A three-mode fusion rainfall tendency identification method, system, terminal and medium based on a humidity adaptive optical storage device

Abstract

The application discloses a three-mode fusion rainfall tendency identification method and system based on a humidity adaptive optical storage device, a terminal and a medium, and the method comprises the following steps: collecting light signals, humidity signals and wind speed signals, converting the light signals into ultraviolet light pulse intensity, converting the wind speed signals into ultraviolet light pulse frequency and pulse width, and combining the wind speed signals to convert into a standardized ultraviolet light pulse sequence; projecting the standardized ultraviolet light pulse sequence to a reservoir array, obtaining the current response curve output by the reservoir array, and obtaining a high-dimensional feature vector; converting the high-dimensional feature vector into a digital feature vector based on an FPGA; inputting the digital feature vector into the readout layer of the FPGA for binary classification identification of rainfall and no rainfall, and outputting a rainfall tendency identification result. The application breaks through the architecture and performance bottleneck of a traditional rainfall prediction system, realizes real-time collection of optical, humidity and wind speed three-mode signals, adaptive feature fusion and accurate rainfall tendency identification.

Description

Technical Field

[0001] This invention relates to the field of rainfall tendency recognition technology, and in particular to a three-modal fusion rainfall tendency recognition method, system, terminal and medium based on humidity adaptive optical storage device. Background Technology

[0002] Rainfall trend identification is an indispensable core technology in meteorological monitoring, disaster prevention and mitigation, agricultural production, and urban operation and maintenance. Accurate rainfall trend prediction can provide scientific decision-making basis for flood control scheduling, data support for farmland irrigation planning, and risk warning for urban traffic management. It plays an irreplaceable role in reducing casualties and economic losses caused by meteorological disasters and improving social productivity. With the rapid development of IoT and edge computing technologies, the traditional centralized rainfall forecasting model relying on large meteorological stations can no longer meet the needs of distributed and scenario-based real-time monitoring. Portable, low-power, and high-response-speed rainfall trend identification devices have become the core direction of industry development.

[0003] Currently, mainstream rainfall forecasting technologies mainly rely on a combination of traditional sensor arrays and digital algorithms. This involves collecting key environmental parameters such as light intensity, ambient humidity, and wind speed, and then using traditional neural networks and machine learning models for data processing and trend analysis. Optical sensing technology, due to its advantages of fast response speed, high detection accuracy, and strong resistance to electromagnetic interference, has become a commonly used method for environmental parameter collection. Humidity, as a direct precursor to rainfall, is crucial for improving rainfall recognition accuracy through its synergistic analysis with light intensity and wind speed. However, existing rainfall forecasting technologies still suffer from several insurmountable technical shortcomings, severely limiting their application in edge computing scenarios. These shortcomings include: discrete design and fixed device response characteristics, resulting in high decision latency, poor environmental adaptability, high energy consumption, and low multimodal fusion efficiency.

[0004] Therefore, existing technologies still have shortcomings. Summary of the Invention

[0005] To address the aforementioned deficiencies in existing technologies, this invention provides a three-modal fusion rainfall tendency identification method, system, terminal, and medium based on humidity-adaptive optical storage devices. The technical solution adopted by this invention is as follows:

[0006] In a first aspect, the present invention provides a three-modal fusion rainfall tendency identification method based on a humidity-adaptive optical storage device, the method comprising:

[0007] Light intensity, humidity, and wind speed signals are collected, and optical encoding is performed on the light intensity and wind speed signals based on the optical encoding module. The light intensity is converted into ultraviolet light pulse intensity, the wind speed is converted into ultraviolet light pulse frequency and pulse width, and the ultraviolet light pulse intensity, ultraviolet light pulse frequency and pulse width, and humidity are converted into a standardized ultraviolet light pulse sequence.

[0008] The standardized ultraviolet light pulse sequence is projected onto a reservoir array composed of several humidity-adaptive photo-storage devices, and the current response curve output by each humidity-adaptive photo-storage device in the reservoir array is obtained. A high-dimensional feature vector is obtained based on the current response curve.

[0009] The high-dimensional feature vector is converted into an analog voltage feature signal based on the FPGA, and the analog voltage feature signal is converted into a digital signal based on the FPGA to obtain a digital feature vector;

[0010] The digital feature vector is input into the readout layer of the FPGA. The digital feature vector is then subjected to binary classification for rainfall and no rainfall by a pre-trained weight matrix in the readout layer of the FPGA, and the rainfall tendency identification result is output. The weight matrix is ​​pre-trained using a soft maximum cross-entropy optimization algorithm and is used to establish a mapping relationship between the digital feature vector and the rainfall label.

[0011] In one implementation, the fabrication process of the humidity-adaptive optical storage device includes:

[0012] Using a silicon wafer as the substrate, the silicon wafer is pre-treated, and a combined process of ultraviolet lithography and electron beam lithography is used to imprint a double layer of photoresist to define the patterned area for the deposition of the bottom electrode film.

[0013] A bottom electrode film is deposited on the pretreated silicon wafer through a thermal evaporation process. The bottom electrode film is a composite electrode layer of chromium and silver.

[0014] The excess metal layer was removed by N-methyl-2-pyrrolidone separation process to obtain a bottom electrode pattern with regular edges;

[0015] The wurtzite nanowires were transferred onto the prepared bottom electrode pattern, and polyvinyl chloride was selected as a support layer to fix and protect the wurtzite nanowires and the bottom electrode pattern, thus obtaining the assembled sample.

[0016] The assembled sample was immersed in ultrapure water for 25 minutes to remove residual impurities on the sample surface and to fully dissolve the support layer, thus obtaining a humidity-adaptive optical storage device.

[0017] In one implementation, the method further includes:

[0018] After fabricating the humidity-adaptive optical storage device, under a fixed humidity condition, the optical response characteristics of the humidity-adaptive optical storage device under different readout voltages, different ultraviolet light pulse intensities, and different ultraviolet light pulse widths were tested.

[0019] Under the conditions of fixed readout voltage, fixed ultraviolet light pulse intensity, and fixed ultraviolet light pulse width, the photocurrent relaxation dynamics of the humidity adaptive photoretention device were tested in different humidity ranges.

[0020] The humidity-adaptive photo-storage device can transform humidity from a simple input signal into a reservoir dynamic regulation variable, and can achieve environmental adaptation in the calculation process through humidity-dependent photocurrent relaxation characteristics.

[0021] In one implementation, the illumination signal and the wind speed signal are optically encoded based on an optical encoding module, converting the illumination signal into ultraviolet light pulse intensity and the wind speed signal into ultraviolet light pulse frequency and pulse width, including:

[0022] After normalizing the light intensity signal, according to the first mapping relationship, the normalized light intensity signal is mapped to a range of 0.03~26.36 mW. cm -2 The intensity of the ultraviolet light pulse;

[0023] According to the second mapping relationship, the wind speed signal is converted into an ultraviolet light pulse frequency of 0.5~5 Hz and an ultraviolet light pulse width of 0.2~0.01s.

[0024] In one implementation, the first mapping relationship is:

[0025] Ultraviolet pulse intensity = 0.88 × normalized irradiance + 0.03;

[0026] The second mapping relationship is: v = 0.1f + 0.4, t = -0.02f + 0.2;

[0027] Where v is the wind speed, f is the ultraviolet pulse frequency, and t is the ultraviolet pulse width.

[0028] In one implementation, the standardized ultraviolet light pulse sequence is projected onto a reservoir array composed of several humidity-adaptive photoretention devices, and the current response curve output by each humidity-adaptive photoretention device in the reservoir array is obtained. A high-dimensional feature vector is then obtained based on the current response curve, including:

[0029] The standardized ultraviolet light pulse sequence is projected onto the reservoir array, which consists of an 8×8 humidity-adaptive photoretention device;

[0030] The current response curve of each humidity-adaptive photoreceptor in the reservoir array is obtained, and the peak value, relaxation time constant, and current accumulation value of the current response curve are extracted to obtain a 64-dimensional high-dimensional feature vector. At the same time, the humidity signal is used as a control variable to adjust the photocurrent relaxation dynamics of each humidity-adaptive photoreceptor in real time to eliminate humidity interference and complete the feature extraction and fusion of the three-mode signal.

[0031] In one implementation, the high-dimensional feature vector is converted into an analog voltage feature signal based on an FPGA, and the analog voltage feature signal is then converted into a digital signal based on the FPGA to obtain a digital feature vector, including:

[0032] Based on FPGA control of the CMOS switching transistors of the reservoir array, independent addressing of individual devices is achieved, and the high-dimensional feature vector output by each humidity-adaptive photomemory device is converted into an analog voltage feature signal.

[0033] Based on the analog-to-digital converter module in the FPGA, the analog voltage characteristic signal is converted into a digital signal at a sampling frequency of 100kHz to obtain a digital feature vector.

[0034] Secondly, embodiments of the present invention also provide a three-modal fusion rainfall tendency identification system based on a humidity-adaptive optical storage device, wherein the system is used to implement the steps of the three-modal fusion rainfall tendency identification method based on a humidity-adaptive optical storage device as described in any of the above solutions, and the system includes:

[0035] A multimodal signal acquisition and optical encoding module is used to acquire light signals, humidity signals, and wind speed signals, and to perform optical encoding on the light signals and wind speed signals based on the optical encoding module, converting the light signals into ultraviolet light pulse intensity, the wind speed signals into ultraviolet light pulse frequency and pulse width, and converting the ultraviolet light pulse intensity, the ultraviolet light pulse frequency and pulse width, and the humidity signals into a standardized ultraviolet light pulse sequence;

[0036] The HAOR reservoir array module is used to project the standardized ultraviolet light pulse sequence onto a reservoir array composed of several humidity-adaptive photo-storage devices, obtain the current response curve output by each humidity-adaptive photo-storage device in the reservoir array, and obtain a high-dimensional feature vector based on the current response curve.

[0037] An FPGA control module is used to convert the high-dimensional feature vector into an analog voltage feature signal based on the FPGA, and to convert the analog voltage feature signal into a digital signal based on the FPGA to obtain a digital feature vector;

[0038] The FPGA readout layer module is used to input the digital feature vector into the readout layer of the FPGA, and to perform binary classification recognition of rainfall and no rainfall on the digital feature vector through a pre-trained weight matrix in the readout layer of the FPGA, and output the rainfall tendency recognition result. The weight matrix is ​​pre-trained using a soft maximum cross-entropy optimization algorithm and is used to establish the mapping relationship between the digital feature vector and the rainfall label.

[0039] Thirdly, embodiments of the present invention also provide a terminal, wherein the terminal includes a memory, a processor, and a three-modal fusion rainfall tendency identification program based on a humidity-adaptive light storage device stored in the memory and executable on the processor. When the processor executes the three-modal fusion rainfall tendency identification program based on a humidity-adaptive light storage device, it implements the steps of the three-modal fusion rainfall tendency identification method based on a humidity-adaptive light storage device in any of the above-mentioned schemes.

[0040] Fourthly, embodiments of the present invention also provide a computer-readable storage medium, wherein the computer-readable storage medium stores a three-modal fusion rainfall tendency identification program based on a humidity-adaptive optical storage device, the three-modal fusion rainfall tendency identification program based on a humidity-adaptive optical storage device implementing the steps of the three-modal fusion rainfall tendency identification method based on a humidity-adaptive optical storage device as described in any of the above schemes on the computer-readable storage medium.

[0041] Beneficial Effects: Compared with existing technologies, this invention provides a three-modal fusion rainfall tendency identification method based on humidity-adaptive optical storage devices. First, it collects illumination signals, humidity signals, and wind speed signals. Then, based on an optical encoding module, it performs optical encoding on the illumination and wind speed signals, converting the illumination signal into ultraviolet (UV) pulse intensity and the wind speed signal into UV pulse frequency and pulse width. Finally, it converts the UV pulse intensity, frequency, and width, along with the humidity signal, into a standardized UV pulse sequence. Next, it projects the standardized UV pulse sequence onto a reservoir array composed of several humidity-adaptive optical storage devices, obtaining the current response curve output by each device. Based on these current response curves, a high-dimensional feature vector is obtained. Then, based on an FPGA, the high-dimensional feature vector is converted into an analog voltage feature signal, and further converted into a digital signal to obtain a digital feature vector. Finally, the digital feature vector is input into the readout layer of the FPGA. The digital feature vector is then subjected to binary classification for rainfall and no rainfall by a pre-trained weight matrix in the readout layer of the FPGA, and the rainfall tendency recognition result is output. The weight matrix is ​​pre-trained using a soft maximum cross-entropy optimization algorithm and is used to establish a mapping relationship between the digital feature vector and the rainfall label.

[0042] This invention breaks through the architectural and performance bottlenecks of traditional rainfall forecasting systems, realizing real-time acquisition of three-modal signals of optics, humidity, and wind speed, adaptive feature fusion, and accurate identification of rainfall trends. It has broad application prospects in portable meteorological monitoring, IoT edge nodes, and outdoor environmental sensing, and can promote the development of meteorological sensing technology towards "adaptive, low-power, and highly integrated" directions, providing core technical support for disaster prevention and mitigation, smart agriculture, and urban operation and maintenance. Attached Figure Description

[0043] Figure 1 This is a flowchart of a preferred embodiment of the three-modal fusion rainfall tendency identification method based on humidity adaptive optical storage device according to the present invention.

[0044] Figure 2 This is a flowchart illustrating the fabrication process of the humidity-adaptive optical storage device according to an embodiment of the present invention.

[0045] Figure 3 This refers to the photoresponse current under different readout voltages and illumination times in the three-modal fusion rainfall tendency identification method based on humidity adaptive optical storage devices according to an embodiment of the present invention.

[0046] Figure 4The schematic diagram shows the response current distribution under different readout voltages and humidity in the three-modal fusion rainfall tendency identification method based on humidity adaptive optical storage device according to an embodiment of the present invention.

[0047] Figure 5 This is a schematic diagram illustrating the optical encoding principle of wind speed signals in the three-modal fusion rainfall tendency identification method based on humidity adaptive optical storage devices according to an embodiment of the present invention.

[0048] Figure 6 This is a schematic diagram of the three-mode signal fusion response current curve in the three-mode fusion rainfall tendency identification method based on humidity adaptive optical storage device according to an embodiment of the present invention.

[0049] Figure 7 This is a schematic diagram of the meteorological data preprocessing process in the three-modal fusion rainfall tendency identification method based on humidity adaptive optical storage device according to an embodiment of the present invention.

[0050] Figure 8 This is a schematic diagram showing the distribution of 27 environmental category features in the three-modal fusion rainfall tendency identification method based on humidity adaptive optical storage device according to an embodiment of the present invention.

[0051] Figure 9 This is a schematic diagram of the HAOR-FPGA integrated system architecture constructed in this embodiment of the invention for realizing rainfall tendency identification.

[0052] Figure 10 This is the confidence distribution curve of multimodal features in the three-modal fusion rainfall tendency identification method based on humidity adaptive optical storage device according to an embodiment of the present invention.

[0053] Figure 11 This is a normalized comparison chart of rainfall prediction sample results using the three-modal fusion rainfall tendency identification method based on humidity adaptive optical storage device according to an embodiment of the present invention.

[0054] Figure 12 This is a schematic diagram of the structure of a three-modal fusion rainfall tendency identification system based on a humidity-adaptive optical storage device according to an embodiment of the present invention.

[0055] Figure 13 A schematic diagram of the terminal provided in an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0057] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content, operations, or steps, nor does it require execution in the described order. For example, some operations or steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0058] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0059] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. For example, "first control information" and "second control information" are only used to distinguish different control information and do not limit their order.

[0060] Those skilled in the art will understand that the words "first" and "second" do not limit the quantity or the order of execution, and that the words "first" and "second" do not necessarily imply that they are different.

[0061] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0062] The shortcomings of existing rainfall forecasting technologies are specifically manifested in the following aspects:

[0063] 1. Decoupled hardware architecture, resulting in high decision-making latency.

[0064] Traditional rainfall forecasting systems strictly follow the principles of von Neumann. The Neumann architecture employs a separate design for sensors, memory, and processor. Light signals are acquired by optical sensors, humidity signals by humidity sensors, and wind speed signals by wind speed sensors. These three types of signals must be transmitted to the processor via a bus. After processing through multiple stages, including analog-to-digital conversion, data caching, feature extraction, and model inference, the prediction result is output. Taking a typical traditional system as an example, the latency from acquisition to transmission to the processor for a single-modal signal is approximately 5-10ms, the latency for synchronization and data preprocessing of three-modal signals is approximately 15-20ms, and the latency for model inference is approximately 10-30ms. The total latency of the entire process generally exceeds 50ms, which cannot meet the second-level response requirements of sudden weather events such as short-duration heavy rainfall, easily leading to delayed flood control decisions and missing the optimal response time.

[0065] 2. Poor environmental adaptability, with significant humidity interference issues.

[0066] The response characteristics of traditional optical sensors are determined by the physical properties of the materials and the fabrication process, and cannot be dynamically adjusted once fabricated. Humidity, as a core indicator for rainfall prediction, is not only a predictive signal but also interferes with the sensing accuracy of devices through physical adsorption and chemical reactions. Increased humidity causes water films to form on the surface of sensitive materials in traditional optical sensors, altering light absorption efficiency and charge transport characteristics, increasing photocurrent detection errors by up to 30% or more. At the same time, humidity changes can cause baseline drift in devices, leading to noise superposition during multi-parameter fusion and a significant reduction in feature extraction efficiency. In high humidity (relative humidity above 60%) or rapidly fluctuating humidity scenarios, the recognition accuracy drops significantly, making it difficult to adapt to complex and changeable outdoor weather environments.

[0067] 3. The contradiction between energy consumption and integration is difficult to reconcile.

[0068] Traditional rainfall prediction models require independent feature extraction from multimodal data such as illumination, humidity, and wind speed. This involves extracting spatial features through convolutional layers, temporal features through recurrent layers, and then using fully connected layers for feature fusion and classification. The resulting model structure is complex and consumes significant computational resources. For example, a traditional CNN-LSTM (Convolutional Neural Network-Long Short-Term Memory) model consumes approximately 1-5 mJ of energy per sample inference, which is insufficient for the low-power requirements of portable devices (typically requiring less than 1 μJ of energy per inference). Furthermore, the limited functionality of traditional devices necessitates separate packaging of sensor arrays, storage modules, and computing chips, resulting in low integration and device sizes typically on the order of cubic decimeters, making them unsuitable for miniaturized applications such as IoT edge nodes and wearable devices.

[0069] 4. The reservoir computing architecture is rigid, lacking dynamic control capabilities.

[0070] Existing reservoir-based meteorological forecasting systems rely on fixed device structures to determine reservoir dynamics (such as relaxation time and nonlinear response intensity). These systems can only treat environmental parameters like humidity and wind speed as simple input signals for feature mapping, neglecting the adaptive control value of these environmental parameters on the computational process. For example, models trained in traditional reservoir computing systems under low humidity conditions experience a 20%–40% drop in prediction accuracy under high humidity conditions due to changes in device response characteristics. Training models separately for different humidity scenarios further increases storage overhead and model switching latency, reducing system practicality and making it unable to adapt to the dynamic changes in complex meteorological scenarios.

[0071] 5. Low efficiency of multimodal signal fusion

[0072] Traditional multimodal signal fusion systems rely on digital algorithms, requiring the conversion of each modal signal into digital feature vectors before fusion through weighted summation and concatenation. This process suffers from feature loss and computational redundancy. Furthermore, fusing the time-domain features of wind speed signals (such as pulse frequency and duration) and the amplitude features of humidity signals (such as relative humidity values) in the digital domain necessitates complex normalization and standardization processes to eliminate dimensional differences. This not only increases computational complexity but also leads to the loss of cross-modal feature correlation information, affecting prediction accuracy and hindering the full realization of the synergistic effects of multimodal signals.

[0073] It is evident that in existing rainfall forecasting technologies, the separate design of sensors, memory, and processors results in a total latency exceeding 50ms, failing to meet the real-time response requirements of sudden weather events. Humidity, as a core forecasting indicator and source of interference, has a significant negative impact on sensing accuracy, leading to a substantial decrease in recognition accuracy under high humidity conditions. The complex structure and discrete hardware of traditional models result in high system energy consumption, low integration, and an inability to adapt to miniaturized edge applications. Furthermore, the existing reservoir computing architecture is rigid, only accepting environmental parameters as input signals and ignoring their value in regulating the computation process, resulting in insufficient robustness in forecasting under complex weather scenarios.

[0074] To address the problems of existing technologies, this invention provides a trimodal fusion rainfall tendency identification method based on humidity-adaptive optical storage (HAOR) devices. It innovatively reconstructs humidity from a simple input signal into a reservoir dynamics regulation variable, utilizing the optical, humidity, and wind speed trimodal response characteristics of HAOR devices to construct an integrated "sensing-storage-computing" edge architecture. This system achieves physical-level fusion of the three modal signals through optical coding technology, directly performing feature extraction and dynamic optimization within the HAOR device. Combined with an FPGA (Field-Programmable Gate Array) readout layer, it achieves low-latency, low-power rainfall tendency prediction.

[0075] This embodiment innovatively designs a humidity-adaptive optical storage and computing framework within the sensor based on a humidity-adaptive optical storage (HAOR) device. Relying on the light-humidity dual-sensitivity characteristics of wurtzite ZnO (zinc oxide) nanowires and the humidity-tunable reservoir dynamics, it integrates the optical encoding technology of a passive anemometer to construct an edge architecture that integrates "sensing-storage-computing" to achieve real-time, low-power, and high-precision prediction of rainfall trends.

[0076] In practical applications, the humidity-adaptive optical storage (HAOR) device in this embodiment is a planar dual-terminal memristor. Its core innovation lies in precisely controlling the fabrication process to enable wurtzite-phase ZnO (zinc oxide) nanowires to possess both light and humidity-sensitive characteristics, while ensuring stable contact between the electrodes and the nanowires. This provides the hardware foundation for three-mode signal fusion and humidity-adaptive regulation. The fabrication process strictly controls key steps such as substrate cleaning, electrode pattern definition, electrode deposition, nanowire transfer, and post-processing shaping, combined with... Figure 2 The specific steps and technical details are as follows:

[0077] First, a silicon wafer is used as the substrate for pretreatment. In this embodiment, a p-type boron-doped silicon wafer with a 300 nm thermally grown SiO2 coating is used as the substrate. This substrate has both good insulation and mechanical stability, effectively avoiding the impact of substrate leakage on the electrical performance of the device. The pretreatment process strictly follows standardized cleaning procedures: First, the silicon wafer is cut into 1 cm × 1 cm square chips and placed sequentially into ultrasonic cleaning tanks containing anhydrous ethanol and deionized water. Each chip is ultrasonically cleaned for 15 minutes, with the ultrasonic power set at 100 W and the frequency at 40 kHz. The cavitation effect of the ultrasound thoroughly removes contaminants such as oil, dust, and metal ions from the substrate surface. After ultrasonic cleaning, the substrate is placed in an oven and dried at 80°C for 30 minutes to remove residual moisture. Finally, high-purity N2 (purity ≥ 99.999%) is used to purge the substrate surface at a pressure of 0.3 MPa for 5 minutes to ensure a clean, flat surface free of any impurities, providing excellent substrate conditions for subsequent photolithography and electrode deposition.

[0078] Then, a combined process of ultraviolet lithography and electron beam lithography was used to imprint the double-layer photoresist to define the patterned area for the bottom electrode film deposition. First, a double-layer photoresist was uniformly coated onto the pretreated substrate surface. The bottom layer was a 3μm thick SU-82002 photoresist, and the top layer was a 0.5μm thick photoresist. The coating speed was 3000 r / min, and the coating time was 45 seconds. Subsequently, it was soft-baked at 115℃ for 60 seconds to ensure a tight bond between the photoresist and the substrate, preventing photoresist detachment during subsequent processes. Next, ultraviolet lithography was performed. The coated chip was placed in an ultraviolet lithography machine (exposure wavelength 365nm), and the large-area contact pads and probe line patterns were exposed through a mask. The amount of photoresist is 100 mJ / cm². After exposure, the chip is developed in SU-8 developer for 3 minutes to remove the photoresist in the unexposed areas, thus defining the channel width with a 2 μm electrode spacing. Finally, electron beam lithography is performed to transfer the chip to an electron beam lithography machine (accelerating voltage 30 kV, beam current 100 pA). Symmetric Ag source and drain electrode patterns are drawn within the predefined probe lines to obtain the patterned area of ​​the bottom electrode film deposition. The electrode length is 5 μm and the width is 2 μm. The electrode spacing is strictly controlled to 2 μm to ensure that the subsequent nanowires can accurately bridge the electrodes.

[0079] Next, a bottom electrode film is deposited on the pretreated silicon wafer using a thermal evaporation process. The bottom electrode film is a composite electrode layer of chromium (Cr) and silver (Ag). In this embodiment, the photolithographically etched chip is placed in the vacuum chamber of a thermal evaporation coating machine, and the vacuum level is evacuated to 5 × 10⁻⁶. -4 To avoid insufficient vacuum leading to oxidation or contamination during electrode deposition, a Cr transition layer was deposited first. The Cr evaporation source current was set to 80 A, the deposition rate was 0.1 nm / s, and the deposition thickness was 5 nm. The Cr layer can chemically bond with the SiO2 substrate and the Ag layer, significantly improving electrode adhesion and preventing electrode detachment in subsequent processes. Subsequently, an Ag electrode layer was deposited. The Ag evaporation source current was set to 100 A, the deposition rate was 0.5 nm / s, and the deposition thickness was 50 nm. Ag is a highly conductive metal (resistivity 1.6 × 10⁻⁶). -8 Ω (m) ensures low contact resistance of the electrodes, reducing signal transmission loss.

[0080] Next, an N-methyl-2-pyrrolidone (NMP) extraction process is used to remove excess metal layers, resulting in a bottom electrode pattern with regular edges. In this embodiment, the deposited chip is immersed in an NMP solution at 50°C for 3 hours to allow the photoresist to fully swell and decompose. Subsequently, it is ultrasonically cleaned with deionized water (50W power, 5 minutes) to remove residual photoresist and metal debris. Finally, it is dried by blowing with high-purity N2, and the electrode pattern is observed under an optical microscope to ensure that the electrode edges are regular, free of burrs, short circuits, and other defects, and that the electrode spacing deviation is controlled within ±0.1μm.

[0081] Next, the wurtzite nanowires were transferred onto the prepared bottom electrode pattern, and polyvinyl chloride (PVA) was selected as a support layer to fix and protect the wurtzite nanowires and the bottom electrode pattern, thus obtaining the assembled sample. In this embodiment, wurtzite ZnO nanowires (purity ≥99.9%) with a diameter of 50-100 nm and a length of 2-5 μm were selected and dispersed in deionized water. The dispersion was ultrasonically dispersed for 10 minutes to prepare a nanowire dispersion with a concentration of 0.1 mg / mL. PVA (molecular weight 100,000) was dissolved in deionized water to prepare a 5% PVA solution, which was uniformly coated onto a glass slide and dried at 60°C to form a 1 μm thick PVA film. The wurtzite ZnO nanowire dispersion was then added dropwise to the surface of the PVA film using a micropipette and allowed to dry naturally at room temperature, ensuring uniform adhesion of the nanowires to the PVA film. The PVA film with the attached nanowires was then placed over an Ag electrode pattern. The position was observed and adjusted using an optical microscope to ensure precise bridging of the 2 μm spacing Ag source and drain electrodes by the ZnO nanowires. The PVA film was then wetted with deionized water to ensure close adhesion to the substrate and dried at 100°C for 5 minutes to enhance the contact between the nanowires and the electrodes.

[0082] Finally, the assembled sample was immersed in ultrapure water for 25 minutes to remove residual impurities from the sample surface and to fully dissolve the support layer, resulting in a humidity-adaptive optical storage device. In this embodiment, the assembled sample was placed in a beaker containing ultrapure water and immersed for 25 minutes. The PVA film was fully dissolved in the water, releasing wurtzite-phase ZnO nanowires. The ultrapure water was replaced, and the sample was immersed again for 10 minutes to remove residual PVA molecules and impurities from the surface. The sample was then removed, dried by blowing with high-purity N2, and then placed in a vacuum oven for annealing at 100°C for 30 minutes to further improve the contact stability between the wurtzite-phase ZnO nanowires and the Ag electrode and reduce the contact resistance.

[0083] The fabricated humidity-adaptive optical storage (HAOR) device needs to undergo a series of characterization methods to verify its structure and performance, ensuring that it meets the design requirements: For structural characterization, scanning electron microscopy was used to observe the morphology and electrode contact of wurtzite ZnO nanowires, ensuring that the nanowires were free of breakage and had good bridging; X-ray diffraction (XRD) was used to characterize the crystal structure of the wurtzite ZnO nanowires, confirming that it is a pure wurtzite phase (corresponding to 2θ=31.77°, 34.42°, 36.25°). (100), (002), (101) crystal planes); X-ray photoelectron spectroscopy (XPS) was used to analyze the surface chemical state of wurtzite ZnO nanowires, confirming the presence of oxygen vacancy defects (the characteristic peak at 531.5 eV in the O1s spectrum corresponds to an oxygen vacancy); in terms of electrical performance verification, under dark conditions and 50% RH, with a readout voltage of 0.01~1V applied, the device dark current was less than 1nA, exhibiting a high-resistivity state; when a 375nm ultraviolet light pulse (intensity 0.95mW) was applied... cm - ², pulse width 0.1s), peak photocurrent exceeding 100nA, on / off ratio exceeding 3 orders of magnitude, verifying the optical response characteristics of the device; in terms of humidity control performance verification, within the range of 10%~80%RH, with a fixed readout voltage of 1V and unchanged ultraviolet light pulse parameters, the photocurrent relaxation dynamics were tested, confirming that the relaxation time constant decreased from 730.0ms (humidity at 10%RH) to 6.3ms (humidity at 80%RH), with the humidity control range exceeding 1 order of magnitude, verifying the ability of humidity to control reservoir dynamics.

[0084] The HAOR device fabricated using the above process has a pure wurtzite ZnO nanowire with a uniform and symmetrical lattice structure and oxygen vacancy defects on the surface (providing a physical basis for the photo-humidity dual-sensitivity characteristics). The electrodes and nanowires have good contact, the device is in a high-resistivity state in the dark, and the photocurrent response is significant under ultraviolet light irradiation. Furthermore, the photocurrent relaxation dynamics can be precisely controlled by the ambient humidity, providing a hardware basis for the fusion of three-mode signals.

[0085] Furthermore, in other implementations, this embodiment can replace the wurtzite ZnO nanowires with other materials possessing dual photo-humidity sensitivity, such as SnO2 nanowires, TiO nanotubes, organic thin films, or other composite nanomaterials. These materials also possess photoresponse characteristics and humidity-controlled conductivity relaxation characteristics, enabling functional replacement of HAOR devices. The device structure can also be modified into a three-terminal phototransistor structure (such as a gate-controlled Zn nanowire field-effect transistor), further modulating the device's photo-humidity response characteristics through gate voltage, thus enhancing the control dimension of multimodal signals. The Ag electrode can also be replaced with other highly conductive metal electrodes, or a transparent conductive electrode can be used, adapting to light-transmitting sensing scenarios. The replacement of electrode materials does not change the core photo-humidity response characteristics of the device. The traditional silicon substrate can also be extended to a flexible substrate to fabricate flexible HAOR devices, adapting to wearable and portable weather monitoring equipment. The fabrication process of the flexible substrate only requires adjusting the photolithography and deposition parameters, without changing the core structure of the device.

[0086] Furthermore, after the humidity-adaptive optical storage (HAOR) device is fabricated, this embodiment can use a comprehensive performance testing platform consisting of a semiconductor parameter analyzer, an ultraviolet light control module, a self-made humidity control chamber, and a passive mechanical anemometer to systematically test the optical response characteristics, humidity control characteristics, and light-humidity-wind speed three-modal fusion response characteristics of the HAOR device. This will clarify the response law of the device under different operating conditions and provide performance basis for subsequent system construction.

[0087] Specifically, the selection and parameter settings of each component of the test platform in this embodiment strictly follow the experimental requirements to ensure test accuracy and reliability: the voltage measurement range of the semiconductor parameter analyzer is ±10nV~±100V, the current measurement range is ±100fA~±1A, and the measurement accuracy reaches 0.01%, used to accurately apply and read the voltage and acquire the current response signal of the device; the ultraviolet light control module uses a 375nm ultraviolet laser (output power adjustable range 0~30mW). cm -2Equipped with a pulse generator (pulse width adjustable range 0.01~1s, frequency adjustable range 0.1~10Hz) to generate ultraviolet light pulses with different parameters to simulate light intensity and wind speed encoded signals; a self-made humidity control chamber made of plexiglass, with a volume of 10cm×10cm×10cm, houses a humidity sensor (measurement range 0~100%RH, accuracy ±2%RH) and a humidity control module, with a humidity control accuracy of ±5%RH, used to simulate different environmental humidity conditions; the passive mechanical anemometer has a wind turbine diameter of 2cm, and the wind speed measurement range is... The wind turbine rotates at speeds of 0.5~10 m / s, periodically blocking ultraviolet light and converting the wind speed signal into an ultraviolet light pulse sequence. The mapping relationship between wind speed and pulse frequency is: wind speed v (m / s) = 0.1 × pulse frequency f (Hz) + 0.4, and the mapping relationship with pulse width is: pulse width t (s) = -0.02 × pulse frequency f (Hz) + 0.22. The data acquisition and analysis system uses software to write data acquisition programs, with the sampling frequency set to 10 kHz, to record parameters such as voltage, current, humidity, and wind speed in real time. The software is used for data processing and curve fitting.

[0088] After fabricating the humidity-adaptive optical storage device, this embodiment tests its optical response characteristics under different readout voltages, different ultraviolet light pulse intensities, and different ultraviolet light pulse widths under fixed humidity conditions. Specifically, in the test of the effect of readout voltage on optical response, a readout voltage Vread of 0.01~1V is applied, and a 375nm ultraviolet light pulse (intensity 0.95mW) is used. cm -2 The device was excited with a pulse width of 0.1s. The test results showed that the peak photocurrent increased exponentially with the increase of Vread: when Vread=0.01V, the peak photocurrent was about 3nA; when Vread=0.1V, the peak photocurrent increased to 35nA; when Vread=1V, the peak photocurrent reached 110nA, an increase of more than 36 times. This is because the increase in read voltage leads to the increase in the electric field strength inside the device, which promotes the separation and transport of photogenerated carriers.

[0089] In the test of the effect of light intensity on optical response, Vread = 0.2V and pulse width 0.1s were fixed, and the ultraviolet light intensity was adjusted to 0.03~26.36mW. cm -2 The test results showed that the peak photocurrent was linearly related to the light intensity: from a light intensity of 0.03mW... cm - ² Increased to 26.36mW cm -At 220 nA, the peak photocurrent increased linearly from 5 nA to 220 nA, with a linear fitting coefficient R² = 0.996, indicating that the device can accurately encode light intensity signals. Simultaneously, the photoresponse speed is stable. Under different light intensities, the device's photoresponse turn-on time (from the application of the light pulse to the photocurrent reaching 50% of its peak value) is approximately 180 μs, and the turn-off time (from the end of the light pulse to the photocurrent dropping to 50% of its peak value) is 2–5 ms. The response speed is unaffected by light intensity, ensuring the real-time performance of signal encoding.

[0090] In the test of the effect of pulse width on optical response, Vread = 0.2V and light intensity = 0.95mW were fixed. cm -2 Adjusting the pulse width to 0.01–0.5 s, the test results showed that the pulse width can effectively regulate the synaptic plasticity of the device: at a pulse width of 0.01 s, the device exhibits short-term enhancement (STP), with the photocurrent recovering to the dark current level within 4 s after the pulse ends; at a pulse width of 0.1 s, the recovery time is extended to 10 s; at a pulse width of 0.5 s, the device transforms into long-term enhancement (LTP), with a recovery time exceeding 100 s. This is because long-pulse illumination promotes deep ionization of oxygen vacancies in ZnO nanowires, forming more stable conductive channels. Furthermore, the peak photocurrent shows a logarithmic relationship with the pulse width: as the pulse width increases from 0.01 s to 0.5 s, the peak photocurrent increases from 80 nA to 150 nA, with a logarithmic fitting coefficient R² = 0.988, indicating that the pulse width can serve as an auxiliary coding dimension, enriching the coding information of the illumination signal.

[0091] Regarding energy consumption characteristics analysis, based on the energy calculation formula E=V×I×t (V is the read voltage, I is the average photocurrent, and t is the pulse width), the single-pulse energy consumption of the device under different operating conditions is calculated: when Vread=0.2V and the light intensity is 0.95mW cm -2 At a pulse width of 0.1s, the single-pulse energy consumption is as low as 10pJ, which is comparable to the energy consumption level of biological synapses (4~40pJ). Even under extreme conditions of Vread=1V and pulse width of 0.5s, the single-pulse energy consumption is only 55pJ, which is far lower than that of traditional optical sensors (usually at the nJ level), providing a hardware foundation for the ultra-low power operation of the system.

[0092] The optical response characteristics of the HAOR device in this embodiment can accurately encode light intensity signals, and the pulse frequency, pulse width, and readout voltage can all be used as controllable means, providing a response basis for optical encoding of wind speed. The photoresponse current under different readout voltages and illumination times is illustrated in the diagram. Figure 3 As shown.

[0093] This embodiment also tests the photocurrent relaxation dynamics of the humidity-adaptive photodetector within different humidity ranges under fixed readout voltage, fixed ultraviolet light pulse intensity, and fixed ultraviolet light pulse width. Specifically, the fixed readout voltage Vread = 1V and the ultraviolet light pulse intensity is 0.95mW. cm -2 The pulse width is 0.1s. The photocurrent relaxation behavior of the humidity-adaptive photomemory device conforms to the stretching exponential decay model. ,in The dark current is nA. Let be the difference between the peak photocurrent and the dark current (nA), τ be the relaxation time constant (ms), reflecting the photocurrent decay rate, and β be the stretching exponent (0 < β < 1), reflecting the uniformity of the relaxation process. Model parameters under different humidity conditions were obtained through nonlinear fitting.

[0094] At a humidity of 10%RH, I0 = 77.7 nA, ΔI = 564.8 nA, τ = 730.0 ms, and β = 0.39;

[0095] At a humidity of 20%RH, I0 = 48.4 nA, ΔI = 469.0 nA, τ = 421.0 ms, and β = 0.33;

[0096] At a humidity of 30%RH, I0 = 22.2 nA, ΔI = 320.6 nA, τ = 238.3 ms, and β = 0.30;

[0097] At a humidity of 40%RH, I0 = 10.7 nA, ΔI = 284.0 nA, τ = 213.0 ms, and β = 0.20.

[0098] At a humidity of 50%RH, I0 = 3.9 nA, ΔI = 268.1 nA, τ = 25.3 ms, and β = 0.17.

[0099] At a humidity of 60%RH, I0 = 0.98 nA, ΔI = 251.0 nA, τ = 13.7 ms, and β = 0.25;

[0100] At a humidity of 70%RH, I0 = 0.2nA, ΔI = 238.8nA, τ = 10.7ms, and β = 0.56.

[0101] When the humidity is 80%RH, I0=0.19nA, ΔI=226.4nA, τ=6.3ms, β=0.68.

[0102] The regulation of relaxation characteristics by humidity is mainly reflected in three aspects: First, the relaxation time constant τ decreases significantly with increasing humidity, τ=730.0ms at 10%RH and 6.3ms at 80%RH, with a regulation range exceeding one order of magnitude, indicating that humidity can efficiently regulate the reservoir dynamics of the device; Second, the peak photocurrent ΔI gradually decreases with increasing humidity, ΔI=564.8nA at 10%RH and 226.4nA at 80%RH. This is because the increase in humidity causes water molecules adsorbed on the surface of ZnO nanowires to form hydroxyl groups (-OH), which combine with oxygen vacancies, inhibiting the generation and transport of photogenerated carriers; Third, the stretching index β shows a trend of first decreasing and then increasing with increasing humidity, with β being the smallest (0.17) at 50%RH, indicating that the relaxation process is most non-uniform at this time, the nonlinear response of the reservoir is strongest, and it is most conducive to feature extraction.

[0103] In the conductivity state test with humidity-light intensity dual-parameter control, the reading voltage Vread = 1V and the pulse width 0.1s were fixed, while the humidity (50%~80%RH) and light intensity (2.5~4.5mW) were adjusted. cm -2 The peak photocurrent of the device was measured, and the results showed that different combinations of humidity and light intensity corresponded to unique photocurrent peaks, achieving ≥12 distinguishable conductance states (photocurrent difference between adjacent states ≥10 nA), providing a foundation for high-dimensional feature encoding of multimodal signals. Meanwhile, the distinguishability of the conductance states decreased slightly with increasing humidity, but this loss could be compensated for by adjusting the light intensity; for example, at 80% RH, the light intensity could be adjusted from 2.5 mW... cm - ² Increased to 4.5mW cm -2 The peak photocurrent increased from 180nA to 250nA, while still achieving four clearly distinguishable conductance states.

[0104] Therefore, the HAOR device achieves adaptive regulation of reservoir dynamics by humidity, transforming humidity from a simple input signal into a computationally controlled variable, thus overcoming the limitation of fixed response characteristics in traditional devices. The response current distribution under different readout voltages and humidity levels is illustrated in the diagram. Figure 4 As shown. Figure 4 These represent the values ​​of 0.95 mW·cm² under conditions of humidity of 80%, 70%, 60%, and 50%, respectively. -2 The study investigated the response current distribution under different readout voltage and humidity conditions.

[0105] The three-modal fusion rainfall tendency identification method based on humidity-adaptive optical storage devices in this embodiment can be applied to a terminal, which can be an intelligent product terminal such as a computer. Figure 1As shown in the figure, the three-modal fusion rainfall tendency identification method based on humidity adaptive optical storage device in this embodiment specifically includes the following steps:

[0106] Step S100: Acquire light signal, humidity signal and wind speed signal, and perform optical encoding on the light signal and the wind speed signal based on the optical encoding module, convert the light signal into ultraviolet light pulse intensity, the wind speed signal into ultraviolet light pulse frequency and pulse width, and convert the ultraviolet light pulse intensity, the ultraviolet light pulse frequency and pulse width and the humidity signal into a standardized ultraviolet light pulse sequence.

[0107] Furthermore, this embodiment also conducts testing and analysis of the optical-humidity-wind speed three-modal fusion response characteristics. Specifically, a passive mechanical anemometer is coupled with a HAOR device to construct an optical-humidity-wind three-modal input system, and the fusion response characteristics of the device are tested:

[0108] The optical encoding principle of wind speed is based on the rotation of the wind turbine in a passive mechanical anemometer. As the turbine rotates, it periodically blocks ultraviolet light, converting continuous ultraviolet light into a pulsed light sequence. The encoding rules are rigorously calibrated.

[0109] When the wind speed signal v=0.5m / s, the wind turbine rotation speed is 5r / s, the pulse frequency is 5Hz, and the pulse width is 0.12s;

[0110] When the wind speed signal v=3m / s, the wind turbine rotation speed is 26r / s, the pulse frequency is 26Hz, and the pulse width is 0.06s;

[0111] When the wind speed signal v=5m / s, the wind turbine rotation speed is 46r / s, the pulse frequency is 46Hz, and the pulse width is 0.02s;

[0112] The pulse frequency is linearly positively correlated with wind speed (R²=0.992), and the pulse width is linearly negatively correlated with wind speed (R²=0.988), ensuring that the wind speed signal can be accurately encoded into optical pulse parameters. The optical encoding principle of the wind speed signal is illustrated in the diagram. Figure 5 As shown.

[0113] In the three-modal fusion response law test, Vread=0.2V was fixed, and five typical light-humidity-wind combination conditions were set:

[0114] Operating condition 1 is a light intensity of 0.08mW. cm -2 The relative humidity was 20%RH, the wind speed was 0.5m / s (pulse frequency 5Hz, pulse width 0.12s), the peak photocurrent was 45nA, the relaxation time constant was 380ms, and the cumulative current was 5.2nA. s;

[0115] Operating condition 2 has a light intensity of 0.95mW. cm -2 Relative humidity 50%RH, wind speed 3m / s (pulse frequency 26Hz, pulse width 0.06s), peak photocurrent 98nA, relaxation time constant 22ms, cumulative current 5.6nA. s;

[0116] Operating condition 3 has a light intensity of 6.73 mW. cm -2 Relative humidity 80%RH, wind speed 5m / s (pulse frequency 46Hz, pulse width 0.02s), peak photocurrent 185nA, relaxation time constant 8ms, cumulative current 3.5nA. s;

[0117] Operating condition 4 has a light intensity of 0.95mW. cm -2 Relative humidity 20%RH, wind speed 5m / s (pulse frequency 46Hz, pulse width 0.02s), peak photocurrent 120nA, relaxation time constant 350ms, cumulative current 2.3nA. s;

[0118] Operating condition 5 has a light intensity of 6.73 mW. cm -2 Relative humidity 80%RH, wind speed 0.5m / s (pulse frequency 5Hz, pulse width 0.12s), peak photocurrent 210nA, relaxation time constant 10ms, cumulative current 24.8nA. s.

[0119] This embodiment analyzes the current response curves under different operating conditions and derives the three-mode fusion response law: the light intensity dominates the peak photocurrent; the higher the light intensity, the larger the peak photocurrent, such as in operating condition 3 (light intensity 6.73mW). cm -2 The peak photocurrent of condition 1 (185 nA) was significantly higher than that of condition 2 (light intensity 0.95 mW). cm -2 (98nA); Humidity dominates the relaxation time constant. The higher the humidity, the smaller the relaxation time constant. For example, τ=380ms for condition 1 (humidity 20%RH) and τ=8ms for condition 3 (humidity 80%RH); Wind speed controls the current accumulation value through pulse frequency and pulse width. The higher the wind speed, the higher the pulse frequency and the narrower the pulse width, and the smaller the current accumulation value. For example, the current accumulation value for condition 4 (wind speed 5m / s) is 2.3nA. The values ​​of s) are much lower than those of operating condition 1 (wind speed 0.5 m / s, 5.2 nA). (s) The three-mode signals, through the coordinated regulation of the three dimensions of "peak-relaxation-accumulation", enable characteristic current response curves corresponding to different environmental conditions, achieving physical-level feature fusion and providing rich feature information for rainfall tendency identification. The three-mode signal fusion response current curve of this embodiment is shown in the figure. Figure 6 As shown.

[0120] Based on this, this embodiment collects light intensity, humidity, and wind speed signals. Specifically, the three-modal signal acquisition module consists of a 375nm ultraviolet detector, the HAOR device itself, and a passive mechanical anemometer. No additional humidity sensor is required; it relies on the humidity-sensitive characteristics of the HAOR device to collect ambient humidity, achieving synchronous, contactless, and high-precision acquisition of light intensity, humidity, and wind speed signals. The 375nm ultraviolet detector has a response wavelength of 350~400nm, collects ambient light intensity in real time, has a sampling frequency of 100Hz, and a dynamic range of 0.01~30mW. cm -2 This ensures the coverage of the HAOR device's optical response range; the passive mechanical anemometer has a measurement range of 0.5~10m / s and a measurement accuracy of ±0.2m / s. When the wind turbine rotates, it periodically blocks ultraviolet light, converting the wind speed signal into an optical pulse sequence. No additional electrical interface is required, enabling contactless signal acquisition.

[0121] The optical encoding module in this embodiment employs a programmable ultraviolet light driving circuit, enabling unified optical encoding of three-modal signals. Specifically, this embodiment normalizes the illumination intensity signal collected by the ultraviolet detector and, according to the first mapping relationship, maps the normalized illumination intensity to 0.03~26.36 mW. cm -2 The ultraviolet light pulse intensity is mapped as follows: Ultraviolet light pulse intensity = 0.88 × normalized illumination intensity + 0.03, ensuring that the encoded light intensity covers the effective response range of the HAOR device. According to the second mapping relationship, the wind speed signal output by the anemometer is converted into the ultraviolet light pulse frequency and pulse width. The second mapping relationship is consistent with the three-mode fusion response characteristic test, i.e., v = 0.1f + 0.4, t = -0.02f + 0.2; where v is the wind speed, f is the ultraviolet light pulse frequency, and t is the ultraviolet light pulse width. This embodiment also uses an FPGA to generate a synchronous clock signal (frequency 10kHz) to ensure that the illumination intensity encoding, wind speed encoding, and the reading timing of the HAOR device are strictly synchronized, with a synchronization error ≤ 1μs. In other implementations, this embodiment can also transform the ultraviolet light pulse frequency and pulse width encoding of wind speed into pulse amplitude encoding.

[0122] Furthermore, in this embodiment, the ultraviolet pulse intensity, ultraviolet pulse frequency and pulse width, and the actual measured humidity signal are converted into a standardized ultraviolet pulse sequence through an optical encoding module to adapt the signal for projection onto the HAOR reservoir array, ensuring accurate transmission of feature information.

[0123] Step S200: Project the standardized ultraviolet light pulse sequence onto a reservoir array composed of several humidity-adaptive photo-storage devices, obtain the current response curve output by each humidity-adaptive photo-storage device in the reservoir array, and obtain a high-dimensional feature vector based on the current response curve.

[0124] The reservoir array in this embodiment consists of 64 humidity-adaptive optical storage (HAOR) devices (8×8=64 in total). Each device corresponds to one optically encoded signal, enabling high-dimensional feature extraction and fusion. The array adopts a square layout with a device spacing of 100μm to avoid crosstalk between devices. The effective area of ​​each device is 2μm×5μm, ensuring high integration of the array. A row-column addressing architecture is adopted, with each row and column equipped with a CMOS (Complementary Metal-Oxide-Semiconductor) switch (on-resistance ≤100Ω). The FPGA controls the switching on and off of the switches to achieve independent addressing and signal reading of individual devices. Each column output is equipped with a low-noise operational amplifier (input offset voltage ≤10μV, bandwidth ≥1MHz) to convert the weak current signal output by the HAOR device into a voltage signal, facilitating subsequent ADC (Analog-to-Digital Converter) sampling.

[0125] In practical applications, under the combined influence of three modal signals—light, humidity, and wind speed—the current response curve output by each humidity-adaptive photoretention device in the reservoir array of this embodiment can be used to extract the peak value, relaxation time constant, and current accumulation value of the current response curve, resulting in a 64-dimensional high-dimensional feature vector. Simultaneously, the humidity signal is used as a control variable to adjust the photocurrent relaxation dynamics of each humidity-adaptive photoretention device in real time, eliminating humidity interference. The inherent differences between devices further enrich the feature dimensions, completing the feature extraction and fusion of the three modal signals.

[0126] Step S300: Based on the FPGA, the high-dimensional feature vector is converted into an analog voltage feature signal, and based on the FPGA, the analog voltage feature signal is converted into a digital signal to obtain a digital feature vector.

[0127] This embodiment uses an FPGA to control the CMOS switches of the reservoir array to achieve independent addressing of individual devices, converting the high-dimensional feature vectors output by each humidity-adaptive photoelectric storage device into analog voltage feature signals. The FPGA control module and FPGA readout layer module integrate signal acquisition, timing control, readout layer training, and inference recognition functions. Specifically, the FPGA's analog-to-digital converter (ADC) module integrates a 12-bit high-precision ADC (model ADS7951) with a sampling frequency of 100kHz, an input voltage range of 0~3V, and a conversion error ≤±1LSB. It can be used to convert the analog voltage signal output from the operational amplifier into a digital signal, thereby obtaining a digital feature vector. The FPGA's digital-to-analog converter (DAC) module integrates an 8-bit DAC (model DAC5571) with an output voltage range of 0~1V and an output resolution ≤4mV, used to precisely control the readout voltage (0.01~1V) of the HAOR device. The timing control module generates ultraviolet light pulse trigger signals, array addressing signals, and ADC / DAC sampling signals, with a timing accuracy ≤10ns, ensuring coordinated operation of all modules.

[0128] Based on this, this embodiment can use the analog-to-digital converter module in the FPGA to convert the analog voltage characteristic signal into a digital signal at a sampling frequency of 100kHz, thus obtaining a digital feature vector. The converted 64-dimensional digital feature vector is directly stored in the module's 8KB asynchronous FIFO (first in first out) buffer, achieving high-speed data caching and transmission, avoiding data loss, completing the transition of the signal from the analog domain to the digital domain, and adapting to the digital computing requirements of the FPGA.

[0129] Step S400: Input the digital feature vector into the readout layer of the FPGA, and perform binary classification recognition of rainfall and no rainfall on the digital feature vector through the pre-trained weight matrix in the readout layer of the FPGA, and output the rainfall tendency recognition result. The weight matrix is ​​pre-trained using the soft maximum cross-entropy optimization algorithm and is used to establish the mapping relationship between the digital feature vector and the rainfall label.

[0130] In this embodiment, a simple readout layer is built inside the FPGA. The soft maximum cross-entropy optimization algorithm is used to train the output weight matrix. The weight matrix has a dimension of 64×2 (64 HAOR device outputs, 2 types of prediction results). The inference process adopts hardware parallel computing, and the single-sample inference latency is ≤1ms. The data cache module integrates an 8KB asynchronous FIFO cache to store ADC sampling data and intermediate results during the training process to avoid data loss.

[0131] Specifically, in this embodiment, the weight matrix in the FPGA readout layer module is pre-trained using a soft maximum cross-entropy optimization algorithm to establish the mapping relationship between digital feature vectors and rainfall labels. During training, this embodiment uses a professional meteorological dataset containing one year of continuous meteorological monitoring data, with a sampling frequency of once per hour, totaling 8760 samples. Each sample contains light intensity (0.01~30mW). cm -2 The system inputs consist of four fields: relative humidity (0%~100%RH), wind speed (0~10m / s), and rainfall label (rain / no rain). During training, three core meteorological indicators highly correlated with rainfall tendency—light intensity, ambient humidity, and wind speed—were selected as system inputs. The system focused on the 10%~80%RH humidity range, a region with high rainfall incidence, to adapt to actual outdoor meteorological monitoring scenarios. 6570 samples were selected from this range, and outliers (such as light intensity >30mW) were removed. cm -2 (Samples with wind speed > 10 m / s).

[0132] To achieve quantitative analysis and classification of multimodal signals, each core indicator is discretized into three levels: low, medium, and high. The low level of light intensity is 0.03~0.95mW. cm -2 Medium grade is 0.95~6.73mW cm -2 High-grade power consumption is 6.73~26.36mW. cm -2 Humidity levels are defined as follows: low level (10%~30%RH), medium level (30%~60%RH), and high level (60%~80%RH); wind speed is defined as follows: low level (0.5~2m / s), medium level (2~4m / s), and high level (4~5m / s). Twenty-seven different environmental categories were constructed using a full permutation method to achieve full coverage of various meteorological scenarios. From each category, 90% of the samples were randomly selected as the training set (5913 samples in total), and 10% were selected as the test set (657 samples in total). The meteorological data preprocessing workflow is illustrated below. Figure 7 As shown in the diagram, the distribution of characteristics for the 27 environmental categories is as follows. Figure 8 As shown.

[0133] The raw meteorological data needs to undergo targeted preprocessing to accurately match the optical response characteristics of the HAOR device and ensure the accuracy of signal transmission and feature extraction: after normalization, the raw light intensity data is linearly mapped to 0.03~26.36mW. cm -2The intensity of the ultraviolet (UV) light pulses is precisely matched to the optical response range of the HAOR device. After normalizing the raw wind speed data, it is mapped to a UV pulse frequency of 0.5–5 Hz, corresponding to a pulse width of 0.2–0.01 s. Higher wind speeds correspond to higher pulse frequencies and narrower pulse widths. Ambient humidity is directly measured using actual relative humidity values ​​(10%–80% RH) without additional encoding or conversion, serving as the reservoir dynamics control variable for the HAOR device. The preprocessed three-modal data is converted into a standardized UV pulse sequence via an optical encoding module, and stably projected onto the HAOR reservoir array with fixed pulse parameters, ensuring accurate transmission of characteristic information.

[0134] In the HAOR reservoir array, each device corresponds to one optical coded signal. Under the combined action of three modes of signals—illuminance, humidity, and wind speed—the devices output current response curves with unique characteristics based on different parameter combinations. Key parameters such as the peak value, relaxation time constant, and current accumulation value of the curves together constitute a high-dimensional feature vector, realizing the characteristic encoding of the three-mode meteorological signals. At the same time, humidity, as a core control variable, can adjust the photocurrent relaxation dynamics of each device in the array in real time according to the dynamic changes in ambient humidity, achieving environmental adaptation for feature extraction and effectively avoiding the interference of humidity changes on sensing accuracy. In addition, the inherent inter-device differences in the HAOR device array can further enrich the dimension of the feature vector without additional algorithmic control, improving the distinguishability of features, eliminating the need for complex digital feature enhancement algorithms, and reducing computational costs.

[0135] The FPGA control module and the readout layer module rapidly convert the analog current feature vector output by the HAOR array into a digital signal using a high-precision ADC, storing it in an asynchronous FIFO buffer to achieve high-speed data caching and transmission. From a dataset of 27 environmental categories, 90% of the samples are randomly selected as the training set and 10% as the test set for model training and performance verification of the readout layer. The training process employs a soft maximum cross-entropy optimization algorithm, focusing only on the output weight matrix (Wout) of the FPGA readout layer, eliminating the need to train the entire HAOR reservoir array, significantly reducing computational complexity and model training costs. The training process for the weight matrix is ​​as follows:

[0136] 1. Weight initialization: The weight matrix (Wout) is assigned values ​​using a random initialization method, with matrix elements uniformly distributed within the interval [-0.1, 0.1].

[0137] 2. Forward Propagation: The training sample (64-dimensional digital feature vector) X, extracted and transformed by the HAOR reservoir array, is input into the readout layer module. The output vector Y = Wout * X is calculated, and the predicted probability distribution y is obtained after normalization by the softmax function. pred =softmax(Y).

[0138] 3. Loss Function Calculation: The soft maximum cross-entropy loss function is used to calculate the error between the predicted value and the true label. The loss function is: L = -Σ(y true ×ln(y pred In the formula, y true For samples with actual rainfall / no rainfall labels, y pred Output the predicted probability for the model.

[0139] 4. Weight Iterative Update: The weight matrix is ​​updated using the gradient descent algorithm. The update formula is: Wout * =Wout-η× L, the learning rate η is initially 0.01, and decreases linearly to 0.001 as the number of iterations increases, gradually reducing the update step size to improve convergence stability.

[0140] 5. Training Termination and Optimal Weight Saving: When any of the following termination conditions are met, the iteration stops and the current optimal weight matrix is ​​saved, resulting in the trained weight matrix:

[0141] (1) The change in the training set loss function over 100 consecutive iterations is less than 1×10. -4 ;

[0142] (2) The number of iterations reaches the preset maximum value of 1000 times.

[0143] The weight matrix Wout obtained after training * This system is used for real-time rainfall trend inference and determination on the FPGA side. When the 64-dimensional digital feature vector from step S300 is input into the readout layer of the FPGA, the pre-trained weight matrix in the FPGA readout layer can perform binary classification (rainfall or no rainfall) on the digital feature vector, outputting the rainfall trend identification result. Simultaneously, the FPGA can achieve continuous and precise control of the HAOR device's readout voltage from 0.01 to 1V via a DAC, dynamically adjusting the device's light response sensitivity according to the strength of the environmental signal, flexibly adapting to different meteorological monitoring scenarios and improving the system's versatility. Furthermore, in other training methods, this embodiment can also replace the soft maximum cross-entropy optimization algorithm of the FPGA readout layer module with algorithms such as binary neural networks, recurrent neural networks (RNNs), and convolutional neural networks (CNNs) to adapt to more complex meteorological scenario predictions.

[0144] Furthermore, in this embodiment, the rainfall trend identification result is presented and stored by the rainfall trend output module. The rainfall trend output module consists of an LCD display unit, a serial communication unit, and a data storage unit: the LCD display unit uses a 1.5-inch OLED screen to display the current light intensity, humidity, wind speed parameters, and rainfall trend prediction results ("rainfall" / "no rainfall") in real time, with an update frequency of 10Hz; the serial communication unit uses a UART (Universal Asynchronous Receiver / Transmitter) interface (baud rate 9600bps) to transmit monitoring data and prediction results to a host computer for data backup and system debugging; the data storage unit can automatically retain the monitoring data and prediction results for the most recent 30 days at a 1-minute interval for subsequent meteorological analysis.

[0145] In addition, this embodiment is equipped with a power supply module, which uses a low-power linear regulator to provide stable power to each module: providing 3.3V / 1.2V power to the FPGA, ADC, and DAC, with output currents of 500mA / 300mA and ripple voltage ≤10mV respectively; providing an adjustable power supply of 0.01~1V to the HAOR device, with an output current of 100mA and ripple voltage ≤5mV; and providing 5V power to the ultraviolet laser, with an output current of 200mA, supporting pulse width modulation (PWM) dimming. The HAOR-FPGA integrated system architecture used in this embodiment for rainfall tendency recognition is illustrated in the diagram. Figure 9 As shown.

[0146] Furthermore, this embodiment underwent comprehensive performance testing on a meteorological dataset, and both quantitative and qualitative test results verified the excellent performance of the present invention. Specifically, in terms of energy consumption performance, a power analyzer was used to test the energy consumption of each module. When the system is in standby mode (without ultraviolet light excitation, only the data acquisition module is working), the energy consumption is 0.5mW; the energy consumption of a single sample inference, including the entire process of data acquisition, optical encoding, feature extraction, and inference recognition, is only 0.12μJ, of which the HAOR reservoir array accounts for 60% (0.072μJ), the FPGA computing accounts for 30% (0.036μJ), and the other modules account for 10% (0.012μJ); under an environment of 25℃ and 50%RH, the total energy consumption of the system working continuously for 24 hours is 10.368mWh, requiring only a 100mAh lithium battery for power, meeting the low power consumption requirements of portable devices.

[0147] Regarding recognition accuracy, the recognition performance of this invention was tested on a test set (657 samples), achieving an overall recognition accuracy of 98.2%. The recall rate for rainfall samples was 99.5%, and the precision rate for no-rain samples was 97.8%. Of the 27 environmental categories, 25 categories achieved a recognition accuracy ≥95%, with only low humidity (10%~30%RH), low wind speed (0.5~2m / s), and low light intensity (0.03~0.95mW) showing the highest accuracy. cm -2 Combination (1 category) with high humidity (60%~80%RH), high wind speed (4~5m / s), and high light intensity (6.73~26.36mW) cm -2 The recognition accuracy of the combination (1 category) is 92%~93%, mainly because the feature vector similarity between the two scenarios is high. Compared with traditional systems, the recognition accuracy of traditional CNN-LSTM-based systems on the same test set is 89.5%, while the recognition accuracy of this invention is improved by 8.7 percentage points, especially the recall rate of rainfall samples is improved by 12.3 percentage points, making it more suitable for flood control scenarios. The multimodal feature confidence distribution curve is shown in the figure. Figure 10 As shown, the normalization comparison of rainfall prediction sample results is illustrated. Figure 11 As shown.

[0148] In terms of real-time performance, the system's overall latency was tested using an oscilloscope and logic analyzer. The signal acquisition latency was 2ms (0.5ms for light signal acquisition, 1ms for humidity signal backpropagation, and 0.5ms for wind speed signal encoding), the feature extraction latency was 3ms (mainly composed of a 0.1s light pulse excitation time and a 2ms signal readout latency), and the inference output latency was 1ms (0.5ms for FPGA readout layer inference and 0.5ms for result display and storage). The total latency was ≤10ms, which is much lower than that of traditional systems (over 50ms), meeting the requirements for second-level response.

[0149] Regarding environmental adaptability, the recognition performance of this invention was tested under different humidity and temperature environments. Within the RH range of 10% to 80%, the recognition accuracy of the system remained above 95%, with a recognition accuracy of 98.5% in the RH range of 60% to 80% (high rainfall scenario), significantly higher than that of traditional systems (around 85%). Within the RH range of -10℃ to 60℃, the recognition accuracy of the system fluctuated by ≤2%, mainly because temperature changes have little impact on the photoresponse characteristics of HAOR devices (temperature coefficient ≤0.01% / ℃). In an environment of 25℃ and 50%RH, the system operated continuously for 72 hours, maintaining a recognition accuracy of over 98% without significant performance degradation, indicating that this invention has good long-term stability.

[0150] In summary, the core advantages of this invention are: the HAOR device transforms humidity from a simple input signal into a reservoir dynamic regulation variable, and achieves environmental adaptation in the calculation process through humidity-dependent photocurrent relaxation characteristics, effectively eliminating humidity interference and improving feature extraction accuracy in complex meteorological scenarios; the system overcomes the limitations of von Neumann architecture. Overcoming the limitations of the von Neumann architecture, this device achieves integrated sensing, storage, and computation of three-modal signals: light, humidity, and wind speed. Data does not need to be transmitted across modules, the overall latency is <10ms, and the single-pulse energy consumption is as low as 10pJ, which is 1000 to 10000 times lower than that of traditional systems. Multimodal signals are fused within the HAOR device through a physical mechanism, eliminating the need for complex digital fusion algorithms and reducing computational complexity and training costs. The device adopts a planar memristor structure, and its fabrication process is compatible with CMOS, resulting in high integration and enabling on-chip integration, making it suitable for miniaturized applications.

[0151] This invention breaks through the architectural and performance bottlenecks of traditional rainfall forecasting systems, realizing real-time acquisition of light-humidity-wind three-modal signals at the edge, adaptive feature fusion, and accurate identification of rainfall trends. It has broad application prospects in portable meteorological monitoring, IoT edge nodes, outdoor environmental sensing, and other scenarios. It can promote the development of meteorological sensing technology towards "adaptive, low-power, and highly integrated" directions, and provide core technical support for disaster prevention and mitigation, smart agriculture, urban operation and maintenance, and other fields.

[0152] Based on the above embodiments, the present invention also provides a three-modal fusion rainfall tendency identification system based on a humidity-adaptive optical storage device, which is used to implement the steps in the above method embodiments. Figure 12As shown, the system in this embodiment includes: a multimodal signal acquisition and optical encoding module 10, a HAOR reservoir array module 20, an FPGA control module 30, and an FPGA readout layer module 40. Specifically, the multimodal signal acquisition and optical encoding module 10 is used to acquire light signals, humidity signals, and wind speed signals, and optically encode the light signals and wind speed signals based on the optical encoding module, converting the light signals into ultraviolet light pulse intensity, the wind speed signals into ultraviolet light pulse frequency and pulse width, and converting the ultraviolet light pulse intensity, the ultraviolet light pulse frequency and pulse width, and the humidity signal into a standardized ultraviolet light pulse sequence. The HAOR reservoir array module 20 is used to project the standardized ultraviolet light pulse sequence onto a reservoir array composed of several humidity-adaptive optical storage devices, obtain the current response curve output by each humidity-adaptive optical storage device in the reservoir array, and obtain a high-dimensional feature vector based on the current response curve. The FPGA control module 30 is used to convert the high-dimensional feature vector into an analog voltage feature signal based on the FPGA, and then convert the analog voltage feature signal into a digital signal based on the FPGA to obtain a digital feature vector. The FPGA readout layer module 40 is used to input the digital feature vector into the readout layer of the FPGA, and perform binary classification recognition of rainfall and no rainfall on the digital feature vector through a pre-trained weight matrix in the readout layer of the FPGA, and output the rainfall tendency recognition result. The weight matrix is ​​pre-trained using a soft maximum cross-entropy optimization algorithm to establish a mapping relationship between the digital feature vector and the rainfall label.

[0153] The HAOR device of this invention can also be integrated with other meteorological sensors (such as temperature sensors and barometric pressure sensors), incorporating temperature and pressure signals into the system through optical encoding or electrical control to achieve multi-dimensional meteorological parameter fusion prediction, further improving the accuracy of rainfall trend identification. The system of this invention can construct an Internet of Things (IoT) meteorological sensor network, networking multiple integrated HAOR-FPGA nodes to achieve regionalized collaborative prediction of rainfall trends, suitable for large-scale scenarios such as urban flood control and watershed meteorological monitoring. The fabrication process of the HAOR device of this invention is compatible with CMOS technology, enabling on-chip integration. The HAOR reservoir array, FPGA control circuit, and signal processing module can be integrated into a single chip, further reducing device size, increasing integration density, and adapting to micro / nano-scale meteorological sensing devices. The system of this invention can also introduce a self-learning and adaptive update mechanism, collecting meteorological monitoring data and prediction results in real time through the FPGA, continuously optimizing the readout layer weight matrix, achieving online model updates, and improving long-term prediction accuracy.

[0154] The principles of each module and each step in the three-modal fusion rainfall tendency identification system based on humidity adaptive optical storage device in this embodiment are the same, and will not be elaborated further here.

[0155] Based on the above embodiments, the present invention also provides a terminal, the principle block diagram of which can be as follows: Figure 13 As shown. The terminal may include one or more processors 100 ( Figure 13 (Only one is shown in the image), memory 101, and computer program 102 stored in memory 101 and executable on one or more processors 100. For example, a trimodal fusion rainfall tendency identification program based on a humidity-adaptive light storage device. When one or more processors 100 execute computer program 102, they can implement various steps in the embodiments of the trimodal fusion rainfall tendency identification method based on a humidity-adaptive light storage device. Alternatively, when one or more processors 100 execute computer program 102, they can implement the functions of various modules / units in the embodiments of the trimodal fusion rainfall tendency identification system based on a humidity-adaptive light storage device, which is not limited here.

[0156] In one embodiment, the processor 100 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0157] In one embodiment, memory 101 can be an internal storage unit of the terminal, such as a hard disk or RAM. Memory 101 can also be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SM), secure digital card (SD), flash card, etc. Furthermore, memory 101 can include both internal and external storage units. Memory 101 is used to store computer programs and other programs and data required by the terminal. Memory 101 can also be used to temporarily store data that has been output or will be output.

[0158] Those skilled in the art will understand that Figure 13The block diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the terminal to which the present invention is applied. A specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0159] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct memory bus RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A three-modal fusion rainfall tendency identification method based on humidity-adaptive optical storage devices, characterized in that, The method includes: Light intensity, humidity, and wind speed signals are collected, and optical encoding is performed on the light intensity and wind speed signals based on the optical encoding module. The light intensity is converted into ultraviolet light pulse intensity, the wind speed is converted into ultraviolet light pulse frequency and pulse width, and the ultraviolet light pulse intensity, ultraviolet light pulse frequency and pulse width, and humidity are converted into a standardized ultraviolet light pulse sequence. The standardized ultraviolet light pulse sequence is projected onto a reservoir array composed of several humidity-adaptive photo-storage devices, and the current response curve output by each humidity-adaptive photo-storage device in the reservoir array is obtained. A high-dimensional feature vector is obtained based on the current response curve. The high-dimensional feature vector is converted into an analog voltage feature signal based on the FPGA, and the analog voltage feature signal is converted into a digital signal based on the FPGA to obtain a digital feature vector; The digital feature vector is input into the readout layer of the FPGA. The digital feature vector is then subjected to binary classification for rainfall and no rainfall by a pre-trained weight matrix in the readout layer of the FPGA, and the rainfall tendency recognition result is output. The weight matrix is ​​pre-trained using a soft maximum cross-entropy optimization algorithm and is used to establish a mapping relationship between the digital feature vector and the rainfall label.

2. The three-modal fusion rainfall tendency identification method based on humidity-adaptive optical storage device as described in claim 1, characterized in that, The fabrication process of the humidity-adaptive optical storage device includes: Using a silicon wafer as the substrate, the silicon wafer is pre-treated, and a combined process of ultraviolet lithography and electron beam lithography is used to imprint a double layer of photoresist to define the patterned area for the deposition of the bottom electrode film. A bottom electrode film is deposited on the pretreated silicon wafer through a thermal evaporation process. The bottom electrode film is a composite electrode layer of chromium and silver. The excess metal layer was removed by N-methyl-2-pyrrolidone separation process to obtain a bottom electrode pattern with regular edges; The wurtzite nanowires were transferred onto the prepared bottom electrode pattern, and polyvinyl alcohol was selected as a support layer to fix and protect the wurtzite nanowires and the bottom electrode pattern, thus obtaining the assembled sample. The assembled sample was immersed in ultrapure water for 25 minutes to remove residual impurities on the sample surface and to fully dissolve the support layer, thus obtaining a humidity-adaptive optical storage device.

3. The three-modal fusion rainfall tendency identification method based on humidity-adaptive optical storage device according to claim 2, characterized in that, The method further includes: After fabricating the humidity-adaptive optical storage device, under a fixed humidity condition, the optical response characteristics of the humidity-adaptive optical storage device under different readout voltages, different ultraviolet light pulse intensities, and different ultraviolet light pulse widths were tested. Under the conditions of fixed readout voltage, fixed ultraviolet light pulse intensity, and fixed ultraviolet light pulse width, the photocurrent relaxation dynamics of the humidity adaptive photoretention device were tested in different humidity ranges. The humidity-adaptive photo-storage device can transform humidity from a simple input signal into a reservoir dynamic regulation variable, and can achieve environmental adaptation in the calculation process through humidity-dependent photocurrent relaxation characteristics.

4. The three-modal fusion rainfall tendency identification method based on humidity-adaptive optical storage device according to claim 3, characterized in that, The illumination signal and the wind speed signal are optically encoded based on an optical encoding module, converting the illumination signal into ultraviolet light pulse intensity and the wind speed signal into ultraviolet light pulse frequency and pulse width, including: After normalizing the light intensity signal, according to the first mapping relationship, the normalized light intensity signal is mapped to a range of 0.03~26.36 mW. cm -2 The intensity of the ultraviolet light pulse; According to the second mapping relationship, the wind speed signal is converted into an ultraviolet light pulse frequency of 0.5~5 Hz and an ultraviolet light pulse width of 0.01~0.2s.

5. The three-modal fusion rainfall tendency identification method based on humidity-adaptive optical storage device according to claim 4, characterized in that, The first mapping relationship is: Ultraviolet pulse intensity = 0.88 × normalized irradiance + 0.03; The second mapping relationship is: v = 0.1f + 0.4, t = -0.02f + 0.2; Where v is the wind speed, f is the ultraviolet pulse frequency, and t is the ultraviolet pulse width.

6. The three-modal fusion rainfall tendency identification method based on humidity adaptive optical storage device according to claim 5, characterized in that, The standardized ultraviolet light pulse sequence is projected onto a reservoir array composed of several humidity-adaptive photonic storage devices. The current response curve output by each humidity-adaptive photonic storage device in the reservoir array is obtained. Based on the current response curve, a high-dimensional feature vector is obtained, including: The standardized ultraviolet light pulse sequence is projected onto a reservoir array, which consists of an 8×8 humidity-adaptive photoretention device; The current response curve of each humidity-adaptive photoreceptor in the reservoir array is obtained, and the peak value, relaxation time constant, and current accumulation value of the current response curve are extracted to obtain a 64-dimensional high-dimensional feature vector. At the same time, the humidity signal is used as a control variable to adjust the photocurrent relaxation dynamics of each humidity-adaptive photoreceptor in real time to eliminate humidity interference and complete the feature extraction and fusion of the three-mode signal.

7. The three-modal fusion rainfall tendency identification method based on humidity-adaptive optical storage device according to claim 6, characterized in that, The high-dimensional feature vector is converted into an analog voltage feature signal based on the FPGA, and the analog voltage feature signal is then converted into a digital signal based on the FPGA to obtain a digital feature vector, including: Based on FPGA control of the CMOS switching transistors of the reservoir array, independent addressing of individual devices is achieved, and the high-dimensional feature vector output by each humidity-adaptive photomemory device is converted into an analog voltage feature signal. Based on the analog-to-digital converter module in the FPGA, the analog voltage characteristic signal is converted into a digital signal at a sampling frequency of 100kHz to obtain a digital feature vector.

8. A three-modal fusion rainfall tendency identification system based on humidity-adaptive optical storage devices, characterized in that, The system is used to implement the steps of the three-modal fusion rainfall tendency identification method based on humidity adaptive optical storage device according to any one of claims 1-7, the system comprising: A multimodal signal acquisition and optical encoding module is used to acquire light signals, humidity signals, and wind speed signals, and to perform optical encoding on the light signals and wind speed signals based on the optical encoding module, converting the light signals into ultraviolet light pulse intensity, the wind speed signals into ultraviolet light pulse frequency and pulse width, and converting the ultraviolet light pulse intensity, the ultraviolet light pulse frequency and pulse width, and the humidity signals into a standardized ultraviolet light pulse sequence; The HAOR reservoir array module is used to project the standardized ultraviolet light pulse sequence onto a reservoir array composed of several humidity-adaptive photo-storage devices, obtain the current response curve output by each humidity-adaptive photo-storage device in the reservoir array, and obtain a high-dimensional feature vector based on the current response curve. An FPGA control module is used to convert the high-dimensional feature vector into an analog voltage feature signal based on the FPGA, and to convert the analog voltage feature signal into a digital signal based on the FPGA to obtain a digital feature vector; The FPGA readout layer module is used to input the digital feature vector into the readout layer of the FPGA, and to perform binary classification recognition of rainfall and no rainfall on the digital feature vector through a pre-trained weight matrix in the readout layer of the FPGA, and output the rainfall tendency recognition result. The weight matrix is ​​pre-trained using a soft maximum cross-entropy optimization algorithm to establish a mapping relationship between the digital feature vector and the rainfall label.

9. A terminal, characterized in that, The terminal includes a memory, a processor, and a three-modal fusion rainfall tendency identification program based on a humidity-adaptive light storage device stored in the memory and executable on the processor. When the processor executes the three-modal fusion rainfall tendency identification program based on a humidity-adaptive light storage device, it implements the steps of the three-modal fusion rainfall tendency identification method based on a humidity-adaptive light storage device as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a three-modal fusion rainfall tendency identification program based on a humidity-adaptive optical storage device, wherein the three-modal fusion rainfall tendency identification program based on a humidity-adaptive optical storage device implements the steps of the three-modal fusion rainfall tendency identification method based on a humidity-adaptive optical storage device as described in any one of claims 1-7 on the computer-readable storage medium.

Images