This process identifies and extracts patterns, referred to as features, from the input data. The input data may be processed prior to feature extraction by the CNN. The convolutional layer(s) may perform mathematical operations to create a feature map. The CNN may comprise a plurality of pooling layers. A pooling layer may be used to reduce the spatial dimensions of feature map(s). According to the example, the CNN may comprise a plurality of convolutional layers and pooling layers. The convolutional and pooling layers establish spatial hierarchies. The transformer may utilise layers. The transformer layers may leverage multi-head self-attention to refine and contextualize features from the CNN. This process may provide accurate classification of conduit features.

[0149] In the present example, TensorFlow is used to provide compatibility with a graphics processing unit (GPU) acceleration, scalability, and seamless deployment. The CNN-transformer hybrid model may provide fast and accurate feature detection in a conduit, such as underground PVC pipes. The model may balance local feature extraction with global contextual learning, and may be well-suited for handling any challenges posed by real-world underground conduit environments.

[0150] Preprocessing may be used to standardise and format input data. The preprocessing may provide data that is optimised or well suited for machine learning (ML) or deep learning (DL) training. Preprocessing may include any of the following: 2026201558   27 Feb 2026

[0151] 1. Normalisation of received signals, where the received signals may be reflected waves received by the device. This preprocessing step may ensure that variations in amplitude do not bias the model.

[0152] 2. Cross-correlation may be used to align received signals with transmitted signals, e.g. received and transmitted waves. This step may compensate for time delays caused by different conduit conditions.

[0153] 3. A filter and / or other denoising technique may be applied. The applied filter may be a bandpass filter. The filter may remove noise outside a relevant frequency range. This step may preserve only signal components indicative of conduit conditions, including but not limited to blockages, damage, an end of a conduit or bends.

[0154] 4. Segmentation - the filtered signal may be segmented based on a predefined distance range. This step may ensure that the model learns features relevant to one or more specific sections of a conduit.

[0155] Any one or more of steps 1 to 4 may be performed on a raw signal to create a preprocessed signal. A continuous wavelet transform (CWT) may then be used on the preprocessed signal to generate a time-frequency scalogram. This may involve converting time-series signals into image representations. The image representations of a scalogram may be better suited to feature extraction by a convolutional neural network (CNN) compared to a raw wave signal.

[0156] The machine learning model according to the present example uses a CNN as a feature extractor. The feature extractor may capture spatial patterns from input scalogram images. The CNN architecture may include a plurality of convolutional layers. The convolutional layers may have increasing filter depths. For example, the filter depths of the convolutional layers may increase from 32 to 256. A plurality of pooling layers may follow the convolutional layers. The pooling layers may be max-pooling layers. The max pooling layers may be used to progressively reduce spatial dimensions while retaining key features. The final convolutional layer may apply strided convolutions. In a strided convolutional layer, the filter skips some pixels as it moves over an input. This may provide downsampling of the input data. The strided convolutions may refine the extracted features before flattening them into tokens. The tokens may be processed by a transformer. This structured approach may allow both local textures and global frequency distributions to be captured. This may make the 2026201558   27 Feb 2026 architecture effective for differentiating between normal conduit conditions, including but not limited to straight sections of unblocked and undamaged conduits, and abnormal conditions, including but not limited to blockages, damage, bends or an end of a conduit.

[0157] A transformer may receive the extracted features or tokens from the CNN. A transformer may process sequential data by learning relationships between elements within the sequence. Transformers may use a self-attention mechanism, which may enable parallel processing and efficient training.

[0158] In the present example, the transformer is responsible for classifying extracted features into categories. The categories may include the type of condition within the conduit or a subcategory of a type of condition. The CNN output may be reshaped into token sequences. Positional embeddings may be added to retain spatial relationships. The transformer may have a multi-head self-attention mechanism. This mechanism may enable the model to weigh important features differently depending on their relevance. This selfattention mechanism may assist in providing differentiation between categories of condition and / or differentiation between subcategories of condition. Feed-forward layers may further refine these features. In the present example, the resulting classification and / or categorisation is performed using a softmax activation function. The use of a transformer may provide effective recognition of subtle variations and / or long-range dependencies in conduit conditions. Therefore, the use of a CNN-transformer hybrid model may be superior or more accurate than a CNN-only model.

[0159] According to an example of the present disclosure, the machine learning framework may utilise a dual-branch late-fusion architecture. The architecture may combine two independently trained models: (i) a feature-based classifier implemented as a Multi-Layer Perceptron (MLP), for example using a scikit-learn framework; and (ii) an image based classifier implemented as a convolutional neural network (CNN) operating on scalogram representations, for example using TensorFlow or PyTorch.

[0160] In the feature based classifier, numeric features are provided as input to the MLP classifier. Using scikit-learn, or an equivalent classifier framework, the features may be subjected to preprocessing steps including imputation and standardisation prior to classification. The MLP may comprise multiple fully connected layers, for example with layer sizes of 256 and 128 units, and may employ a rectified linear unit (ReLU) activation function. 2026201558   27 Feb 2026 During training and evaluation, performance metrics are generated, including any of accuracy, precision, recall, F1 scores (macro and / or weighted), and macro one-versus-rest area under the receiver operating characteristic curve (AUROC-macro-OVR). Associated evaluation artefacts may include reports, confusion matrices, and metric bars.

[0161] In the image based classifier, scalogram images are input to the CNN. The CNN may be implemented as a lightweight architecture employing depthwise-separable convolutions and Squeeze-and-Excite modules, for example using TensorFlow or PyTorch. The trained CNN model is persisted, for example as a scalogram_cnn.keras or .pt file. The same set of evaluation metrics and visualisations as used for the feature based branch may be generated for the CNN, including any of accuracy, precision, recall, F1 scores, AUROC-macro-OVR, and / or corresponding reports and confusion matrices.

[0162] The outputs of the feature based MLP and the image based CNN are combined using late fusion. In this approach, fixed weighting combinations of the respective model outputs are evaluated, for example to determine an optimal combination. The selected fusion weights for the feature-based and image-based outputs (w_feat, w_img) are stored, for example in a fusion.json file. A corresponding label map is stored, for example in a labels.json file. A training summary card may also be generated, documenting information such as environment versions, dataset counts, and input and output tensor shapes associated with the training process. LOCALISATION

[0163] In addition to being classified, a condition within a conduit may also be localised. In other words, the location of the condition may be determined. In the present examples, localisation occurs in one dimension. In other words, the localisation occurs as a distance along the conduit and does not relate to a transverse position within the conduit.

[0164] Localisation of the condition may be performed by a suitable method. According to some examples, a machine learning or deep learning model may be used to determine the location of the condition. This may be the same model that is used to classify the condition, such as a CNN-transformer hybrid model. Alternatively, localisation of a condition within a conduit may be performed using another method. The method of localisation may use peak detection of the reflected wave / signal received by the device or the preprocessed signal or the scalogram. 2026201558   27 Feb 2026

[0165] According to the present example, localisation is performed by analysing the received signal / wave. The peak or peaks of a waveform may be associated with a condition. The location of the condition may be determined using the peaks and calculations from time-of-flight (ToF) of the emitted signal. A computer programming language may be used to perform part of all of the analysis. According to the example, the Python programming language was used. SciPy is an open-source scientific computing library for Python. A function known as scipy.signal.find_peaks was used on signal regions identified to contain a blockage. Parameters of this function, such as height, threshold, and prominence, were selected based on the nature of the experimental setup. For example, the test bed utilised pipe lengths of about 3 to 4 m. The output of this function for a given input may be an array of peaks. The distance d from the testing device to each peak in the array may be measured using known ToF calculations. d = —^-^--Equation 5 2 x mampl

[0166] An example equation for distance calculation is provided by Equation 5. Here, n is the sample index, e.g. number of samples taken between emission of the wave to the peak. The speed of the wave is denoted by v. Where, the speed of sound in air at 20oC is about 343 ms-1. The sampling rate of the receiver of the signal is / sample ,which may be defined in the device's configuration file. According to an example, the sampling rate is 48kHz. This approach may accurately locate a detected condition within a conduit.

[0167] Figure 2 represents the method of classifying and localising conditions within a conduit the example as described above. An acoustic wave may be emitted by a speaker of the device. A raw acoustic wave 501 is received by a microphone of the device. The raw acoustic wave 501 undergoes data preprocessing 502. The data preprocessing 502 may include any one or more processes to format and / or standardise the raw acoustic signal 501. Data preprocessing 502 may involve any one or more of normalisation, cross-correlation, a bandpass filter, and segmentation. The resulting preprocessed signal 503 may be passed to a continuous wavelet transform (CWT) 505. The CWT 505 is used to generate a timefrequency scalogram 506. The time-frequency scalogram 506 undergoes machine learning or deep learning modelling process. The time-frequency scalogram 506 is processed by convolutional layers and pooling layers of a convolutional neural network (CNN) 507. The CNN 507 may be used to extract features of the input time-frequency scalogram 506. The extracted features may be flattened into tokens. The extracted features or tokens from the 2026201558   27 Feb 2026 CNN 507 are input to a transformer 508. The transformer is used to classify and / or categorise one or more conditions detected in the signal through this process. Simultaneously with the steps required to classify the signal, the preprocessed signal undergoes peak detection 504. According to other examples, peak detection may occur before or after classification. The peak detection 504 is used to localise a detected condition. The result of these processes is classification and localisation 510 of a condition within a conduit. MODEL TRAINING

[0168] The machine learning or deep learning model to be used for conduit condition assessment should be trained before it can be employed in real world situations. This may be referred to as the training pipeline. The CNN-transformer hybrid model according to the present example was trained using TensorFlow and Keras. An input shape of 256 x 256 x 3 pixels was used to reflect the size of the scalogram images. As described, the CNN extracts spatial-frequency features. As also described, the transformer refined feature relationships for classification. Hyperparameters included: a batch size of 32; 300 training epochs; a learning rate of 0.001 (Adam optimizer); and a dropout rate of 0.1 to prevent overfitting.

[0169] The type of condition within conduit is known during training. This may be a blockage of a certain type and of a certain size, positioned at a known location within a test pipe of a test bed. The known type of condition may be input into the training pipeline. Iterations of training are performed with different types of condition, different size of condition and at different locations. The model may learn which extracted features are associated with certain conditions and may learn through many iterations of testing commonalities between types of condition within a conduit, sizes of condition within a conduit and / or locations of condition within a conduit. The example shown in Figures 3 and 4 utilised four sizes of blockage 401, 402, 403, 404 within test conduits 350, 351, 352 within the test bed. This test bed was used to perform preliminary training and testing of the model. A model trained on the four blockage sizes used (which are of the same type of blockage being a planar blockage perpendicular to the length of the pipe), when correctly trained, will be able to classify a condition within a conduit to be one of these four blockage sizes. However, this same trained model cannot be used to classify other conditions on which it has not been trained. Optimally, a model that is used in real world conditions may be thoroughly trained in test situations on a large number of conditions. Preferably, the conditions on which the model is trained will cover a full range of conditions that may be present in real world underground conduit scenarios. In 2026201558   27 Feb 2026 addition to data from test pipes with known conditions within a test bed being used to train the model, it is also possible to train the model using real world data. In this way the model can continuously evolve after it has been deployed, which may lead to a more accurate model that may be able to classify a larger number of conduit conditions.

[0170] Throughout the training pipeline, key metrics may be monitored to evaluate performance. The trained model may be saved locally for deployment in the inference pipeline. The inference pipeline being the use of the trained model to make inferences or predictions about conditions within a conduit. The inference pipeline can be used for testing the model and / or real world deployment. ACOUSTIC DATA COLLECTION - EXAMPLE

[0171] In the given example, preliminary testing determined the frequencies that were to be used for detecting different conditions within the conduits 350, 351, 352 in the laboratory test bed 300. The transmitted signal specifications were a sinusoid with a linearly increasing frequency from 2 kHz to 6 kHz for 2 milliseconds. This signal being a chirp signal emitted by a speaker of device 100. Signals generated by the application software on the device 100 and audio recordings of reflections received by the device 100 were sampled at 48 kHz.

[0172] A synthetic blockage 401,402, 403, 404 was placed 100 mm from the start of a first pipe 350,. The start of the pipe 350 being the end 356 adjacent to the device 100. Data collection was initiated by the software application of the device 100. At least 20 signals were transmitted and received with the same blockage at the same location in the same pipe. The blockage was incrementally moved 100 mm further until it was 3.5 m from the start of the first pipe 350. At each 100 mm interval data collection was initiated by the software application of the device 100 at least 20 times. This process was repeated for all four synthetic blockages 401, 402, 403, 404. The signals were split into training and validation datasets, with a roughly 80% and 20% split respectively. The training dataset comprised 1918 individual signals, while the validation dataset comprised 482 individual signals. The training dataset was used to train the model. The validation dataset was used to test performance of the model, which may include the model's accuracy. EVALUATION OF MODEL 2026201558   27 Feb 2026

[0173] The trained model may be evaluated quantitatively and / or qualitatively. Evaluation of the trained model before it is deployed in a real world scenario may ensure that the model, when used for the inference pipeline, is fit for task and accurate and may prevent or reduce the likelihood of incorrect classifications. Poor qualitative or quantitative performance(s) may indicate that the model requires further or more specific training.

[0174] Quantitative evaluation of the model may utilise any suitable metrics. Quantitative examination of the model according to the present example evaluated precision, recall and F1-Score. F1-Score is a metric for measuring object classification performance. The F1-Score represents the harmonic mean of precision and recall. F1-Score values range from 0.0 to 1.0, with a value approaching 1.0 signifying high accuracy in both precision and recall.

[0175] Validation tests were performed on using the system according to the present example. The trained model was used to classify a validation set of the four blockage sizes 401, 402, 403, 404 at varying locations. In this case the model classifies the condition within the conduit rather than learning from input condition information. The CNN-transformer model of this example achieved exceptionally high F1-scores across all four blockage sizes (401, 402, 403, 404), as shown in Table 1. An overall F1-Score of 0.99 was achieved, which suggests that the model has a highly effective blockage classification capability.

[0176] Though the results of the validation tests of this example are positive, they are limited to classification of the specific conditions assessed (four blockage sizes 401, 402, 403, 404). This means that, while the model will be highly effective at classifying blockages of these type and size, other conditions with different parameters, such as datasets featuring pipes with degrading interiors or blockages of irregular shapes and sizes, may not be classified by the same model. These challenges may be addressed through additional training on different conduit conditions and more diverse datasets. Thus, extensive training in a variety of different conditions may be necessary to ensure the conduit condition assessment model is robust in real world scenarios. Blockage size Precision Recall F1 Score 25% 0.99 0.99 0.99 50% 0.96 1.00 0.98 75% 0.98 0.97 0.98 100% 1.00 0.98 0.99 Table 1: Performance metrics of trained model 2026201558   27 Feb 2026

[0177] A qualitative assessment of the model may be conducted by analysing output visualisations. The inference process of the model may generate graphical outputs. The graphical outputs may help a user to interpret the reflected acoustic signal to determine conditions in the conduit. The graphical output may be a plot of the signal in a graph. The x axis may show distance, e.g. the distance along the conduit. A distance of 0 indicating the end of the conduit into which the wave is emitted. The y axis may show amplitude or normalised amplitude or correlation between transmitted and received signals. The signal may be plotted on the graph as a sinusoidal waveform. Other visualisations or graphical outputs may be provided by the system. For example, according to other embodiments a scalogram generated by the system representative of the detected signal may be displayed to a user. A report of the classified condition may be provided to the user. The report may be a written summary.

[0178] Examples of graphical outputs are shown in Figures 5, 7 and 9. These graphical outputs include markers. The markers are located at peaks on the graph. The markers may indicate classified conditions predicted by the system. In the examples of Figures 5, 7 and 9 the markers indicate blockages and class of blockage, e.g. size of blockage as used in the training examples. The markers may represent predictions made by the system on condition type (e.g. blockage) and also indicate the calculated distance of the condition from the first end of the conduit. Text or other visual indication may be provided that indicates the type of condition, the location / distance of the condition and / or a class or sub-class of the condition.

[0179] Figure 5 presents an output plot 600 of a processed signal from a validation data set, e.g. used to test the predictions made by the system. The output plot 600 includes an x-axis showing the distance from a first end of the conduit in metres. The output plot 600 includes a y-axis showing the normalised amplitude reading. The signal 610 is shown as a waveform. In the example shown, the signal 610 has undergone preprocessing, however according to other examples the plot may show an unprocessed signal. The waveform of the signal 610 includes a cluster of peaks. Marker 620 is provided on the plot 600 to show the location of a predicted condition. In this example, the system predicts a 25% blockage at a distance of 2.51m along the conduit. This compares favourably to the actual blockage which was a 25% blockage at 2.50m along the conduit. Thus, the system correctly classified the blockage, and provided the location at an error of just 0.4%. In the output 600, the blockage indicated by marker 620 is distinctly visible within a cluster of peaks, while the rest of the processed signal remains substantially free from noise. 2026201558   27 Feb 2026

[0180] The output plot 600 shown in Figure 5 includes a pipe representation 630. The pipe representation 630 may reflect the length of the pipe being analysed or a section of the pipe being analysed. A blockage representation 640 may represent the actual location of a blockage within the pipe being analysed. The pipe representation 630 and blockage representation 640 may be provided in test cases where the pipe length and precise location(s) of a blockage are known, such as where test bed 300 is used. In the example shown in Figure 5, the representations 630, 640 relate to an experimental test with a 25% blockage positioned at 2.50m from the first end of the conduit. Representations of other conditions other than blockages may also be included on the pipe representation 630 where they are present in the pipe being analysed. The blockage representation 640 or other condition representation may include a visual indication of the class, type, size and / or subclass of condition. For example, the condition or blockage may be shown in different colours and / or shapes to represent different the class, type, size and / or sub-class of condition represented. The inclusion of a pipe representation 630 and blockage representation 640 (or representation of any other condition) may enable a user of the device to quickly compare the position of a marker signifying a condition with a representation of the condition as a rapid tool for assessing the accuracy of the predicted condition. The clarity provided by a plot of this type may simplify condition classification and localization for a user.

[0181] Figure 6 shows a scalogram 650 derived from the signal 610 shown in the output plot 600 of Figure 5. The 25% blockage indicated by marker 620 in Figure 5 is reflected by the bright region 655 in the scalogram 650. The scalogram 650 may be generated in colour or greyscale with respective amplitude the signal represented by change of colour or brightness. In one example, the parts of the signal without any significant changes in amplitude are shown as a dark background region. In the example, the parts of the signal that are associated with increased amplitudes or peaks in the signal, such as the region 655, appear brighter and / or have a change of colour from the background regions. The scalogram is generated by applying a CWT to the preprocessed signal. The scalogram 650 may be fed into the machine learning model (e.g. CNN) to make predictions about the condition of the conduit under test. The predictions of the system may be shown in the output plot 600, for example a class of 0.25 (25% blockage) at a distance of 2.51m as shown at marker 620 in Figure 5.

[0182] Output plot 700 for a signal 710 of the validation dataset is shown in Figure 7. A scalogram 750 derived from the signal 710 is shown in Figure 8. The pipe representation 730 includes a blockage representation 740 showing a blockage located at 1.50m along the 2026201558   27 Feb 2026 conduit. The blockage under test was a 100% blockage. The scalogram 750 includes the bright region 755 which reflects the marker 720 in the plot 700. The elongated nature of region 757 shown in scalogram 750 reflects the noisy signal section 725 in the plot 700. As indicated by the marker 720, the system correctly classified this blockage as a 100% blockage (class 1.00). The marker 720 predicted the blockage to be at 1.55m along the conduit, providing a localisation with error of 3.3%.

[0183] The example of Figure 7 includes an additional secondary reflection which was incorrectly classified as shown by marker 721 as another 100% blockage (class 1.00) located at 3.09 m along the conduit. The secondary reflection represented by marker 721 is also shown in the scalogram 750 at region 765. The region 765 appears similar in shape to region 755, although is much fainter . This secondary reflection, with a lower magnitude than the primary reflection, appeared at almost exactly twice the predicted distance of the initial blockage (1.55 m), indicating that it was caused by the initial signal reflecting twice within the short-length pipe. This demonstrates that further training of the system may be necessary to improve classification accuracy. For example, the system could be trained to distinguish between a primary reflection which should be classified from a secondary or other later reflection which does not require classification. Refinements of this type may help prevent false predictions of conditions in underground pipes, ensuring more reliable detection and classification. Similarly, training the system to classify and differentiate between different types of conditions within a conduit may also assist in ensuring the system is more reliable.

[0184] In order to further evaluate the machine learning system according to the present example, test data from different pipe configurations was fed into the system. Figure 9 shows an output plot 800 for a pipe configuration that included different conditions than blockages, namely bends in the conduit. The specific pipe configuration depicted by the output plot 800 included three 90-degree pre-fabricated bends. Figure 10 shows a scalogram 850 derived from the signal 810 shown in output plot 800.

[0185] Pipe representation 830 shows a representation of the configuration of the pipe. The pipe representation 830 includes a first bend representation 840, a second bend representation 842 and a third bend representation 844. The first and second bend representations 840, 842 are rectangle shapes in the output plot 800. The first and second bend representations may be depicted in a first colour. The first and second bends depicted by first and second bend representations 840, 842 have a radius of 105 mm. The third bend 2026201558   27 Feb 2026 representation 844 is depicted as an elongated rectangle and denotes a third bend with a radius of 305 mm. The third bend representation may be depicted in a second colour different to the first colour. The type of feature, and / or differences in respective features, such as differences in radius of bends, may be visually distinct from one another, for example by being denoted by different colours displayed to a user.

[0186] In the test bed 300 used in the shown example, the conduits 350, 351,352 are formed of prefabricated parts. The bends in the examples shown are pre-fabricated parts that slide into straight pipe sections. A small overhanging lip being present at each joint between bend and straight section of pipe. Nevertheless, a respective peak was detected near the start of each of the first and second bends, denoted by markers 820, 821. The markers 820 and 821 are respectively representative of the bright regions 855 and 856 shown in scalogram 850. In addition, peaks were detected near the start and near the end of the third bend, denoted by markers 822, 823. The peaks denoted by markers 822 and 823 have a lower amplitude than the peaks at markers 855 and 856, and this is also shown in the corresponding fainter regions 857 and 858 in the scalogram 850. The locations on signal 810 relating to the first and second bends denoted by first and second bend representations 840, 842 respectively include a cluster of peaks over the duration of the smaller bends. These clusters may be indicative of the start and end of the first and second bends each being located 165 mm apart, which may result in greater noise surrounding the reflected signal. Marker 824 on signal 810 is not located in proximity to any of the bends. The marker 824 may denote the second end of the pipe or a feature at or proximal to the second end of the pipe. The location of marker 824 corresponds to the bright region 859 in the scalogram 850.

[0187] The system had not been trained to classify bends or the end of a pipe. Each of the markers 820, 821, 822, 823, 824 predicted a 25% blockage (class 0.25) in the conduit at the locations of the bends. This may be symbolic of the nature of the physical transition from pipe to bend, e.g. small overhanging lip, being very minimal and most similarly comparable to a 25% blockage of the four blockage sizes assessed. Preferably, the model used by the system will be trained on a plurality of test conduits including bends of varying size and / or angle of known angle, radius and position to enable it to make predictions of conduits having bends with similar features.

[0188] The time-frequency scalograms 650, 750, 850 shown in Figures 6, 8 and 10 provide an intuitive visualization of how a signal’s frequency content evolves over time. One 2026201558   27 Feb 2026 observation is the distinct difference in shape between the 25% blockage feature indicated by marker 620 in Figure 6 (derived from signal 610 of Figure 5) and the 100% blockage in Figure 8 (derived from signal 710 of Figure 7). This contrast highlights the dynamic nature of scalograms. The scalograms may offer finer resolution at lower frequencies and coarser resolution at higher frequencies, which may mirror the behaviour of natural signals. Additionally, the features shown as scalogram 850 in Figure 10 appear larger and less distinguishable than the features in scalograms 650, 750. This may be due to increased noise introduced by multiple physical components in the pipe, i.e. bends. Further preprocessing of the signals may improve feature clarity and may also be particularly helpful when training a model to classify multiple blockage types.

[0189] The example discussed above trains a machine learning model to classify and / or localise conditions within a conduit using physical simulations, such as that of test bed 300. It may also be possible to obtain signal data from simulated sources. In other words, accurate virtual simulations of conduit conditions could be used to generate signals representative of a real work sound wave reflections. One example of a mathematical modelling software that could be used for such a purpose is COMSOL. METHODS OF USE OF THE DEVICE(S) AND SOFTWARE

[0190] The general principles of use of the device shown in Figure 1 will be similar to the use of the device shown in Figure 11. One difference being that, in Figure 1, the emitting and receiving device is also a processing / display device, whereas in Figure 11, the device that emits and receives acoustic signals is separate or remote from a master device. According to embodiments herein, as described, a device for acoustic inspection of a conduit comprises an interface to engage with an end of the conduit. The interface may be as discussed in any embodiment herein, including and not limited to the adapter 1500 with extension 1530 described with reference to Figures 11-17. It is understood that the term interface may reference any suitable apparatus for engaging with an end of a conduit and conveying soundwaves.

[0191] A main device may be used to display signals, particularly processed signals, to a user. The main device may be the same device that transmits and receives soundwaves, e.g. mobile device 100 as in Figure 1, or may be a master device for a separate device or devices 1000 that transmit and / or receive soundwaves shown in Figure 11. According to embodiments, the main device comprises software according to the present disclosure. The 2026201558   27 Feb 2026 software enables the user to review signals representative of features and / or conditions within a conduit and determine locations of the features and / or conditions.

[0192] To perform an acoustic inspection of a conduit, optionally a user may input an estimate of the conduit length and / or diameter. This information, where provided, may enable the application to define the parameters of the soundwaves to be emitted and / or may assist with selecting bounds of axes of plots later provided to the user.

[0193] The interface is inserted into the end of the conduit. The user interacts with the main device to instruct it to begin interrogation of the conduit. According to embodiments, the device emits a series of chirps through the interface into the conduit over a predetermined time period. In examples, the number of chirps emitted is 1 to 100, 1 to 50, 10 to 50, 5, 10, 20, 30, 40 or 50 chirps. The chirps may be emitted over a period of 10 seconds to 10 minutes, 20 second to 5 minutes, 30 seconds to 2 minutes, 30 seconds to 1 minute, or 50-60 seconds, as suitable for the number of chirps. In examples, exponential chirps may be configured based on the length or estimated length of the conduit and / or based on the diameter or an estimated diameter of a conduit. In examples, the chirps may have different configurations for different distances. For example, the chirps may change in frequency and / or duration. The selected frequency and duration of each chirp may be selected or calculated based on conduit diameter, selected distance range and by utilising pipe acoustic theory, including the cutoff frequency for plane-wave propagation and signal attenuation over distance.

[0194] In one example, the lowest and highest frequencies used are constrained by the usable acoustic range of the speaker employed. In examples, the lowest frequency is approximately 180 Hz and, separately, the highest frequency is approximately 20 kHz. The highest practical frequency is limited by the acoustic cutoff frequency of the conduit, which depends on conduit diameter and can be estimated using established theory as follows: •  P20 (23.3 mm): ~8,700 Hz •  P35 (39.65 mm): ~5,000 Hz •  P50 (53.0 mm): ~3,800 Hz •  P100 (104.9 mm): ~1,900 Hz 2026201558   27 Feb 2026

[0195] The cutoff frequency is used as an upper bound so that only plane-wave propagation is utilised. This simplifies signal analysis and reduces sensitivity to attenuation effects associated with higher-order modes that occur above the cutoff frequency.

[0196] Examples of chirp durations, start and end frequencies shown in the table below. These were calculated based on acoustic attenuation theory, taking into account the start and end distances associated with each search range. Search Range Conduit Size Start Frequency (Hz) End Frequency (Hz) Chirp Duration (s) 0-10m P20 3570 5720 0.0052 P35 1910 4050 0.0052 P50 890 3030 0.0052 P100 180 1530 0.0052 10-30m P20 1910 3050 0.0087 P35 1910 3050 0.0087 P50 1890 3050 0.0087 P100 390 1530 0.0087 30-60m P20 1430 2290 0.0124 P35 1430 2290 0.0124 P50 1430 2290 0.0124 P100 680 1530 0.0124 60-100m P20 950 1520 0.0175 P35 950 1520 0.0175 P50 950 1520 0.0175 P100 950 1520 0.0175 Table 2: Example chirp start and end frequency and duration selected for different distance range and conduit size.

[0197] There may be a plurality of distance ranges and a subset of chirps may be associated with each distance range. In one specific example, four distance ranges are: 0- 2026201558   27 Feb 2026 10m; 10-30m; 30-60m; and 60-100m, although any desired ranges of distances may be used depending on the likely length and / or diameter of a conduit under test. At least one chirp may be emitted for each set of distance ranges. A plurality of chirps, e.g. a subset of chirps, may be emitted by an emitter for each set of distance ranges. For example, where five chirps are emitted per distance range, and there are four distance ranges, the total number of chirps would be 20. Each chirp is emitted by an emitter into the conduit and then received by a receiver from the conduit. The signals received can be considered as echoes of the signals emitted into the conduit. The signals associated with each chirp are processed separately and the data is combined and averaged. If a single chirp was used, interference or other factors could lead to an inaccurate or noisy signal. The use of a series of chirps may avoid issues associated with individual signals affected by interference by averaging out the data over a set of values.

[0198] The receiver sends the signals associated with received chirps to a processor. The processor may be part of the same device as the receiver, e.g. mobile device 100, or may be part of a separate device. In the present example, the main device comprises a processor. The main device may also comprise communication means, such as Wi-Fi or cellular connectivity or some other network access. The main device may communicate with a server. The main device may communicate the received signals to the server. According to examples, the main device pre-processes and / or processes the received signals. The server may preprocess and / or process the received signals. Both the main device and the server may perform exactly the same preprocessing and / or processing steps to ensure synchronisation between data at the main device and the server. The preprocessed signals may be displayed to a user in a plot 2000 on a display screen of the main device, e.g. as shown in Figure 17. Each received signal may be preprocessed in a desired manner to normalise the signal for display.

[0199] According to one example, the received signals are recorded in a 16-bit pulse code modulation format (PCM16). The signal is sampled at a suitable rate, in the present example the sampling rate is 48 kHz. The formatted signal is filtered. The filter may be a bandpass filter around a band of the chirp signal. In the present example, the first bandpass filter is a Butterworth filter centred on the emitted chirp frequency band with guard margins to suppress out-of-band noise. The Butterworth filter uses guard ±500 Hz; order=1. 2026201558   27 Feb 2026

[0200] Filtered signals may be cross-correlated with the emitted chirp using Fast Fourier Transform (FFT) based circular convolution. The convolution enhances echo responses and improves time-of-flight estimation. The resulting cross-correlation traces are normalised by signal energy to reduce sensitivity to absolute gain and coupling variability. Distance mapping may be performed using standard acoustic time-of-flight principles. The time-of-flight principles use a nominal speed of sound of 343 m / s and account for the two-way propagation path. Distance mapping may occur using Equation 5 herein, for example (343 / 2)x(number of samples / Fsample).

[0201] Additional smoothing and denoising steps may be applied. The smoothing and denoising may stabilise peak structure prior to peak detection. Optional Savitzky-Golay filtering may be applied. The preprocessing may be performed by a software on the main device. The software may be an app for a smartphone device. Where data is also transferred to a server or other external processor, this preprocessing pipeline is identical for data visualisation in the software of the main device and server or external processor. The server or external processor may be used for downstream machine-learning inference. Applying the same preprocessing may ensure consistency between outputs displayed to a user and results of machine learning feature classification.

[0202] Peaks are detected in the preprocessed signal. The peaks may be representative of a condition or feature within the conduit. Peak detection may be performed by the processor prior to using a trained model to classify features. One method of peak detection may use smoothed Z-Score peak segmentation with lag of about 100-250, threshold of about 4, influence of about 0.01, and peakMin of about 0.003. The segmentation may generate windows with tail for visibility. The windows may in include a visual indicator for display on a display device. The visual indicator may be a colour or colour coding. An example uses tooltips to report distance (m); autoscale adjusts the y-axis to 1.25xmax|y|.

[0203] Candidate reflections may be identified from the preprocessed and smoothed cross-correlation trace. This may use the scipy.signal.find_peaks algorithm. Peak detection parameters may be defined relative to local and global signal characteristics to robustly detect meaningful acoustic reflections while suppressing spurious responses. Characteristics of peaks may include prominence and minimum separation in distance.

[0204] Detected peaks are converted into distance-based windows and visualised on a display device. Figure 17 shows an example of a plot 2000 of a preprocessed signal. The plot 2026201558   27 Feb 2026 has an x-axis that reflects a determined distance along the conduit from the end at which the interface is engaged. The y-axis reflects the normalised amplitude of the signal received. The display screen may provide an interactive plot 2000 with which a user may interact. The user may rescale the amplitude and distance axes to improve readability. The display screen could be a touchscreen. The user interaction with the display screen may involve any one or more of taps, swipes, held (long) presses and pinches, or any other suitable touch gestures. For example the user may pinch in the x or y direction to adjust the scale. The user may also tap on a peak to select that peak.

[0205] In the plot of Figure 17, the region with a distance near to zero, e.g. close to the interface where signals are emitted and received, represents a signal transmission period 2010. The signal transmission period may be shaded in a first colour to highlight this area to a user. The parts of the signal that are determined to be peaks by the processor are shown as peak regions 2020. The user may use the shape of the plot and peaks displayed to assist in making an initial determination of conditions, features and / or objects in the conduit. For example: an area of flat signal may indicate that no obstructions or features are present and the conduit is clear in that area; a sharp or distinct peak may indicate that an object or feature is present in the conduit at a location of the peak; a rough or noisy signal may indicate that there is possible congestion and / or multiple objects in the conduit at a location.

[0206] The display screen may include one or more graphic control elements or buttons to enable a user to navigate a menu or options on the device. Before or after reviewing the plot on the display screen, the user may input observations about the conduit on the main device. The input may include a keypad, touchpad, keyboard or other suitable input. The displayed plot 2000 of Figure 17 includes a note icon 2040 which a user may select to input observations. Examples of observations to be input by a user include: the number of cables in the conduit; visible water, mud, congestion or objects in the conduit; conduit appears clear; and / or possible locations of bends in the pipe. Observations of the user may be input into the machine learning model to provide feedback and improve future accuracy.

[0207] The preprocessed or processed signal or the received signals in recorded format, e.g. PCM16, may be provided to a server or other device having a trained model. In other examples the main device includes the trained model. The trained model may be trained in a suitable manner, such as via the methods described herein. The signal may undergo peak detection in a similar manner as above to detect peaks. Peaks may be detected using the 2026201558   27 Feb 2026 scipy.signal.find_peaks function applied to the preprocessed cross-correlation signal. Detection parameters are defined in relative terms (e.g. prominence fractions, minimum separation in one-way distance, and maximum peaks per shot) to maintain robustness across varying conduit conditions and signal amplitudes. An adaptive thresholding mode may be applied, in which the smoothed cross-correlation is normalised using local magnitude statistics prior to peak detection. According to an example, peaks may be detected using Python scipy find_peaks and optimised using parameters threshold_frac and prominence_frac (relative to local / global scale), minimum separation (on one-way metres), and caps on max peaks per shot. An adaptive thresholding mode may normalise the smoothed cross-correlation by a bin-wise percentile of local magnitude.

[0208] The resulting peak centres may define candidate object locations and may be used consistently for distance-based window extraction, feature computation, and machinelearning inference. This approach ensures that object candidates detected by the trained model may align directly with those displayed in a plot on the display screen. The software may appropriately label the peaks or other features on the plot.

[0209] The signal is analysed by the trained model and features are correlated to peak locations on the plot. The plot including the trained model inferences may be displayed on the user display. This allows a user to review insights or inferences of the trained machine learning model. Figure 18 shows a plot 2500 with inferences from the trained model on a display screen. The device or display screen may include an input or button that allows a user to switch between a processed signal plot, for example as shown in Figure 17, and a plot showing inferences, for example as shown in Figure 18. The trained model may highlight peaks or other part of the signal that it has determined to be associated with a particular feature. Features highlighted by the model may include blockage, estimated percentage of pipe blocked, e.g. 25%, 50%, 75% or 100%, bend, opening, or any other feature that may be determined based on the training of the model. In Figure 18, the model has highlighted three predicted features 2510, 2520 and 2530. The model may provide a confidence score for each determined feature. The score may be a percentage or value between 0 and 1. For example, at a distance of about 7.8m in the example of Figure 18, the model has determined that the feature 2530 may be an opening and displays a confidence 2535 of 0.50. The user may use inputs or interactions with the display screen to select features or cycle from one highlighted feature to the next. 2026201558   27 Feb 2026

[0210] The model may be continuously trained in a feedback loop. Continuing to train the model may improve the accuracy of future predictions. Where a user does not agree with the prediction made by the model about a particular feature type they may insert feedback or a correction of the feature type. For example, Figure 19 shows the plot 2500 where a user is interacting to select from a menu 2560 a correction to a feature type. In this example, the model predicted that the feature was an opening but the user may select from a different feature if they believe the prediction is incorrect.

[0211] The data associated with the plots, inferences, locations of peaks or features, observations and / or corrections may be saved to a memory. The saved file may be uploaded to a server. The server may have a database of conduit inspections performed and / or may be provided into session folders. The data may be used to further train the model.

[0212] The predicted conditions and locations of peaks on the display screen may be used by the user to determine whether any action is required and / or what action to take. For example, a determination of a 100% blockage may require excavation at the detected location. Localising the determined feature or condition within the conduit may assist the user to take appropriate action, e.g. excavate, replace a pipe section, clear blockage, etc, in a specific location which may make the task more labour and time efficient than previous methods of manual rodding, etc.

[0213] Figure 20 shows an example of a training process 3000 associated with training a machine learning model according to the present disclosure. The process 3000 begins with receiving a raw acoustic data 3010. The raw acoustic data 3010 being a raw signal associated with a soundwave received from a conduit, as described herein. In the training process, the location and classification of a condition within the conduit are known and these relate to the raw acoustic data 3010. The raw acoustic data 3010 is preprocessed 3020 by a suitable means, for example using filtering and normalisation as provided herein. The preprocessed data / signal is analysed to detect windows 3030, which involves detection of peaks in the preprocessed data, each peak associated with a respective window. For each window or peak, the preprocessed data undergoes feature extraction 3040, for example using a CNN or MLP. Scalogram generation 3050 is also carried out on the preprocessed data. The scalograms may be input into a CNN. Each window may be labelled 3060, for example using the known classification and / or distance information associated with the known condition in the conduit. The training information, including extracted features, scalogram and / or labels, 2026201558   27 Feb 2026 may be saved into a storage, particularly the storage may store various types of classification of the condition within a conduit in respective folders 3070. For example, a folder may contain the training information associated with 25% blockage conditions, another folder may contain the training information associated with 50% blockages, and so on for all of the classification types input into the model. The datasets may be split 3080 and analysed by respective artificial neural networks. In one example herein, both feature based MLP and image based CNN are used and combined, although these are merely examples and in other embodiments other artificial neural networks could be employed.. The combination of the respective model outputs, e.g. MLP and CNN outputs, are evaluated and, based on the evaluation, a weighting 3100 is applied to the combination, e.g. fusion weights.

[0214] Figure 21 shows an example of an inference process 4000. Some of the initial steps of the process 4000 are similar to those in the training process 3000. For example, the process begins with a raw acoustic data 4010, which is associated with a soundwave received from a conduit. In the inference process 4000, the classification and distance of one or more conditions within the conduit are unknown before testing begins. The raw acoustic data 4010 is preprocessed 4020 by a suitable means, for example using filtering and normalisation as provided herein. The preprocessed signal is analysed to detect windows 4030 associated with peaks. For each window, the preprocessed data undergoes feature extraction 4040 and scalogram generation 4050. At this point, the extracted features and scalogram are fed into the trained machine learning model 4060, for example as trained via training process 3000. The machine learning model makes a prediction 4070 of a classification of a condition in the conduit associated with each window / peak. The location of the condition may be estimated 4080, for example using time-of-flight calculations as described herein. The location may be provided as a distance from the end of the conduit. In other examples, a distance scale may be provided on a plot of the preprocessed signal, and the distance scale may be used to determine the location of a condition. The results, classification prediction(s) and distance estimate(s) and any other data used, such as the preprocessed signal, scalograms etc, may be stored electronically, for example in a folder 4090. Data for display to a user may be sent to a display device 4100. The results may be displayed to the user of as a plot 4110, for example as shown in Figures 17 to 19.

[0215] While the invention has been described in conjunction with a limited number of embodiments, it will be appreciated by those skilled in the art that many alternative, modifications and variations in light of the foregoing description are possible. Accordingly, the 2026201558   27 Feb 2026 present invention is intended to embrace all such alternative, modifications and variations as may fall within the spirit and scope of the invention as disclosed.

[0216] Any reference to or discussion of any document, act or item of knowledge in this specification is included solely for the purpose of providing a context for the present invention. It is not suggested or represented that any of these matters or any combination thereof formed at the priority date part of the common general knowledge, or was known to be relevant to an attempt to solve any problem with which this specification is concerned.

[0217] For the avoidance of doubt, in this specification, the terms ‘comprises’, ‘comprising’, ‘includes’, ‘including’, or similar terms are intended to mean a non-exclusive inclusion, such that a method, system or apparatus that comprises a list of elements does not include those elements solely, but may well include other elements not listed. List of drawing references: 100 Mobile device 110   Display 120   Input 130   First end of device 140   Second end of device 200   Cradle 210   Body 220   End wall 230   Adjustor 232   Thread 234   Knob 240   Waveguide 242   Outlet 300   Test bed 310   Side of test bed 320   Protective material 330   Debris 350, 351, 352 Conduit 353   Straight section 354   First bend 2026201558   27 Feb 2026 355   Second bend 356   First end of conduit 357   Second end of conduit 360   Line 401   Synthetic 25% blockage 402   Synthetic 50% blockage 403   Synthetic 75% blockage 404   Synthetic 100% blockage 501   Raw acoustic wave 502   Data preprocessing 503   Preprocessed signal 504   Peak detection 505 Continuous wavelet transform (CWT) 506   Time-frequency scalogram 507   Convolutional neural network (CNN) 508   Transformer 510   Classification and localisation 600, 700, 800  Output plot 610, 710, 810  Signal 620, 720, 721, 820, 821, 822, 823, 824     Marker 630, 730, 830  Pipe representation 640, 740       Blockage representation 650, 750, 850  Scalogram 655, 755, 757, 765, 855, 856, 857, 858, 859       Scalogram region 725   Noisy signal section 840   First bend representation 842   Second bend representation 844   Third bend representation 1000  Device 1100  Body 1110  Backplate 1115  Screw 1120  Speaker aperture 1130  Microphone aperture 2026201558   27 Feb 2026 1140 Tapered section 1150 Neck 1160 Interior volume 1170 Charging aperture 1180 Alignment indicator 1190 Engagement surface 1200 Speaker 1250 Control actuators 1300 Microphone 1320 Strap 1350 Wire 1400 Handle 1410, 1420 Support 1450 Handgrip 1460 First opening 1470 Second opening 1500 Adapter 1520 Funnel section 1530 Extension 1540 Cut-out section 1560 Inner volume 1580 Alignment indicator 1590 Engagement surface 1600 Conduit end 2000 Plot 2010 Signal transmission period 2020 Peak regions 2040 Note icon 2500 Plot with inferences 2510, 2520, 2530 Predicted feature 2535 Displayed confidence of predicted feature 2560 Menu 3000 Training process 3010 Raw acoustic data 2026201558   27 Feb 2026 3020  Preprocessing 3030  Detect windows 3040  Extract features 3050  Generate scalogram 3060  Label window 3070  Store in class folders 3080  Split dataset 3090  Train model 3100  Trained weights 4000  Inference process 4010  Raw acoustic data 4020  Preprocessing 4030  Detect windows 4040  Extract features 4050  Generate scalogram 4060  Model inference 4070  Class prediction 4080  Distance estimation 4090  Store in test folder 4100  Send to display device 4110  Display results on plot

Claims

1. A device for use in assessing conditions in a conduit, the device comprising:an emitter configured to emit soundwaves;a receiver configured to receive soundwaves; andan interface comprising an inner volume acting as a wave guide for soundwaves from the emitter and soundwaves to the receiver, the interface configured to engage an open end of the conduit such that the inner volume is in communication with an interior of the conduit,wherein the device is configured such that, in use, the emitter emits a soundwave through the inner volume of the interface into the conduit, and the receiver receives a reflected soundwave from the conduit through the inner volume.

2. The device of claim 1, wherein the device comprises a body, and the emitter and receiver are housed within an interior of the body.

3. The device of claim 1 or 2, wherein the interface is formed integrally with the body.

4. The device of claim 1 or 2, wherein the interface is in the form of an adapter that is removably attachable to the body, and wherein the device comprises a plurality of interfaces, each interface configured to engage with an open end of a conduit of a different respective size than the other interfaces, and wherein the body is configured to removably attach to one interface at a time.

5. The device of claim 4, wherein the interface comprises first engagement surfaces and the body comprises second engagement surfaces, and wherein the first engagement surfaces and second engagement surfaces are configured to interact with one another, wherein the interface and the body are releasably lockable together via a relative rotation of the interface and the body when brought into contact with one another resulting in interactions between the first engagement surfaces and the second engagement surfaces.

6. The device of any one of the preceding claims, wherein the or each interface comprises an extension, the extension being configured for insertion into a conduit of internal diameter that is substantially the same as the external diameter of the extension.2026201558   27 Feb 20267. The device of claim 6, wherein the extension has an external diameter sized for insertion into a conduit of internal diameter of 20-100 mm.

8. The device of claim 6 or 7, wherein the device is configured to be held in position adjacent to the open end of the conduit via a friction engagement between an exterior surface of the extension and an internal surface of the conduit when the extension is inserted into the conduit.

9. The device of claim any one of claims 6 to 8, wherein the interface comprises a funnel section, the funnel section being attached at a first end to the extension, wherein the funnel section tapers from a first internal diameter at a second end to a second diameter at the first end.

10. The device of any one of claims 6 to 9, wherein the interface comprises a cut out region, the cut out region being located at least in the extension of the interface, wherein the cut out region is configured such that extension may be inserted into a conduit which contains at least one cable, wire or fibre.

11. The device of any one of the preceding claims, wherein the emitter is a speaker and the receiver is a microphone, wherein the emitter and the receiver are each configured to communicate with a master device respectively via a wired or wireless connection, and wherein the master device is located remotely to one or both of the emitter and the receiver.

12. The device of any one of the preceding claims, wherein the device comprises a handle, and wherein the device is portable and configured to be handheld prior to and after the interface engages a conduit.

13. A system for assessing conditions in a conduit, wherein the system comprises: the device of any one of the preceding claims;a trained machine learning model that receives data, the data comprising a raw signal from the receiver and / or the data comprising a preprocessed signal, processed signal, image or scalogram derived from the raw signal, wherein the trained machine learning model makes a classification prediction based on the data to classify at least one condition within the conduit.2026201558   27 Feb 202614. The system of claim 13, further comprising a localisation module configured to estimate a distance of the at least one condition from the end of the conduit at which the interface is engaged.

15. The system of claim 14, wherein the localisation module estimates the distance using at least the equation:d = (v x n) / (2 x Fsample)where d is a distance of the condition from the open end of the conduit, v is the speed of the wave, n is a sample index, and Fsample is the sampling rate of the receiver.

16. The system of any one of claims 13 to 15, further comprising a processor configured to process the raw signal received by the receiver into a preprocessed signal.

17. The system of claim 16, wherein the preprocessed signal is analysed to determine at least one window associated with a peak or peaks in the preprocessed signal, the at least one window corresponding to the at least one condition within the conduit.

18. The system of claim 16 or 17 when dependent from claim 14 or 15, further comprising a display screen configured to display the preprocessed signal, the at least one classified condition, and the estimated distance(s) to a user.

19. The system of claim 18, wherein the at least one classified condition is colour coded to represent its respective classification, wherein different classifications are represented in different colours.

20. The system of any one of claims 13 to 19, further comprising an input configured to receive inputs of the user, wherein the input is configured to receive a new classification of a condition from the user when the user disagrees with a predicted classification, and wherein the new classification of the condition is used to further refine the trained machine learning model.

21. The system of any one of claims 13 to 20, wherein the trained machine learning model comprises a convolutional neural network (CNN) and / or a Multi-Layer Perceptron (MLP) used as feature extractor(s) to extract features from the signal or preprocessed signal or an image or scalogram based on the preprocessed signal.2026201558   27 Feb 202622. The system of claim 13 or 21, wherein the trained machine learning model has been trained on known condition data to classify one or more of the following conditions within the conduit: a bend in the conduit; a shape and / or radius of the bend; an opening in a conduit; a second end of the conduit; a blockage in the conduit; a type of blockage; a size of the blockage and / or a percentage or proportion of the conduit diameter that is blocked; a damage to the conduit; a type of the damage to the conduit; a clear length of pipe free from blockage; and / or a liquid or water in the conduit.

23. A method of assessing conditions in a conduit, comprising:engaging an interface with an open end of a conduit, the interface having an inner volume acting as a waveguide;emitting by an emitter a soundwave through the inner volume of the interface into the conduit;receiving by a receiver a reflected soundwave from the conduit through the inner volume of the interface;transferring a data to a trained machine learning model, the data comprising a raw signal representative of the reflected wave from the receiver and / or the data comprising a preprocessed signal, processed signal, image or scalogram derived from the raw signal, wherein the trained machine learning model makes a classification prediction based on the data to classify at least one condition within the conduit; anddisplaying details or a representation of the classification of the at least one condition in the conduit to a user.

24. The method of claim 23, wherein the method comprises emitting a plurality of soundwaves, each soundwave being a chirp having a respective frequency and duration, wherein a plurality of reflected soundwaves representative of each chirp are received by the receiver, and wherein a processor processes each reflected soundwave separately to form processed data, and the processed data is combined and averaged, wherein a signal representative of the averaged processed data is displayed to the user.

25. The method of claim 24, wherein the frequency and duration of each chirp is selected based on at least one of a diameter of the conduit, and an estimated length of the conduit, and / orwherein subsets of the chirps have frequency and duration selected in respect of a distance range, each subset having a different frequency and duration than each other2026201558   27 Feb 2026subset, and each subset corresponding to a different distance range than each other subset.

Images