A flow measurement probe

Abstract

A probe for measuring fluid velocity comprises a sting portion and a body. The sting portion comprises three or more holes for pressure sensing. The body includes a respective pressure sensor associated with each of the three or more holes. Each pressure sensor is configured to generate a respective sensor output. The body further includes a data processing system comprising: computer-readable storage storing parameters which define an artificial neural network, wherein the artificial neural network is trained for use in measuring fluid velocity, and one or more processors for generating a fluid velocity measurement output based on the output of the artificial neural network. Generating the fluid velocity measurement output comprises: providing an input to the artificial neural network, wherein the input is derived from at least the sensor outputs, and processing the input with the artificial neural network to produce an output of the artificial neural network.

Description

[0001] A Flow Measurement Probe

[0002] Field

[0003] This disclosure relates to a flow measurement probe. In particular, it relates to a probe for measuring fluid velocity.

[0004] Background

[0005] Known multi-hole probes produce pressure sensor output data which may be used to determine the velocity of a fluid flow (e.g. a flow of air or water) at the probe tip. This velocity measurement is performed using a calibration grid, which may be obtained in a wind tunnel or open jet experiment.

[0006] In known approaches, the calibration grid is stored in the computer memory of a computing system external to the probe, which performs the calculation of the velocity vector. For a calibration grid of -44°...+44° in both pitch and yaw with steps of 2.2°, about 60516 bytes of memory is required to store the required data. In practice, the calibration grid is typically resampled (interpolated) onto a much finer grid to reduce the step size (or the calibration itself done on a finer grid), such that more memory is typically needed. That is, the requirement of 60516 bytes for a calibration grid represents a lower bound.

[0007] Summary

[0008] This specification describes a probe for measuring fluid velocity, comprising a sting portion and a body. The sting portion comprises three or more holes for pressure sensing. The body includes a respective pressure sensor for each of the three or more holes. Each pressure sensor is configured to generate a respective sensor output. The body also includes a data processing system comprising computer-readable storage storing parameters which define an artificial neural network, wherein the artificial neural network is trained for use in measuring fluid velocity. The data processing apparatus further comprises one or more processors for generating a fluid velocity measurement output based on the output of the artificial neural network. Determining the one or more measurements comprises: providing an input to the artificial neural network, wherein the input is derived from at least the one or more sensor outputs, and processing the input with the artificial neural network to produce an output of the artificial neural network.

[0009] The sting portion may include a tip region which includes the one or more holes. The sting portion may comprise an elongate region which terminates at the tip region. The elongate region may be straight, or it may comprise one or more bends or curves. The elongate region may extend, at least initially, in a direction outward of the body. In some examples, the elongate region may include one or more bends, e.g., a rightangle bend.

[0010] Thus, the sting portion may comprise an elongate region and a tip region. One or more of the holes may be disposed on the tip region. However, in some examples, one or more of the holes may be disposed on the elongate region.

[0011] The elongate region may be uniform or substantially uniform in cross section. The elongate region may have a circular cross section. The tip region may taper towards its distal end.

[0012] In some examples, the sting portion may comprise a modular portion of the probe. This allows the sting portion to be replaced when needed.

[0013] More generally, the present specification describes a flow measurement probe comprising one or more holes for pressure sensing and one or more pressure sensors associated with the one or more holes. Each pressure sensor is configured to generate a respective sensor output. The probe further comprises a data processing system comprising computer-readable storage storing parameters which define an artificial neural network, wherein the artificial neural network is trained for use in measuring one or more quantities. The data processing apparatus further comprises one or more processors for determining one or more measurements of the one or more quantities. Determining the one or more measurements comprises: providing an input to the artificial neural network, wherein the input is derived from at least the one or more sensor outputs, and processing the input with the artificial neural network to produce an output of the artificial neural network.

[0014] The one or more quantities may comprise fluid velocity. For example, the artificial neural network may be trained for use in measuring three components of fluid velocity. The one or more processors may determine a measurement of the three components of fluid velocity.

[0015] The one or more measurements of the one or more quantities may comprise the output of the artificial neural network. For example, the output of the artificial neural network may comprise a measurement of the three components of fluid velocity. Alternatively, the output of the artificial neural network may be processed by the one or more processors to obtain the one or more measurements of the one or more quantities. The one or more measurements may be provided in engineering units.

[0016] In some examples, the one or more quantities may further comprise one or more of pressure altitude, local static pressure, or dew point. In some examples, the one or more quantities may include Mach number.

[0017] In some examples, the input to the artificial neural network may include a measure of fluid temperature.

[0018] The artificial neural network may include an output layer having at least three nodes. In some implementations, the output layer comprises three nodes, each corresponding to a different velocity component. Alternatively, the three nodes may respectively correspond to velocity magnitude, yaw angle and pitch angle. More generally, the output layer may comprise three nodes for three respective values which together provide a measure of fluid velocity.

[0019] The computer-readable storage may comprise one or more memories, wherein the one or more memories store the parameters which define the artificial neural network, and wherein the one or more memories have a total storage capacity of 50 kilobytes or less.

[0020] The one or more memories may have a total storage capacity of 2048 bytes or less.

[0021] The parameters which define the artificial neural network may be stored in less than 1000 bytes of memory.

[0022] The data processing system may comprise: a microprocessor, the microprocessor comprising a central processing unit (CPU), and an integrated non-volatile memory. The processor may comprise the central processing unit (CPU) of the microprocessor. The computer-readable storage may comprise the integrated non-volatile memory of the microprocessor.

[0023] The probe may comprise only one microprocessor.

[0024] The artificial neural network may include an input layer having a number of nodes equal to the number of holes of the probe. In some examples, the input layer may include one or more additional nodes to receive data relating to one or more other quantities (e.g., fluid density and / or fluid temperature).

[0025] The artificial neural network may include no more than two hidden layers. However, in some embodiments, the artificial neural network may include two or more hidden layers.

[0026] The artificial neural network may comprise one or more hidden layers comprising a plurality of neurons having an activation function comprising a hyperbolic tangent function, and an output layer comprising one or more neurons having an activation function comprising an identity function.

[0027] The artificial neural network may be trained on a training data set comprising calibration data. The memory required to store the parameters which define the artificial neural network may be smaller than the memory required to store the calibration data.

[0028] The calibration data may comprise a set of pressure values and, for each pressure value, corresponding values for speed and direction.

[0029] The computer-readable storage may further store parameters which define a preprocessing artificial neural network, wherein the preprocessing artificial neural network is configured for generating a pre-processed output based on a respective sensor output. The input derived from the one or more sensors outputs may comprise the pre-processed output.

[0030] The preprocessing neural network may be configured by its training to compensate for a difference between pressure at a hole and pressure at a respective pressure sensor.

[0031] The probe may be a multi-hole probe.

[0032] Brief Description of the Figures

[0033] So that the invention may be more easily understood, embodiments thereof will now be described, by way of example only, with reference to the accompanying drawings, in which:

[0034] Figure 1(a) and 1(b) illustrate a multi-hole probe for measuring fluid velocity in accordance with an example embodiment; Figure 2 illustrates various example multi-hole probes having different body and sting geometries and configurations;

[0035] Figure 3 illustrates the tip region of a seven-hole probe.

[0036] Figure 4 is a schematic illustration of a microprocessor which is embedded in the probe.

[0037] Figure 5 is a visualisation of an example artificial neural network which may be used to determine fluid velocity based on pressure measurement data, and Figure 6 shows an example of how an artificial neural network (ANN) may be represented as data in memory.

[0038] Detailed Description

[0039] Embodiments described in this specification permit a multi-hole probe with an embedded microprocessor having a relatively small amount of memory to determine a fluid velocity vector on board the probe, in real-time, without the use of any external computing system. The described approach is particularly suitable for use in small probes in which the embedded microprocessor size, and therefore also its memory resources, are limited due to the size of the probe. In various embodiments, the probe may have a total memory capacity of less than 60516 bytes, for example 50 kilobytes or less, 10 kilobytes or less, or 2048 bytes or less.

[0040] Figure 1(a) and 1(b) illustrate a multi-hole probe 100 for measuring the velocity (both speed and direction) of a flow of fluid in accordance with an example embodiment. It will be understood that the term "fluid", as used herein may refer to a gas (e.g. air), or liquid (e.g. water).

[0041] As indicated by the dimensions (which are in mm) in Figure 1(a), the probe has a compact footprint, which is beneficial for various applications such as unmanned aerial vehicle (UAV) flight controls, automotive applications (e.g. motorsport instrumentation), process control, and wind tunnel testing.

[0042] It will be understood that the probe 100 measures the velocity of the flow impinging on the probe in the frame of reference of the probe 100. If the probe is moving (e.g. if it is integrated with a UAV which is in motion in the air), then the velocity of the probe relative to the fluid (e.g. air) that it is moving in can thus be determined.

[0043] As shown, the probe 100 comprises a sting assembly 110 and a cylindrical body 120 which is connected to the sting assembly 110 by a gasket 130 and retainer 140. The sting assembly 110 includes a sting 150 having a tip 160 in which a number of holes are formed for pressure sensing. In the example of Figure 1(a) and 1(b), the tip 160 includes seven holes.

[0044] The retainer 140 permits the sting assembly 110 to be readily interchanged with another type, e.g. with a sting assembly having a different number of holes in the tip. Furthermore, while the sting 150 shown in Figure 1(a) and 1(b) includes a 90 degree bend, in some cases the sting may be straight or may include a bend of a different angle. By way of illustration Figure 2 illustrates various example multi-hole probes having different body and sting geometries and configurations. As shown, the sting generally comprises an elongate region, which may be straight, or may comprise one or more bends or curves. The elongate region terminates at the tip, wherein the tip includes the pressure-sensing holes. In some examples, the tip is tapered towards its distal end.

[0045] Figure 3 illustrates the tip region of a sting assembly having a seven-hole probe by way of example. However, those skilled in the art will appreciate that a sting assembly having fewer holes than seven (e.g. 3 or 5 holes), or more holes than seven (e.g. 19, 21 or 37 holes), may be employed in some examples, depending on use case and other factors. The configuration of the holes and physical geometry of the tip (i.e. hemispherical, conical, elliptical, etc.) is also not constrained by the invention.

[0046] In general, each hole in the tip of a multi-hole probe forms the inlet of a tubular channel which extends through the inside of the sting assembly 110 such that the hole is in fluid communication with a respective sensor that is housed in the body 120. Hence, the body includes a separate pressure sensor for each hole of the multi-hole probe, thereby permitting an independent pressure measurement for each hole. As is known in the art, each pressure sensor may be a differential pressure sensor which is in fluid communication with its respective hole as well as a static pressure reference, thereby to measure dynamic pressure (i.e. a pressure that has the local static pressure subtracted off it). The static pressure reference may for example be obtained from holes on the side of the sting, positioned at sufficient distance downstream of the tip to give a good approximation of static pressure.

[0047] Returning to Figure 1(a) and 1(b), body 120 houses an embedded data processing system in the form of an on-board microprocessor 400, which is illustrated schematically in Figure 4. As described in more detail below, the microprocessor 400 is configured for use in processing the output of the pressure sensors to obtain a fluid velocity measurement. As shown in Figure 4, the microprocessor 400 includes a central processing unit 410 and an integrated memory 420. The central processing unit may for example be an 8- bit, 32-bit or 64-bit processor. The integrated memory may comprise EEPROM flash memory.

[0048] A small size of probe requires a small microprocessor which consequently has quite limited memory resources but on the other hand low power requirements (which may be advantageous for various applications). For example, the microprocessor 400 of Figure 4 may measure about 10 x 10 mm (13.25 x 13.25 mm including legs), and may have 2048 bytes of integrated memory. However, it will be appreciated that the various other sizes of microprocessor may be used which may accommodate somewhat more memory, while still being compact enough to be included within the probe. For example, various embodiments may have 50 kilobytes of memory or less, 10 kilobytes or less, 5 kilobytes or less of memory, or 2048 bytes or less of memory.

[0049] The memory 420 stores parameters 430 which define a feed-forward artificial neural network. In particular, the memory 420 stores values 440 which define a neural network topology such as the width, depth and number of hidden layers, and optionally further parameters such as the activation function type. The memory 420 also stores the weights and biases 450 of the artificial neural network.

[0050] The memory 420 further comprises instructions in the form of firmware code 460 to build the artificial neural network using the stored parameters 420. The firmware code 460 is further configured to cause the central processing unit 410 to interface with the pressure sensors by way of a data acquisition subroutine which acquires real-time digital pressure measurement values (in suitable units) from the pressure sensors.

[0051] The obtained pressure measurement data may, in some examples, be provided directly to the artificial neural network for processing.

[0052] Figure 5 is a visualisation of an example artificial neural network 500 which may be used to determine fluid velocity based on the pressure measurement data. As shown, the artificial neural network 500 comprises an input layer 510, first and second hidden layers 520, and an output layer 530. The input layer has seven nodes, one for each of the seven pressures sensors of the probe. Each of these seven nodes may receive sensor data from its respective sensor directly from the data acquisition subroutine. Each hidden layer 520 comprises seven nodes, each node being a nonlinear unit in the form of a neuron 540. The output layer comprises three nodes, each also comprising a respective neuron 540.

[0053] It will be understood that the illustrated topology is shown by way of example, and that other examples may have different numbers of layers or different numbers of nodes per layer. Whilst various configurations are possible, in a typical example, the input layer includes one node for each hole of the sensor, and the output layer includes three nodes for three respective values (e.g., components of a 3D vector) which together provide a measure of velocity (speed and direction), expressed with respect to a particular coordinate system (any suitable coordinate system may be used).

[0054] As will be understood by those skilled in the art, each neuron 540 receives, as its input, the output of the nodes of the preceding layer. Each neuron 540 is associated with a weight for each of these inputs, and an overall bias value. The neuron is configured to determine a weighted combination of the outputs of the preceding layer using the weights, add or subtract the bias, and then apply a function known as an activation function to the result. In this way, the neural network processes the values of the nodes of the input layer through the layers of the neural network so as to generate an output (i.e. the output of the neurons of the output layer). The firmware code 460 includes instructions to cause the central processing unit 410 to perform this processing.

[0055] Compared to other types of activation function, use of a hyperbolic tangent function as the activation function for the hidden layers, and an identity function (i.e. direct output) for the output layer neurons was found to be advantageous for fluid velocity measurement.

[0056] The output of the microprocessor may in various examples comprise the output of the three nodes of the output layer, or alternatively these values may be processed (e.g. transformed into values with respect to another coordinate system) by the central processing unit 410 before being output by the microprocessor 400.

[0057] For example, the output of the three nodes of the output layer 540 may provide values for speed, pitch and yaw respectively. As will be well understood by the person skilled in the art, these values may be converted into cartesian coordinates (e.g. a velocity vector comprising three orthogonal directions of speed), and in some examples, the firmware code 460 may include instructions to perform this transformation.

[0058] The microprocessor 400 may output a serial data stream comprising its output in engineering units (e.g. m / s for the speed, degrees for the pitch and yaw angles). Consequently, no post-processing is required; the data is ready for use immediately by an end user or by the system in which the probe is integrated (e.g. UAV).

[0059] Taken together, the weights and biases of the neurons 540 form the learned parameters of the artificial neural network 500. The weights and biases of the neurons 540 are determined by a training process. In the training process, the weights and biases of the neurons 540 are adjusted iteratively in order to reduce the "error" in the output of the neural network compared to what the true answer should be (ground truth). The ground truth data is provided by a calibration grid generated in a wind tunnel test. The calibration involves sweeping the probe 100 through a set of (known) pitch and yaw angles, for a known wind speed and direction. In the case of an n-hole probe (e.g., 7 hole probe), the output is a set of n (e.g., 7) pressure values for each combination of yaw, pitch and wind speed. Thus, a training data set is obtained comprising a set of pressure value inputs, and the corresponding true values for speed, pitch and yaw.

[0060] Various training algorithms may be used to determine the weights and biases of the neurons using the training data set. It has been found that in the present context, training using a Particle Swarm Optimisation (PSO) approach is advantageous. Training neural networks using PSO is known per se to those skilled in the art and will not be described in detail here. Briefly, in PSO, the hyperspace positions of virtual "particles", each representing a potential solution (values of weights and biases) are adjusted in the parameter hyperspace to minimise a global error. In the present context, the global error may be given by the sum of the differences between the ANN output and that of the true answer over at least a subset of the training data set. It will be understood that various alternative approaches, for example the backpropagation algorithm, may alternatively be used to train the artificial neural network using the training data set.

[0061] In various embodiments, the memory requirements for storing the parameters defining the ANN are significantly less than the memory requirements for storing the calibration grid on which the ANN was trained. For example, a trained ANN having the architecture shown in Figure 5 may be fully parameterised and stored in only 578 bytes of memory. In this case, the parameters which define the trained artificial neural network occupy just 30% of the available 2048 bytes of EEPROM space, leaving plenty of memory for other variables whilst allowing the possibility of larger ANNs, e.g. for more holes. Indeed, on a smaller 8-bit microprocessor processor that typically has 1024 bytes of available EEPROM, this ANN could still be stored in EEPROM, with memory to spare.

[0062] Figure 6 shows one example of how an ANN may be represented as data in EEPROM (shown in hexadecimal). In Figure 6, a CRC-16 checksum is also included at the end of the ANN data block to ensure data integrity and so that, should any of the bytes become corrupted, the firmware can prevent use of the network and report the error and act accordingly. The above bytes encapsulate everything required to construct the ANN - the topology (input layers, hidden layers depth and width, output layers), hidden layer activation function type, output layer activation function type, every neuron's weights and bias values (to 32-bit precision) and the integrity checksum.

[0063] The ANN-based approach described above is suitable for a wide range of applications, permitting computation of the velocity vector on board the probe in real-time using a minimal Al on-board that can fit in the limited memory resources of a low power microprocessor. This makes the probe self-contained and able to independently determine its speed and direction through air which may be particularly beneficial when the probe is integrated in an autonomous vehicle such as an autonomous survey craft.

[0064] In some examples, pressure values obtained by the data acquisition subroutine may be processed prior to input to the ANN to compensate for differences between the pressure at a hole in the tip and the pressure at the position of the corresponding sensor. The tubing connecting the sensing holes at the tip of the probe and the corresponding pressure sensors can alter the dynamic response of the pressure measurement system, resulting in a frequency-dependent gain and phase shift. The gain can be particularly low at high frequencies, resulting in poor signal-to-noise ratio and a pronounced error especially in conditions of rapid flow fluctuation (e.g. high turbulence conditions), and so compensating for this effect is particularly advantageous for probes intended to be used in such conditions.

[0065] In an embodiment, this effect may be compensated for through the use of a preprocessing ANN which may also be stored in the embedded memory of the probe. Real-time data for each respective sensor may be input into the pre-processing ANN to generate an output which is then provided as input to the ANN that is used to determine fluid velocity. Thus, the embedded memory of the probe may store two ANNs, one (the pre-processing ANN) for dynamic compensation of pressure sensor values, and the other for generating the velocity vector (as described above with reference to Figure 5). The pre-processing ANN may be trained using known training techniques using a training data set. The training data set may be obtained by applying a set of pressure signals of known frequencies and amplitudes to the probe tip within a reference chamber that has its own time-resolved independent pressure measurement capability. Alternatively, the pre- preprocessing ANN may be trained using an analytical model of dynamic compensation.

[0066] Although the sensors of the example probe 100 described above may be differential sensors, absolute pressure sensors may be used as an alternative to differential pressure sensors in some embodiments. In this case, the probe may include a further sensor dedicated to measuring the static pressure, which may be subtracted from the independent absolute pressure measurement for each hole in order to obtain dynamic pressure. Furthermore, as will be understood by those skilled in the art, in some cases, rather than having a static pressure reference on the probe itself, an external static reference port may for example be used, which may for example be connected to a pressure tap on a wind tunnel wall at a reference station.

[0067] In some examples, the probe may be provided with one or more additional sensors, e.g. one or more inertial measurement sensors (IMU) to measure acceleration and gyration vectors and / or environmental sensors (absolute atmospheric pressure, temperature and humidity).

[0068] Various implementations described above relate to a probe for measuring fluid velocity. The probe comprises three or more holes for pressure sensing, and a respective pressure sensor associated with each of the three or more holes, each pressure sensor being configured to generate a respective sensor output. The probe includes a data processing system comprising computer-readable storage storing parameters which define an artificial neural network, wherein the artificial neural network is trained for use in measuring fluid velocity. The data processing system further comprises one or more processors for generating a fluid velocity measurement output based on an output of the artificial neural network. Generating the fluid velocity measurement output comprises: providing an input to the artificial neural network, wherein the input is derived from the sensor outputs, and processing the input with the artificial neural network to produce the output of the artificial neural network. This specification is not limited to a probe for measuring fluid velocity. More generally, it relates to a flow measurement probe for measuring a flow to extract one or more measurements of at least one quantity of interest. The one or more measurements may comprise a time series of measurements based on real-time values of the one or more sensors outputs.

[0069] A quantity of interest may comprise a property of the flow, for example speed, direction and / or Mach number of the flow. Alternatively, it may comprise a property of the probe such as the airspeed, angle of attack, sideslip angle and / or Mach number of the probe (or equivalently of a vehicle such as an aerial vehicle to which the probe is mounted). Alternatively, it may comprise a quantity relating to the environment of the probe e.g. pressure altitude, local static pressure or dew point.

[0070] The on-board ANN may be trained for use in determining the one or more measurements of the one or more quantities of interest. While in some cases, the one or more measurements may be provided directly by the output of the ANN, in other cases the ANN output may be processed by the data processing system to provide the one or more measurements.

[0071] The one or more measurements of the one or more quantities may be determined by providing the artificial neural network with an input comprising data derived from one or more pressure sensors included in the probe. In some cases, additional information (e.g. a measure of temperature) may also be used in determining the measurement.

[0072] In some examples, the probe may be configured for or capable of measuring supersonic flow. In such a case, the probe may output the three components of velocity as Mach numbers (Mx, My, Mz). Alternatively, the ANN may be trained to output three components of velocity together with Mach number.

[0073] In these cases, the ANN input may comprise pressure measurements derived from the pressure sensors and a measure of temperature from an external temperature sensor. The training data for the ANN may be based on a calibration map between pressure sensor and temperature values on the one hand, and Mach number (and flow direction relative to the probe in the case of three or more holes) on the other.

[0074] Alternatively, the output of the ANN may comprise three components of velocity only, and the Mach number may be determined by the data processing apparatus of the probe using the ANN output and a measure of temperature from an external temperature sensor.

[0075] In one particular example, an ANN may be trained to determine Mach number based on the output of the pressure sensor of a probe having a single hole for pressure sensing, together with a signal from an external temperature sensor.

[0076] In some examples, local static pressure may be alternatively or additionally measured. As will be understood by those skilled in the art, static pressure is not measured directly by the pressure sensors of the probe if the probe is yawed and / or pitched, but can be inferred from the pressure sensor output using a calibration map for static pressure. An ANN may thus be trained on such a calibration map to determine static pressure based on the measured pressure values. As will be understood by the person skilled in the art, static pressure may then be used to determine pressure altitude.

[0077] This specification includes the subject matter of the following clauses:

[0078] 1. A flow measurement probe comprising: one or more holes for pressure sensing; one or more pressure sensors associated with the one or more holes, each pressure sensor being configured to generate a respective sensor output, and a data processing system comprising: computer-readable storage storing parameters which define an artificial neural network, wherein the artificial neural network is trained for use in measuring one or more quantities, and one or more processors for determining one or more measurements of the one or more quantities, wherein determining the one or more measurements comprises: providing an input to the artificial neural network, wherein the input is derived from at least the one or more sensor outputs, and processing the input with the artificial neural network to produce an output of the artificial neural network.

[0079] 2. A probe according to clause 1, wherein the one or more quantities comprise one or more of flow speed, flow direction, probe speed, probe direction, pressure altitude, local static pressure, or dew point. 3. A probe according to clause 1 or clause 2, wherein the one or more quantities comprise fluid velocity.

[0080] 4. The probe of any one of clauses 1 to 3, wherein the input to the artificial neural network includes a measure of temperature.

[0081] 5. A probe according to any one of clauses 1 to 4, wherein the one or more quantities comprise Mach number.

[0082] 6. A probe according to any one of clauses 1 to 5, wherein the artificial neural network includes an output layer having at least three nodes, wherein each of the at least three nodes corresponds to a different velocity component.

[0083] 7. A probe according to any one of clauses 1-6, wherein the computer-readable storage comprises one or more memories, wherein the one or more memories store the parameters which define the artificial neural network, and wherein the one or more memories have a total storage capacity of 50 kilobytes or less.

[0084] 8. The probe of clause 7 wherein the one or more memories have a total storage capacity of 2048 bytes or less.

[0085] 9. A probe according to any one of clauses 1 to 8, wherein the parameters which define the artificial neural network are stored in less than 1000 bytes of memory.

[0086] 10. A probe according to any one of the clauses 1 to 9, wherein the data processing system comprises: a microprocessor, the microprocessor comprising a central processing unit, CPU, and an integrated non-volatile memory, wherein: said processor comprises the central processing unit, CPU, of the microprocessor, and said computer-readable storage comprises the integrated non-volatile memory of the microprocessor.

[0087] 11. The probe of clause 10, wherein the probe comprises only one microprocessor. 12. A probe according to any one of clauses 1 to 11, wherein the artificial neural network includes an input layer having a number of nodes equal to the number of holes of the probe.

[0088] 13. A probe according to any one of the clauses 1 to 12, wherein the artificial neural network includes no more than two hidden layers.

[0089] 14. A probe according to any one of clauses 1 to 13, wherein the artificial neural network comprises: one or more hidden layers comprising a plurality of neurons having an activation function comprising a hyperbolic tangent function, and an output layer comprising one or more neurons having an activation function comprising an identity function.

[0090] 15. A probe according to any one of clauses 1 to 14, wherein the artificial neural network is trained on a training data set comprising calibration data, wherein the memory required to store the parameters which define the artificial neural network is smaller than the memory required to store the calibration data.

[0091] 16 A probe according to clause 15, wherein the calibration data comprises a plurality of sets of pressure values and, for each set of pressure values, corresponding values for speed and direction.

[0092] 17. A probe according to any one of clauses 1 to 16, wherein the computer- readable storage further stores parameters which define a preprocessing artificial neural network, wherein the preprocessing artificial neural network is configured for generating a pre-processed output based on a respective sensor output, wherein the input derived from the one or more sensors outputs comprises the pre-processed output.

[0093] 18. A probe according to clause 17, wherein the preprocessing neural network is configured by its training to compensate for a difference between pressure at a hole and pressure at a respective pressure sensor.

[0094] 19. A probe according to any one of clauses 1 to 18, wherein the probe is a multihole probe.

[0095] 20. The probe of any one of clauses 1 to 19, comprising: a sting portion comprising the one or more holes, and a body which includes the one or more pressure sensors and the data processing system.

[0096] 21. The probe of clause 20, wherein the sting portion comprises a tip region, wherein one or more of the holes are disposed on the tip region.

[0097] 22. The probe of clause 21, wherein the tip region tapers towards its distal end.

[0098] 23. The probe of any one of clauses 20 to 22, wherein the sting portion comprises an elongate region.

[0099] 24. The probe of clause 23, wherein the elongate region has a uniform cross section.

[0100] 25. The probe of clause 23 or 24, wherein one or more of the holes are disposed on the elongate region.

[0101] 26. The probe of any one of clauses 20 to 25, wherein the sting portion is a modular portion of the probe.

[0102] 27. The probe of any one of clauses 1 to 19, wherein the probe comprises an elongate region which terminates at a tip region.

[0103] 28. The probe of clause 27, wherein one or more of the holes are disposed on the tip region.

[0104] 29. The probe of clause 27 or clause 28, wherein one or more of the holes are disposed on the elongate region.

[0105] 30. The probe of any one of clauses 27 to 29, wherein the elongate region has a uniform cross section.

[0106] 31. The probe of one of clauses 27 to 30, wherein the tip region tapers towards its distal end.

[0107] 32. The probe of any one of clauses 27 to 31, further comprising a body, wherein the body includes the one or more pressure sensors and the data processing system. 33. The probe of any one of clauses 27 to 32, wherein the elongate region is a modular portion of the probe. 34. The probe of any one of clauses 1 to 33, wherein the one or more holes comprise at least three holes.

[0108] 35. The probe of any one of clauses 1 to 34, wherein the one or more sensors comprise a respective sensor for each of the one or more holes.

[0109] Many further variations and modifications are possible, that fall within the scope of the following claims:

Claims

Claims1. A probe for measuring fluid velocity, comprising: a sting portion, the sting portion comprising three or more holes for pressure sensing, and a body, the body including: a respective pressure sensor associated with each of the three or more holes, each pressure sensor being configured to generate a respective sensor output, and a data processing system comprising: computer-readable storage storing parameters which define an artificial neural network, wherein the artificial neural network is trained for use in measuring fluid velocity, and one or more processors for generating a fluid velocity measurement output based on the output of the artificial neural network, wherein generating the fluid velocity measurement output comprises: providing an input to the artificial neural network, wherein the input is derived from at least the sensor outputs, and processing the input with the artificial neural network to produce an output of the artificial neural network.

2. A probe according to claim 1, wherein the artificial neural network includes an output layer having three nodes for three respective values which together provide a measure of fluid velocity.

3. A probe according to claim 1 or claim 2, wherein the computer-readable storage comprises one or more memories, wherein the one or more memories store the parameters which define the artificial neural network, and wherein the one or more memories have a total storage capacity of 50 kilobytes or less.

4. The probe of claim 3 wherein the one or more memories have a total storage capacity of 2048 bytes or less.

5. A probe according to any one of the preceding claims, wherein the parameters which define the artificial neural network are stored in less than 1000 bytes of memory.

6. A probe according to any one of the preceding claims, wherein the data processing system comprises: a microprocessor, the microprocessor comprising a central processing unit, CPU, and an integrated non-volatile memory, wherein: said processor comprises the central processing unit, CPU, of the microprocessor, and said computer-readable storage comprises the integrated non-volatile memory of the microprocessor.

7. The probe of claim 6, wherein the probe comprises only one microprocessor.

8. A probe according to any one of the preceding claims, wherein the artificial neural network includes an input layer having a number of nodes equal to the number of holes of the probe.

9. A probe according to any one of the preceding claims, wherein the artificial neural network includes no more than two hidden layers.

10. A probe according to any preceding claim, wherein the artificial neural network comprises: one or more hidden layers comprising a plurality of neurons having an activation function comprising a hyperbolic tangent function, and an output layer comprising one or more neurons having an activation function comprising an identity function.

11. A probe according to any preceding claim, wherein the artificial neural network is trained on a training data set comprising calibration data, wherein the memory required to store the parameters which define the artificial neural network is smaller than the memory required to store the calibration data.

12. A probe according to claim 11, wherein the calibration data comprises a plurality of sets of pressure values and, for each set of pressure values, corresponding values for speed and direction.

13. A probe according to any one of the preceding claims, wherein the computer- readable storage further stores parameters which define a preprocessing artificial neural network, wherein the preprocessing artificial neural network is configured forgenerating a pre-processed output based on a respective sensor output, wherein the input derived from the sensors outputs comprises the pre-processed output.

14. A probe according to claim 13, wherein the preprocessing neural network is configured by its training to compensate for a difference between pressure at a hole and pressure at a respective pressure sensor.

15. A probe according to any one of the preceding claims, wherein the probe is a multi-hole probe.

16. The probe according to any one of the preceding claims, wherein the sting portion comprises a tip region, wherein one or more of the holes are disposed on the tip region.

17. The probe of claim 16, wherein the tip region tapers towards its distal end.

18. The probe according to any one of the preceding claims, wherein the sting portion comprises an elongate region.

19. The probe of claim 18, wherein the elongate region has a uniform cross section.

20. The probe of claim 18 or claim 19, wherein one or more of the holes are disposed on the elongate region.

21. The probe of any one of the preceding claims, wherein the sting portion is a modular portion of the probe.

Images