System for discriminating the size of supercooled droplets for an aircraft and method of using the system
A laser-based system with a convolutional neural network classifies supercooled droplets by size, addressing the expense and solar sensitivity of existing systems, ensuring reliable icing condition differentiation.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- SAFRAN AEROSYST
- Filing Date
- 2025-11-24
- Publication Date
- 2026-06-04
AI Technical Summary
Existing aircraft icing detection systems are expensive, sensitive to solar radiation, and struggle to accurately distinguish between supercooled droplets based on size, particularly for differentiating between droplets under Annex O and C conditions.
A system using a laser source to generate optical back-injection interference signals, combined with a convolutional neural network for analyzing amplitude variations, classifies droplets based on size, eliminating the need for separate detectors and reducing sensitivity to ambient light.
The system provides accurate droplet size discrimination without solar interference, using compact and cost-effective components, enabling reliable differentiation between different icing conditions.
Smart Images
Figure FR2025051096_04062026_PF_FP_ABST
Abstract
Description
[0001] Description
[0002] TITLE: SYSTEM FOR DISCRIMINATING DROPLET SIZE IN SUPERFLUID FOR AN AIRCRAFT AND METHOD FOR USING THE SYSTEM
[0003] TECHNICAL FIELD
[0004] The invention relates to the field of analyzing the environmental flight conditions of an aircraft. It relates in particular to a system for discriminating the size of supercooled droplets for an aircraft, as well as a method for using this system.
[0005] PREVIOUS TECHNIQUE
[0006] The prior art includes in particular the documents US-A-20180209892, CN-A-102564909, CN-A-118794848 and FR-A1-3145343.
[0007] In the aeronautical field, regulations define icing conditions (likely to impair flight) in which certain aircraft (airplanes, helicopters, drones, etc.) are certified to fly.
[0008] These icing conditions (or freezing conditions) exist when the air contains droplets in a supercooled state (or supercooled droplets), that is, droplets of liquid water at a negative temperature.
[0009] The "Code of Federal Regulation" 1 The document defines three categories of icing conditions in its annexes. Annex C describes the case of supercooled water droplets with a diameter less than 100 µm, for which the average diameter is generally around 10 µm. Annex O concerns supercooled water droplets with a diameter greater than 100 µm. Finally, Annex D concerns ice crystals.
[0010] In addition, the differentiation of these categories by droplet diameter is linked to the area where the frost layer appears: drops with a diameter less than 100pm form frost on the leading edge, those with a diameter greater than 100pm cause it to appear further along the aerodynamic surface of the aircraft (which deteriorates its aerodynamic performance), and the crystals of Annex D can clog sensors such as pitot tubes or cool the engines which deteriorates flight performance.
[0011] Thus, the existence of these categories presupposes the ability to equip aircraft with systems capable of distinguishing each of these icing conditions within their environment. In particular, to distinguish between icing conditions falling under Annex O and those falling under Annex C, it is necessary to use systems capable of discriminating between supercooled water droplets based on their size (i.e., their diameter).
[0012] There are currently two main types of sensors: atmospheric and accretion. The first detects icing conditions by detecting droplets or crystals in the atmosphere. The second detects the accretion of frost onto a surface.
[0013] These sensors rely on various measurement principles (mechanical, acoustic, thermal, optical, etc.), but despite their diversity, only two technologies have demonstrated the ability to distinguish between conditions covered by Annex O and those covered by Annex C: optical interferometry and light scattering. Both technologies perform detection on a small volume, allowing them to isolate the signal from large individual droplets. They rely on the use of fast, high-frequency, and sensitive optoelectronic components and therefore involve expensive systems. Furthermore, they use a detector and a light source that are separate, making them highly sensitive to solar radiation, which can interfere with the measurement.
[0014] SUMMARY OF THE INVENTION
[0015] The present invention offers a solution to these drawbacks.
[0016] Thus, one objective of the invention is to propose a simple, economical system capable of detecting small scattering signals specifically from supercooled droplets without being dazzled by ambient light.
[0017] To this end, the invention, according to a first aspect, relates to a system for discriminating the size of supercooled droplets for an aircraft, said system comprising at least one laser source configured to emit a laser beam and a focusing device, intended to focus the laser beam into a cloud of supercooled droplets,
[0018] said system being characterized in that the laser source is configured to generate an optical back-injection type interference signal from a reflection of the laser beam in the cloud of supercooled droplets and in that it further comprises a processing unit configured to perform the acquisition of the interference signal and the analysis of the amplitude variations of said interference signal by a convolutional neural network so as to classify the droplets of the cloud according to their size,
[0019] The system according to the invention may comprise one or more of the following features, taken individually or in combination with each other:
[0020] ■ The laser source includes a power supply unit and a laser diode, the wavelength of which is preferably between 1520nm and 1580nm.
[0021] - the system further includes an amplifier, positioned between the laser source and the processing unit, and configured to amplify the interference signal from the laser source.
[0022] - The convolutional neural network has an architecture comprising a stack of one-dimensional convolutional layers with pooling layers between two convolutional layers.
[0023] ■■ The convolutional neural network has an "inception" type architecture comprising several layers of convolutional networks, each layer comprising convolutions at different spatial scales.
[0024] The system further comprises a second laser source, associated with a second focusing device and configured to generate a second optical back-injection interference signal from a second reflection of a second laser beam emitted by said second laser source into the supercooled droplet cloud. The processing unit is configured to acquire all interference signals and analyze the amplitude variations of said interference signals using a convolutional neural network in order to classify the droplets in the cloud according to their size. The convolutional neural network comprises several channels corresponding respectively to the different interference signals, and the processing unit is configured so that a first convolutional layer operates on all channels and merges the resulting kernels into a single output space.
[0025] ■ The convolutional neural network has an architecture comprising a stack of one-dimensional convolutional layers with pooling layers between two convolutional layers.
[0026] - the convolutional neural network (presents an "inception" type architecture comprising several layers of convolutional networks, each layer comprising convolutions at different spatial scales.
[0027] The invention, according to a second aspect, also relates to a method of using a supercooled droplet size discrimination system for an aircraft, according to the first aspect, said method comprising the following steps:
[0028] - acquisition, by the processing unit, of the optical back-injection interference signal obtained from a reflection of the laser beam in the supercooled droplet cloud; and,
[0029] - analysis, by the processing unit, via a convolutional neural network, of the variations in amplitude of said interference signal in order to classify the cloud droplets according to their size.
[0030] BRIEF DESCRIPTION OF THE DRAWINGS
[0031] The invention will be better understood with the aid of the following description, given solely by way of example and made with reference to the accompanying drawings in which:
[0032] Figure 1 is a schematic representation of a supercooled droplet size discrimination system for an aircraft according to a first embodiment of the invention; Figure 2 is a functional diagram illustrating the analysis performed by the processing unit of the system according to the first embodiment of the invention;
[0033] Figure 3 is a schematic representation of a supercooled droplet size discrimination system for an aircraft according to a second embodiment of the invention;
[0034] Figure 4 is a functional diagram illustrating the analysis performed by the system's processing unit according to the second embodiment of the invention;
[0035] Figure 5 is a set of optical back-injection type interferometry signals according to examples of signals acquired by a processing unit of a system according to the invention;
[0036] Figure 6 is an example of the classification result of supercooled droplets based on their size, as obtained by a system according to the invention; and,
[0037] Figure 7 is a step diagram of a method of use according to one implementation of the invention.
[0038] DESCRIPTION OF IMPLEMENTATION METHODS
[0039] With reference to Figure 1, we will now describe an embodiment of a 101 system for discriminating the size of supercooled droplets for an aircraft according to the invention.
[0040] Such a system can be installed in any type of aircraft, such as an airplane, a drone, or a helicopter.
[0041] In the example shown, the system 101 includes a laser source 103 configured to emit a laser beam 105 and a focusing device 107 designed to focus the laser beam 105 into a cloud 109 of supercooled droplets. For example, the focusing device could be a lens with a numerical aperture between 0.5 and 0.8, for instance, 0.68. Such a cloud could, for instance, include mist and a mixture of supercooled droplets of varying sizes (i.e., diameters), or only supercooled droplets of uniform size. The term "droplets" used here and in what follows refers to water droplets. The term SLD (Supercooled Large Droplets) is also used to refer to supercooled water droplets with a diameter greater than 50 µm.
[0042] In the non-limiting example shown in Figure 1, the laser source 103 includes a power supply unit 103a and a laser diode 103b, the wavelength of which can be, for example, between 1520nm and 1580nm.
[0043] Advantageously, this type of laser source is compact, inexpensive, and allows the generation of a Self-Mixing type interference signal (i.e. an interference signal obtained by optical back-injection, also called optical back-injection type, which is described in more detail later) which corresponds to a voltage and which is directly usable by a processing unit.
[0044] Indeed, in the invention, the laser source 103 is configured to generate a Self-Mixing type interference signal from a reflection of the laser beam 105 in the cloud 109 of supercooled droplets.
[0045] In particular, the system uses Self-Mixing Interferometry (SMI), the principle of which is based on the dynamic regime of a laser source subjected to feedback. Specifically, in the example shown, a portion of the laser beam 105 that is reflected, in this case by a droplet, is reinjected into the cavity of the laser source 103 and disturbs it, generating interference. This interference, which corresponds to fluctuations in the voltage across the laser diode 103b in the example shown, depends directly on the relative displacements of the droplet with respect to the laser source 103 and makes it possible to determine the size of this droplet through its analysis.
[0046] In the non-limiting example shown in Figure 1, the system 101 further includes a processing unit 113 which is configured to
[0047] The acquisition of the interference signal and the analysis of its amplitude variations using a convolutional neural network are performed in order to classify the droplets in cloud 109 according to their size. Figure 2 illustrates in more detail the processing of an interference signal 115, which is carried out by the processing unit 113 using a convolutional neural network 117, whose output data 119 corresponds to a classification of the droplets in cloud 109 according to their size.
[0048] In particular, this classification makes it possible to establish for each droplet whether its diameter is less than or greater than 100pm and therefore whether it falls under Annex C or Annex O of the "Code of Federal Regulation".
[0049] Advantageously, using a convolutional neural network eliminates the need to identify rising or falling edges in the interference signal (a method whose reliability is variable) to discriminate droplet size. Indeed, through learning based on known data (i.e., specific signals associated with droplets of known size), the processing unit (e.g., a computer) can deduce rules for identifying droplet size from a new interference signal. In other words, training the convolutional neural network allows for the development of a statistical model capable of predicting new results based on new input data (i.e., new interference signals).
[0050] As a non-limiting example, clouds of water droplets analogous to those of icing conditions can be generated. The signals from the interferometer at the location of these clouds are recorded and labeled according to the observed scene with a unique label:
[0051] ■■ no icing condition: no cloud in front of the laser beam which is labeled 0; - Annex C: cloud composed of drops with a diameter less than 100pm which is labeled 1;
[0052] - Annex O freezing drizzle: cloud composed of water droplets between 0 and 500 pm in diameter which is labeled 2; and,
[0053] - Annex O freezing rain: cloud composed of water droplets between 0 and 3mm in diameter which is labeled 3.
[0054] The signals and labels thus form two corresponding sets of data used to train the neural network. As another example, Figure 5 shows four self-mixing interference signals from the same laser source, each corresponding to a different scene (i.e., a different environment in which the measurement is performed).More specifically, from top to bottom, Figure 5 shows a first interference signal corresponding to the measurement by system 101 of a cloud comprising fog and a first set of droplets having a first diameter, a second interference signal corresponding to the measurement by system 101 of a cloud comprising fog and a second set of droplets having a second diameter different from the first diameter, a third interference signal corresponding to the measurement by system 101 of fog alone and a fourth interference signal corresponding to the measurement by system 101 of air.
[0055] Figure 6 (described in more detail later) shows the results of classification of supercooled droplets according to their size obtained by the convolutional neural network 117 (implemented by the processing unit 113) for each of the interference signals shown in Figure 5.
[0056] Furthermore, in the non-limiting example described here, the interference signals 115 acquired by the processing unit 113 correspond to the voltage across the laser diode 103b, which has been amplified by an amplifier 121. Indeed, the system 101 further includes an amplifier 121, which is positioned between the laser source 103 and the processing unit 113, and which is configured to amplify the interference signal 115 from the laser source 103. The amplifier 121 can, for example, have a bandwidth of several megahertz and a gain of approximately 80 dB. The amplifier thus allows the processing unit 113 to acquire an interference signal with an amplitude sufficient for processing.In a particular embodiment of the invention, the convolutional neural network 117 has an architecture comprising a stack of one-dimensional convolutional layers with sharing layers between two convolutional layers. Advantageously, the low computing power required to implement such processing allows the use of a low-power, embedded component-type processing unit.
[0057] In another particular embodiment, the convolutional neural network 117 features an "inception" type architecture comprising several layers of convolutional networks, each layer of which includes convolutions at different spatial scales. Such processing requires greater computing power and therefore the use of a more expensive component, but in return allows for better performance in the implementation of the convolutional neural network and consequently in the classification of droplets according to their size.
[0058] Figure 3 shows another embodiment of the system 101 according to the invention. In this embodiment, the system 101 comprises a plurality (i.e., at least two) of laser sources 103. Each laser source 103 is also associated with a focusing device 105 and is configured to generate a self-mixing interference signal 115 from a reflection of the laser beam 105 emitted by the laser source 103 in the cloud 109 of supercooled droplets.
[0059] The system is therefore similar to that shown in Figure 1, with the difference that several interference signals 115 are generated by several laser sources 103 (which are all focused in the cloud 109) to then be analyzed by the processing unit 113.
[0060] Figure 4 illustrates in more detail the processing of a plurality of interference signals 115 which is carried out by the processing unit 113 via the use of a convolutional neural network 117 whose output data 119 corresponds to a classification of the cloud droplets 109 according to their size.
[0061] Thus, the processing unit 113 is configured to acquire all the interference signals 115 and to analyze the amplitude variations of these interference signals 115 by a convolutional neural network 117 in order to classify the cloud droplets 109 according to their size.
[0062] Advantageously, this embodiment allows the convolutional neural network 117 to extract useful information (related to droplet size) from several interference signals. The potential variation in backscattering conditions for each droplet, which can have an unpredictable effect on the acquired interference signal, does not impact the system's performance. In particular, the speckles, resulting from the scattering of the laser beam 105 by the droplets and seen by the laser source 103, are different for each laser source 103. Thus, based on its training, the convolutional neural network 117 can exploit only the interference signals 115 relevant to the impact of these speckles. Furthermore, this embodiment also avoids the impact of a potential failure of a laser source whose interference signal is not taken into account.
[0063] In one particular embodiment, the convolutional neural network 117 comprises several channels which correspond respectively to the different interference signals and the processing unit 113 is configured so that a first convolutional layer operates on all the channels and merges the resulting kernels into a single output space.
[0064] Once this processing step is completed, as in the embodiment described with reference to Figure 1, the convolutional neural network 117 can exhibit an architecture comprising a stack of one-dimensional convolutional layers with pooling layers (also called maxpooling layers) between two convolutional layers, or an "inception" type architecture comprising several layers of convolutional networks, each layer comprising convolutions at different spatial scales. As mentioned above, Figure 6 shows an example of classification results for supercooled droplets based on their size (i.e., output data from the convolutional neural network) as obtained by the convolutional neural network for each of the interference signals shown in Figure 5. These different signals (i.e.Those in Figure 5) were obtained previously from a system such as the one shown in Figure 1 for different predetermined icing conditions (i.e., different scenes). In particular, these results are shown in the form of a confusion matrix 119 in which each row corresponds to a real scene (among the four scenes corresponding to the four interference signals 115 in Figure 5) and each column corresponds to a scene determined by the convolutional neural network. This matrix illustrates the reliability of the prediction model established by the convolutional neural network.
[0065] Finally, Figure 7 shows a step diagram of a method 701 for using the system 101 (regardless of its embodiment). Thus, the method 701 includes a step 703 of acquisition, by the processing unit 113, of the Self-Mixing type interference signal 115 from the laser source 103 and a step 705 of analysis, by the processing unit 113, via a convolutional neural network 117, of the variations in amplitude of the interference signal 115 in order to classify the droplets of the cloud 109 according to their size.
[0066] Finally, thanks to the invention, it is possible to classify supercooled droplets according to their size in a given environment without solar radiation altering the performance of the system and without the back-diffusion conditions of the droplets also altering the performance of the system.
Claims
DEMANDS 1. System (101) for discriminating the size of supercooled droplets for an aircraft, said system (101) comprising at least one laser source (103) configured to emit a laser beam (105) and a focusing device (107) for focusing the laser beam (105) into a cloud (109) of supercooled droplets, said system (101) being characterized in that the laser source (103) is configured to generate an interference signal (115) of the optical back-injection type from a reflection of the laser beam (105) in the cloud (109) of supercooled droplets and in that it further comprises a processing unit (113) configured to perform the acquisition of the interference signal (115) and the analysis of the amplitude variations of said interference signal (115) by a convolutional neural network (117) so as to classify the droplets of the cloud (109) according to their size.
2. System according to claim 1, wherein the laser source (103) comprises a power supply unit (103a) and a laser diode (103b), the wavelength of which is preferably between 1520nm and 1580nm.
3. System according to claim 1 or claim 2, further comprising an amplifier (121), positioned between the laser source (103) and the processing unit (113), and configured to amplify the interference signal (115) from the laser source (103).
4. System according to any one of claims 1 to 3, wherein the convolutional neural network (117) has an architecture comprising a stack of one-dimensional convolutional layers with pooling layers between two convolutional layers.
5. A system according to any one of claims 1 to 3, wherein the convolutional neural network (117) has an "inception" type architecture comprising several layers of convolutional networks, each layer comprising convolutions at different spatial scales.
6. System according to any one of claims 1 to 3, further comprising a second laser source (103), the second laser source (103) being associated with a second focusing device (107) and being configured to generate a second interference signal (115) of the optical back-injection type from a second reflection of a second laser beam (105) emitted by said second laser source (103) in the cloud (109) of supercooled droplets, the processing unit (113) being configured to perform the acquisition of all the interference signals (115) and the analysis of the amplitude variations of said interference signals (115) by the convolutional neural network (117) so as to classify the droplets of the cloud (109) according to their size.
7. System according to claim 6, wherein the convolutional neural network (117) comprises several channels corresponding respectively to the different interference signals and the processing unit (113) is configured such that a first convolutional layer operates on all channels and merges the resulting kernels into a single output space.
8. System according to claim 7, wherein the convolutional neural network (117) has an architecture comprising a stack of one-dimensional convolutional layers with pooling layers between two convolutional layers.
9. System according to claim 7, wherein the convolutional neural network (117) has an "inception" type architecture comprising several layers of convolutional networks, each layer comprising convolutions at different spatial scales.
10. A method (701) for using a supercooled droplet size discrimination system (101) for an aircraft according to any one of the preceding claims, said method (701) comprising the following steps: acquisition (703), by the processing unit (113), of the optical back-injection interference signal (115) obtained from a reflection of the laser beam (105) in the cloud (109) of supercooled droplets; and, analysis (705), by the processing unit (113), via a convolutional neural network (117), of the variations in amplitude of said interference signal (115) so as to classify the droplets of the cloud (109) according to their size.