Detection of red-brown weevils by applying machine learning to signals detected by fiber optic distributed acoustic sensors

Abstract

A fiber optic distributed acoustic sensing (DAS) system (500) for detecting red brown weevils (RPWs) includes an optical fiber (100) configured to be wound around a tree and a DAS box (502) connected to the optical fiber (100). The DAS box (502) includes a processing unit (530) configured to receive a filtered Rayleigh signal (520) reflected by the optical fiber (100) and process the filtered Rayleigh signal (520) via a neural network system (1100, 1500) to determine the presence of RPWs in the tree.

Description

[0001] Cross-references to related applications

[0002] This application claims priority to U.S. Provisional Patent Application No. 62 / 994,502, filed March 25, 2020, entitled "COMBINING ARTIFICIAL INTELLIGENCE AND DISTRIBUTED FIBER OPTIC SENSING TO DETECT RED PALM WEEVIL", and U.S. Provisional Patent Application No. 63 / 139,008, filed January 19, 2021, entitled "COMBINING ARTIFICIAL INTELLIGENCE AND DISTRIBUTED FIBER OPTIC SENSING TO DETECT RED PALM WEEVIL", the disclosure of which is incorporated herein by reference in its entirety. background Technical Field

[0004] The embodiments of the subject matter disclosed in this application generally relate to a system and method for detecting red thorn weevils using fiber optic distributed acoustic sensing, and more specifically, to an enhancement system that processes data recorded using fiber optic distributed acoustic sensing through machine learning algorithms to separate the true RPW signal from ambient noise. Background Technology

[0006] The red palm weevil (RPW) (Rhynchophorus ferrugineus) is a snout pest originating in tropical Asia. Over the past few decades, it has spread to many parts of the world, including North Africa, the Middle East, and the Mediterranean region. This pest has devastated numerous palm farms in various countries and is therefore considered a very serious problem. In the Gulf states and the Middle East, millions of dollars are spent annually just to remove infested palm trees. The cost of treating infested palm trees can be even higher. Furthermore, it is estimated that by 2023, the costs and revenue losses from RPW control in Italy, Spain, and France will reach millions of dollars.

[0007] The problem with this pest is that, although techniques exist to cure palm trees infested with RPW (pig weevil) are available, detecting the threat in its early stages (the first two to three weeks of the weevil larval stage) is challenging. This is because by the time a palm tree shows visible signs of distress (such as drooping crowns), it usually means the RPW infestation is already severe, and it's too late to save the tree. Therefore, governments in many countries are working to develop reliable and effective methods for early detection to address this issue.

[0008] Several methods have been reported to address this serious danger. For example, trained dogs have been used to sniff the gases released by contaminated palm trees during fermentation. Unfortunately, this gas-sensing process is not precise because its efficiency is affected by other volatile products. Alternatively, computer-based tomography systems could be used to screen for contaminated trees. However, this technique is impractical due to the slow and expensive nature of such screening methods.

[0009] The first detectable signal of an infested tree originates from the noise produced by weevil larvae feeding inside the trunk. Therefore, the most promising early detection methods rely on the use of acoustic sensors. More specifically, existing methods using acoustic sensors primarily involve inserting an acoustic probe (e.g., a microphone) into a hole drilled into the palm tree trunk, and then the probe records the sounds produced by the beetles in real time. The sound is recorded on a computer connected to the acoustic probe. The main difference between methods using acoustic sensors lies in the signal processing techniques implemented to process the recorded sound. However, all of these methods require on-site monitoring.

[0010] The disadvantages of the aforementioned acoustic methods include the damage to the trees caused by drilling holes in the trunks to insert acoustic probes, the labor required for drilling each tree, the impact of the holes on palm growth, and the possibility of other insects nesting in the holes in the trunks. Furthermore, since a red palm orchard may have thousands of such trees, providing each tree with an acoustic sensor with a wireless communication interface for continuous monitoring would significantly increase the cost of the entire RPW monitoring system.

[0011] A more advanced solution is proposed in international patent application PCT / IB2020 / 057865, filed on August 21, 2020 [2], and this solution is assigned to the assignee of this application (the entire contents of which are incorporated herein by reference). According to that application, the optical fiber 100 can be distributed along one or more trees 110, such as Figure 1 As shown. More specifically, if multiple trees are to be monitored, the distributed acoustic sensor (DAS) box 120 is connected to a single optical fiber 110, and the fiber extends to multiple trees 110-II, where I is any integer equal to or greater than 1. The same optical fiber 110 can be wound around each tree 110-I, for example at a height of about 1 m from the ground to the trunk, where the probability of detecting RPW larvae is highest. However, other heights can be used. Between the trees, depending on environmental conditions, the fiber optic cable can be laid on the ground or buried in the soil.

[0012] This RPW detection technology may still face challenges in the feasibility study phase on actual farms. This is because there may be numerous noise sources inside and around RPW-infested trees, which could trigger false alarms. For example, other insects, including scale-bearing and borer insects, also produce sound, which may interfere with the RPW detection. Furthermore, sounds and mechanical vibrations from birds, digging, watering, etc., will also be captured by the DAS system. These noise sources can cause false alarms, a major drawback of any sensing system.

[0013] Therefore, a new DAS system is needed that can not only detect the presence of RPW larvae, but also distinguish RPW sounds from ambient sounds with minimal infrastructure investment and support, in order to reduce the number of false alarms. Summary of the Invention

[0014] According to an embodiment, there is a fiber optic distributed acoustic sensing (DAS) system for detecting red brown weevils (RPWs), and the DAS system includes an optical fiber configured to be wound around a tree and a DAS box connected to the optical fiber. The DAS box includes a processing unit configured to receive a filtered Rayleigh signal reflected by the optical fiber and process the filtered Rayleigh signal through a neural network system to determine the presence of RPWs in the tree.

[0015] According to another embodiment, there is a method for detecting red palm weevils using an optical fiber DAS system, and the method includes transmitting a modulated signal through an optical fiber wrapped around a tree; receiving a reflected Rayleigh signal at a DAS box connected to the optical fiber, the reflected Rayleigh signal being a reflection of the modulated signal by the optical fiber; filtering the reflected Rayleigh signal using a system configured to receive and filter the reflected Rayleigh signal to generate a filtered Rayleigh signal; and processing the filtered Rayleigh signal using a neural network system to determine the presence of RPWs in the tree.

[0016] According to another embodiment, there is a DAS box for detecting red-brown weevils, and the DAS box includes a light source configured to generate continuous wave light, an optical modulator configured to modulate the amplitude of the continuous wave light emitted by the light source to generate modulated light, a circulator configured to receive the modulated light and inject the modulated light into an optical fiber, a processing unit configured to receive a filtered Rayleigh signal reflected by the optical fiber, and a fully connected artificial neural network (ANN) or convolutional neural network (CNN) system configured to process the filtered Rayleigh signal to determine the presence of RPWs in the tree. Attached Figure Description

[0017] To gain a more complete understanding of the present invention, reference is now made to the following description in conjunction with the accompanying drawings, wherein:

[0018] Figure 1 This is a schematic diagram of a distributed acoustic sensor system used to monitor trees;

[0019] Figure 2 This is a schematic diagram of a distributed acoustic sensor box configured to transmit modulated optical signals into an optical fiber;

[0020] Figures 3A to 3B Various implementations of the distributed acoustic sensor system are illustrated;

[0021] Figure 4 The configuration of a distributed acoustic sensor system for evaluating the effect of ambient noise on a recorded Rayleigh signal is shown.

[0022] Figure 5 This is a schematic diagram of a distributed acoustic sensor box configured to apply convolutional neural network processing to distinguish between signals associated with RPW larvae and signals associated with ambient noise.

[0023] Figure 6 The Rayleigh trajectory recorded using a distributed acoustic sensor box is shown;

[0024] Figure 7A and 7B The acoustic spectrum produced by RPW larvae is shown;

[0025] Figure 8A and 8B The signal recorded for an infested tree using a distributed acoustic sensor box is shown when different diameter sheaths are used on the optical fiber and there is no ambient noise.

[0026] Figure 9A and 9B The signal recorded for an infested tree using a distributed acoustic sensor box is shown when sheaths of different diameters are used on optical fibers with simulated wind-generated noise.

[0027] Figure 10A and 10B The signal recorded for an infested tree using a distributed acoustic sensor box is shown when sheaths of different diameters are used on optical fibers with simulated bird noise.

[0028] Figure 11 The configuration of an artificial neural network used in conjunction with a distributed acoustic sensor box is illustrated schematically.

[0029] Figure 12A and 12B The accuracy / loss of the artificial neural network and the confusion matrix are shown respectively when using a time dataset without environmental noise;

[0030] Figure 13A and 13B The accuracy / loss of the artificial neural network and the confusion matrix are shown separately when using a spectrum dataset without ambient noise.

[0031] Figure 14 The performance of artificial neural networks under various environmental conditions is shown;

[0032] Figure 15 The configuration of a convolutional neural network used with a distributed acoustic sensor box is illustrated schematically.

[0033] Figure 16A and 16B The accuracy / loss of the convolutional neural network and the confusion matrix are shown separately when using a time dataset without environmental noise;

[0034] Figure 17A and 17B The accuracy / loss of the convolutional neural network and the confusion matrix are shown separately when using a spectrum dataset without ambient noise.

[0035] Figure 18 The performance of convolutional neural networks under various environmental conditions is shown;

[0036] Figure 19 This is a flowchart of a method for determining when a tree is infected based on DAS boxes and convolutional neural networks. Detailed Implementation

[0037] The following description of the embodiments refers to the accompanying drawings. The same reference numerals in the different drawings denote the same or similar elements. The following detailed description does not limit the invention. Rather, the scope of the invention is defined by the appended claims. For simplicity, the following embodiments of a DAS system having a machine learning algorithm for separating noise generated by RPW larvae from ambient noise are discussed. However, the embodiments discussed below are not limited to determining the presence of RPW larvae or using only machine learning algorithms, but can be used to detect other organisms.

[0038] Throughout this specification, the reference to "one embodiment" or "an embodiment" means that a particular feature, structure, or characteristic associated with an embodiment is included in at least one embodiment of the disclosed subject matter. Therefore, the phrases "in one embodiment" or "in an embodiment" appearing in various places throughout the specification do not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any suitable manner in one or more embodiments.

[0039] According to an embodiment, a fiber-optic DAS is introduced, which is also programmed to run machine algorithms to separate ambient noise from noise generated by RPW (Reactive Power Rendering). Before discussing the new system, a possible component of a DAS box in an application is introduced. The underlying operating concept of a fiber-optic DAS relies on using a coherent (narrow linewidth) laser source to emit optical pulses into an optical fiber. As the optical pulse propagates along the fiber, it can be scattered within the fiber, generating a backscattered pulse that propagates in the opposite direction of the original optical pulse relative to the original optical pulse. When the interaction between the initial optical pulse and the fiber is elastic, the backscattered pulse can be Rayleigh scattering, and / or when the interaction is inelastic, the backscattered pulse can be Stokes-Raman and / or anti-Stokes-Raman. This backscattered pulse (Rayleigh, Stokes-Raman, or anti-Stokes-Raman) propagates backward along the fiber and is received at the fiber input port of the DAS box for storage and processing. By monitoring the intensity-time evolution of the recorded backscattered pulses, it is possible to (1) remove the portion of the signal associated with ambient noise, and (2) accurately calculate the location along the fiber affected by the acoustic signal and determine its frequency. Based on these responses, the processing section of the DAS system can be configured to determine the presence of the RPW in the tree. In the following text, for simplicity, the backscattered pulses are considered Rayleigh pulses. However, the embodiments discussed herein apply to any backscattered pulse.

[0040] in this regard, Figure 2 A fiber optic DAS system 200 is shown, capable of measuring strain applied to an optical fiber by changes in pressure, temperature, and / or acoustic noise. System 200 has two main components: a DAS housing 120 and an optical fiber 100 connected to the DAS housing 120. The DAS housing 120 includes all electronics for generating a light beam, transmitting the beam into the optical fiber, receiving reflected light from the optical fiber, and processing the reflected light to detect the RPW (Repeated Continuous Wave W). More specifically, the DAS housing 120 includes a light source 202 configured to generate coherent continuous wave (CW) light 204. For example, the light source 202 may be a laser or a light-emitting diode. The CW light 204 is converted into optical pulses 206 by an optical modulator 208. The optical modulator 208 is connected to a function generator 210. The function generator 210, controllable by a computing device 211, is configured to generate a mathematical function that is applied to the modulator to modulate the light 204. For example, Figure 2 The modulator 208 is shown applying a rectangular pulse 209 to light 204 to obtain an optical pulse 206 (or modulated light). Other shapes may be used for the pulse 209. The computing device 211 is also connected to an input / output module 213, which is capable of communication, for example, sending messages and receiving instructions / commands wirelessly or wiredly with a smartphone, personal computer, or any other electronic device.

[0041] Optionally, system 200 includes amplifier 212 for amplifying the modulated light 206 before it is transmitted into optical fiber 100 via circulator 214. Figure 2 An optical fiber 100 is schematically shown being guided through multiple trees 110-I. The circulator 214 can be, for example, a three-port or four-port optical device designed such that light entering from any port exits from the next port. This means that if light enters at the first port 214A, it is emitted from the second port 214B. However, if some of the emitted light 216 is reflected back to the circulator 214, it does not exit from the first port 214A, but rather from the third port 214C. This allows the reflected Rayleigh signal 222 to be directed to the photodetector 224 after reaching the circulator 214, rather than being sent to the amplifier 212.

[0042] As the optical pulse 216 propagates along the optical fiber 100, a Rayleigh signal 222 is backscattered from tree 110-I. In the reverse direction, the Rayleigh signal is recorded by photodetector 224 and then sampled using analog-to-digital converter (ADC) 226. Digital signal processing (DSP) 228 can be used to filter out the RPW sound in the frequency domain and accurately identify the location of the infected palm tree 110-I using, for example, a time-domain signal.

[0043] Fiber 100 can be a single-mode fiber (SMF). At the fiber input port, the trajectory of continuous Rayleigh backscattering is recorded in the time domain. Due to coherent interference of the signal reflected from the scattering center over the duration of the injected pulse, each Rayleigh trajectory has a similar spot-like profile. In the absence of intrusion along the fiber, i.e., no refractive index perturbation, the recorded Rayleigh trajectories are ideally identical. When an acoustic signal (e.g., the sound of a weevil larva) is applied at a certain location along the fiber, the effective refractive index of the fiber changes at that location, and therefore, intrusion can be sensed by observing the intensity fluctuations of the corresponding spots in the recorded trajectories.

[0044] By monitoring the intensity-time evolution of the recorded Rayleigh signal 222, the location along fiber 100 affected by the acoustic signal emitted by the RPW can be accurately calculated, thus determining the location of the RPW. This is for the purpose of early RPW detection. Figure 2The system 200 shown is superior to existing acoustic sensors in the literature for at least one of the following reasons: 1) it will provide uninterrupted monitoring of palm trees at a relatively low cost; 2) the typical fiber optic DAS has a sensing length of approximately 10 kilometers, which can cover a wide farm area; 3) multiple fibers can be connected to the same DAS box by using optical switches and time division multiplexing (TDM) in case a larger farm area needs to be monitored; 4) no intrusive sensing is required because the fiber optic cable would be wrapped around the palm from the outside; and 5) the fiber optic cable used for acoustic sensing can simultaneously monitor ambient temperature with a resolution of less than 0.1°C, which is crucial for controlling farm fires, another major problem worldwide.

[0045] In one embodiment, all components of system 200, except for fiber optic 100, can be housed in a single housing 240. This means that all optical components, such as lasers and photodetectors, are concentrated in a DAS box located, such as a control master station, while only fiber optic 100 is wound around palm tree 110-I in the form of an optical network.

[0046] Fiber 100 can be like Figure 1 As shown, it wraps itself around the trunk of tree 110-I, or as... Figure 3A and 3B As shown, it is protected by a cover layer. More specifically, Figure 3A The optical fiber 100 is shown to be completely surrounded by a filler material 310 (e.g., cotton or a similar fibrous material) and an outer shielding layer 312. The optical fiber assembly 300 may have a fabricated outer shielding layer 312 to completely encapsulate the filler material and the optical fiber. The outer shielding layer 312 may be made of a rigid material (e.g., a steel tube or pipe, a plastic tube) or a flexible material (e.g., a flexible metal or plastic tube, etc.). In one embodiment, the outer shielding layer 312 is rigid for one part of the optical fiber and flexible for another part.

[0047] For example, such as Figure 3BAs shown, the fiber optic assembly 300 has a first portion 302 completely buried underground 301 and a second portion 304 wrapped around the trunk 231 of the tree 110-I. The first portion 302 can be made rigid, while the second portion 304 can be made flexible to allow it to wrap around the trunk. The purpose of the filler material and the outer shielding layer 312 is to protect the fiber optic 100 from mechanical or thermal damage during field deployment. In this regard, note that there is heavy equipment moving around trees on farms to perform various agricultural procedures. In one application, the depth D1 of the first portion 302 buried underground is 0.5 to 2 m. For optimal efficiency, the height H1 at which the fiber optic assembly begins to wrap around the trunk 231 is approximately 0.5 to 3 m, as weevil larvae tend to attack the trunk at these heights. The length of the fiber optic 100 present in the second portion 304 is 1 to 10 m. Other values ​​of the parameters discussed here can be used. However, one or more portions of the fiber optic may be directly exposed to the environment at desired locations between trees to allow for the detection of ambient temperature and humidity.

[0048] The DAS system 200 may not be well-equipped enough to separate the noise of RPW larvae from ambient noise. Therefore, the DAS system 200 was modified and used in the laboratory to detect actual RPW larvae in palm trees while artificially generating ambient noise. Figure 4 The overall configuration of this embodiment is shown. The optical and electronic components of the DAS system 200 are assembled within the DAS box 120. The output light from the DAS box 120 is emitted into a single-mode fiber (SMF) 100, which is wound around a tree trunk 110. Inside the trunk of the tree 110, a speaker 410 is placed, which continuously plays sounds similar to the feeding sounds produced by a 12-day-old weevil larva. At a distance of approximately 1 meter from the tree 110, a fan 420 is provided to blow air toward the fiber 100 and the tree 110 to simulate wind in an actual palm tree farm. Another speaker 430 is installed away from the tree 110 to simulate birdsong. The ambient noise, consisting of wind and birdsong, along with the sound generated by the RPW, is recorded by the fiber 100.

[0049] The DAS box includes a 100-fiber interrogation system using a phase-sensitive optical time-domain reflectometer. [1] Design the inquiry system. In this embodiment, the following is used: Figure 5The system 500 shown includes a DAS box 502 and an optical fiber 100. The DAS box 502 is configured such that a narrow-linewidth laser generates a 100Hz linewidth continuous-wave (CW) light 204 with 40mW optical power. The laser 204 is then converted into optical pulses 206 using an acousto-optic modulator (AOM) 208, which generates pulses 206 with a width of 50ns and a repetition rate of 5kHz. The selected pulse width gives the DAS system 500 a spatial resolution of 5m. The power of the optical pulses 206 is amplified using an erbium-doped fiber amplifier (EDFA) 212, while its output light 213 is emitted through a circulator 214 into an approximately 2km long SMF 100. At a distance of approximately 1km from the SMF input port, a 5m long section of optical fiber is wound around the trunk of a tree 110. The backscattered signal 222 from the SMF is amplified by another EDFA 512 and then sent to a second circulator 514. The signal is then sent to a fiber Bragg grating (FBG) 510 to discard amplified spontaneous emission (ASE) noise. The filtered Rayleigh signal 520 is detected by a photodetector (PD) 522 and sampled by a digitizer 524 with a sampling rate of 200 MHz. The sampled Rayleigh signal 528 is recorded at 1-second intervals (5000 tracks per interval) and stored in a storage device 526. A processing unit 530 (e.g., a microprocessor) can be connected to the components discussed herein to coordinate their operation. Although Figure 5 The processing unit 530 is shown as part of the DAS box 502, but in one embodiment, the processing unit can be external to the DAS box, for example, as a remote server performing the calculations required by the DAS box. This embodiment uses two separate standard SMFs, 100 and 100', protected with different sheaths of 900 μm diameter (denoted as "JKT1") and 5 mm diameter (denoted as "JKT2"), respectively. In one embodiment, the thicknesses of JKT1 and JKT2 can vary by + / -20% and still achieve the same effect. Note that only one SMF is needed to determine the presence of RPW larvae. Fibers with either a JKT1 or JKT2 configuration are used as alternatives to identify the fiber best suited for detecting RPW larvae.

[0050] Figure 6An example of a Rayleigh track 600 recorded by an optical fiber DAS system 500 under ideal conditions (i.e., no ambient noise) is shown. The high-power signal 602 located at the beginning of the SMF corresponds to a Fresnel reflection from the front face of the SMF. Under these ideal conditions, when there is no refractive index perturbation along the SMF, the shape of the Rayleigh track 600 remains stationary in the time domain for all spatial points along the entire fiber. Therefore, the difference between the time-subsequent Rayleigh track and the initial reference signal is ideally zero. Conversely, the presence of larval sounds within the tree trunk can modulate the fiber refractive index at the tree location, causing a change in the corresponding time-based Rayleigh signal 610 only at that tree location. By applying a normalized differential method and a Fast Fourier Transform (FFT) to the time-based Rayleigh track, the location of the affected tree and the frequency of the larval sounds can be identified separately. However, when ambient noise is present, many other signals 620 appear in addition to the Rayleigh signal corresponding to the tree location, such as... Figure 6 As shown, these signals can mask signals produced by larvae or generate false alarms.

[0051] This section discusses possible ways to mitigate environmental noise (such as wind and birdsong) that may degrade the performance of fiber optic DAS systems during RPW testing. One proposed noise reduction technique involves applying a spectral bandpass filter to reduce noise levels in the recorded signal and further exploring various fiber optic sheaths that may sway due to wind.

[0052] First, we explored the spectral components of actual larval sounds. Specifically, we implanted a commercial recorder inside a genuinely infected tree trunk, next to approximately 12-day-old larvae, and recorded the noises they produced, such as... Figure 7A and 7B As shown. The age of the larvae could be well controlled through a human intervention process conducted in a safe research facility to avoid spreading RPW to other healthy trees. The larvae were observed to emit almost continuous sounds while chewing the trunk. Figure 7A and 7B Two representative examples of the larval sound power spectrum are shown, each corresponding to a 0.5-second recording interval. Based on the recordings in the figure, it can be reasonably assumed that the majority of the larval sound's optical power has frequencies below 800 Hz. Furthermore, it is important to discard low vibrational frequency components below 200 Hz to compensate for unavoidable mechanical vibrations and wind-induced tree swaying in a laboratory setting. Therefore, for the use of... Figure 4 and Figure 5 The configuration shown performs measurements by applying a [200Hz-800Hz] bandpass filter to the temporal vibration data collected using fiber optic DAS to enhance the system's signal-to-noise ratio (SNR).

[0053] Next, Figure 4The configuration shown has the fan 420 and noise speaker 430 turned off, while only the larval sounds produced by the speaker 410 implanted inside the trunk 110 remain on. Figure 8A and 8B Two representative examples of normalized differential time-domain signals 800 and 810 recorded using a DAS system 500 are shown, followed by the application of [200Hz-800Hz] bandpass filters when the SMF100 is used with JKT1 and JKT2 configurations, respectively. The two optical fibers accurately pinpointed the location of the infected tree approximately 1 km from the fiber input port. Additional noise signals sometimes appearing at the beginning of the SMF are a result of reflections from the fiber front end.

[0054] Next, turn off both speakers 410 and 430, and turn on only fan 420 to determine the effect of wind on the SMF100. Wind is considered a major noise source in open-air farms, especially since fiber optic cable 100 is wrapped around tree trunks, meaning the fiber is directly affected by the wind. Even with a [200Hz-800Hz] bandpass filter applied, the SMF100 with JKT1 is still affected by the wind, producing noise such as... Figure 9A The temporal vibrations shown are as follows. Low-frequency vibrations caused by trees swaying in the wind can be discarded by filtering out frequencies below 200Hz. However, when wind blows directly onto fiber 100, it causes the fiber to sway, the frequency of which depends on the thickness and material of the fiber sheath. For example... Figure 9A As shown, SMFs with JKT1 have a relatively small diameter (900 μm) and generate wind-induced vibration signals 900, which may resemble signals produced by larvae. This behavior could confuse machine learning algorithms when distinguishing between healthy and infected trees.

[0055] However, when the DAS box 502 is used with an SMF with a JKT2 configuration, the fiber rarely experiences jitter in the [200Hz-800Hz] range due to wind, because the JKT2 is relatively thick (5mm diameter). Figure 9B The recorded signal 910 is shown. Comparing two optical fibers with different sheaths helps determine the appropriate fiber optic cable for use in real farms in terms of reducing wind-generated noise. Furthermore, JKT2 has the added advantage of being more durable enough to withstand harsh agricultural conditions compared to the JKT1 configuration, and the internal SMF of JKT2 cannot be easily damaged, for example, by a farmer stepping on the fiber.

[0056] Next, the effect of bird noise on the optical fiber was investigated. Specifically, the larval sound speaker 410 and fan 420 were turned off, while the external speaker 430 was continuously played to produce bird sounds at a volume approximately equal to that present on the farm. Note that the two SMFs with JKT1 and JKT2 could not "hear" the bird sounds, as shown below. Figure 10A and 10B As shown. This is because the air between the speaker 430 and the fiber optic sheath significantly attenuates the vibrational energy of the bird sounds.

[0057] Based on these observations, a machine learning method trained through supervised learning has now been implemented in the DAS box 502 to distinguish between infected and healthy trees. Machine learning is able to reveal patterns associated with larval sounds while processing the large amounts of data generated by the DAS box 502. In this embodiment, the efficiency of classifying healthy and infected trees using time-domain and frequency-domain data as separate inputs to a neural network designed using a fully connected artificial neural network (ANN) and convolutional neural network (CNN) architecture was compared. Given the aforementioned advantages of the SMF with JKT2, this configuration has been used to classify healthy and infected trees using machine learning methods.

[0058] First, we discuss how to organize and label the time-domain and frequency-domain data of the ANN. As mentioned earlier, a 5-meter section of fiber optic cable 100 is wrapped around the tree, and the DAS box samples the data at a frequency of 200 MHz. Therefore, given the time of flight within the sensing system 502, the fiber optic section around the tree is represented by 10 spatial points. For each point, the digital converter reads at a 1-second cycle, meaning that 5000 readings are taken in the time domain per reading cycle due to the pulse repetition rate of 5 kHz. Since digital bandpass filters typically distort the time-domain signal by a short interval at the beginning, the first 250 time-domain readings for each spatial point are discarded. Therefore, the time data collected in each trial is organized into a vector of length 47500 (4750 consecutive time-domain readings × 10 spatial points). Conversely, by applying an FFT to the time-domain data of each spatial point, 2375 frequency components are obtained. Subsequently, the spectral data for each trial is organized into a vector of length 23750 (connecting the 2375 frequency components × 10 spatial points).

[0059] Based on the SNR value of the acoustic signal at the tree location, the data is labeled as either "infected" or "healthy" trees. In this embodiment, SNR is defined as the ratio between the root mean square (RMS) value of the time-domain signal at the tree location and the time-domain value at a reference fiber segment with a length of 5 m. The ability of the machine learning algorithm to classify infected and healthy trees was evaluated under both windless and windy conditions. Considering the first case, windless, only the speaker 410 inside the tree trunk is playing, while the external speaker 430 and fan 420 are stopped to generate a signal corresponding to an infected tree. If the SNR > 2 dB, which is the minimum acceptable SNR for the DAS system, the system records the signal and labels it as "infected." In this embodiment, 2000 instances of infected signals were collected, with the volume of the larval speaker 410 set to different values ​​below and above the level at which human hearing larval sounds is possible under acceptable ambient noise. Conversely, when the larval speaker 410 and fan are off, another 2000 samples are recorded for "healthy" signals. Regardless of whether the SNR value is higher or lower than the 2dB threshold, it will be recorded as an example of a "healthy" signal.

[0060] Next, data needs to be labeled when considering the presence of wind. Therefore, for this stage, the larval speaker 410 and fan 420 are simultaneously turned on to record examples of "infected" signals. When the SNR value exceeds a 2dB threshold, an additional 2000 distinct samples are recorded. Next, the larval speaker 410 is turned off, while the fan 420 remains on regardless of the SNR value to record an additional 2000 samples corresponding to healthy trees.

[0061] ANN models for processing time-domain and frequency-domain data have similar architectures

[1100] , such as Figure 11 As shown, the structure includes an input layer 1110, two hidden layers 1120 and 1130, and an output layer 1140. The number of nodes 1112 in the input layer 1110 matches the number of elements in the data vector, which is 47500 for the time domain data and 23750 for the frequency domain data. Furthermore, the first hidden layer 1120 contains 500 nodes, and the second hidden layer 1130 contains 50 nodes. The number of nodes was determined through trial and error to maximize classification accuracy. At the end of the fully connected ANN system, there is a single-node output layer 1140 for binary classification (infected or healthy). Regarding the activation function for each layer, this embodiment uses a rectified linear unit (ReLU) for the hidden layers and a sigmoid function for the output layer.

[0062] When wind is ignored (fan 420 off), the collected time / spectral data is split into a 60% (2400 examples) training dataset, a 20% (800 examples) validation dataset, and a 20% (800 examples) test dataset. In this case, Figure 12A This shows the evolution of training / validation accuracy and loss over time (epochs) when using temporal data. Figure 13A The same situation is shown for the spectral data. At the end of the training period, the validation accuracy values ​​for the time domain and frequency domain data (1200 and 1300) were 82.0% and 99.8%, respectively. When using, as... Figure 12A When using time-varying data, the final validation accuracy was lower than the accuracy during training (1210), indicating that the model failed to generalize. Conversely, as... Figure 13A As shown, for spectral data, the validation accuracy of 1300 and the training accuracy of 1310 completely overlap, which confirms that the ANN model learns the features well, rather than just memorizing the input data.

[0063] After the training and validation processes, the test dataset is used to estimate the performance of the two models. Figure 12B and 13B The confusion matrices are shown for both time-domain and frequency-domain data. Generally, the confusion matrix includes four main metrics, denoted as True Negative (TN), False Negative (FN), False Positive (FP), and True Positive (TP), which compare the actual target value with the value predicted by the machine learning model. In addition, the confusion matrix also includes other important performance metrics (accuracy, precision, recall, and false positives), and these are defined as follows:

[0064]

[0065]

[0066] as well as

[0067]

[0068] like Figure 12B and 13B As shown in the confusion matrix, the temporal data provides an overall classification accuracy of 83.6%, while the spectral data provides 99.3%. The performance parameters of the ANN in windless conditions are summarized in... Figure 14The table shown includes rows two and three. To achieve high discrimination accuracy between infected and healthy trees, it was found that using an ANN with larval sound spectrum data was optimal. This is attributed to the fact that the chewing sounds of larvae can shift within a 1-second recording frame, making it difficult for ANN models to learn with limited datasets. However, shifted temporal acoustic signals produce similar spectra, which facilitates the classification process using frequency domain data. Given these observations, in this embodiment, when considering the effects of wind, it was decided to rely on spectral components with the ANN to analyze subsequent, more complex scenarios.

[0069] The spectral data collected when fan 420 was on was divided as follows: 60% (2400 examples) for the training dataset, 20% (800 examples) for the validation dataset, and 20% (800 examples) for the test dataset. After the training and validation processes, the spectral test dataset was used to check the performance of the trained model. In this case, Figure 14 The third row of the table shows a summary of the ANN performance results. The ANN model provides an overall classification accuracy of 99.6%, slightly higher than the accuracy achieved without wind. Precision, recall, and false positives also show minor improvements compared to the windless case. These results demonstrate that the ANN model can perfectly learn the spectral patterns of larval sounds in both windy and windy conditions, while the small perturbations caused by wind slightly increase the model's robustness and generalization ability.

[0070] A more complex scenario involves combining two spectral datasets, one with wind and one without, as the noise source. This approach is reasonable given that airflow is intermittent in a real farm. Therefore, the two datasets were merged, resulting in a total of 8000 examples of infected and healthy trees. Similarly, the entire dataset was divided into three parts: 60% (4800 examples) for the training dataset, 20% (1600 examples) for the validation dataset, and 20% (1600 examples) for the test dataset. Using the combined data resulted in improvements in classification accuracy, precision, recall, and false positive rate compared to the previous two scenarios. Figure 14 As shown in the fourth row of the table, these results demonstrate that the performance of the ANN model improves with large amounts and diverse training data. Therefore, it can be concluded that the ANN model performs exceptionally well when using combined spectral data representing more realistic farm scenarios; however, the ANN model exhibits relatively poor performance with temporal input data.

[0071] CNNs are popular deep neural network architectures designed to be spatially invariant. In other words, they are insensitive to the location of features, which effectively handles temporal larvae sounds moving in the time domain. Furthermore, compared to fully connected ANNs, CNNs require relatively fewer parameters to train, making them easier and more efficient to train using the same amount of data. Because CNNs have proven highly effective for image classification, the temporal and spectral data are arranged in a two-dimensional matrix. Specifically, the time-domain and frequency-domain examples are arranged as 10 (spatial points) × 4750 (time readings) and 10 (spatial points) × 2375 (spectral components), respectively.

[0072] in this regard, Figure 15 The architecture of a CNN system 1500 for processing temporal and spectral input data is shown. System 1500 has an input layer 1502, two pairs of convolutional layers 1510 and a max-pooling layer 1520, a flattening layer 1530, a fully connected layer 1540, and an output layer 1550. The first convolutional layer 1512 uses the ReLU activation function and includes 32 filters of size 3×50 and stride 1×25, while the second convolutional layer 1522 uses the ReLU activation function and includes 64 filters of size (3×5) and stride (1×5). The two max-pooling layers 1514 and 1524 have the same 2×2 pooling size and the same 2×2 stride. After the flattening layer 1530, the fully connected layer 1540 uses the ReLU activation function and consists of 50 nodes. Similar to ANN 1100, the output layer 1550 of CNN system 1500 also has a node with a sigmoid activation function for binary classification (healthy or infected).

[0073] Now consider the data labeling and splitting of the CNN system 1500, using the same techniques and data volume as the fully connected ANN 1100. For the ideal case where there is no wind (i.e., fan 420 is off), Figure 16A and 17A The evolution of training / validation accuracy and loss over time is shown for both time and spectral data. After completing the training cycle, validation accuracy values ​​of 100% for time-domain data and 99.5% for frequency-domain data are obtained, as shown by curves 1600 and 1700, respectively. Furthermore, the two confusion matrices using the time and spectral test datasets are shown below. Figure 16B and 17B As shown. The confusion matrix results indicate that the CNN system performs exceptionally well on temporal data (1500 pairs), achieving 100.0% accuracy, while its performance on spectral data is slightly lower at 99.3%. When these results are compared with… Figures 12A to 13BWhen comparing the ANN configuration results shown, the CNN configuration significantly improves classification efficiency in the time domain. This demonstrates the two main advantages of the CNN model over the fully connected ANN model mentioned above: the spatial invariance of CNNs and the fewer parameters required for training. These results indicate that RPW can be detected in real time using CNNs without applying FFT to the time-domain data.

[0074] Figure 18 The table summarizes the performance of CNNs using temporal and spectral data, corresponding to scenarios (1) no wind, (2) wind, or (3) a mixture of both. The table shows that the CNN model exhibits superior performance in all scenarios, with a minimum classification accuracy of 98.3%. Considering that temporal data is easier to process than spectral data requiring an additional FFT step, in one embodiment, a CNN model and temporal data are preferably used to detect RPWs. For this implementation, the CNN model with temporal data provides 99.7% accuracy, 99.5% precision, 99.9% recall, and 0.5% false positives. Figure 18 As shown in the third row of the table, high accuracy and low false positives confirm the reliability of the CNN model in distinguishing between healthy and infected trees. On the other hand, high recall represents the CNN model's strong ability and sensitivity to separate "infected" signals from a mixed set of "healthy" and "infected" trees.

[0075] In an application Figure 5 The DAS box 502 shown stores the data recorded from the SMF100 in the memory 526, and uses the processing unit 530 to implement an ANN or CNN model. Therefore, Figure 15 The system 1500 shown can be implemented in the processing unit 530 and stored in the memory 526 to determine whether the recorded signal is associated with the RPW or with ambient noise.

[0076] Now about Figure 19 A method for detecting reactive power trees (RPWs) using a fiber optic DAS system 500 is discussed. The method includes the following steps: Step 1900: transmitting a modulated signal through an optical fiber wrapped around a tree; Step 1902: receiving a reflected Rayleigh signal at a DAS box connected to the optical fiber, the reflected Rayleigh signal being a reflection of the modulated signal by the optical fiber; Step 1904: filtering the reflected Rayleigh signal using a fiber Bragg grating configured to receive and filter the reflected Rayleigh signal to generate a filtered Rayleigh signal; and Step 1906: processing the filtered Rayleigh signal using a convolutional neural network (CNN) system to determine the presence of RPWs in the tree.

[0077] The method may further include the steps of modulating the amplitude of continuous-wave light emitted by the light source to generate a modulated signal, and / or recording the light intensity of the filtered Rayleigh signal using a photodetector, and digitizing the recorded light intensity using a digital converter. In one application, the CNN system has an input layer, a first convolutional layer and a second convolutional layer, a first pooling layer and a second pooling layer, a flattened layer, a fully connected layer, and an output layer. Each of the first and second convolutional layers uses a rectified linear unit (ReLU) activation function. The fully connected layer uses a ReLU activation function. The output layer uses a sigmoid activation function. In one application, the sheath diameter of the optical fiber is approximately 5 mm. The method may further include the steps of applying a Fourier transform to the filtered Rayleigh signal to obtain a spectrum, and using this spectrum with the CNN system to detect the presence of the RPW.

[0078] The disclosed embodiments provide an optical DAS system for monitoring tree infection and using a neural network to distinguish the sounds produced by actual larvae from environmental noise. It should be understood that this description is not intended to limit the invention. Rather, the embodiments are intended to cover alternatives, modifications, and equivalents included within the spirit and scope of the invention as defined by the appended claims. Furthermore, numerous specific details are set forth in the detailed description of the embodiments to provide a full understanding of the claimed invention. However, those skilled in the art will appreciate that various embodiments can be implemented without these specific details.

[0079] Although the features and elements of this embodiment are described in specific combinations in the embodiments, each feature or element may be used alone without the other features and elements of the embodiment, or in combination with or without the other features and elements disclosed herein.

[0080] This written description uses examples of the disclosed subject matter to enable any person skilled in the art to practice the examples of the disclosed subject matter, including making and using any device or system and performing any incorporated methods. The patent scope of this subject matter is defined by the claims and may include other examples that would occur to a person skilled in the art. Such other examples are intended to be within the scope of these claims.

[0081] References

[0082] [1]Bao, X.; Zhou, DP; Baker, C.; Chen, L.

[0083] [2] International Patent Application PCT / IB2020 / 057865.

Claims

1. A fiber optic distributed acoustic sensing (DAS) system (500) for detecting red-brown weevils (RPW), the DAS system (500) comprising: Optical fiber (100), configured to be wound around the tree; as well as DAS box (502), which is connected to the optical fiber (100), the DAS box includes: A light source (202) is configured to generate continuous wave light; An optical modulator (208) is configured to modulate the amplitude of continuous wave light emitted by the light source (202) to generate modulated light; A circulator (214) configured to receive the modulated light and inject the modulated light into an optical fiber (100); The processing unit (530) is configured to receive the filtered Rayleigh signal (520) reflected by the optical fiber (100); The processing unit (530) is configured as follows: Receive the filtered Rayleigh signal (520) reflected by the optical fiber (100), and The presence of RPW in the tree is determined by processing the filtered Rayleigh signal (520) through a neural network system (1100, 1500). The DAS box further includes: A fiber Bragg grating is configured to receive and bandpass filter reflected Rayleigh signals to generate a filtered Rayleigh signal, wherein the bandpass filter allows reflected Rayleigh signals from 200 Hz to 800 Hz to pass through, and The neural network system labels data as "infected" or "healthy" trees based on the signal-to-noise ratio of the acoustic signal at the tree location, wherein the signal-to-noise ratio is the ratio between 1) the root mean square value of the time-domain signal at the tree location and 2) the time-domain signal value at the reference fiber segment.

2. The fiber optic DAS system according to claim 1, wherein, The DAS box also includes: A photodetector configured to record the light intensity of a filtered Rayleigh signal; and A digital converter configured to digitize the recorded light intensity.

3. The fiber optic DAS system according to claim 1, wherein, The neural network system is a convolutional neural network (CNN) system, which has an input layer, a first convolutional layer and a second convolutional layer, a first pooling layer and a second pooling layer, a flattened layer, a fully connected layer and an output layer.

4. The fiber optic DAS system according to claim 3, wherein, Each of the first and second convolutional layers uses the Rectified Linear Unit (ReLU) activation function, the fully connected layer uses the ReLU activation function, and the output layer uses the sigmoid activation function.

5. The fiber optic DAS system according to claim 1, wherein, The neural network system is a fully connected artificial neural network (ANN) with one input layer, two hidden layers, and one output layer.

6. The fiber optic DAS system according to claim 5, wherein, Each of the two hidden layers uses a rectified linear unit as the activation function, and the output layer uses the sigmoid function.

7. The fiber optic DAS system according to claim 1, wherein, The diameter of the sheath of the optical fiber is approximately 5 mm.

8. A method for detecting red-brown weevil RPWs using a fiber optic distributed acoustic sensing (DAS) system (500), the method comprising: A modulated signal (206) is transmitted (1900) through an optical fiber (100) wrapped around the tree; A reflected Rayleigh signal (222) is received (1902) at a DAS box (502) connected to an optical fiber (100), the reflected Rayleigh signal (222) being a reflection of the modulation signal (206) by the optical fiber (100); The reflected Rayleigh signal (222) is filtered using a fiber Bragg grating bandpass filter (1904), the fiber Bragg grating being configured to receive and filter the reflected Rayleigh signal (222) to generate a filtered Rayleigh signal (520), wherein the bandpass filter allows the reflected Rayleigh signal from 200 Hz to 800 Hz to pass through; and The filtered Rayleigh signal (520) is processed (1906) by using a neural network system (1100, 1500) to label the data as "infected" or "healthy" trees based on the signal-to-noise ratio of the acoustic signal at the tree location, wherein the signal-to-noise ratio is the ratio between 1) the root mean square value of the time-domain signal at the tree location and 2) the time-domain signal value at the reference fiber segment.

9. The method according to claim 8, further comprising: The amplitude of a continuous wave of light emitted by a light source is modulated to generate a modulated signal.

10. The method of claim 8, further comprising: The light intensity of the filtered Rayleigh signal is recorded using a photodetector; as well as The recorded light intensity is digitized using a digital converter.

11. The method according to claim 8, wherein, The neural network system is a convolutional neural network (CNN) system, which has an input layer, a first convolutional layer and a second convolutional layer, a first pooling layer and a second pooling layer, a flattened layer, a fully connected layer and an output layer.

12. The method according to claim 11, wherein, Each of the first and second convolutional layers uses the Rectified Linear Unit (ReLU) activation function, the fully connected layer uses the ReLU activation function, and the output layer uses the sigmoid activation function.

13. The method according to claim 8, wherein, The neural network system is a fully connected artificial neural network (ANN), which has one input layer, two hidden layers, and one output layer.

14. The method according to claim 13, wherein, Each of the two hidden layers uses a rectified linear unit as the activation function, and the output layer uses the sigmoid function.

15. The method according to claim 8, wherein, The diameter of the sheath of the optical fiber is approximately 5 mm.

16. The method of claim 8, further comprising: The Fourier transform is applied to the filtered Rayleigh signal to obtain the spectrum; as well as The spectrum was used in conjunction with a CNN system to detect the presence of RPW.

17. A distributed acoustic sensing (DAS) box (502) for detecting red-brown weevil RPW, the DAS box (502) comprising: A light source (202) is configured to generate continuous wave light; An optical modulator (208) is configured to modulate the amplitude of continuous wave light emitted by the light source (202) to generate modulated light; A circulator (214) configured to receive the modulated light and inject the modulated light into an optical fiber (100); The processing unit (530) is configured to receive the filtered Rayleigh signal (520) reflected by the optical fiber (100); as well as A fully connected artificial neural network (ANN) or convolutional neural network (CNN) system (1500) is configured to process a filtered Rayleigh signal (520) to determine the presence of RPWs in the tree. The DAS box further includes: A fiber Bragg grating is configured to receive and bandpass filter reflected Rayleigh signals to generate a filtered Rayleigh signal, wherein the bandpass filter allows reflected Rayleigh signals from 200 Hz to 800 Hz to pass through, and The fully connected artificial neural network (ANN) or convolutional neural network (CNN) system labels data as "infected" or "healthy" trees based on the signal-to-noise ratio (SNR) of the acoustic signal at the tree location, wherein the SNR is the ratio between 1) the root mean square value of the time-domain signal at the tree location and 2) the time-domain signal value at the reference fiber segment.

Images