GB2585821A - Non-destructive testing method and apparatus - Google Patents

Abstract

A method of non-destructive testing of a target object, such as a train wheel, comprises: reconstructing a forward projection operator for the target object relating an excitation source to a target object response, the reconstruction comprising iteratively determining an objective function based on the forward projection operator, wherein the forward projection operator comprises separate forward projection operator components for two or more modes of measurement and also cross correlations between respective pairs of said two or more modes of measurement. Non-uniform features in the target object are determined using the objective function. The measurements may be: surface video image capture, surface video image gradient capture, surface geometry, ultrasound, magnetic field, impedance or capacitance. Components of the forward projection operator may be weighted according to the measurement mode and the correlations between the modes. The objective function may be minimized using a Newton-Raphson iterative solver.

Description

NON-DESTRUCTIVE TESTING METHOD AND APPARATUS

The present invention relates to a non-destructive testing method and apparatus. Previously, a number of non-destructive testing techniques have been used to probe target objects such as train wheels.

In some cases such separate techniques are complementary, and thus can be used in parallel to improve understanding of the state of the target object.

However, is still desirable to further improve this understanding, by forming a richer picture of material conditions of the target object and improving the detectability of localised features of it The present invention aims to address or mitigate this problem.

In a first aspect, a method of non-destructive testing of a target object is provided in accordance with claim 1.

In another aspect, a non-destructive testing apparatus for non-destructive testing of a target object is provided in accordance with claim9.

Further respective aspects and features of the invention are defined in the appended claims.

Embodiments of the present invention will now be described by way of example with reference to the accompanying drawings, in which: Figure 1 is a schematic diagram of non-destructive testing techniques.

Figure 2 is a schematic diagram of non-destructive testing sensors.

- Figure 3 is a schematic diagram illustrating inter-dependencies between measured data sets. Figure 4 is a photograph of a machined model aluminium wheel.

- Figure 5 is a schematic diagram of a map of the surface and sub-surface defects of the model wheel.

Figure 6 is a photograph of the model wheel in structured light.

- Figure 7 is a schematic diagram of an ultrasound response signal. Figure 8 is a schematic diagram of a plurality of recorded measurement data.

Figure 9 is a schematic diagram of SVD eigenvalues derived from a plurality of recorded measurement data.

Figure 10 is a flowchart of a method of non-destructive testing of a target object.

Figure 11 is a schematic diagram of a non-destructive testing apparatus.

A non-destructive testing method and apparatus are disclosed. In the following description, a number of specific details are presented in order to provide a thorough understanding of the embodiments of the present invention. It will be apparent, however, to a person skilled in the art that these specific details need not be employed to practice the present invention. Conversely, specific details known to the person skilled in the art are omitted for the purposes of clarity where appropriate.

Referring now to Figure 1, non-destructive testing (NDT) or non-destructive evaluation and sub-surface imaging methods are typically based on the re-construction of internal physical material properties of a target object from surface measurements. Typically the aim is to determine the mapping, forward projection or transformation function, 53, of the IS object, a, from measurements M on the object's boundary or surface, an, caused by a source excitation, S. In continuous mathematical form, this relation is represented as shown below and in Figure 1.

JvC(c712)=.73(Q) S(012) Figure 1 illustrates a foward model of a source excitation, S a target object ifs surface traversing through and along the object having a physical medium/material distribution P to reach measurement sensors M. The source can be located anywhere on the surface of SI, as can the measurement probes. Multiple sources and multiple probes can be active and/or recording data at the same time.

Typically a finite number of discrete sources are used to generate a finite number of discrete responses, and the continuous object space is thus approximated on a discretised numerical space.

With M being the vector of discrete measurements made (the effects), g being the source signal (the cause) and IP denoting the object's discretized forward projection operator. In general, IP is a function of material properties such as electric, magnetic, mechanic or acoustic transmission and reflection characteristics, and is dependent on object topology. In a few applications this is a linear operator, however, in most cases it will be non-linear. IP may even be time-dependent for objects whose properties change over time.

In this approach, IP is represented as a function of time-independent material density, (12): In non-destructive testing, the source signals and measurements are typically known on the object 1/ domain's surface 011 and 1P is the sought-after object property that is typically non-accessible through other means (e.g. the internal and occluded properties of objects).

To reconstruct 43, the inverse problem solution can utilise the application of the transpose 10 TT to equation (1) to obtain the normal equation relating the measurement vector ri to the source excitation vector §: To solve for material parameters, an objective function, 0, is formed as L2-norm representing the difference of the measured data, fi, and expected projection of the excitation source, §: Then minimisation techniques may be applied to this equation to find a local minimum for the sought material vector resulting in equation (5) below, which ultimately corresponds to equation (3) if invariant numerical techniques have been applied. 0 1,

This problem is typically solved numerically with an iterative Newton-Raphson approach or a linear conjugate gradient method that aim to minimise the difference between calculated and observed measurements with respect to iterative updates to the sought model parameters p(14) in The Newton Raphson method is highly dependent on a number of factors including sensitivity to the chosen initial starting vector of the material parameters, the closeness of local roots and thus the convergence rate and accuracy of the algorithm, and the ratio of Eigenvalues of the square matrix, i.e. the ill-conditioning of the problem.

With more data available, the sensitivity of the problem on numerical discretisation and error sources can be reduced. Preferably, an over-determined system with a least-square best-fit solution is sought. However, in in practice the problem is often under-determined as there are many more unknown material parameters than surface measurements and the projection P is usually considered non-linear. In addition, for real measurements, measurement noise has to be taken into account which may affect the stability of the algorithm.

Consequently to regularise the ill-conditioning of the problem, additional constraints can be imposed, such as for example the smoothness of the image in Electrical Impedance Tomography or the application of a generalised weighted Tikhonov regularisation.

Embodiments of the present invention seek to extend the objective function, and thus regularise and improve the reconstruction of material defects in a target object such as solid metal wheel model, with the application of visual, gradient, and geometry data and ultrasound data from electromagnetic-acoustic transducers (EMATs).

Advantageously this can lead to a better determinedness of the ill-conditioned problem and a more accurate complex model reconstruction. For the purposes of explanation, the embodiments and examples are presented in the non-limiting context of the nondestructive testing of railway wheels and therefore the experimental data will refer mostly to rotational wheel geometries, but it will be appreciated that the principles and techniques herein are not limited to this particular target object.

Referring now also to figure 2, in embodiments of the present invention a plurality of data sources may be considered to reconstruct the surface and subsurface material defects in a target object such as an aluminium wheel.

Accordingly a corresponding plurality of sensors may be combined into a single data capture sensor bar device, including but not limited to: * Video -providing surface data; * Video with gradient image -providing edge detection of surface features; * Ultrasound -providing sub-surface acoustic data; and * Structured light pattern -providing object geometry data.

Other examples of potentially suitable sensors include magnetic field measurements and/or electric field measurements (e.g. impedance and/or capacitance). The equations presented herein may use any plurality of such sensor types in any suitable combination.

It will also be appreciated that whilst a single data capture sensor bar device is convenient, it is not necessary and separate sensors may be used. For example, it may be appropriate for a video camera to be placed separately to an ultrasound sensor due to their conventional operational distances from a target object.

Figure 2 shows an example combination of sensor data components that allow the capture of surface, sub-surface and geometry data.

Each of the data sets on their own provides full characteristics relevant to the inspection and measurement/ gauging of the model wheel as a target object. However, in embodiments of the present invention, correlating these data points may provide a much richer picture than the individual elements on their own: Consider the following functional abbreviations where x, y, and z represent two-and three-dimensional (projected) data capture respectively at time t: V (x,y; t) Video capturing a 2D projection of a 3D surface at time t D(x,y; t) Delta/Gradient surface information 2D projection at time t U(x,y,z; t) Ultrasound reflection in 3D material at time t, captured for example with

EMAT

L(5;d; t) Structured light source, origin (3, direction cl of pattern at time t G(x,y,z; t) Geometry -structured light reconstructed 3D model shape & geometry As can be seen, optionally a time variable may be included when there is a time varying property to the target object such as when a wheel will be rotating around its central axis while the data is captured. However, as continuous data is captured, it is possible to project this into a static overall image, in particular for example where the speed of ultrasound in the material (6320 m/s in the case of the wheel) is much bigger compared to the rotational speed of the wheel (3 m/s).

Interdependency Referring now to Figure 3, in embodiments of the present invention, two or more of these data streams are combined into richer dataset as follows.

Using an interdependency approach for models relating to each data stream, is possible to deduce an algorithm to extract and take into account the missing parameters for each function/model from one or multiple of the others as shown in Figure 3, where: V -> projects image information onto -> G D -> projects differential image gradient information onto G U -> internal defect/feature reconstruction -> correlates with D, V G -> utilises L, and feeds geometry/boundary information into -. U Figure 3 illustrates inter-dependencies between measured data sets to reconstruct surface and sub-surface properties of the object. In the Figure, indicates correlation effects combining data streams into subsequent analysis stages.

This interdependency approach provides a sequential processing of captured data into the overall object representation. Geometry extraction leads to a better informed ultrasound boundary capture and video correlation. However, potentially noise or errors on the data will also be incorporated into the process and may lead to slightly higher error margins.

Mathematically, this interdependency approach results in a similar correlating matrix representation to extending the objective function % from equation (4) to take into account the additionally available constraints and data obtained from supplementary data streams/sensor types, which leads to a more regularised form of the objective reconstruction as shown below: Hence alternatively, or in addition (for example in the case of crosschecking) to the interdependency approach above, the objective function 0 may be extended as follows.

Extended objective function Ft/ and C' are the measured and source data points respectively and A, represent possible different regularisation weighting factors towards the overall objective. These can be chosen to allow for different weighting of the sensor data towards the overall objective; for example, if a higher accuracy video data is available compared to a differential gradient data contribution, then optionally the Av factor could be chosen to be higher than AD.

Taking the gradient to find a local minimum then leads to: (7) By way of example, in this equation Sv is the source for the video detection (e.g. the visible surface properties of the object). This may include changes in colour, shading, (caused by scratches, oxidisation, deformation, surface roughness, etc). Hence 1Pv is the mapping of the 3D object's surface appearance to the 2D video frames. This can be measured against expectation (for new wheels) or against an average across the wheel (for used wheels).

SG is the source for the geometry (which is typically the expected geometry of -in this case -the wheel shape). is therefore the mapping of the expected real 3D geometry (roundness, thickness, width, height) to the measured geometry in the captured 2D frames.

SD is the gradient information from edge detection. This can take the form of a refined or processed sub-set of Sv and may be included as it may contain more information about localised defects after gradient processing, in particular directional changes (e,g, along the wheel or across or diagonal).

Similarly, Su may the source excitation of the ultrasound, and Tumay be the mapping to the measured ultrasound.

The transformation matrices W, are not independent from each other as they rely on the physical structure of the object itself. Thus re-writing the problem into a linearized vector-matrix model, an over-determined system can be achieved where matrix elements zpo represent correlations between parameter indices a and 13 E{V,G,U,D}.

Hence whilst Eq. 7 may be considered as a standard objective function, where the assumption is that the parameters are uncorrelated, by contrast Eq.8 is an extended objective function, where correlations between the different parameters are taken into account in the numerical solution through the cross-correlation factors Wad.

Using equation (8) in the Newton-Raphson gradient approach, and abbreviating AaMa = Ma' obtains: and thus: Performing a transposition of the matrices, the equation yields: Evaluating the square system matrix before inversion, the correlation functions between the plurality of sensor parameters (in this example, four) and their contributions to the overall inverse problem become apparent: 1p,g are functions representing the product moment correlation between parameters a and /3 and their properties include operator symmetry tpap=tppa.

In embodiments of the present invention, where two moments exhibit a high correlation (for example between video and ultrasound data) it is inferred that there is a high probability of a feature in the target object such as a surface or sub-surface (deeper) defect, whereas if there is a large diagonal matrix entry with low side correlation, it is inferred that there can either be a sub-surface defect without apparent damage to the surface or a surface feature such as paint without the equivalent ultrasonic depth response.

Thus a system matrix has been created that adds substantial information to the inverse problem through accounting for effects of multiple collected sensor data points and the inter-dependencies between them.

The solution to equation (8) can thus represent / correspond to the difference between expected model parameters and measured parameters. It can give a measure of difference that can then be presented as numerical value or image to the operator of the device, who can then determine whether to classify the difference(s) as a defect.

Alternatively or in addition to user classification, automatic classification (e.g. from a dictionary of differences/defects) can be performed, for example using an Al algorithm or any other suitable technique for fuzzy classification.

From the interdependency approach, it is known that the geometry G is a function of the object's topology, 11, and hence so is its material parameter distribution, (14 This can be written as: Expression 9 assumes the target object is surrounded by air and has sufficiently different material properties that the boundary is effectively binary. Other boundary conditions may be assumed for cases where the target object is surrounded by a gas or liquid with more comparable material properties, such as water or oil.

Similarly, the ultrasound response is a function of geometry (and thus the material distribution). Again, a binary distribution of material is assumed for a target object in air, but again may vary for different conditions.

Applying an ultrasound B-Scan algorithm on the shear wave front reflections, the nearest boundaries where the ultrasound wave reflects back from can be determined using a time-of-flight determination. Due to the complex wave front propagation (due to multiple internal reflections and wave dispersion) optionally only the first two major reflections of minimum distances travelled may be analysed, with others being discarded from the reconstruction to allow for a clearer image of internal and external material boundaries: The surface gradient, D, was used to detect sharp gradient differences in surface properties through lighting and pattern differences, leading to a more detailed detection of cracks or indentations. However, should there be discolouration on the surface without it affecting the geometry or topology of the wheel, this represented a potential source with the capability to introduce spurious errors that do not correspond to a real geometric defect. In these circumstances, the determination of the weighting factor for surface data, AG should preferably be carefully selected or reduced. This selection may be empirically based, for example on large-scale experiential experiments.

Alternatively one may assume to a first approximation that gradients correspond to geometric or defect effects and thus do not reduce the contribution of D to the objective function: It will be appreciated that advantages of expanding on the original reconstruction algorithm with additional sensor measurement data as presented above include: Firstly, by using actual data for regularisation of the ill-conditioned problem, there is less of a need to impose an artificial normalisation or smoothing term that would result in loss of information and/or detail.

Secondly, recording/streaming additional sensor data in parallel results in the same data capture time as for a single data entity.

Thirdly, a richer image of surface and sub-surface data can be obtained. Worked example A worked example of the methods and techniques described herein is now described with reference to Figure 4, which shows a machined model aluminium wheel to which a variety of surface and subsurface features had been introduced.

The material used in the example target object is a commercial grade Aluminium Alloy 6082 T6 cylinder of 15" diameter (381 mm diameter), a cylinder-thickness of 100 mm and material hardness of 95 HB.

A variable shape V-shaped groove was machined into the surface of the cylinder, enabling detection tests of shutter speed, lighting conditions and framerate in the video recordings to resolve fine details with the video camera while the wheel was spinning. Also, specific profile faults have been added, such as an area of 0.8 mm removed surface material, corresponding to a developing wheel flat area in its early stages, an area with pitting defects and two gouges on the rounded rim edge. Furthermore, scratched lines in the spin direction and normal to this were added on the surface with depths up to 0.5 mm as well as a range of controlled drilled circular pits that range from 1.5-4 mm in diameter and at three distinct depth levels of 0.5, 1.5, and 3 mm.

In addition to these surface faults, three subsurface defects were added: Two small 2.5 mm diameter holes on either side of the model and a deeper 6mm diameter hole drilled from within the side overhang gap.

The depth of the side profile gap has been varied to four different levels to allow for the detection of depth differences. These have been kept deliberately close within a range of 5 mm.

Accordingly, Figure 5 provides a map of the resulting surface and sub-surface defects of the model wheel, combining machined and innate defects.

Video and gradient capture In the worked example, surface video data was recorded using an 8 megapixel image sensor, but other resolutions may be considered. To obtain a clear image at wheel surface speeds up to 3 m/s, the framerate was set to 60 fps (frames-per-second) and the shutter speed to 400 us, although again this may vary for different lighting conditions, tagets, sensors, and optics. In the worked example, the optics were adjusted to focus to a clear image at a fixed distance of 14 cm (in straight-on view) of the model wheel. Diffuse sunlight was sufficient to illuminate the aluminium to an acceptable level even at this high shutter speed. At times when it was dark or cloudy, a controlled 3W focussed LED lamp was used instead.

The video streams were recorded at a resolution of 1280 x 720 pixels to cover the full width of the wheel (10 cm). Thus a theoretical detection resolution of features as small as 80 pm could be achieved: In practice, the visible camera frame covered a larger area than just the wheel thickness as it was not fully centred and thus required more than the 10 cm width to be recorded. This translated to a detection sensitivity of about 100 um/pixel: p The same resolution and framerate was achieved for the gradient capture when processing the captured video stream through an embossing gradient algorithm such as that provided by the in-built Broadcom graphics chipset of a Raspberry Pi (1) 3 Model B Version 2.

Despite using such off-the shelf components, it is possible to carry out data capture and recording at high speeds, resolution and accuracy. Features were resolved that could not be detected with the naked eye -this was in particular true when the wheel was spinning at up to 3 m/s surface speed.

The recorded video streams (for example in H264-format) were subsequently converted to individual frame images to be stitched together to form a single long picture of the wheel's surface using the ffmpeg libraries.

Such a process can result in in a striped image, depending on lighting conditions. Optionally this can be addressed by normalisation of light gradients across each image frame, although this can lead to loss of contrast.

Geometry capture Video data was captured using a second camera module, capturing a portion of the rotating model illuminated by structured light, as shown in photograph of figure 6.

A number of geometric patterns were trialled, and it was found that a parallel line projection with an offset angle of about 60 degrees between projection and camera plane resulted in particularly good geometry reconstruction for in-time photogrammetric analysis of the patterns on a turning wheel (at a speed of up to 3 m/s) with a practical processor.

Using structured light, the resolution of the resulting geometry capture is itself dependent on the pattern density/resolution. In the case of the interdependency/extended objective function approaches, the main purpose is to achieve a general course geometric representation to be used for ultrasound wave reflection calculations and potential deviations against set limits (for example, through where a lot of materials). Accordingly, a pattern density/resolution of similar scale to the desired resolution of the course geometric representation is appropriate. Hence for example a light/dark banding repeating every 5, 10, 15 or 20 mm is likely to be appropriate, although for larger scale structures then a proportionately larger scale pattern may be considered.

Ultrasound capture A plurality of electro-magnetic acoustic transducers (EMATs) were provided (as a non-limiting example, four EMATs made with high-performance 6 mm diameter, 50 mm long cylindrical N42 Neodymium magnet cores with 1.6 kg pull force, surrounded with 1000 coil windings of 0.15 gauge copper wire with 0.005 gauge insulating polyester sheathing to the height of 20 mm). The EMATs preferably achieve high detection sensitivity of shear waves in the material and allow for short-pulse high current activation.

For the non-limiting EMATs describe above, non-magnetic end caps were applied to ensure a constant distance ('air gap') from the object. The EMAT sensors were then potted in a non-magnetic block to stabilise their positions. The installed EMAT permanent magnetic field configuration is in alternating N-S-N-S order. These example transducers were found to have an optimal resonance operating frequency of 2 MHz.

In the example apparatus, driving power was supplied from a dedicated function generator at 20Vpeak and a controllable current up to 2A, corresponding to a maximum peak power delivery of 40W. The transmitter coil was driven with a short single waveform pulse. This created a strong enough signal to be picked up by the same type coils used in receiver mode over a distance up to 7.8 cm away from the transmitter. Beyond this distance, the signal-to-noise ratio decreased to -11 dBm (,--8%) and thus became comparable to the noise floor of the prototype equipment.

Typically in larger-scale EMAT applications, such as testing steel pipes used in the gas and water industries, power requirements of over 1 kW are not uncommon for the transmitter to address the low transmission and detection sensitivity due to the air gap. However, embodiments of the present invention are directed to localised shallow sub-surface defects in the target object, furthermore in combination with other sensor data. Consequently the reflection data analysis should take into account the short travel times of wave fronts in the region of micro-seconds which would overlap a multi-burst actuation signal that would have been longer than the peak echo receive time.

The speed of sound in aluminium is approximately 6320 m/s, and is of a similar order to Steel. In the example apparatus, an NI Elvis II+ 100 MegaSamples per second (MS/s) data acquisition interface board was used to record the response data. The 4 channels were sampled concurrently at a maximum 25 MS/s which corresponded to a maximum spatial resolution of 6320 m/s / 25MS/s = 0.25 mm wave propagation per sample point. At 2 MHz wave frequency a single wave period was thus sampled over 8 points which satisfied the Nyquist criteria for analogue to digital conversion of the wave signal. After two samples a short 14.8 psec recovery time was required by the equipment.

For short distances travelled in high density materials, reflected wave depth extraction was carried out through detection of echo response hull peak signal analysis (PSA). Figure 7 shows the drawn-out immediate response signal for very short travel times of the sample EMAT transducer operated in single waveform burst activation mode across a 5 mm thick metal sample plate.

In figure 7, the peak of the yellow response Gaussian hull occurs at 1.68 p,s (after reflection). A much lower-amplitude elongated secondary Gaussian response is overlayed with peak at about 3.20 as, which corresponds to twice the internal reflection path length as expected for a higher order reflection.

Analysing the response for each of the EMAT receivers in the matrix arrangement along the rotating wheel surface results in depth and distance maps for boundary surface reflections. As the dampening effects of the airgap above the material can be significant, the raw captured data may be processed through a hull peak detection algorithm. The result may then be subjected to a threshold level filter to normalise for visualisation and then assigned for example to a 64 depth-level logarithmic pixel map.

Due to multiple internal reflections of the produced shear-wavefront optionally only the first one or two clear peak-signals may be analysed, with further weaker signals being discarded from the calculations.

Ultrasound reflections from opposite material surfaces, side-located material surfaces within the propagation cone and abnormal crack or hollow subsurface features may thus be represented by the detected peaks as distances of material boundary from the EMAT actuation source.

For other sensor captures, such as magnetic field capture and electrical field capture (e.g. for impedance and/or capacitance) then corresponding capture sensors (e.g. one or more magnetic field sensors, and/or one or more electric field sensors) may be employed, together with any suitable associated pre-processing if required.

Measurements Figure 8 shows the data recorded for the example target object and apparatus, processed into processed into individual data streams across the full circumference of the wheel.

Figure 8A shows a complete and continuous version of the defect map of claim 5, for reference. Figure 8B shows the captured video image (the stripes are an artefact of stitching the images together for illustration, and are not part of the captured image). 30 Figure 8C is a corresponding gradient image. Figure 8D shows smoothed geometry / topology information, in the example case showing off-axis wheel position and variation in groove width. Finally, Figure 8D shows a reflection map for the first two peak reflection points of ultrasound wave for EMAT sensor 1 (for location see Figure 8b).

It can be noted that in the ultrasound level graph, depth or distances of 7 and 14-15 mm respectively were recorded along the whole wheel profile (where no superior peak signal was received). These correspond to the distance of the outer boundary of the wheel at about 10 mm from the centre of the receiving EMAT and a 12 mm depth of the overhanging Aluminium substrate, respectively. On approaching sub-surface defects with the sensors, a typical circumflex 'A' reflection pattern can be observed.

In embodiments of the present invention, to address the problem of varying wheel speed during a rotation, optionally all data points can be mapped on to a discrete millimetre grid, for example of size 100 width x 1197 length x 64 depth (=2^7), corresponding to the measurements of the test wheel, and hence more generally a grid corresponding to the measurements of the test object. In addition a dual point synchronisation and Hall-effect speed sensor may be added to the data capture apparatus to allow for automated wheel speed determination at the start and end of measurements.

Analysis During data capture and after analysis the following observations were made: * Video data (14): A high-framerate (60 fps) high-resolution video of the rotating model wheel is recorded in real-time, allowing a detection of sub-mm surface defects.

Defects smaller than 0.5 mm can be observed on the.h264 compressed video as well as on individual frames auto-stitched together to obtain the full circumferential image. A pixel resolution of 100 pm is achievable.

Analysis resulted in three types of detectable features: o Shallow surface defects (scratches, pitting, depth < 0.5 mm) o Surface defects (deep scratches, pitting, depth > 0.5 mm) o Discolouration (paint, dirt, shiny/reflective surface areas, etc) Gradient data (ft): High framerate (60 fps) high-resolution video of the rotating model wheel is recorded (optionally corresponding to the video data above) and converted through a gradient algorithm.

- With specific side-angle lighting conditions, deeper surface defects may be more easily observable.

- Discolouration of the surface (paint, dirt, reflective surfaces) led to feature detection by the gradient algorithm. This can either be considered as useful data or a false positive when looking for material defect, depending on the purpose of the nondestructive testing.

Geometry (1. ): - Multiline projected geometry reconstruction may be influenced by surface features that do not correspond to actual (large-scale) geometric differences. Feature shading, discolouration and surface features influence this.

The location of target object edges, such as the rounded and chamfered edge, may be usefully extracted and thus a measurement of width of the target object as well as the position of persistent features such as the wheel groove in the example may be identified within the overall surface geometry.

This measurement mode can also result in the unexpected detection of an off-round, off-axis mounting that is not otherwise observable with the naked eye (as seen in the variability of the geometry in Figure 8D).

Ultrasound (7.): High speed data capture using one transducer as source and three as receivers allowed for a detailed capture of ultrasonic characteristics of the material. it will be appreciated however, that one, two or more receivers may be used.

Subsurface features (holes drilled from the sides) were detectable to a distance of about 7.5 cm before the SNR became too small for analysis in the example apparatus, although in principle a higher power source, or other variables, could be changed to increase this range. Optionally the single-shot transducer response data may be processed with a peak-detection algorithm to mitigate the short distances involved.

Defects smaller than 1 mm were not detectable with ultrasound alone when located on the surface of the model wheel, potentially due to the air gap and the small feature size and distance. Again in principle a higher power source, or other variables, could be changed to improve this; however it will be appreciated that the subsequent combination of data sources, as discussed previously herein, serves to mitigate this issue as well.

Analysed wave reflections in the example corresponded to internal and external boundaries that could be reconstructed into several possible locations. This was due to the omni-directional nature of the acoustic wave in the material. Unique distance and thus shape reconstructions from ultrasound require the knowledge of the outside boundaries of the material to enable predictive reconstructions.

In summary, the example apparatus enable the discernible detection of surface and sub-surface features in combined sensor modes, such as pitting (in video, gradient, ultrasound), surface scratches (in video, gradient), depth defects (in ultrasound, geometry), flats and paint or dirt covering (in video, gradient, geometry).

Newton-Raphson iterative convergence As noted previously herein, the success of an iterative solution to model the target object properties depends on how ill-conditioned the problem is. Embodiments of the present invention seek to improve this by combining data from multiple different and complementary sources, as described above for pitting, surface scratches, depth defects, flats or coatings, and the like.

Turning now to Figure 9, this this shows singular-value-decomposition eigenvalues for sample 4 modal system matrices on single depth layer. In particular, it shows the singular value decomposition (SVD) eigenvalues of the extended square matrix functions limp used in the Newton-Raphson iterative solver.

This technique is comparable with principal component analysis (PCA) to extract the most significant image features and their corresponding eigenvectors to reduce dimensionality through orthogonal subspace transforms. Typically in physical NDT reconstructions, about a third to half of eigenvalues are significant and therefore determine the local convergence characteristics of the algorithm. However, by using the techniques disclosed herein, about two thirds of eigenvalues for individual data sets were significant for single mode data capture.

Hence in Figure 9, while the two standard single data SVDs.ipu and ipv (B and C) show the strong ill-conditioning through high-to-low SVD ratios, the combined sensor SVDs Tuctp. and IPTuctPvp (A and D) show a much improved conditioning ratio, resulting in better reconstruction convergence.

It will be appreciated that embodiments of the present invention increases the number of significant eigenvalues in the SVD, and thus improves the conditioning of the problem, by expanding the matrix size through inclusion of additional sensor data of different kinds.

This approach allows the inclusion of additional information while the actual number of reconstructed material voxels (7.) remains constant.

Advantageously this leads to better convergence of the reconstruction algorithm.

Hence it will be appreciated that the above techniques serve to reduce the ill-conditioning of the inverse problem in non-destructive material testing through the use of multiple physical parameter measurements.

In particular, multi-modal sensor data has been combined to achieve a better quality reconstructed material distribution and thus a better defect detection from a large multi-modal data set.

Measurements have been combined from multiple sensors for example including video data capture, image gradient analysis, geometry capture through projected line analysis and ultrasound testing through a small EMAT sensor array. This is considered suitable for data capture in applications such as non-destructive testing of target objects such as for example a railway wheel.

By using the captured data to over-determine the material parameters where possible, the techniques described herein show a better convergence to the actual model parameters through a larger number of significant singular values contributing to the iterative 25 algorithm.

In addition, a richer picture of the surface and subsurface condition has been obtained compared to a single sensor mode reconstruction which ultimately leads to a richer analysis of target object defects.

Implementations of these techniques may use current off-the-shelf hardware, in combination with bespoke software implementing the enclosed techniques, to enable the build of a system that is of compact size, which makes it suitable for automation applications (for example to be moved by a robot arm) while providing enough computing power for a fast data capture for non-destructive testing applications.

Turning now to Figure 10, in a summary embodiment of the present invention, a method of non-destructive testing of a target object comprises: In a first step s1010, reconstructing a forward projection operator II' for the target object, relating an excitation source to a target object response, the reconstruction in turn comprising: In a second step s1020, iteratively determining an objective function based on the forward projection operator, wherein the forward projection operator comprises separate forward projection operator components for two or more modes of measurement and also cross correlations between respective pairs of said two or more modes of measurement; and In a third step s1030, identifying one or non-uniform features in the target object responsive to the determined objective function.

It will be apparent to a person skilled in the art that variations in the above method corresponding to operation of the various embodiments of the method and/or apparatus as described and claimed herein are considered within the scope of the present disclosure, including but not limited to: - the forward projection operator components for respective modes of measurement are respectively weighted by a factor responsive to both the forward projection operator for the respective mode of measurement, and the cross correlation between that mode of measurement and one or other modes of measurement; -the objective function 0 being where M is measured data, S" is the excitation source, and T is the forward projection operator, and where the gradient of the objective function is responsive to respective weighting factors /la of the components Pa of the forward projection operator for respective modes of measurement; the weighting factors for forward projection operator components for respective modes of measurement being respectively weighted by a factor Ac, wherein o though where not all these modes, or different modes, are used then equivalent matrices would be used; the two or more modes of measurement comprising at least one mode of surface measurement, and at least one mode of sub-surface measurement; the modes of measurement comprise one or more selected from the list consisting of surface video image capture; surface video image gradient capture; surface geometry capture; and ultrasound capture; and which the step of at least locally minimising an objective function comprising using a Newton-Raphson iterative solver.

It will be appreciated that the above methods may be carried out on conventional hardware suitably adapted as applicable by software instruction or by the inclusion or substitution of dedicated hardware.

Thus the required adaptation to existing parts of a conventional equivalent device may be implemented in the form of a computer program product comprising processor implementable instructions stored on a non-transitory machine-readable medium such as a floppy disk, optical disk, hard disk, solid state disk, PROM, RAM, flash memory or any combination of these or other storage media, or realised in hardware as an ASIC (application specific integrated circuit) or an FPGA (field programmable gate array) or other configurable circuit suitable to use in adapting the conventional equivalent device. Separately, such a computer program may be transmitted via data signals on a network such as an Ethernet, a wireless network, the Internet, or any combination of these or other networks.

Hence in another summary embodiment of the present invention, a non-destructive testing apparatus 1100 for non-destructive testing of a target object comprises two or more measurement devices (1112, 1114, 1116) respectively providing distinct modes of measurement; a processor (1130) adapted (e.g. under suitable software instruction) to reconstruct a forward projection operator 111 for the target object, relating an excitation source to a target object response, the processor being adapted (e.g. under suitable software instruction) to iteratively determine an objective function based on the forward projection operator, the forward projection operator comprising separate forward projection operator components for two or more modes of measurement and also cross correlations between respective pairs of said two or more modes of measurement, the processor being adapted (e.g. under suitable software instruction) to identify one or nonuniform features in the target object responsive to the determined objective function; and a user interface (1140) operable to identify the one or non-uniform features and the target object to a user.

In an instance of this summary embodiment, the two or more measurement devices (1112, 1114, 1116) may comprise at least one device for surface measurement, and at least one device for sub-surface measurement. Optionally, the output of any of these devices may be provided to an appropriate pre-processor 1120, for example to provide image processing. It will appreciated that such a pre-processor may be distinct from the main processor 1130, or may be the main processor also operating under suitable software instruction as the pre-processor.

In an instance of this summary embodiment, the one or more measurement devices may be selected from the list consisting of a video capture device; an image processor operable to generate gradient data from captured video images; a structured light projector and video capture device; and one or more electromagnetic acoustic transducers. It will be appreciated therefore that the pre-processor 1120 may act as a secondary measurement device where it extracts additional distinct information from a primary sensor.

In an instance of this summary embodiment, the forward projection operator components for respective modes of measurement may be respectively weighted by a factor responsive to both the forward projection operator for the respective mode of measurement, and the cross correlation between that mode of measurement and one or other modes of measurement.

In an instance of this summary embodiment, the objective function 0 may be 2.3 where ri is measured data, S is the excitation source, and IP is the forward projection operator, and where the gradient of the objective function is responsive to respective weighting factors Acr of the components Wa of the forward projection operator for respective modes of measurement Further, in an instance of this summary embodiment, the weighting factors for forward projection operator components for respective modes of measurement may be respectively weighted by a factor Aa wherein The foregoing discussion discloses and describes merely exemplary embodiments of the present invention. As will be understood by those skilled in the art, the present invention may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. Accordingly, the disclosure of the present invention is intended to be illustrative, but not limiting of the scope of the invention, as well as other claims. The disclosure, including any readily discernible variants of the teachings herein, defines, in part, the scope of the foregoing claim terminology such that no inventive subject matter is dedicated to the public.

Claims (14)

1. CLAIMS1. A method of non-destructive testing of a target object comprises: reconstruction of a forward projection operator 1P for the target object, relating an excitation source to a target object response, the reconstruction in turn comprising: iteratively determining an objective function based on the forward projection operator, wherein the forward projection operator comprising separate forward projection operator components for two or more modes of measurement and also cross correlations between 10 respective pairs of said two or more modes of measurement; and identifying one or non-uniform features in the target object responsive to the determined objective function.
2. 2. The method of claim 1, in which forward projection operator components for respective modes of measurement are respectively weighted by a factor responsive to both the forward projection operator for the respective mode of measurement, and the cross correlation between that mode of measurement and one or other modes of measurement.
3. 3. The method of claim 1 or claim 2, in which the objective function 0 is where IV/ is measured data, S is the excitation source, and 'V is the forward projection operator, and where the gradient of the objective function is responsive to respective weighting factors 2a of the components W of the forward projection operator for respective modes of measurement.
4. 4. The method of any one of claims 1 to 3, in which the weighting factors for forward projection operator components for respective modes of measurement are respectively weighted by a factor.la wherein
5. 5. The method of any one of the preceding claims, in which the two or more modes of measurement comprise at least one mode of surface measurement, and at least one mode of sub-surface measurement.
6. 6. The method of any one of the preceding claims, in which the modes of measurement comprise one or more selected from the list consisting of: i. surface video image capture; ii. surface video image gradient capture; iii. surface geometry capture; and iv. ultrasound capture;v. magnetic field capture;vi. impedance capture; and vii. capacitance capture.
7. 7. The method of any one of the preceding claims, in which the step of at least locally minimising an objective function comprising using a Newton-Raphson iterative solver.
8. 8. A computer program comprising computer executable instructions adapted to cause a computer system to perform the method of any one of the preceding method claims.
9. 9. A non-destructive testing apparatus for non-destructive testing of a target object, comprising; two or more measurement devices respectively providing distinct modes of measurement; Apo :PDS WDU a processor adapted to reconstruct a forward projection operator '1-1 for the target object, relating an excitation source to a target object response, the processor being adapted to iteratively determine an objective function based on the forward projection operator, the forward projection operator comprising separate forward projection operator components for two or more modes of measurement and also cross correlations between respective pairs of said two or more modes of measurement, the processor being adapted to identify one or non-uniform features in the target object responsive to the determined objective function; and a user interface operable to identify the one or non-uniform features and the target object to a user.
10. 10. The nondestructive testing apparatus of claim 9, in which the two or more measurement devices comprise at least one device for surface measurement, and at least one device for sub-surface measurement
11. 11. The nondestructive testing apparatus of claim 9 or claim 10, comprising one or more measurement devices selected from the list consisting of: i. a video capture device; ii. an image processor operable to generate gradient data from captured video images; iii. a structured light projector and video capture device; iv. one or more electromagnetic acoustic transducers;v. a magnetic field sensor; andvi. an electric field sensor.
12. 12. The nondestructive testing apparatus of any one of claims 9 to 11, in which forward projection operator components for respective modes of measurement are respectively weighted by a factor responsive to both the forward projection operator for the respective mode of measurement, and the cross correlation between that mode of measurement and one or other modes of measurement
13. 13. The nondestructive testing apparatus of any one of claims 9 to 12, in which the objective function 4) is where c/ is measured data, g' is the excitation source, and 'V is the forward projection operator, and where the gradient of the objective function is responsive to respective weighting factors Aa of the components IP, of the forward projection operator for respective modes of measurement.
14. 14. The nondestructive testing apparatus of any one of claims 9 to 13, in which the weighting factors for forward projection operator components for respective modes of measurement are respectively weighted by a factor Aa wherein