A heavy fog detector and method for eliminating the influence of light difference

Abstract

This invention discloses a fog detector and method for eliminating the influence of illumination differences. The device includes a multispectral parallel acquisition module, a dedicated data processing chip, an adaptive band selection module, a self-calibrating optical path, a MEMS active alignment module, and a digital twin maintenance platform. The multispectral parallel acquisition module includes a wide-angle incident optical lens, a beam splitter, and a multi-detector array. The beam splitter distributes the incident light transmitted through the wide-angle incident optical lens through different bands to the multi-detector array. The multi-detector array includes several independent CMOS image sensors, each corresponding to a specific band and equipped with a bandpass filter. This invention relates to the field of fog detection technology. This invention solves the problems of insufficient adaptability of traditional fog detection equipment to illumination differences, leading to signal distortion under complex lighting conditions, the need for frequent manual calibration, and susceptibility to misjudgment and missed detection in critical scenarios such as transportation and aviation.

Description

Technical Field

[0001] This invention relates to the field of fog detection technology, and in particular to a fog detector and method for eliminating the effects of light difference. Background Technology

[0002] In the field of fog detection, dynamic changes in lighting conditions, such as day-night cycles, changes in weather conditions, and variations in the angle of light during sunrise and sunset, are the core interference factors affecting detection accuracy. Existing fog detection equipment generally suffers from insufficient adaptability to differences in lighting conditions.

[0003] Traditional equipment often employs single-band or fixed-band optical detection schemes, with detection wavelengths typically concentrated in the visible light range, only suitable for fog scattering characteristics under specific lighting conditions. However, in practical applications, the optical characteristics of fog vary significantly under different lighting scenarios—at dawn and dusk, the proportion of long-wavelength components in sunlight increases, altering the scattering intensity and spectral distribution of fog; on cloudy days or at night, ambient light intensity is significantly reduced and the spectral composition is homogeneous, making traditional fixed-band detection susceptible to background light noise interference, thus hindering fog concentration signal extraction.

[0004] This sensitivity to differences in illumination directly leads to a series of problems: First, the detection signal is distorted. Under complex illumination conditions, the fog concentration data output by the equipment deviates significantly and cannot accurately reflect the actual fog conditions. Second, the adaptability is narrow. The detection stability of the same equipment is poor at different times and under different weather conditions, requiring frequent manual calibration to match changes in illumination, which is cumbersome and costly to maintain. Third, the application is limited. In scenarios such as transportation and aviation, where the real-time and reliability requirements for detection are extremely high, the risk of misjudgment and missed detection caused by illumination interference may lead to serious safety hazards.

[0005] Therefore, how to effectively eliminate the interference of light difference on fog detection and improve the adaptability and detection accuracy of the equipment in complex lighting environments has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the problems of traditional fog detection equipment's insufficient adaptability to lighting differences, resulting in signal distortion under complex lighting conditions, the need for frequent manual calibration, and the susceptibility to misjudgment and missed detection in critical scenarios such as transportation and aviation, the present invention aims to provide a fog detector and method that eliminates the influence of lighting differences.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a fog detector that eliminates the influence of illumination differences, comprising a multispectral parallel acquisition module, a dedicated data processing chip, an adaptive band selection module, a self-calibrating optical path, a MEMS active alignment module, and a digital twin maintenance platform;

[0008] The multispectral parallel acquisition module includes a wide-angle incident optical lens, a beam splitter, and a multi-detector array. The beam splitter distributes the incident light transmitted through the wide-angle incident optical lens to the multi-detector array through different wavelengths. The multi-detector array includes several independent CMOS image sensors, each corresponding to a specific wavelength band and equipped with a bandpass filter.

[0009] The dedicated data processing chip integrates a multi-channel data parallel processing module, a pipeline architecture, and a neural network accelerator for real-time parallel processing of multi-band data.

[0010] The self-calibrating optical path integrates a reference light source group;

[0011] The MEMS active alignment module includes a MEMS micromirror array, which is a MEMS micromirror integrated into each optical path.

[0012] In the multispectral parallel acquisition module, the data output terminals of each CMOS image sensor are respectively connected to the multi-channel data input interface of the dedicated data processing chip, which is used to transmit the image data acquired in each band to the dedicated data processing chip in parallel.

[0013] The preprocessed data output terminal of the dedicated data processing chip is connected to the data input interface of the adaptive band selection module to provide real-time spectral analysis data; the decision instruction output terminal of the adaptive band selection module is connected to the control instruction input terminal of the dedicated data processing chip to provide feedback on dynamic band selection instructions.

[0014] The calibration control output terminal of the dedicated data processing chip is connected to the control signal input terminal of the self-calibrating optical path for sending calibration trigger commands; the calibration data output terminal of the self-calibrating optical path is connected to the correction calculation unit of the dedicated data processing chip for feedback of channel response coefficients and optical distortion correction coefficients.

[0015] The image data output terminal of the dedicated data processing chip is connected to the image monitoring input terminal of the MEMS active alignment module to provide real-time image data for each channel; the control signal output terminal of the MEMS active alignment module is connected to the driving terminal of each MEMS micromirror to output optical path compensation control signals.

[0016] The status data output terminal of the dedicated data processing chip is connected to the data access interface of the digital twin maintenance platform for uploading equipment operating status data; the prediction result output terminal of the digital twin maintenance platform is connected to the maintenance command input terminal of the dedicated data processing chip for issuing equipment maintenance early warning information.

[0017] The communication interface of the dedicated data processing chip is connected to an external monitoring center to output fog concentration detection results.

[0018] The wide-angle incident optical lens is a fisheye lens, and the beam splitting component is a beam splitter with a hybrid structure of prism, grating, and prism, which divides the spectrum in the range of 400-1000nm into 8 specific bands: 400-450nm, 450-500nm, 500-550nm, 550-600nm, 600-650nm, 650-700nm, 700-800nm, and 800-1000nm, with a half-width of 15±2nm for each band; the corresponding CMOS image sensors are set to 8, and a global shutter is used to achieve hardware synchronous exposure.

[0019] Preferably, the reference light source group consists of 8 TO-Can packaged laser-stabilized LEDs, each with a center wavelength aligned with the center of the detection band and matched to the corresponding detection band, and a half-width at half-maximum of 10 nm; the self-calibrating optical path also integrates a high-speed optical shutter, a fiber-coupled lens group, a fiber bundle, and a uniform light integrating sphere.

[0020] The fiber-coupled lens group includes three lenses, which are used in sequence to initially converge light from the high-speed optical shutter, correct chromatic aberration, and finely focus it;

[0021] The fiber bundle is a 19-core multimode fiber bundle, which uniformly couples the reference light to each detection channel.

[0022] The light-emitting surface of the reference light source group is sequentially connected to the high-speed optical shutter, the fiber-coupled lens group, the incident end of the fiber bundle, the emitting end of the fiber bundle, and the incident port of the integrating sphere homogenizing system in the optical path; the light-emitting port of the integrating sphere homogenizing system is coupled to the optical path of the main optical system, thereby injecting the standard reference light into the multi-detector array in the main optical path.

[0023] Preferably, the multi-channel data parallel processing module includes several independent data preprocessing units, each preprocessing unit is connected to the data output terminal of a CMOS image sensor, and synchronously receives raw image data of each band through a dedicated parallel data bus, and performs black level correction, non-uniformity correction and Bayer decoding operations in parallel.

[0024] The pipeline architecture employs a three-stage pipeline. The first-stage pipeline is directly connected to the output of the multi-channel data parallel processing module, responsible for receiving preprocessed multi-channel image data and performing image registration and alignment operations. The second-stage pipeline receives the output data from the first-stage pipeline and performs multispectral data fusion and feature extraction operations. The third-stage pipeline receives the output data from the second-stage pipeline and performs image enhancement and normalization processing.

[0025] The neural network accelerator is connected to the third-stage pipeline output of the pipeline architecture via a high-speed on-chip network to receive standardized feature data. It contains multiple parallel convolutional computation units and attention mechanism computation units, and uses weight preloading and data pipeline technology to achieve overlapping execution of convolutional computation and data transmission.

[0026] The output of the neural network accelerator is fed back to the control unit of the multi-channel data parallel processing module, and the data throughput priority of each channel is dynamically adjusted through an adaptive bandwidth allocation algorithm to form a closed-loop optimization control system.

[0027] The neural network accelerator adopts a spiking neural network architecture, which supports dynamic weight updates and sparse computation.

[0028] Preferably, the adaptive band selection module includes a feature extraction and fusion unit, a signal-to-noise ratio evaluation unit, a fog sensitivity analysis unit, a multi-objective optimization decision-making unit, an intelligent decision engine, and an instruction generation unit;

[0029] The feature extraction and fusion unit's data input terminal is connected to the preprocessed data output terminal of the dedicated data processing chip, and is used to receive real-time multispectral data cubes and environmental status data, and to perform fusion processing on multi-source data;

[0030] The input of the signal-to-noise ratio (SNR) evaluation unit is connected to the output of the feature extraction and fusion unit, and is used to calculate the SNR of the image in each band. ,in The mean value of the λ-band image. The standard deviation of the λ-band image;

[0031] The input of the fog sensitivity analysis unit is connected to the output of the feature extraction and fusion unit, and is used to calculate the fog sensitivity index for each band. ,in This represents the rate of change in fog intensity caused by changes in fog concentration. This represents the standard deviation of noise in the λ-band;

[0032] The input terminal of the multi-objective optimization decision unit is connected to the output terminals of the signal-to-noise ratio evaluation unit and the fog sensitivity analysis unit, respectively, and is used to perform multi-objective optimization calculations by integrating the signal-to-noise ratio and fog sensitivity indicators;

[0033] The input end of the intelligent decision engine is connected to the output end of the multi-objective optimization decision unit, and it has a built-in algorithm model based on the fusion of attention mechanism and reinforcement learning to generate the optimal band selection strategy.

[0034] The input terminal of the instruction generation and output unit is connected to the output terminal of the intelligent decision engine, and is used to send the band selection instruction to the control unit of the dedicated data processing chip.

[0035] Preferably, the MEMS active alignment module further includes an image feature extraction unit, a registration error calculation unit, a control algorithm unit, a drive signal generation unit, and a closed-loop feedback unit;

[0036] The input of the image feature extraction unit is connected to the image data output of the dedicated data processing chip, and is used to extract SIFT / ORB feature points from the real-time images of each channel; the input of the registration error calculation unit is connected to the output of the image feature extraction unit, and is used to calculate the translation error Δx, Δy and rotation error Δθ between the images of each channel; the input of the control algorithm unit is connected to the output of the registration error calculation unit, and calculates the adjustment amount required for the MEMS micromirror based on the registration error; the input of the drive signal generation unit is connected to the output of the control algorithm unit, and is used to generate a high-precision drive voltage signal for the MEMS micromirror; the closed-loop feedback unit includes a high-precision temperature sensor and a capacitive position sensor integrated on the MEMS micromirror substrate, and is used to collect the actual deflection angle and temperature sensor data of each MEMS micromirror in real time through the closed-loop feedback unit, and feed them back to the control algorithm unit to form a closed-loop control system.

[0037] A method for detecting fog using a fog detector that eliminates the influence of illumination differences, comprising the following steps:

[0038] S1. Synchronous acquisition and preprocessing of multispectral data:

[0039] The multispectral parallel acquisition module synchronously acquires optical signals in multiple bands, and the acquisition process follows an atmospheric scattering physics model.

[0040] ;

[0041] in, represent The original acquired signal of the band at pixel (x,y). For target signals in an ideal fog-free scenario, Indicates atmospheric transmittance. Atmospheric light intensity, This is dark current noise;

[0042] S2. Adaptive Band Selection and Data Optimization:

[0043] Key bands are dynamically selected based on signal-to-noise ratio and fog sensitivity index. The fog sensitivity index is calculated as follows:

[0044] ;

[0045] in, The original acquired signal in the λ band With fog concentration The absolute value of the rate of change; It is the noise standard deviation in the λ band. Signal-to-noise ratio; The highest signal-to-noise ratio among all candidate bands;

[0046] The arithmetic mean of the key band fog sensitivity index N=3~5;

[0047] S3. Dedicated chip parallel processing and fog concentration inversion:

[0048] A dedicated data processing chip performs parallel processing on multi-band data, employing a fog concentration inversion algorithm based on dark channel prior optimization.

[0049] ;

[0050] in, For the estimated fog concentration, For adaptive band weighting, It is a wavelength-dependent scattering coefficient. To detect distance, The total variational regularization coefficient;

[0051] S4. Periodic self-calibration and data correction:

[0052] The system automatically initiates the calibration process every 10 minutes, correcting the data for each channel using the following formula:

[0053] ;

[0054] The correction coefficients are obtained through reference light source calibration. For channel response coefficients, This is the optical distortion correction factor;

[0055] S5. MEMS Micromirror Active Optical Path Alignment:

[0056] By monitoring the feature points of each image channel in real time, the micromirror adjustment amount is calculated using an adaptive control algorithm based on Lyapunov stability.

[0057] ;

[0058] in, Let be the deflection angle of the MEMS micromirror in the x-direction. Let be the deflection angle of the MEMS micromirror in the y-direction. The current deflection angle of the micromirror in the x and y directions at the nth step at the current moment; This represents the deflection angle of the objective function J with respect to the x and y directions of the MEMS micromirror. The gradient; This is the temperature drift compensation term, calculated using real-time temperature sensor data; K is the control gain coefficient.

[0059] S6. Digital twin-assisted predictive maintenance:

[0060] The digital twin platform uses a deep survival analysis model to predict device health status:

[0061] ;

[0062] in, Let be the risk function at time t. As the benchmark risk function, For the weight vector, For Transformer encoders; when the predicted value exceeds the safety threshold, a maintenance warning is triggered in advance;

[0063] S7. Results Output and Decision Support:

[0064] Output fog concentration detection results and confidence levels. The confidence level is calculated based on the following formula:

[0065] ;

[0066] in, Maximum value of FSI in key band, pixel Model predicts signal strength at [location] , The scattering coefficient is wavelength-dependent. That is, the distance from the detector to the target pixel;

[0067] When the confidence level falls below a set threshold, a recalibration process is automatically triggered, and the results are securely transmitted to the monitoring center via quantum key distribution technology.

[0068] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0069] 1. This invention, by dividing the spectrum into multiple specific bands and using multiple global shutter CMOS synchronous acquisition, significantly improves the spectral sensitivity coverage and can accurately capture the scattering characteristics of fog under complex lighting conditions such as morning fog, sunset, and cloudy days.

[0070] 2. This invention dynamically selects 3-5 key bands, discards invalid bands with strong light interference, and has strong anti-interference ability through dynamic band selection, achieving all-weather indiscriminate detection and fully meeting the usage needs of different scenarios such as day and night and sunny and cloudy days.

[0071] 3. This invention eliminates sequence acquisition delay through a multispectral parallel acquisition architecture and utilizes a dedicated ASIC chip's parallel preprocessing unit, a three-stage pipeline, and a pulse neural network accelerator in synergy. This results in high efficiency in the entire process of multi-band data processing, ensuring timely fog concentration warnings and mitigating safety risks.

[0072] 4. This invention reduces the amount of invalid data processing through an adaptive band selection algorithm, further improving the response speed without sacrificing detection accuracy, and achieving a synergy between high precision and high real-time performance.

[0073] 5. This invention automatically starts calibration through a self-reference optical path and eliminates channel deviation through laser-stabilized LEDs and correction formulas, avoiding the tedious operation of manual calibration and reducing labor costs and downtime losses.

[0074] 6. In this invention, the MEMS active alignment technology achieves high-precision optical path registration through closed-loop feedback, completely solving the fusion error caused by temperature and vibration, and ensuring the consistency of multi-channel data.

[0075] 7. This invention eliminates optical path deviations through MEMS active alignment and suppresses noise by combining it with a dark channel prior optimization algorithm. The two work together to significantly improve the accuracy of fog concentration inversion.

[0076] 8. This invention evaluates the validity of detection results in real time through confidence calculation, and automatically triggers recalibration when the confidence level is low, avoiding invalid or erroneous data output and ensuring the reliability of decision-making basis.

[0077] This invention achieves comprehensive advantages such as all-weather adaptability, millisecond-level response, low operation and maintenance costs, and high data security through synergistic optimization of spectral sensitivity, real-time performance, and stability. It can be widely applied in fields such as transportation, aviation, meteorology, and ports, providing highly reliable technical support for safety management and early warning decision-making in foggy weather, and has significant practical value and market prospects. Attached Figure Description

[0078] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0079] Figure 1 This is a schematic diagram of the fog detector system of the present invention that eliminates the influence of light difference;

[0080] Figure 2 This is a schematic diagram of the optical path of the fog detector of the present invention to eliminate the influence of illumination differences;

[0081] Figure 3 This is a schematic diagram of the MEMS active alignment module of the present invention;

[0082] Figure 4 This is a schematic diagram of the adaptive band selection mode of the present invention. Detailed Implementation

[0083] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification.

[0084] Please see Figures 1 to 4 It should be understood that the structures, proportions, sizes, etc., illustrated in the accompanying drawings are merely for illustrative purposes to aid those skilled in the art and to facilitate understanding and reading. They are not intended to limit the scope of the invention and therefore have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, without affecting the effectiveness and purpose of the invention, should still fall within the scope of the technical content disclosed in this invention. Furthermore, the terms such as "upper," "lower," "left," "right," "middle," and "one" used in this specification are merely for clarity and not intended to limit the scope of the invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention's implementation.

[0085] This invention provides a technical solution: a fog detector that eliminates the influence of light difference, comprising six core units: a multispectral parallel acquisition module, a dedicated data processing chip, an adaptive band selection module, a self-calibrating optical path, a MEMS active alignment module, and a digital twin maintenance platform. Each unit achieves data interaction and command transmission through high-speed interfaces including an LVDS parallel data bus, a Gigabit Ethernet interface, an I²C interface, and an SPI interface. The overall device measures 300mm × 200mm × 150mm, weighs ≤5kg, operates in temperatures ranging from -30℃ to 60℃, has an IP65 protection rating, and can withstand the influence of complex outdoor environments.

[0086] The multispectral parallel acquisition module consists of a wide-angle incident optical lens, a beam splitter, a multi-detector array, and a parallel data transmission link. Its core function is to simultaneously acquire atmospheric scattered light signals in eight specific bands.

[0087] The wide-angle incident optical lens adopts a large field-of-view fisheye structure with a field of view of 180°, a distortion rate of less than 1%, and a focal length of 2.8mm. The lens material is anti-ultraviolet optical glass, and the surface is coated with an anti-reflection film in the 400-1000nm band with a transmittance of more than 95%, which can realize all-round fog coverage detection without monitoring blind spots.

[0088] The beam splitter is a hybrid structure of "optical glass prism + holographic grating" with dimensions of 30mm×20mm×15mm. It can split incident light in the 400-1000nm range into 8 specific bands: 400-450nm, 450-500nm, 500-550nm, 550-600nm, 600-650nm, 650-700nm, 700-800nm, and 800-1000nm. Each band has a half-width of 15±2nm and a beam splitting efficiency greater than 85%.

[0089] The multi-detector array contains eight independent global shutter CMOS image sensors, each with a resolution of 512×512 pixels, a pixel size of 3.45μm, a frame rate of 30fps, and supports hardware synchronous exposure with an adjustable exposure time of 10μs~100ms. Each sensor integrates a custom bandpass filter at the front end, with a cutoff depth of OD6 and temperature stability of ±0.5nm / ℃, allowing only the corresponding target wavelength light signal to enter the detection unit.

[0090] The parallel data transmission link uses an LVDS parallel data bus with a transmission rate of 1.2Gbps, ensuring synchronous transmission of raw image data from each CMOS sensor with a transmission delay of less than 10μs.

[0091] The light-emitting end of the wide-angle incident optical lens is connected to the incident end optical path of the beam splitter. The eight light-emitting ends of the beam splitter correspond one-to-one with the light-receiving surfaces of the eight CMOS image sensors in the multi-detector array. The data output end of each CMOS image sensor is directly connected to the multi-channel data input interface of the dedicated data processing chip through the LVDS parallel data bus to realize the parallel transmission of eight channels of image data.

[0092] The dedicated data processing chip is manufactured using a 28nm CMOS process, with a core frequency of 1.2GHz and a power consumption of less than 5W. It integrates a multi-channel parallel data processing unit, a three-stage pipeline architecture, a neural network accelerator, and multiple types of communication interfaces, and is the core of the detector's data processing and control.

[0093] The multi-channel parallel data processing unit contains eight independent data preprocessing units, each corresponding to one of the eight CMOS sensors in the multi-detector array. Each unit can perform black level correction, non-uniformity correction, and Bayer decoding operations in parallel. Black level correction relies on dark current reference values ​​pre-stored in the chip ROM. These reference values ​​are calibrated in 10 levels at temperatures ranging from -30℃ to 60℃. Non-uniformity correction is based on bright / dark field data from a reference light source, and the correction coefficients are updated in real time. Bayer decoding can convert RAW format data to RGB format with a decoding error of less than 1%.

[0094] Three-stage pipeline architecture:

[0095] Level 1: The input end directly receives the output data of the multi-channel parallel processing unit, and uses the phase correlation method to perform translation / rotation alignment on the 8-channel image, with a registration accuracy of less than 0.1 pixels and a processing delay of less than 20μs;

[0096] The second stage receives the aligned image output from the first stage, fuses multi-band data using a weighted average method, with the fusion weights dynamically adjusted by the signal-to-noise ratio of each channel, and uses the Sobel operator to extract droplet scattering features with an accuracy greater than 98%.

[0097] The third stage receives the fused feature data output from the second stage, uses adaptive histogram equalization (CLAHE) to enhance fog details, and standardizes the feature data to the 0-1 range with a standardization error of less than 0.5%.

[0098] The neural network accelerator adopts a spiking neural network architecture, supports dynamic weight updates with an update period of 100ms, supports sparse computation with a sparsity rate of 60%, and includes 8 parallel convolutional computation units with adjustable kernel sizes of 3×3 / 5×5, as well as 2 attention mechanism computation units and 16MB SRAM weight storage units with access latency of less than 1ns. It has a built-in adaptive bandwidth allocation algorithm that can adjust the data throughput according to the signal-to-noise ratio of each channel, and increase the priority of the corresponding channel when the SNR is less than 20dB.

[0099] The communication interface integrates a gigabit Ethernet interface, an I²C interface, and an SPI interface. The gigabit Ethernet interface is used to output fog concentration detection results, the I²C interface is used to connect to external environmental sensors, and the SPI interface is used to interface with the digital twin maintenance platform.

[0100] The multi-channel data parallel processing unit has eight input interfaces, which correspond one-to-one with the LVDS bus output of the multispectral parallel acquisition module.

[0101] The three-stage pipeline architecture is connected in series via internal high-speed data links, and the output of the third stage is connected to the input of the neural network accelerator via an on-chip network.

[0102] The output of the neural network accelerator is fed back to the control port of the multi-channel data parallel processing unit to form a closed-loop optimization control.

[0103] The preprocessed data output terminal of the dedicated data processing chip is connected to the adaptive band selection module via the SPI interface, the calibration control output terminal is connected to the self-calibration optical path via the I²C interface, the image data output terminal is connected to the MEMS active alignment module via the LVDS interface, the status data output terminal is connected to the digital twin maintenance platform via the SPI interface, and the communication interface is connected to the external monitoring center via Gigabit Ethernet.

[0104] The self-calibration optical path consists of a reference light source group, a high-speed optical shutter, a fiber-coupled lens group, a fiber bundle, a uniform light integrating sphere, and a calibration data feedback link. It is used to eliminate optical errors caused by device aging and temperature drift in real time.

[0105] Reference light source group: It contains 8 TO-Can packaged laser stabilized LEDs, the center wavelength of each LED is aligned with the center of the multispectral detection band, and the aligned wavelengths are 425nm, 475nm, 525nm, 575nm, 625nm, 675nm, 750nm and 900nm respectively. Each LED has a half width at half maximum (WHM) of 10nm, output power stability of ±1% / ℃ and lifetime greater than 5000h.

[0106] The high-speed optical shutter has a response time of less than 10μs and a light-transmitting aperture of 8mm. It can receive trigger signals from a dedicated data processing chip and triggers once every 30 minutes. It will also trigger when the illumination changes by more than 20%.

[0107] The fiber-coupled lens group is a three-element combination structure. The first element is made of optical glass with a focal length of 10mm for initial light convergence. The second element is an achromatic cemented doublet with a focal length of 15mm for chromatic aberration correction. The third element is an aspherical structure with a focal length of 8mm for fine focusing. The overall coupling efficiency is greater than 80%.

[0108] The fiber bundle uses 19-core multimode fiber with a core diameter of 50μm and a numerical aperture of 0.22, which can evenly distribute the reference light to 8 detection channels.

[0109] The uniform light integrating sphere is made of PTFE with a diffuse reflectance greater than 98%, has an inner diameter of 50 mm and a light output uniformity greater than 98%, and is used to convert reference light into a uniform surface light source.

[0110] The calibration data feedback link consists of a signal acquisition end and a data transmission interface, which can feed back the light intensity detection data during the calibration process to a dedicated data processing chip.

[0111] The light-emitting surface of the reference light source group is sequentially connected to the incident end of the high-speed optical shutter, the incident end of the fiber-coupled lens group, and the incident end of the fiber bundle in the optical path. The emitting end of the fiber bundle is connected to the incident port of the uniform light integrating sphere. The emitting port of the uniform light integrating sphere is coupled to the main optical path of the multispectral parallel acquisition module. The main optical path is located between the wide-angle incident optical lens and the beam splitter to realize the injection of standard reference light. The control end of the high-speed optical shutter is connected to the calibration control output end of the dedicated data processing chip. The calibration data output end of the uniform light integrating sphere is connected to the calibration calculation unit interface of the dedicated data processing chip through the signal acquisition link to provide feedback channel response coefficients and optical distortion correction coefficients.

[0112] The MEMS active alignment module consists of a MEMS micromirror array, an image feature processing unit, a control algorithm unit, a drive signal generation unit, and a closed-loop feedback unit, and is used to compensate for optical path offset caused by temperature / vibration in real time.

[0113] The MEMS micromirror array contains eight single-axis MEMS micromirrors, which correspond one-to-one with eight detection channels. Each micromirror has a deflection angle range of ±5°, a resolution of 0.1μrad, a response time of less than 1ms, and a driving voltage of 0-5V, and is integrated into the optical path of each detection channel.

[0114] The image feature extraction unit is used to extract SIFT / ORB feature points from real-time images of each channel, with an extraction quantity of ≥50 and a feature matching rate of greater than 95%.

[0115] The registration error calculation unit calculates the translation error Δx / Δy and rotation error Δθ between each channel image and the reference channel based on the least squares method. The reference channel is in the 550-600nm band, with a translation error accuracy of 0.01 pixels and a rotation error accuracy of 0.01°.

[0116] The control algorithm unit uses a PID control algorithm to calculate the adjustment amount of the MEMS micromirror based on the registration error;

[0117] The drive signal generation unit contains a 12-bit DAC with a resolution of 0.1mV, which can convert the adjustment amount into a high-precision drive voltage signal.

[0118] The closed-loop feedback unit includes a high-precision temperature sensor and a capacitive position sensor integrated on the MEMS micromirror substrate. The high-precision temperature sensor has an accuracy of ±0.5℃, and the capacitive position sensor has a resolution of 0.1μrad, which are used to collect the actual state data of the micromirror.

[0119] The input of the image feature extraction unit is connected to the image data output of the dedicated data processing chip to receive real-time images from each channel; the output of the image feature extraction unit is connected to the input of the registration error calculation unit, the output of the registration error calculation unit is connected to the input of the control algorithm unit, and the output of the control algorithm unit is connected to the input of the drive signal generation unit; the output of the drive signal generation unit is connected one-to-one with the drive ends of the eight MEMS micromirrors to output optical path compensation control signals; the output of the closed-loop feedback unit is connected to the feedback input of the control algorithm unit to form a closed-loop control system that corrects the micromirror adjustment amount in real time.

[0120] The adaptive band selection module consists of a feature extraction and fusion unit, a signal-to-noise ratio evaluation unit, a fog sensitivity analysis unit, a multi-objective optimization decision-making unit, an intelligent decision engine, and an instruction generation unit. Its core function is to dynamically select the optimal detection band based on illumination and fog conditions.

[0121] The feature extraction and fusion unit is used to receive and fuse multispectral data cubes and environmental status data. The multispectral data cubes are 512×512×8 pixels in size, and the environmental status data includes temperature and humidity, with a temperature range of -30℃ to 60℃ and a humidity range of 10% to 90%.

[0122] The signal-to-noise ratio evaluation unit is based on the formula ,in The mean value of the λ-band image. The standard deviation of the λ-band image;

[0123] Fog sensitivity analysis unit based on formula ,in This represents the rate of change in fog intensity caused by changes in fog concentration. This represents the standard deviation of noise in the λ-band;

[0124] The multi-objective optimization decision unit generates three sets of candidate band combinations with the objectives of "maximizing signal-to-noise ratio" and "maximizing fog sensitivity".

[0125] The intelligent decision engine incorporates an attention mechanism and a reinforcement learning DQN algorithm model. The training dataset contains 100,000 samples with varying lighting and fog concentrations, ranging from 10 lux to 10 lux. 4 The lux and fog concentration range is 0~500 mg / m³, and the optimal band selection strategy can be output.

[0126] The instruction generation unit converts the optimal strategy into binary control instructions. Each instruction is 8 bits long, where 1 indicates that the corresponding band is enabled and 0 indicates that the corresponding band is disabled.

[0127] The input of the feature extraction and fusion unit is connected to the preprocessed data output of the dedicated data processing chip; the output of the feature extraction and fusion unit is connected to the input of the signal-to-noise ratio evaluation unit and the fog sensitivity analysis unit, respectively; the outputs of the signal-to-noise ratio evaluation unit and the fog sensitivity analysis unit are both connected to the input of the multi-objective optimization decision unit; the output of the multi-objective optimization decision unit is connected to the input of the intelligent decision engine, and the output of the intelligent decision engine is connected to the input of the instruction generation unit; the output of the instruction generation unit is connected to the control instruction input of the dedicated data processing chip through the SPI interface, and feeds back the dynamic band selection instruction.

[0128] The digital twin maintenance platform consists of a data receiving module, a digital twin model, a fault early warning unit, and a maintenance command issuing module. It constructs a virtual model of the detector based on 3D modeling technology to realize equipment status monitoring and fault early warning.

[0129] The data receiving module is used to receive device operating status data uploaded by the dedicated data processing chip. This data includes the temperature, voltage, current of each module, CMOS frame rate, MEMS alignment error, and reference light source power.

[0130] The digital twin model replicates the physical structure of the detector in a 1:1 ratio and can synchronize the operating status of the physical device in real time. For example, it can visualize the deflection angle of MEMS micromirrors and highlight the error in red when the error exceeds the threshold.

[0131] The fault early warning unit uses an LSTM neural network algorithm to predict equipment lifespan based on historical operating data, with an early warning accuracy rate of over 95%.

[0132] The maintenance instruction distribution module can generate maintenance instructions, such as "trigger forced calibration" and "adjust MEMS drive parameters", and push them to the operation and maintenance terminal via the wireless module.

[0133] The input of the data receiving module is connected to the status data output of the dedicated data processing chip via an SPI interface to receive device status data; the output of the data receiving module is connected to the data source interface of the digital twin model to provide real-time data support for the virtual model; the status output of the digital twin model is connected to the input of the fault early warning unit, and the output of the fault early warning unit is connected to the input of the maintenance command issuing module; the output of the maintenance command issuing module is connected to the maintenance command input of the dedicated data processing chip via an SPI interface to issue device maintenance early warning information and control commands.

[0134] A method for detecting fog using a fog detector that eliminates the influence of illumination differences, comprising the following steps:

[0135] S1. Synchronous acquisition and preprocessing of multispectral data:

[0136] The multispectral parallel acquisition module synchronously acquires optical signals in multiple bands, and the acquisition process follows an atmospheric scattering physics model.

[0137] ;

[0138] in, represent The original acquired signal of the band at pixel (x,y). For target signals in an ideal fog-free scenario, Indicates atmospheric transmittance. Atmospheric light intensity, This is dark current noise;

[0139] S2. Adaptive Band Selection and Data Optimization:

[0140] Key bands are dynamically selected based on signal-to-noise ratio and fog sensitivity index. The fog sensitivity index is calculated as follows:

[0141] ;

[0142] in, The original acquired signal in the λ band With fog concentration The absolute value of the rate of change; It is the noise standard deviation in the λ band. Signal-to-noise ratio; The highest signal-to-noise ratio among all candidate bands;

[0143] The arithmetic mean of the key band fog sensitivity index N=3~5;

[0144] S3. Dedicated chip parallel processing and fog concentration inversion:

[0145] A dedicated data processing chip performs parallel processing on multi-band data, employing a fog concentration inversion algorithm based on dark channel prior optimization.

[0146] ;

[0147] in, For the estimated fog concentration, For adaptive band weighting, It is a wavelength-dependent scattering coefficient. To detect distance, The total variational regularization coefficient;

[0148] S4. Periodic self-calibration and data correction:

[0149] The system automatically initiates the calibration process every 10 minutes, correcting the data for each channel using the following formula:

[0150] ;

[0151] The correction coefficients are obtained through reference light source calibration. For channel response coefficients, This is the optical distortion correction factor;

[0152] S5. MEMS Micromirror Active Optical Path Alignment:

[0153] By monitoring the feature points of each image channel in real time, the micromirror adjustment amount is calculated using an adaptive control algorithm based on Lyapunov stability.

[0154] ;

[0155] in, Let be the deflection angle of the MEMS micromirror in the x-direction. Let be the deflection angle of the MEMS micromirror in the y-direction. The current deflection angle of the micromirror in the x and y directions at the nth step at the current moment; This represents the deflection angle of the objective function J with respect to the x and y directions of the MEMS micromirror. The gradient; This is the temperature drift compensation term, calculated using real-time temperature sensor data; K is the control gain coefficient.

[0156] S6. Digital twin-assisted predictive maintenance:

[0157] The digital twin platform uses a deep survival analysis model to predict device health status:

[0158] ;

[0159] in, Let be the risk function at time t. As the benchmark risk function, For the weight vector, For Transformer encoders; when the predicted value exceeds the safety threshold, a maintenance warning is triggered in advance;

[0160] S7. Results Output and Decision Support:

[0161] Output fog concentration detection results and confidence levels. The confidence level is calculated based on the following formula:

[0162] ;

[0163] in, Maximum value of FSI in key band, pixel Model predicts signal strength at [location] , The scattering coefficient is wavelength-dependent. That is, the distance from the detector to the target pixel;

[0164] When the confidence level falls below a set threshold, a recalibration process is automatically triggered, and the results are securely transmitted to the monitoring center via quantum key distribution technology.

[0165] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A large fog detector which eliminates the influence of light difference, characterized in that, The multispectral parallel acquisition module, the special data processing chip, the adaptive waveband selection module, the self-calibration optical channel, the MEMS active alignment module, and the digital twin maintenance platform are connected. The multispectral parallel acquisition module is connected with the multi-channel data input interface of the special data processing chip, and is used for parallel transmission of image data collected in each waveband to the chip; the special data processing chip provides real-time spectral analysis data to the adaptive waveband selection module, and the adaptive waveband selection module feeds back a dynamic waveband selection instruction to the chip; the special data processing chip sends a calibration trigger instruction to the self-calibration optical channel, and the self-calibration optical channel feeds back a channel response coefficient and an optical distortion correction coefficient to a correction operation unit of the chip; the special data processing chip outputs real-time image data of each channel to the MEMS active alignment module of the special data processing chip; the special data processing chip uploads equipment operation state data to the digital twin maintenance platform, and the digital twin maintenance platform issues equipment maintenance early warning information to the chip; and the special data processing chip is connected with an external monitoring center through a communication interface. The multispectral parallel acquisition module comprises a wide-angle incident optical lens, a light splitting assembly and a multi-detector array; the light splitting assembly simultaneously distributes incident light transmitted through the wide-angle incident optical lens to the multi-detector array through different wavebands; the multi-detector array comprises a plurality of independent CMOS image sensors, each sensor corresponds to a waveband, and is equipped with a band-pass filter; The adaptive waveband selection module comprises a feature extraction and fusion unit, a signal-to-noise ratio evaluation unit, a fog sensitivity analysis unit, a multi-objective optimization decision unit, an intelligent decision engine and an instruction generation unit. The feature extraction and fusion unit is connected with the pre-processing data output end of the special data processing chip through the data input end, and is used for receiving real-time multispectral data cubes and environmental state data, and performing fusion processing on the multi-source data; The signal-to-noise ratio evaluation unit is connected with the output end of the feature extraction and fusion unit, and is used for calculating the signal-to-noise ratio of each waveband image wherein is the mean value of the λ waveband image, is the standard deviation of the λ waveband image; The mist sensitivity analysis unit is connected with the output end of the feature extraction and fusion unit, and is used for calculating the mist sensitivity indexes of each wave band wherein represents the intensity change rate caused by the change of the mist concentration, represents the noise standard deviation of the λ wave band; The multi-objective optimization decision unit is connected with the output ends of the signal-to-noise ratio evaluation unit and the fog sensitivity analysis unit through the input ends, and is used for multi-objective optimization calculation based on the signal-to-noise ratio and the fog sensitivity index; The intelligent decision engine is connected with the output end of the multi-objective optimization decision unit through the input end, and is internally provided with an algorithm model based on the fusion of attention mechanism and reinforcement learning, and is used for generating an optimal waveband selection strategy; The instruction generation and output unit is connected with the output end of the intelligent decision engine through the input end, and is used for issuing a waveband selection instruction to the control unit of the special data processing chip.

2. The heavy fog detector capable of eliminating the influence of light difference according to claim 1, wherein: The multispectral parallel acquisition module is connected with the multi-channel data input interface of the special data processing chip through the data output ends of the CMOS image sensors. The wide-angle incident optical lens is a fisheye lens, and the light splitting component is a prism, a grating, or a prism hybrid structure light splitter, which splits the light spectrum in the range of 400-1000 nm into eight specific wavebands: 400-450 nm, 450-500 nm, 500-550 nm, 550-600 nm, 600-650 nm, 650-700 nm, 700-800 nm, and 800-1000 nm, and the half width of each waveband is 15±2 nm; the CMOS image sensors are correspondingly arranged as eight, and a global shutter is adopted to realize hardware synchronous exposure.

3. The large fog detector of claim 1, wherein: The self-calibration optical channel integrated reference light source group; The reference light source group is composed of eight TO-Can packaged laser spectrum stabilized LEDs, the central wavelength of each LED is aligned with the center of the detection waveband, and is matched with the corresponding detection waveband, and the half width is 10 nm; the self-calibration optical channel is also integrated with a high-speed optical shutter, a fiber-coupled lens group, a fiber bundle, and an integrating sphere; The fiber-coupled lens group includes three lenses, which are used for preliminary convergence, chromatic aberration correction, and fine focusing of light from the high-speed optical shutter in sequence; The fiber bundle adopts a 19-core multi-mode fiber bundle to uniformly couple the reference light to each detection channel; The light emitting surface of the reference light source group is connected to the high-speed optical shutter, the fiber-coupled lens group, the incident end of the fiber bundle, the emission end of the fiber bundle, and the incident port of the integrating sphere light homogenization system in sequence on the optical path; the light emitting port of the integrating sphere light homogenization system is coupled with the optical path of the main optical system, so that the standard reference light is injected into the multi-detector array in the main optical path.

4. The large-fog detector of claim 1, wherein: The special-purpose data processing chip integrates a multi-channel data parallel processing module, a pipeline architecture, and a neural network accelerator, which are used for real-time parallel processing of multi-waveband data; The multi-channel data parallel processing module includes a plurality of independent data preprocessing units, each of which is connected to the data output end of a CMOS image sensor, synchronously receives the original image data of each waveband through a dedicated parallel data bus, and performs black level correction, non-uniformity correction, and Bayer decoding operations in parallel; The pipeline architecture adopts a three-stage pipeline, the first stage pipeline is directly connected to the output end of the multi-channel data parallel processing module, is responsible for receiving the preprocessed multi-channel image data and performing image registration and alignment operations; The second stage pipeline receives the output data of the first stage pipeline, performs multi-spectral data fusion and feature extraction operations; The third stage pipeline receives the output data of the second stage pipeline, performs image enhancement and standardization processing; The neural network accelerator is connected to the output end of the third stage pipeline of the pipeline architecture through a high-speed on-chip network, receives the feature data after standardization processing, and includes a plurality of convolution calculation units and attention mechanism calculation units working in parallel inside, adopts weight preloading and data pipeline technology, and realizes overlapping execution of convolution calculation and data transmission. An output end of the neural network accelerator is feedback connected to a control unit of the multi-channel data parallel processing module, and a data throughput priority of each channel is dynamically adjusted through an adaptive bandwidth allocation algorithm to form a closed-loop optimization control system. The neural network accelerator adopts a spiking neural network architecture and supports dynamic weight updating and sparse calculation.

5. The large-fog detector that eliminates the influence of light difference according to claim 1, characterized in that: The MEMS active alignment module includes a MEMS micromirror array, and the MEMS micromirror array is a MEMS micromirror integrated in each optical path. A control signal output end of the MEMS active alignment module is correspondingly connected to a driving end of each MEMS micromirror, and is used for outputting an optical path compensation control signal. The MEMS active alignment module further includes an image feature extraction unit, a registration error calculation unit, a control algorithm unit, a driving signal generation unit and a closed-loop feedback unit. An input end of the image feature extraction unit is connected to an image data output end of the special-purpose data processing chip, and is used for extracting SIFT / ORB feature points from real-time images of each channel; an input end of the registration error calculation unit is connected to an output end of the image feature extraction unit, and is used for calculating a translation error Δx, Δy and a rotation error Δθ between images of each channel; an input end of the control algorithm unit is connected to an output end of the registration error calculation unit, and is used for calculating an adjustment amount required by the MEMS micromirror according to the registration error; an input end of the driving signal generation unit is connected to an output end of the control algorithm unit, and is used for generating a high-precision driving voltage signal of the MEMS micromirror; and the closed-loop feedback unit includes a high-precision temperature sensor and a capacitive position sensor integrated on a substrate of the MEMS micromirror, and is used for collecting actual deflection angles and temperature sensor data of each MEMS micromirror in real time through the closed-loop feedback unit, and feeding back to the control algorithm unit to form a closed-loop control system.

6. A detection method of a large-fog detector capable of eliminating the influence of light difference, the large-fog detector being the large-fog detector of any one of claims 1-5, and the detection method comprising the following steps: S1. Multi-spectral data synchronous acquisition and preprocessing: synchronously acquiring light signals of multiple wavebands by using a multi-spectral parallel acquisition module; S2. Adaptive waveband selection and data optimization: dynamically screening key wavebands according to a signal-to-noise ratio and a fog sensitivity index; S3. Special-purpose data processing chip parallel processing and fog concentration inversion: the special-purpose data processing chip implements parallel processing on the multi-waveband data; S4. Periodic self-calibration and data correction: the system automatically starts a calibration process every 10 minutes; S5. MEMS micromirror active optical path alignment; S6. Digital twin assisted predictive maintenance; S7. Outputting a fog concentration detection result and a confidence level.

7. The method according to claim 6, wherein the method is characterized by: The acquisition process in S1 follows an atmospheric scattering physical model: ； wherein, represents the original acquisition signal of the waveband at the pixel point (x, y), is the target signal under the ideal haze-free scene, denotes the atmospheric transmittance, is the atmospheric light intensity, is the dark current noise; The fog sensitivity index in S2 is calculated in the following manner: ； in, This is the original acquired signal in the λ band. With fog concentration The absolute value of the rate of change; It is the noise standard deviation in the λ band. Signal-to-noise ratio; The highest signal-to-noise ratio among all candidate bands; the arithmetic mean of the key band fog sensitivity indices N = 3-5; The fog concentration inversion algorithm in S3 is based on dark channel prior optimization: ； wherein, is an estimated fog concentration, is an adaptive band weight, is a wavelength-dependent scattering coefficient, is a detection distance, is a total variation regularization coefficient; In S4, each channel data is corrected by the following formula: ； wherein the correction coefficient is obtained by reference light source calibration, is a channel response coefficient, is an optical distortion correction coefficient; In S5, by monitoring image feature points of each channel in real time, an adaptive control algorithm based on Lyapunov stability is used to calculate the adjustment amount of the micromirror: ； wherein, is the deflection angle of the MEMS micro mirror in the x direction, is the deflection angle of the MEMS micro mirror in the y direction, is the current deflection angle of the nth micro mirror in the x, y directions at the current moment; represents the gradient of the target function J with respect to the deflection angles of the MEMS micro mirror in the x, y directions ; and is a temperature drift compensation term, which is calculated by real-time temperature sensor data; K is a control gain coefficient. The S6 adopts a deep survival analysis model to predict the health status of the equipment: ； wherein, is a risk function for time t, is a baseline risk function, is a weight vector, is a Transformer encoder; an early maintenance warning is triggered when the predicted value exceeds a safety threshold; The S7 outputs the mist concentration detection result and the confidence level, and the confidence level is calculated based on the following formula: ； wherein, the maximum value of the FSI in the critical band, at the pixel point predicted by the model , is the wavelength-dependent scattering coefficient, i.e. the distance from the detector to the target pixel point; When the confidence level is lower than the set threshold, the recalibration process is automatically triggered, and the result is safely transmitted to the monitoring center.

Images